Summary

• Counties and municipalities should be competing for construction of data centers in appropriate locations within their boundaries.

• Electricity is the biggest cost and constraint on local resources, and that bottleneck will exist independent of data center construction.

• DataCenterAtlas.org web tool highlights the significant benefits of data centers, especially in terms of property tax revenue and jobs.

• The most common reasons for local opposition to data centers can be overcome with better information, more transparent decision processes, and a focus on aesthetics.

Subscribe now

“Prosperity is energy-intensive.”
(Josh Smith, Testimony, House Oversight Committee, 2025)

“...[W]e are entering a very new age... which has made clear the critical importance of both digital infrastructure and connectivity in every sector of the economy and every corner of the world, as well as energy.”
(Satya Nadella, CEO, Microsoft, “G7 Summit Remarks,” 2024)

The current wave of American economic infrastructure is power and energy infrastructure followed closely by digital infrastructure and data centers. And data center investment likely has the strongest growth trajectory.1

Across the United States, billions of dollars are being invested to build the computing infrastructure that powers artificial intelligence, cloud computing, scientific research, financial markets, healthcare, manufacturing, and almost every other industry. It is not a question of if America will build them. It is a question of where.

The data center decision process is one that is largely being decided at the county and municipal level. However, a recent uptick in state-legislature-led data center building restrictions and proposed limitations is redirecting much of the investment.2 State legislatures are setting tax policy and utility commissions are governing rates, but the zoning approvals, conditional use permits, and community negotiations that determine whether a project happens at all mostly sit with county commissions, boards of supervisors, and municipal planning departments.

This is a serious responsibility, and local officials are being asked to carry it amid a noisy, often misleading national conversation. Residents arrive at public hearings with claims pulled from viral posts; developers arrive with rosy projections; and officials are left to referee without neutral, accessible data. The Abundance Institute built the DataCenterAtlas.org web application to fill that information gap.3

This article lays out the framework for how county leaders can capture the substantial benefits of data center development, address legitimate community concerns with transparency, and avoid policy mistakes that leave communities with neither the revenue nor the resolution of the concerns.

Localities Across America Are Already Competing

Many counties and municipalities have already concluded that data centers represent an attractive economic development opportunity. Communities across Texas, Illinois, Virginia, Ohio, and Arizona have actively competed for data center investment because these facilities can generate significant local economic activity.4

Montana, Ohio, South Carolina, and New Hampshire have passed, are currently debating, or introduced “Right to Compute Act” legislation, which makes it harder for state governments to enact restrictive regulations on computational resources including data centers. Right to compute laws require governments to justify restrictions under a heightened standard, while still allowing narrowly tailored public safety regulations. Table 1 shows the current status of Right to Compute laws in US states.5

Data center benefits to localities generally fall into four categories:

• Thousands of construction jobs during the building phase.

• High-paying permanent operational phase jobs.

• Large property tax revenues for counties and school districts and sales and use tax for the states.

• Additional economic activity as suppliers, contractors, restaurants, housing, and local businesses expand to serve a growing workforce.

Let’s start with the fiscal reality. Data centers are among the most favorable land uses a county can host from a budget perspective: enormous capital investment concentrated on a modest footprint, generating property tax revenue far out of proportion to the public services the facility consumes. A data center does not add students to schools, cars to rush hour, or meaningful demands on police and fire. It does add assessed value, often billions of dollars of it, once buildings, servers, and electrical equipment are counted.6

Loudoun County, Virginia, which is home to the world's densest concentration of data centers, reports that the facilities occupy roughly 4 percent of just the county's commercial land yet produce 38 percent of its general fund revenue and nearly half of all property tax revenue, returning about $26 in taxes for every $1 the county spends serving them.7

But the story is no longer confined to Northern Virginia. Georgia's state economic analysis found new data center complexes transforming parcels worth tens of millions into assessed properties worth more than two billion dollars, and state officials report that metro Atlanta counties are planning new school construction and long-deferred water infrastructure on the strength of that revenue.8

Rural counties, in particular, have found that a single campus can anchor a tax base that decades of conventional economic development never delivered. One example comes from Richland Parish, Louisiana, which has a population of about 20,000, where the local school board issued bonus checks worth nearly $51,000 to every teacher based on the amount of sales tax revenue that was generated above expectations during just the initial construction phase of Meta’s Hyperion data center.9 Local businesses have seen increased sales, and neighboring Ouachita Parish also issued much larger than normal bonus checks to their teachers.10

The DataCenterAtlas.org web tool from the Abundance Institute lets officials see accurate county-specific benefits and costs from data centers for themselves. It maps data center capacity, tax revenue, and water demand across the United States at the county level, and translates the raw numbers into equivalences that residents can evaluate. When a constituent asks “what will this do to our water supply?”, the DataCenterAtlas.org tool gives an official a grounded answer based on reasonable assumptions grounded in publicly available sources.

The Numbers Can Be Surprisingly Large

Consider a representative example. A 300-megawatt data center located in a county with a 1.0 percent effective property tax rate—similar to Reagan County, Texas—could generate roughly $13.3 million in annual county property tax revenue, even if the county granted a 50 percent property tax exemption.

To put that into perspective, that annual revenue could finance approximately:

• construction of more than three neighborhood parks each year, or

• salaries for roughly 140 teachers.

The employment effects of this hypothetical data center are similarly significant.

Construction of a facility this size is estimated to support approximately 1,800 jobs during the construction period before stabilizing at roughly 168 permanent high-paying operational jobs, or nearly the same amount of people employed at two average grocery stores. Those workers then generate additional spending throughout the local economy.

Data Centers Are Not Right for Every Locality

We fully understand that not every locality should pursue data centers. Some communities may reasonably conclude that the costs outweigh the benefits. But too often the debate begins and ends with overstated headlines, without carefully weighing the substantial economic opportunities these facilities can create.

There are legitimate reasons why a county may decide that recruiting data centers is not the best strategy.

If electricity is already scarce or prohibitively expensive, the additional electrical demand may not be practical. Localities with very limited undeveloped land or exceptionally high land values may find that alternative uses generate greater economic value. And ultimately, local governments should respect the preferences of their residents if a community simply does not want this type of development.

Those are all reasonable considerations. But they should be weighed against the benefits, and not blanket restricted across all the counties of entire states.11 Data center developers, like those in every other industry, are also likely to avoid areas where the costs are the highest.

Jasmine Sun, a writer for The Atlantic, took a 10-day road trip through Wisconsin and Michigan interviewing activists, local officials, union leaders, politicians, and residents to understand why opposition to data centers has become such a potent political issue. In her article, Sun (2026) concludes that local opposition to data centers is driven by four major factors.

• Visible local costs: data centers can be large, noisy, energy-intensive industrial facilities that can disrupt communities during construction.

• Broken decision-making processes: Residents often feel projects are negotiated in secret (through NDAs or limited public input), leaving communities feeling excluded.

• Weak perceived local benefits: Many people doubt promises of jobs, tax revenue, or economic development, particularly after past disappointments like Foxconn.

• Deep distrust of elites: The friction stems from a rising populist sentiment across America, where residents suspect that political leaders and large tech firms prioritize corporate gains over the welfare of local citizens.

Data center producers and the localities where they reside need to do a better job of communicating the true benefits and costs, including the public in a more transparent decision process. Sun suggests that one reason for states superseding localities in regulating data centers is that local officials have become overwhelmed by increasingly contentious local battles.

In a recent interview with Josh Einiger and Olympia Sonnier (Aug. 6, 2026) of MS NOW, Shark Tank celebrity investor Kevin O’Leary admitted that transparency and communication were severely lacking in the approval process for the proposed Project Stratos data center in Box Elder County, Utah.

If there is one thing O’Leary and his Utah detractors finally agree on, it is that his company failed to provide enough information and reach out to locals before announcing the gargantuan plans.

“We really screwed this up,” O’Leary admitted to MS NOW. “We have not communicated what we do.”
(Einiger and Sonnier, “Kevin O’Leary wants to atone for his data center sins,” Aug. 6, 2026)

More transparency on the part of city councils, zoning and planning commissions, and county commissions could come in the form of mandatory disclosure of tax exemption terms between the data center operator and locality. Other city Mayors, such as Craig Johnson, Mayor of Elk Grove Village, Illinois, have found that mandating a level of aesthetic design into data centers can go a long way toward helping local citizens be accepting of the project.12 The image at the top of this article is a great example of a data center that is designed to be a positive visual contribution to the Eatonville, Florida community, where that data center is schedule to be operational by Q3 2026.

The Environmental Picture is Changing

Much of the public discussion surrounding data centers focuses on environmental concerns, and these deserve serious consideration. Fortunately, the technology is evolving rapidly.

Many of the newest data center facilities employ closed-loop liquid cooling systems that dramatically reduce water consumption compared with earlier generations of data centers. Electricity demand remains substantial, but computing hardware continues to become more efficient, while utilities across the country are investing heavily in new generation and transmission capacity.

In fact, counties seeking sustained economic growth will likely need expanded electrical infrastructure regardless of whether they pursue data centers. Reliable and abundant electricity is becoming a prerequisite for modern economic development of nearly every kind.

As our former colleague at the Abundance Institute said in his testimony before the House Oversight Committee:

“Prosperity is energy-intensive…. America’s economic growth, technological progress, and improved quality of life all require more energy, not less. There is a clear relationship between energy use and wealth. Harnessing energy is both a cause and a consequence of rising living standards.”
(Josh Smith, Testimony, House Oversight Committee, Subcommittee on Economic Growth, Energy Policy, and Regulatory Affairs, July 22, 2025)

From the Electric Power Research Institute:

“And meeting this demand with energy supply does not need to drive up costs for county residents. Much of a utility’s generation, transmission, and distribution expense is fixed, so adding a data center customer that uses infrastructure consistently can improve asset utilization and reduce the fixed-cost component of each kilowatt-hour.”
(EPRI, “The Economics of High Load Factor Customers: How AI Datacenters Can Reduce System-Wide Electricity Rates,” Nov. 2025)

There are many variables that go into the energy cost equation, but Loudon County residents pay at or below the national average electricity bills despite being the worldwide epicenter of the data center buildout.

Don’t Sit Out This Infrastructure Boom

America is in the early stages of a once-in-a-generation expansion of computing capacity. The facilities will be built somewhere. The question facing county commissioners and municipal leaders is whether their communities should compete for those investments or watch them go elsewhere.

That decision deserves more than slogans or social media headlines. It deserves careful analysis of both the benefits and the costs. If your city or county is considering its economic future, we invite you to explore DataCenterAtlas.org. Better information leads to better decisions, and better decisions lead to stronger communities.

References

Brandsaas, Eirik Eylands, Daniel Garcia, Robert Kurtzman, Joseph Nichols, and Adelia Zytek, “Estimating Aggregate Data Center Investment with Project-level Data,” Finance and Economics Discussion Series, Federal Reserve Board of Governors (Dec. 17, 2025).

Bobrowsky, Meghan and Ben Eisen, “The Teachers Getting $50,000 Bonuses Thanks to a Massive Meta Data Center,” The Wall Street Journal (Jun. 11, 2026).

Einiger, Josh and Olympia Sonnier, “Kevin O’Leary wants to atone for his data center sins,” MS NOW (Aug. 6, 2026).

EPRI, “The Economics of High Load Factor Customers: How AI Datacenters Can Reduce System-Wide Electricity Rates,” Electric Power Research Institute (Nov. 2025).

Evans, Richard W., Jared Lambert, and Bryce N.Y. Chinault, “Introducing DataCenterAtlas.org,” Econosseur (June 9, 2026).

Glock, Judge, “The Surprising Heart of the Data-Center Boom,” City Journal (Spring 2026).

Griffin, Greg S., “Tax Incentive Evaluation: Georgia Data Center Sales & Use Tax Exemption,” Georgia Department of Audits and Accounts (Dec. 2025).

Justice, Jessie S. and Tommie L. Shepherd, “Tax Incentive Evaluation: Georgia Data Center Sales & Use Tax Exemption,” Carl Vinson Institute of Government, University of Georgia (Dec. 2025).

Metzger, Daniel J., “Local Moratoria Against Data Center Construction: Considerations for Municipal Governments,” Climate Law, A Sabin Center Blog, Sabin Center for Climate Change Law, Columbia Law School (May 27, 2026).

Nadella, Satya, “Remarks at the 2024 G7 Summit,” Microsoft Corporate Blogs (June 13, 2024).

NCSL, “Subsidizing Servers: How States Are Competing to Attract Data Centers,” Report, National Conference of State Legislatures (April 1, 2026).

Smith, Josh, “The New Atomic Age: Advancing America’s Energy Future,” Testimony, House Oversight Committee, Subcommittee on Economic Growth, Energy Policy, and Regulatory Affairs (July 22, 2025).

Sun, Jasmine, “no data centers in my backyard,” @jasmine’s substack (Aug. 3, 2026).

Tagami, Ty, “State report says data centers a boon to economy despite tax giveaway,” The Current (Jan. 7, 2026).

Truong, Thanh, “Meta’s $27 billion AI data center is transforming rural Louisiana,” Fox 8 Local First News (May 12, 2026).

Those who do not engage in the traditional arts might be wary of calling themselves artists. They might perceive creativity as something extraordinary or beyond their capabilities. A calling for the special few who are born with these gifts.
Fortunately, this is not the case.
Creativity is not a rare ability. It is not difficult to access. Creativity is a fundamental aspect of being human. It’s our birthright. And it’s for all of us.
Creativity doesn’t exclusively relate to making art. We all engage in this act on a daily basis.
To create is to bring something into existence that wasn’t there before. It could be a conversation, the solution to a problem, a note to a friend, the rearrangement of furniture in a room, a new route home to avoid a traffic jam.
What you make doesn’t have to be witnessed, recorded, sold, or encased in glass for it to be a work of art. Through the ordinary state of being, we’re already creators in the most profound way, creating our experience of reality and composing the world we perceive. (Rick Rubin, The Creative Act: A Way of Being, 2023, pp. 1-2)1

This is my first time writing anything on the combination of music and economics, despite both of these areas representing the majority of my daily time. Music has always been a love and leisure of mine. And economics is my profession, which also is something I love and think about when I am relaxing. But I have never formally combined the two… until last week.

I spent the first half of last week at the Abundance Institute’s Creative Frontiers Summit in Deer Valley, Utah, hosted by my colleagues John Hardin and Kayla Harris. The Summit program focused on the music industry, incentives, and hurdles for musicians, new music technologies, and AI. (John Hardin also curates the Creative Frontiers Substack, commenting at the crossroads of innovation, culture, and art.)

Subscribe now

I remember when the Abundance Institute ran its inaugural Creative Frontiers program in 2024, with me thinking, “Why is my energy-tech-AI-focused policy organization running a music workshop?” But since that initial presumption, I have realized that the Abundance Institute was ahead of the curve. The three most influential cultural media today are probably music, movies, and social media. And questions about AI, technology, and regulation abound in all three. The Abundance Institute was ahead of the curve to recognize that the music industry was a key cultural and policy touchstone for how we incorporate AI productively into our lives.

It was a refreshing change of perspective and insights to spend two days last week talking, eating, and jamming with some of the coolest musicians and creators.

The following is a summary of five of the main economic and policy insights I learned spending two days talking about music, technology, regulation, and AI with these musicians and creators.

1. Make music that people like, and the algorithms become secondary

2. Attribution and copyright laws govern who gets paid and how

3. Music compensation has changed since the 1950s and influenced what music gets made

4. All creative output is based on training on previous work

5. Measure the health of the music industry with dollars and entrants

1. Make music that people like, and the algorithms become secondary

I have heard recent experts across platforms lament how important the algorithms are at driving traffic and consumption. But I am convinced that artists and service providers should spend most of their effort on the quality and consistency of their content, and then spend a left-over smaller amount of effort on marketing and algorithm maximization. The following three recent stories showcase the importance of platforms and algorithms. But they also beg a question of whether a better outcome could be realized with better content.

• On July 16, Jon Lewis of SportsMediaWatch.com analyzed the ratings trend after Major League Baseball debuted its Home Run Derby portion of the All Star events on Netflix. He noted the viewership was down 8% from the 2025 event numbers on ESPN, which represented the lowest TV audience for the event since 2003. He proposes that much of the decline in viewership came from switching platforms from the more traditionally accessed ESPN cable network to the Netflix streaming platform. However, I think it is likely that many more people would watch baseball’s home run derby if they were more focused on the content.

  Netflix had over 325 million paid subscriptions as of the end of 2025.2 The highest viewership years for the MLB Home Run Derby came in 1998 and 2008, each year with approximately 9.2 million viewers. The 1998 Derby was at the height of the McGwire-Sosa home run record chase, with Ken Griffey, Jr. also participating. And the 2008 Derby was the final one at the original Yankee Stadium, the same stadium where Babe Ruth, Lou Gehrig, Joe DiMaggio, Mickey Mantle, and Yogi Berra hit home runs. Only four of the top-10 home run hitters participated in the 2026 Home Run Derby, with the most popular player, Shohei Ohtani, notably missing.3

• Olivier Blanchard is the rare combination of a very good economist with a long and renown career in both academia and policy who also has a relatively active and insightful X account.4 He had an interesting post on July 17 noting that he was happy his X account had just passed 300,000 followers. But he noticed something strange. The algorithm can greatly influence the number of followers of an account by a factor of 20, even if there is no change in the quality or frequency of the posts.

  …But there is something interesting and worrisome here. Since March, I have seen the number of new followers per day being about 20 times what it was earlier. It is not because I had great tweets; I had fewer than usual. It is not because of bots: most new followers seem legit. It seems that it is due to a change in the X algorithm, which puts more weight on tweets which appear to trigger good discussions (I do not know the exact criteria).

  Now, obviously, I am delighted by the change. But the scary part is how an algorithm, depending on what it wants to promote, can multiply (and presumably divide…) visibility by a factor of 20. (Olivier Blanchard, X post, Jul. 17, 2026)

• YouTube music commentator Rick Beato posted an interesting video on July 17, entitled, “Why Paul McCartney Won’t Be On My Channel”. In this post, he quotes Bob Lefsetz, the author of a long-running online music newsletter speaking about the marketing of the Rolling Stones’ new album, Foreign Tongues, who said the following about Beato’s channel:

  The biggest mistake the Stones made was not going on Rick Beato’s show. Beato is not subject to the algorithm, his fans are dedicated. They tune in just for him. And Beato specializes in the music itself, how it was created. That’s the kind of information that needs to be spread if you want people to actually stream a song. Go DEEPER! (Bob Lefsetz, “The Stones’ New Album,” Jul. 12, 2026).

  My conclusion is that the necessary condition for consistent attention seems to be consistent, focused, quality content. The various objectives and algorithms for the media platforms are a significant set of incentives for marketing to navigate. But the primary focus of the artist should be on the content first, then then marketing.

In contrast to focusing on algorithms over content, I proposed the following maxim to a number of participants at the Creative Frontiers Summit. And they all gave me an energetic agreement. This maxim was first posed to me at the Summit for case of musicians by my brother, Romney Evans, who is a tech founder and CMO of True Fit, and who has also been in a band, produced an album, and used Suno to create a working example of a title piece of music for a forthcoming movie. I asked the attendees to tell me if they thought the following was true:

The hardest job for an artist is to create a work that other people love. (Romney Evans, Jul. 14, 2026)

One attendee who produces music and provides touring services to bands grabbed my shoulder, looked me in the eyes, and said, “Yes, that is the only thing.” The point is to create algorithm-proof or algorithm-independent content that people are willing to come and get, wherever it is available. Only after that, the marketing job is to make the product as available as possible.

Rick Rubin’s order of priorities for creating great music surprisingly places the audience last.

In terms of priority, inspiration comes first. You come next. The audience comes last. (Rick Rubin, The Creative Act: A Way of Being, 2023, p. 130)

In a similar vein, Steve Jobs famously focused on producing products that he and his friends would want to use as a winning strategy for gaining customers and marketshare in business.

…every good product that I’ve ever seen… is because a group of people cared deeply about making something wonderful that they and their friends wanted…. That’s how almost everything I know that’s good has come about. It didn’t come about because people were trembling in the corner worried about some big company stomping on them. Because if a big company made the product that was right, then most of these things wouldn’t have happened…. (Steve Jobs, Fireside Chat Q&A Session, Apple WWDC, 1997)

I think the focus of the artist, as well as anyone who wants to make a living producing services (myself included with my economics output), should be to put most of his focus on trying to make a product that he loves and that he thinks other people will love. Then use the media and information platforms as effectively as possible to market the work. And lastly let the market speak. Because it is so hard to predict what the market would want, the most successful creators in the business recommend focusing on what you would want yourself.

If content quality and consistency are the primary focus of creators, the roles of marketing, algorithms, platform, and AI’s contribution become secondary. Creators focusing on quality make consumers focus on the product quality, regardless of the tools a creator uses.

2. Attribution and copyright laws govern who gets paid and how

Attribution, copyright, and the institutions and conventions that ensure the creator of a product owns that product and is paid for that product represented a large part of the discussions at the Abundance Institute Creative Frontiers Summit. Of particular interest was the relationship and liability among AI models that were trained on previous works, the creators of those previous works on which the models were trained, the owners of the AI models, the output from the AI models, and the AI users who prompted the models to create the output.

2.1. Current US copyright law

Current copyright law in the United States says that AI-assisted works can receive copyright protection when there is sufficient human creative contribution, while purely AI-generated expressive material is not copyrightable.5 However, there is still debate about what constitutes a human contribution. The current question to answer to determine the legal threshold is “Did a human determine the expressive elements of the work?” Table 1 below shows the US Copyright Office’s positions summarized as a spectrum.6

One of the current debates is whether elaborate AI prompts and human iteration with an AI should represent a copyrightable human contribution. Table 1 shows that this case would likely not be copyritable under current law. However, one can make a strong case that the iterative process with an AI is analogous and similar to the iterative process of recording takes in the studio.

2.2. Sampling rules, drum discrimination, and AI

On the first night of the Creative Frontiers Summit, a dinner panel discussion was held that included John Hardin and Virginia Postrel of the Abundance Institute, Charles C. Mann of The Atlantic Monthly, and music producer Om’Mas Keith. and Discussion of the dinner panel discussion with Om’Mas about sampling law and how it changed music, copyright, and attribution.

Music sampling, the practice of taking segments of sound from another artists work and including it in your work, has gone through four distinct eras: (i) 1970s to mid-1980s, sampling emerges from DJs extending jams from funk, soul, and disco records, (ii) mid-1980s to 1991, affordable sampling technology gives rise to massive use in hip-hop with loose licensing requirements, (iii) 1991 to mid-1990s, Grand v. Warner legal decision requires artists to license samples, and (iv) 2000s to present, modern licensing era in which any digital sample or likeness must be licensed.7 Similar to the line between AI output copyrightability and non-copyrightability from Table 1, sampling law is also an area of active policy debate regarding how similar a current work can be to a previous work and not require licensing.

The introduction of AI into music generation is the most recent augmentation of the sampling debate. Many musicians and producers can now use products like Suno, Udio, and Splice to generate sounds “in the style of” published music on which their AI models have trained, separate stems from recordings, recreate vintage instruments, and imitate production techniques without copying the original recording.

An example of some of the contradictions that exist in musical purism arguments include those who detest AutoTuned vocals, the technology that digitally adjusts the tone of a vocal to an exact pitch.8 One of the attendees at the Creative Frontiers Summit highlighted the hypocrisy of objectively opposing AutoTuned vocals on the basis of artificial musical production saying the following.

If you are against AutoTune, you can only listen to classical violin and cello or Jaco Pastorius. That means any stringed instrument with a fret board can’t be music. (Attendee at the Creative Frontiers Summit, Deer Valley, Utah, Jul. 14, 2026)

This attendee’s argument also applies to the large numbers of digital pedals, amplifiers, and software effects that have sampled other instruments and allow the user to modify them in customized ways.

As a follow-up rant to the quote above, the same Creative Frontiers attendee, who was also a drummer, highlighted an arbitrary rule in US sampling law that drum beats, fills, and patterns cannot be copyrighted in the same way chord progressions, melodies, and guitar solo segments can be copyrighted. I think I agree that drum beats and fills should not be copyrightable. But I don’t know why that instrument should be differentiated from the output of a guitar, piano, or vocal.

Another high profile contradiction among some of AI’s most ardent advocates comes from those involved in the music industry in Nashville. Nashville is famous and somewhat unique from its east coast and west coast music industry colleagues in the degree to which Nashville separates the music writers from the musicians.9 At lunch on the second day of the Creative Frontiers Summit, a music producer highlighted to me how the Nashville music industry is one of the biggest and loudest opponents to AI in music. Yet this producer told me it is common practice in Nashville-produced music to prepare a song with real drums played by a musician, then layer in commercial drum samples that are quantized, compressed, EQ’d, saturated, and heavily processed.

AI is just the most recent tool that polishes and emits music.

And one final note on the topic of US copyright law. Because the sampling and likeness law is so strict in the US, many artists will give precautionary writing credits to other artists just because their previous work has a vague similarity to the current artist’s new work.

At lunch the same afore-mentioned lunch on the second day of the Summit, I asked two music producers a question about The Strokes’ song, “Bad Decisions” on their 2020 album The New Abnormal. When you look at the Wikipedia page for the song, the writing credit goes to the band, a number of engineers and to Billy Idol and Tony James. After seeing this when the song came out, I could see a similarity between the chord progression in the Strokes’ song “Bad Decisions” and in Billy Idol’s song “Dancing with Myself”. But the Strokes song seemed significantly different.

I asked these two producers why the Strokes would give away album points and writing credits for this vague similarity. They both said that an artist will do that if there is even a slight possibility of a law suite. This precautionary assignment of writing credits preempts other artists from suing for rights. The fact that US copyright law imposes a disincentive on artists whose creative products vaguely resemble other works (an increase in cost) must certainly reduce the amount of music that gets published.

2.3. GitHub for music production

In the field of software development, GitHub is the most widely used platform for open source and proprietary version control and software collaboration. Attribution is one of the superpowers of this collaboration engine. You can literally pull up any page of code from a GitHub project and use the “git blame” function to see exactly which lines of code were contributed by which collaborators and when. Software developers rarely rely on resumes for job interviews but rather refer potential employers to their GitHub pages, where the employer can see every line of code that developer has contributed to any GitHub projects.

What if music projects had a GitHub-like platform that showed every individual who contributed to a work of music and what and how big was their contribution? Tom Stingemore is trying to solve that problem with his company, Alloy, which allows a user to quickly and efficiently see all the stakeholders for a creative work and reach out to them for licensing rights.

Detailed and precise attribution for works of music is a key first step toward getting copyright law and regulation right.

2.4. Rick Rubin’s take on abundance vs. scarcity mindset and AI

It is interesting to me that Rick Rubin has a chapter entitled “The Abundant Mindset” in his 2023 book. As one of the giants of the history of the music industry, he is a surprisingly strong advocate for removing barriers that slow down the creation of art.

A river of material flows through us. When we share our works and our ideas, they are replenished. If we block the flow by holding them all inside, the river cannot run and new ideas are slow to appear.
In the abundant mindset, the river never runs dry. Ideas are always coming through. And an artist is free to release them with the faith that more will arrive.
If we live in a mindset of scarcity, we hoard great ideas. A comedian may be presented with a perfect opportunity to tell a favorite new joke they’ve written, but instead will hold it back waiting for a more high-profile occasion. When we use our material, new content comes through. And th more we share, the more our skills improve.
Choosing to live in scarcity leads to stagnation. If we work on one project forever, we never get to make another. The fear of drought and the impulse for perfectionism prevent us from moving on and block the river’s flow.
…In the abundant world, we have a greater capacity to complete and release our work. When there are so many ideas available and so much great art to make, we are compelled to engage, let go, and move forward.
If there is only one work to do, and we intend to retire when it’s done, there is to impetus to finish.
…Our life’s work is far greater than any individual container. The works we do are at most chapters. There will always be a new chapter, and another after that. (Rick Rubin, The Creative Act: A Way of Being, 2023, pp. 201-203)

I don’t think Rubin was an advocate of creative commons fair use for all creative works. I’m sure he would defend copyright law to some degree and the need to pay artists for use of their work unless permission were expressly granted. But he is clearly advocating against the mindset of holding on to intellectual property with a zero-sum mindset of scarcity. I highlighted and italicized his statements, “Choosing to live in scarcity leads to stagnation…. In the abundant world, we have a greater capacity to complete and release our work.”

It is also interesting to note that Rubin’s comments over the last year, from one of the giants of the music industry with likely hundreds of millions of dollars of album points and recording and publishing royalties,10 have been squarely in the pro-AI camp of AI being a legitimate tool in the music making process which lowers the costs of producing and making music and that human involvement is what makes something copyrightable art. In the first part of an interview on the Daily Stoic podcast in June 2025, Rubin describes AI as a disruptive force comparable to punk rock because it challenges established assumptions about who gets to create and how creativity happens.11

Any service industry that is overregulated and overprotected will stagnate, be it music, movies, construction, aviation, or information technology.

3. Music compensation has changed since the 1950s and influenced what music gets made

The way musicians are compensated is likely one of the biggest influences on what kind of music gets made. And musicians’ revenue streams have changed dramatically since the 1950s. Table 2 shows the era, main source of revenue, and corresponding typical musician business model.

The main trend of the last 75 years is moving the main sources of revenue from radio plays and album sales to streaming and touring revenues. In concert with those changes, the music industry includes more independent artists, and all artists rely more on touring, marketing, merchandising, and customized fan interactions. The successful artists of today operate like businesses.

Some of the characteristics of music and the music industry have improved over the years.

• Recording professional-quality music is far cheaper. Software and digital tools and emulators are better.

• Artists can distribute music globally without a major label. Social media has massively reduced the cost of marketing music.

• Social media allows easier direct fan engagement.

• Independent artists can keep a larger share of their rights and revenues if they self-release.

Other aspects of the music industry have become more challenging.

• Streaming pays much less per listener than record or CD sales did.

• Competition is vastly greater because barriers to entry are lower.

• Marketing has become essential, and musicians often need to create content continuously.

• Income is less predictable and is increasingly concentrated among top performers.

The changes in music compensation described in Table 2 have changed the kind of music that is published and sold.

• Current songs have become shorter. Four- and five-minute songs were common in the 1970s and 1980s. Current songs are rarely more than three minutes. For example, a listener who streams two 2.5-minute songs generates roughly twice as many paid streams as someone listening to one 5-minute song.

• Hooks arrive much earlier in songs today. During the album era, some songs had long introductions, gradual buildups, and extended solos. Because listeners can skip streaming songs instantly, the artists must hook them early.

• Individual songs matter more than albums. Albums of past decades often had thematic unity, narrative progression, experimental tracks, and songs that only made sense in the context of other songs on the album. Today, artists release more singles, EPs, and standalone tracks.

The importance of streaming music rather than radio more quickly exposes listeners to many musical genres simultaneously. For this reason, today’s music regularly blurs the boundaries of musical genres. Taylor Swift moved from country to pop. Coldplay went from independent to pop. The killers went from rock to country. And Jelly Roll, Post Malone, and Beyoncé went from hip hop to country.

With all these trends in current music, AI is further reducing the barriers to creating and marketing music. In the words of Rick Rubin’s June 2025 Daily Stoic podcast prediction, “AI Isn’t Stealing Creativity, It’s Supercharging It.”

4. All creative output is based on training on previous work

Nothing begins with us.
The more we pay attention, the more we begin to realize that all the work we ever do is collaboration.
It’s a collaboration with the art that’s come before you and the art that will come after. It’s also a collaboration with the world you’re living in. With the experiences you’ve had. With the tools you use. With the audience. And with who you are today.
…What’s considered art is simply an agreement. And none of it is true.
What is true is that you are never alone when you’re making art. You are in a constant dialogue with what is and what was, and the closer you can tune in to that discussion, the better you can serve the work before you. (Rick Rubin, The Creative Act: A Way of Being, 2023, pp. 89, 91)

All bands start out as cover bands. And every artist’s original work is based on his or her training on the previous works of the previous artists works he or she has loved and also on the ones he or she has disliked. New works of art being based on previous works of art is not only acceptable, it is universal. It is also true that artists should be able to copyright their work and be paid for its use.

One of the biggest arguments in music and AI today is whether AI music companies like Suno, Udio, and Splice should have to license or otherwise pay the creators of the musical works on which their models were trained. Up to this point litigation brought by major record companies against AI music companies has resulted in licensing payments rather than prohibition and restriction.

In June 2024, the three major record companies Universal Music Group, Sony Music Entertainment, and Warner Music Group filed separate federal lawsuits against AI music generators Suno and Udio. The labels alleged that the companies copied millions of copyrighted recordings to train their models without permission, calling it infringement “on an almost unimaginable scale.” They sought injunctions and statutory damages of up to $150,000 per infringed recording.12

Universal Music Group and Warner Music Group settled with the AI companies in late 2025, deciding on licensing and revenue-sharing for Universal’s music catalogues and the associated artists and songwriters.13 Sony has not yet broadly settled and is applying significant pressure to the AI music companies.14

What seems to be happening in the Universal and Warner settlements is indicative of the likely the direction of a longer run steady-state between the traditional music industry and AI music. The AI companies will have to make some kind of licensing and revenue sharing payments to train their models on the music libraries of the major labels.

This is exactly what we have seen with the larger AI model companies like OpenAI, Anthropic, Google, and X. Owners of textual information content like local newspaper companies are making it difficult for AI companies to crawl and scrape their content without permission. And the AI companies are trying to license unique sources of information from these local newspaper companies. It looks like the record companies will find a price that is acceptable to the AI music companies to license their libraries for AI training. AI music is not going away. It has already changed the music landscape and trajectory.

5. Measure the health of the music industry with dollars and entrants

What is the metric for the health of the music industry? Rubin has an interesting take on success that completely and deliberately ignores what comes after a creative work is published and focuses entirely on the contribution from within the individual.

How shall we measure success?
It isn’t popularity, money, or critical esteem. Success occurs in the privacy of the soul. It comes in the moment you decide to release the work, before exposure to a single opinion. When you’ve done all you can to bring out the work’s greatest potential. When you’re pleased and ready to let go.
Success has nothing to do with variables outside yourself. (Rick Rubin, The Creative Act: A Way of Being, 2023, p. 219)

Rubin’s definition of success, however true it may be of individual wellness and happiness, cannot be the metric for assessing the health of the music industry. This question, “What is the metric for the health of the music industry?” was asked of the music producer Om’Mas Keith in the dinner panel discussion on the first evening of the Summit. And I am not sure that I heard a satisfactory answer from him that night.

One set of measures of the health of the music industry is total dollars spent by consumers on music, total dollars invested by creators and producers, and the total of the two. These aggregate measures neglect questions of the depth and number of artists and works.

One could imagine a world in which one artist produces all the music for the world to consume. It is hard to make the case that one artist is the optimal number, even if the aggregate measures of total expenditure are large from that music. But one could also imagine a world in which creators are heavily subsidized, and every individual has the opportunity to create art. This world would likely result in more creators than is optimal.

Aside from the number of creators, one would expect a healthy music industry to exhibit a large disparity in the quality and popularity across the population of creators. Music quality is subjective, and one can see a large disparity in preferences across music consumers. So rather than a measure of the total number of artists, a good measure of the health of the music industry might include the number of new artists. A healthy industry is one in which new artists can easily enter and try to offer their product to the market.

I propose a measure of music industry health that is a weighted average of two objectives—the annual percent change in total consumption of music industry output and the annual percent change in the number of new artists (suppliers) in the industry.

I propose using the percent change in aggregate consumption and the percent change in new artists because the percent change puts both values in the same units. And the weighted average parameter 𝛼 between 0 and 1 represents the importance of the aggregate consumption measure relative to the new artists measure. In the end, this Music Industry Utility Index gives a number that represents the health of the industry in terms of both total spending and the flow of new artists into the industry.

Conclusion

AI is lowering the barriers to music production, music creation, and writing. It is a tool that can create, enable, enhance, and accelerate the creation, production, and marketing or art. Many musicians today are using AI in their creative process, some publicly and some privately. And there is no sure way to reliably identify if AI was used at all in the production of a creative product.

Despite voices protesting the use of AI in the music industry, a significant number of the true greats among musicians, writers, and producers are using AI to make more music that people want to hear. And rather than reducing the number of artists, AI seems to be enabling increased numbers to join the ranks of the creators.

There are policy debates that need to be resolved in the music industry with regard to AI.

• When is a work of music similar enough to another work of art that it must obtain a license from the other work of art?

• What types of music should be copyrightable, including music produced by AI models?

• How should AI companies compensate musicians for the works on which their models are trained?

• How should we measure whether AI is helping or hurting the music industry?

Similar to general AI models, AI music has already shown its value and is here to stay. The music industry seems to be moving towards an equilibrium in which AI companies have to pay to license the songs on which their models train. And it will be interesting to see how the United States chooses to answer the policy questions above in this age of AI, and how those decisions will influence the music that gets made in the future. For music lovers around the globe, our hopes are for an increase in innovation and output of that musical medium that has the power to move us to joy and tears.

If one wishes to know whether a kingdom is well governed, examine the character of its music. (Paraphrase of Confucius, 551-479 BC, see Legge, 1885)15

References

Beato, Rick, “Why Paul McCartney Won’t Be On My Channel,” YouTube video (Jul. 17, 2026).

Blanchard, Olivier, X post, Jul. 17, 2026.

Bridgeport Music, Inc. v. Dimension Films, 410 F.3d 792, 6th Cir. (2005).

Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569 (1994).

Chan, Kelvin, “Universal Music and AI song tool Udio settle lawsuit and partner on new platform, sparking backlash,” AP News (Oct. 30, 2025).

Davies, Hugh. “A History of Sampling,” Organised Sound, 1:1, pp. 3-11 (Apr. 1 1996).

Grand Upright Music Ltd. v. Warner Bros. Records Inc., 780 F. Supp. 182, S.D.N.Y. (1991).

Jobs, Steve, Fireside Chat Q&A Session, Apple Worldwide Developers Conference (WWDC), San Jose, California (May 13, 1997).

Lanquist, Edward D. and Benjamin West Janke, “Copyright Law in 2025: Courts begin to draw lines around AI training, piracy, and market harm,” Reuters (Mar. 16, 2026).

Lefsetz, Bob, “The Stones’ New Album,” The Lefsetz Letter: First in Music Analysis (Jul. 12, 2026).

Legge, James, trans. The Li Ki (Book of Rites). Vol. 2 of The Sacred Books of China. Clarendon Press, Oxford, 1885.

Lewis, Jon, “MLB All-Star Game hits high; Derby at low in Netflix debut,” SportsMediaWatch.com (Jul. 16, 2026).

Mcleish, Claire E.A., “Hip-hop sampling aesthetics and the legacy of Grand Upright v. Warner,” Popular Music, 42:1, pp. 79-103 (Feb. 2023).

McLeod, Kembrew and Peter DiCola, Creative License: The Law and Culture of Digital Sampling, Duke University Press (2011).

Netflix, “Fourth Quarter 2025 shareholder letter,” (Jan. 20, 2026).

Padmanabhan, Anjana, “Inside the Copyright Office’s Report, Copyright and Artificial Intelligence, Part 2: Copyrightability,” blog, US Library of Congress (Feb. 6, 2025).

Perlmutter, Shira, “Copyright and Artificial Intelligence Part 1: Digital Replicas,” A Report of the Register of Copyrights, United States Copyright Office (Jul. 31, 2024).

Perlmutter, Shira, “Copyright and Artificial Intelligence Part 2: Copyrightability,” A Report of the Register of Copyrights, United States Copyright Office (Jan. 29, 2025).

RIAA, “Record Companies Bring Landmark Cases for Responsible AI Against Suno and Udio in Boston and New York Federal Courts, Respectively,” Press Statements, Recording Industry Association of America (Jun. 24, 2024).

Rose, Tricia, Black Noise: Rap Music and Black Culture in Contemporary America, Wesleyan University Press (1994).

Roth, Emma, “Here are the 30,000 songs Sony is suing Udio’s AI music generator over,” The Verge (Jul. 20, 2026).

Rubin, Rick, The Creative Act: A Way of Being, Penguin Press (2023).

Schloss, Joseph G., Making Beats: The Art of Sample-Based Hip-Hop, Wesleyan University Press (2014).

Sewell, Amanda, A Typology of Sampling in Hip-Hop, dissertation, Indiana University (May 2013).

Sewell, Amanda, “How Copyright Affected the Musical Style and Critical Reception of Sample-Based Hip-Hop,” Journal of Popular Music Studies, 26:2-3 pp. 295-320 (Jun.-Sep., 2014).

• Data centers should be assessed based on their benefits relative to their costs, including externalities imposed on other businesses and residents. And data centers should be evaluated by the same criteria as all other businesses and investments across industries.

• The DataCenterAtlas.org web tool provides a transparent, county-level tool that allows policymakers, local officials, and residents to compare the economic benefits, tax revenues, electricity demand, water use, and employment effects of data centers using publicly available data.

• Property tax revenue from data centers can be sizable to counties and municipalities, even when significant tax exemptions are granted.

• Data center water and electricity use are significant, but declining. And current data center water and electricity use are comparable to or lower than many other common uses across industries.

Subscribe now

Figure 1. Screenshot of www.DataCenterAtlas.org

1. Introduction

“The AI breakout is supercharging an already booming data center industry, with estimates that US data center investments will exceed $1 trillion over the next five years. Even this may not be enough: one industry estimate suggests that the largest players in the industry—the so-called ‘hyperscalers’—will need to deploy an additional $1.8 trillion in capital by 2030 to keep up with demand. These investments are coming. The only question is where they are located, and state tax policy choices can play a meaningful role in that decision.” (Jared Walczak. See Walczak, 2025; Martin, et al, 2025; Lee, et al, 2025)

Data centers are the current engine of tech hot takes. Across the country, local planning boards, state legislatures, and even federal agencies are being asked to make consequential decisions about whether and how to allow new data center development.1 Unfortunately, those decisions are increasingly being shaped by a mix of outdated assumptions, cherry-picked statistics, and outright misinformation.2

This article has a simple purpose to lay out the facts on data centers–on both the costs and the benefits. We want to provide a grounded look at what data centers actually do, what they cost, and what false claims to watch out for. We also introduce our tool, DataCenterAtlas.org, that demonstrates these benefits and costs for different sized data centers in different US counties.

The DataCenterAtlas.org app allows you to choose a state and a county within that state and calculate the estimated property tax revenue to that county from a data center of a given size.3 You can see how the property tax revenue from the data center changes as you change the size of the data center and the degree of tax incentives. You can then explore how that data center compares in electricity consumption, water consumption, and employment to other common industries.

Data centers power hospital and banking services, weather forecasts, most of the internet, and AI applications. It is true that data centers use water and electricity, take up space, can make noise, and emit heat. Our goal in this article is to document the true costs and benefits of data centers so local residents and officials can make the most informed decisions about data centers in their areas.

Data centers should be judged by the same metrics and requirements as other industries. What are the technologies enabled by data centers, and are we willing to give them up? What are the tradeoffs? Essential technologies are often initially faced with fear and uncertainty, such as electricity, cars, and the radio.

Other good articles in this space highlighting the benefits of data centers and the way to overcome any relevant resource constraints include Loesel and Smith (2025), Springman and Beard (2026), and Barkley and Edmonds (2026).

2. What Data Centers Do

Long before data centers were the workhorses of the AI arms race, these computational factories were providing the backbone of the IT services our cultures and economies have come to depend on. Since the 1940’s, they have been storing, processing, and transmitting data. Data centers predate barcodes, credit cards, and even velcro.

Here are some examples of what is currently running through data centers every day:

• Cloud computing services (business operations running on services like Amazon Web Services, Microsoft Azure, Google Cloud)

• Financial transactions (banks, credit cards, trading platforms)

• Healthcare systems (electronic medical records, imaging storage, telemedicine)

• Government services (tax filing, social services, public safety systems)

• Telecommunications and internet traffic (email, video calls, messaging, browsing)

• Streaming and content delivery (Netflix, YouTube, Spotify)

• Enterprise software (Enterprise resource planning systems, logistics, supply chains)

• Scientific computing (climate models, genomics, national labs)

• AI training and inference (ChatGPT, Claude, Gemini, Grok, AlphaFold, etc)

Different data centers provide different services to the general compute infrastructure. But if you removed data centers tomorrow, you wouldn’t just lose chatbots. You would lose payroll systems, hospital records, logistics networks, and most of the modern economy. Put differently, data centers are to the digital economy what roads, ports, and factories are to the physical one.

3. Data Center Equivalences

Data centers are not a new industry. The first data centers in the US were built in the 1940s and housed the first mainframe computers.4 Over the years, their use cases have broadened in scope and in intensity. In our tool, we equate data center water use and electricity use to those of more familiar industries, like golf courses and grocery stores.

3.1. Property tax revenue to counties and municipalities

In addition to state sales tax, state corporate income tax, and federal corporate income tax, data centers provide significant property tax revenue to the counties where they are located.5 Loudoun County, Virginia, boasts the highest concentration of data centers in the country—200 data facilities, 50 million square feet—with specific zoning ordinances and tax incentives to facilitate entrants. The data centers in Loudoun County generate almost half of the county’s tax revenue.6 In the words of Loudoun County’s chief of economic development, Buddy Rizer:

“The reason Loudoun, and now other places, wanted data centers is clear. Rizer noted that the centers ‘provide tax revenue, and they don’t put a drain on services. It’s really a perfect industry for a fast-growing community like ours.’” (Buddy Rizer, Loudoun County, Virginia, Chief of Economic Development. See Glock, 2026 and Verge, 2014)

Not only do property tax rates vary within counties and sometimes across types of businesses, but data centers also often enjoy some degree of tax exemption on their property taxes.7 Despite these exemptions, data centers are a significant source of revenue to the county and municipal revenues where they are located.

In our DataCenterAtlas.org tool, you can choose a data center of a given size—ranging from 0.9 megawatts (MW) to 2.0 gigawatts (GW; 1 GW = 1,000 MW). We used a dataset of publicly available data center sizes and data center valuations to estimate the relationship between data center electrical capacity as a measure of size and its market valuation.8 We then use this market valuation with the average effective property tax rate in each county, including the percent of property tax exemption for that county, in order to calculate the estimated property tax revenue from the data center.

There has been a lot of recent debate, and some protests, around the proposed construction of a 1.5 GW data center in Box Elder County in northern Utah.9 A 1.5 GW data center in Box Elder County, Utah would have an estimated fair market value of $14.7 billion. And with a 75% property tax exemption, would still bring in $37.1 million in property tax revenue every year to the county. That is 49% of the 2025 total budget for Box Elder County.10

3.2. New power demand

Peak electricity usage capacity is a valuable metric of the resources consumed by data centers and of their size. We compare this to the new power demand of other large-scale industrial or commercial facilities, such as factory complexes or pumping plants. Primary aluminum smelters consume on average 889 MW of continuous power, and the Edmonston, CA pumping plant runs up to 835 MW.11 Like data centers, these facilities are power hungry and provide essential services.

In our example of a 1.5 GW data center in Box Elder County, the DataCenterAtlas.org tool shows that this data center’s electricity usage would be 9.86 TWh/year—equivalent to 1.3 primary aluminum smelters or to 11 large factory complexes. This size data center would require significant electrical capacity resources.12

The good news on the electrical front is that capacity is increasing across all electricity production methods: nuclear, coal, natural gas, geothermal, wind, solar, and hydroelectric. Some states are experiencing an energy renaissance, with new nuclear reactors coming online in Utah and Idaho by July 2026.13 The Abundance Institute’s 2026 Annual Gala in association with the Operation Gigawatt Summit in May 2026 brought politicians, policy makers, and private sector builders together to discuss electrical capacity and its necessity for sustaining progress.14

A recent study by Ryu and Hiatt at the University of Southern California Marshall Zage Business of Energy Initiative finds that new data centers drive up wholesale energy prices by a small amount among electric cooperatives and municipalities, but have almost no effect on the investor owned wholesale market utilities, which make up 68% of the wholesale market. Further, they find that new data centers have no significant effect on retail consumer electricity prices.15

3.3. Cooling water

The two most scarce resources that data centers use are electricity and water. In our assessment of data center water usage, we make a conservative assumption, which we detail below. However, this conservative assumption reflects the current average water usage of data centers in operation today rather than those currently being built or in the planning phases. As such, we are providing an upper bound of the water that new data centers will use now in the future.

A Department of Energy report from the Energy Analysis and Environmental Impacts Division (Shehabi, et al, 2024) estimated that average data center cooling water usage efficiency is 0.45-0.48 L/kWh (liters per kilowatt hour) or 450-480 L/MWh (liters per megawatt hour). However, more recent advances in data center cooling suggest that newly constructed data centers will be more in the 0.15-0.25 L/kWh range.16 We assume a water usage efficiency (WUE) for new data centers at a conservative 0.2 L/kWh. This estimate is conservative because most systems being constructed now have closed loop liquid cooling, which can have much lower WUE than 0.2 L/kWh.

Our assumption of WUE of 0.2 L/kWh for new data centers results in an estimate that the 1.5 GW data center proposed in Box Elder County, Utah, would use 521 million gallons of water per year for its cooling. This would be equivalent to the water usage of 13.2 golf courses.17 The initial water application for this data center was for 1,900 acre-feet of water per year (619 million gallons) from the Salt Wells Spring Stream.18

General non-data-center Industrial facilities can consume much more water per year, such as the retired Paducah Gaseous Diffusion Plant (uranium enrichment), which consumed nearly 9 billion gallons of water per year.19 And the US Steel Granite City Works in Illinois used 5.8 billion gallons per year.20 And, as mentioned above, a modern data center is likely to use a closed loop liquid cooling system that uses much less water than our 0.2 L/kWh assumption.

3.4. Construction phase and ongoing local jobs supported

During construction, large data centers will have as many as 1,688 employees, primarily working on construction of the facility for 18-to-24 months. After the construction phase, data centers have relatively low labor demands. The typical large data center will support 157 ongoing jobs to maintain the facility.21 A large supermarket will have around between 100-200 full-time and part-time employees.22 More industrial facilities such as the US Steel Granite City Works have about 1,300 employees.23

4. Conclusion

Data centers are not a passing trend. Rather, they are the continuation and direct outgrowth of the computing, information technology, and internet revolutions that began in the 1940s. They are foundational infrastructure for a modern economy that runs on data and computation.

This does not mean that data centers are costless. They use power, water, and land. But so does almost every form of economic activity. The relevant question is not whether data centers have costs. The correct approach is to assess every industry’s benefits minus the costs including any externalities imposed on other businesses and residents.

If data centers were evaluated using the same criteria of other industries, with accurate information about their true costs and true benefits, we would likely see less opposition and a better mix of investment and development that promote the biggest advances in our standard of living.

Appendix

Methodology for calculating data center estimated property tax revenue

In the DataCenterAtlas.org tool, we let the user choose a state, then choose a county in that state with a data center size in terms of electrical capacity.

Figure 2. Screenshot of www.DataCenterAtlas.org state page for Utah with data center facility size

The formula for calculating the state revenue from a data center of a given size is the following:

In most cases, the effective property tax rate is equivalent to the statutory property tax rate, or the officially posted property tax rate. These rates are sometimes posted as percent of assessed value (a property tax rate) and are sometimes posted as mill rates or the dollars per $1,000 of assessed value.

The assessed data center value is always a function of the market value of the data center. In some cases, like Colorado, there is an explicit assessment ratio for which they take a fraction of the market value of the property as the assessed value.24 But in every case, we need the market value of a data center to calculate the assessed value in order to estimate the property tax revenue from the data center.

The most often used characteristic of a data center to describe its size is not its market value but its peak electrical capacity. The newly proposed data center in Box Elder County is expected to begin as a 1.5 gigawatt (gW) data center, which means it will have a 1.5 gW peak electrical capacity.25 Most current data centers in the US are much smaller than 1.5 gW.

In order to estimate the property tax revenue from a given data center, we need to know its fair market value. To do this, we compiled a list of 22 data centers across the United States for which we could find data on the size of the data center in terms of peak electrical capacity. For 11 of those data centers, we also had data on fair market value. Table 1 and Figure 3 show the data for those 11 data centers, their peak electrical capacity, and their fair market value.

Figure 3. Eleven US data centers by electrical capacity and fair market value

As can be seen from Figure 3, the relationship between peak electrical capacity and fair market value is well approximated by a linear relationship, at least for the ranges of values for which we have data. The estimated values for the blue linear regression line in Figure 3 for electrical capacities greater than 95 MW are the following.

FMV_i = 0.010018082083149032 ✕ MW_i − 0.3496020387036265

Because the predicted linear regression line predicts negative fair market values for peak electrical capacity less than 34.9 MW, we fit a three-parameter exponential function to all values with peak electrical capacity less than 95 MW.

You can see how we solve for these values in the code fig3tab1_MWtoFMV.py in the open source GitHub repository for this article.26

The function in Figure 3 maps the data center size in peak electrical capacity (MW) into the fair market value of the data center ($ billions). We then calculate the effective tax rate in each county based on the local assessment rules.

We also include a slider for property tax exemption percent. This allows the user to assess how much property tax revenue a county could earn given different levels of tax exemption. We set the default property tax exemption value at 50%, which is roughly representative of what most data centers get. Tax exemption and tax incentives are common for data centers across the US.27

It is important to note that total property tax liability for a given parcel of land is a sum of applicable county and local property tax rates across taxing jurisdictions, such as county rates, municipal rates, special taxing districts such as fire, police, and education.28

In the DataCenterAtlas.org web tool, we average the total effective property tax rate across each county in the states for which we have data. Below are the details for the calculation of average effective property tax rates by county.

Florida. The Florida Department of Revenue publishes an annual “Millage Rates by County” spreadsheet.29 The most recent data are from November 2025. The “Millage Rates” sheet in this spreadsheet is entitled “Millage Rates by County.” It shows the total millage rate for each county as a sum of all the different county, municipal, and special district property tax rates.

The millage rate is the amount of property tax liability for $1,000 of taxable property value. So a total millage rate of 15.4115 for Holmes County, Florida represents an average effective 0.0154115 (or 1.54%) property tax rate. The Florida county property tax rates we use in the DataCenterAtlas.org tool are the total millage rates by county, divided by 1,000.

Florida taxes commercial properties, including data centers, and 100% of their fair market value, including tangible personal property. Residential property is also taxed at 100% of fair market value, but commercial properties are assessed much more aggressively than residential properties.30

Maine. The Department of Administrative and Financial Services at the Maine Revenue Services publishes “Full value tax rates for years 2013-2024” in an Excel spreadsheet.31 These property tax rates are equalized or effective tax rates calculated by dividing the 2024 Municipal commitment by the 2026 state valuations. The tax rates are also adjusted to account for homestead exemptions, business and equipment tax exemption (BETE) and tax increment financing (TIF) adjustments.

The BETE exemption and Dirigo Business Incentive Program were recently removed from data center property tax liability calculation by LD 713 in Maine, which increases the effective property tax burden on data centers in Maine.32 This recent change is not reflected in the 2024 average tax rates by county. So a user could capture this by decreasing the default property tax exemption slider in the DataCenterAtlas.org web tool. Before this change, Maine commercial businesses, including data centers, were assessed at 100% of their fair market value, minus the BETE exemption. Now data centers are assessed at 100% of their fair market value including the value of their business personal property.

North Carolina. In North Carolina, commercial properties and data centers are assessed at 100% of the value of the land, buildings, site improvements, and business personal property after application of depreciation schedules. The North Carolina Department of Revenue publishes the 2025-2026 County and Municipal Tax Rates and Effective Tax Rates, which combines the county and municipal effective tax rates for each municipality in each county for the fiscal year 2025-2026.33 We then calculate the county average tax rate as the equally weighted average of each municipality’s statutory combined property tax rate within the county. We use the statutory rate instead of the effective rate because most of the reductions in effective rate from the statutory rate likely come from residential properties and not commercial properties.

Some municipalities in the North Carolina tax data showed no municipal property tax rate in which cases the combined property tax rate would default to “None”. We changed this to be the county property tax rate rather than “None”.34

A better approach would be to have weight values for each municipality in calculating the county average tax rate. But we could not find any good weights for each tax area. The ideal weighting data to get more accurate county averages would be total property tax revenue per municipality.

North Carolina offers major sales-and-use tax exemptions for qualifying data centers, but not broad statewide property tax exemptions comparable to those found in states such as Virginia. Qualifying North Carolina data centers can receive exemptions on electricity and certain equipment purchases if investment and employment requirements are met.35

Utah. The Utah State Tax Commission (USTC) publishes annual tax rates by area.36 This document for 2025 (taxarearates2025.pdf) shows the final adopted property tax rate for each tax area within each county as the sum of county, municipal, and special district property tax rates. We then calculate the county average tax rate as the equally weighted average of each tax area total tax rate within the county.

A better approach would be to have weight values for each tax area in calculating the county average tax rate. But we could not find any good weights for each tax area. The ideal weighting data to get more accurate county averages would be total property tax revenue per tax area.

We have posted our CSV file summary of the USTC PDF, as well as a summary .csv file of that document (data/UT/rates_ut_2025.csv), a more highly summarized data file of average Utah 2025 property tax rates by county (data/UT/avg_proptax_rate_by_cnty_ut_2025.csv) and the code that created it (code/UT/ut_analyses.py) in the open source GitHub repository for this article (https://github.com/OpenSourceEcon/DataCenterAtlas).

In Utah, commercial properties and data centers are assessed at 100% of the value of the land, buildings, site improvements, and business personal property after application of depreciation schedules. This is in contrast to residential properties which are assessed at 55% of market value, a 45% residential property tax exemption.37

References

Auer, Emma, “Athens-Clarke County data center ban lifted, replaced with ordinance,” WUGA, University of Georgia (Apr. 9, 2026).

Barkley, Taylor and Jason Edmonds, “From Atoms to Algorithms: Positioning Tennessee to be America’s Energy Leader,” Beacon Center of Tennessee (Apr. 27, 2026).

Bashe, Charles, IBM’s Early Computers, MIT Press (1986).

Cox, Spencer J., “Establishing a Higher Bar for Data Center Development in Utah,” Executive Order, Governor’s Office of the State of Utah (May 29, 2026).

Day, Tim and Nam D. Pham, “Data Centers: Jobs and opportunities in Communities Nationwide,” Technology and Engagement Center, US Chamber of Commerce (May 1, 2017).

Cabrera, Alixel, “Company withdraws its bid for data center water, but not because it’s backing down,” Utah News Dispatch (May 7, 2026).

Carroll County Board of Commissioners, “Carroll County Board of Commissioners Approve 100-day Moratorium on Data Centers & Battery Energy Storage Systems”, Carroll County Board of Commissioners Resolution, WLBB (Mar. 4, 2026).

Dodge, Blake and Harris Sockel, “The Data Center Water Crisis Isn’t Real,” Pirate Wires (Dec. 18, 2025).

DOE, “Water Pump Project Improves Efficiency, Sustainability at Paducah Site,” blog post, Office of Environmental Management, US Department of Energy (Aug. 8, 2023).

Edmonds, Lauren, “Utah’s governor has made it harder for Kevin O’Leary to build his data center,” Business Insider (May 10, 2026).

EPRI, “Annual and Peak Electricity Use: History and Future Projections,” Powering Intelligence 2026, Main Report, Electric Power Research Institute (Feb. 24, 2026).

Fox 5 Atlanta Digital Team, “New data center applications paused in Rockdale County,” Fox 5 Atlanta (May 12, 2026).

Gibbs, Jabari, “Camden becomes first county on the Georgia coast to adopt data center moratorium,” WABE, NPR, PBS (May 7, 2026).

Glock, Judge, “The Surprising Heart of the Data-Center Boom,” City Journal (Spring 2026).

Granskog, Anna, Diego Hernandez Diaz, Jesse Noffsinger, Lorenzo Moavero Milanesi, and Pankaj Sachdeva, with Arjita Bhan and Sofia vo Schantz, “The role of power in unlocking the European AI revolution,” Article, McKinsey & Company (Oct. 24, 2024).

Hansen, Zachary, “Second Atlanta Area Government Hits Pause on Data Centers,” The Atlanta Journal-Constitution (May 8, 2025).

Hiatt, Shon R., “Data Centers Are Not Driving Up Your Electric Bill,” Real Clear Energy (Jun. 8, 2026).

Lee, Vivian, Pattabi Seshadri, Clark O’Niell, Archit Choudhary, Braden Holstege, and Stefan A. Deutscher, “Breaking Barriers to Data Center Growth,” Boston Consulting Group (Jan. 20, 2025).

Loesel, Turner and Josh Smith, “D⁠i⁠g⁠i⁠⁠t⁠al Founda⁠t⁠⁠i⁠ons: The Essen⁠t⁠⁠i⁠al Gu⁠i⁠de ⁠t⁠o Da⁠t⁠a Cen⁠t⁠ers and The⁠i⁠r Grow⁠t⁠h,” Center for Technology and Innovation, James Madison Institute (Jul. 15, 2025).

Martin, Marc S., Michael C. Herrera, Dania Assas, and Ashley Connely, “Trends in the Growth of Investment in US Data Centers Under the New Administration,” Perkins Coie (Feb. 4, 2025).

NCSL, “Policy Snapshot: Data Center Incentives,” Snapshot, National Conference of State Legislatures (Apr. 17, 2026).

Nelson, Rett, “First nuclear reactor built on DOE land in 50 years unveiled at Idaho National Lab,” East Idaho News (Mar. 20, 2026).

Phillips, Nelson, “County passes 2025 budget with smaller tax increase,” Box Elder News Journal (May 27, 2026).

Ryu, Angela and Shon R. Hiatt, “Data Center Entry and U.S. Electricity Prices: From the Wholesale Market to the Power Bill,” Energy Brief, USC Marchall Zage Business of Energy Initiative (May 22, 2026).

Schneider, Ryan, “Growth vs. Grid: Bulloch County Hits Pause on Data Centers with 90-Day Moratorium,” PIER Commercial Real Estate (Feb. 26, 2026).

Shaddox, Travis W., J. Bryan Unruh, Josh Tapp, Clark D. Brown, Greg Stacy, and Emily Fuger, “Survey of Water Use and Management Practices on US Golf Courses from 2005 to 2024,” HortTechnology, 35:5, pp. 848-857 (Sep. 2025).

Shehabi, Arman, Sarah J. Smith, Alex Hubbard, Alex Newkirk, Nuoa Lei, Md Abu Bakar Siddik, Billie Holecek, Jonathan Koomey, Eric Masanet, and Dale Sartor, “2024 United States Data Center Energy Usage Report,” Energy Analysis and Environmental Impacts Division, Lawrence Berkeley National Laboratory (Dec. 2024).

Solomon, Steve, “Sustainable by design: Next-generation datacenters consume zero water for cooling,” Microsoft AI (Dec. 9, 2024).

Springman, Jen and William Beard, “Data Centers: A Free Market Model for the Digital Future,” Goldwater Institute (Jan. 28, 2026).

Tan, Eli, “Their Water Taps Ran Dry When Meta Built Next Door,” The New York Times, (Published Jul. 14, 2025; updated July 16, 2025).

Terry, Ted, “Data Centers in DeKalb County,” Current Issues, Ted Terry Commissioner, Super District 6 (Dec. 17, 2025).

USEIA, Electric Power Monthly, U.S. Energy Information Administration, Independent Statistics and Analysis (Feb. 2026).

Verge, Jason, “In Loudoun, Rizer Works to Help Data Centers Rock the Economy,” Data Center Knowledge (Feb. 18, 2014).

Walczak, Jared, “State Taxation of Data Centers,” Tax Foundation (Dec. 19,2025).

Warnke, Lucinda, “Monroe County considering new rules for data center construction. What are they?” The Telegraph (Feb. 26, 2026).

I estimate that the annual cost of SB 482 in increased Florida government outlays in the first year of activity 2027 would be $4.1 million, followed by increased annual costs of $3.3 million per year thereafter…. Much more significant are the estimated $6.6 billion in extra costs in 2027 on Florida businesses, and $3.1 billion in extra private business costs every year thereafter.

Florida’s SB 482, more commonly known as the Florida AI Bill of Rights, has unanimously passed the Senate Commerce and Tourism Committee and the Senate Appropriations Committee and has been placed on the Special Order calendar for vote by the full Senate.1 The companion bill in the Florida House (HB 1395) exists, but has not progressed as consistently as the Senate Bill.2 The most recent cost estimate from the Florida Appropriations Committee staff is that SB 482 would increase government outlays by $987,062.3

In this article, I estimate both the government cost and private economic cost on Florida businesses of the current version of SB 482. I only estimate the components of the bill that have direct and explicit costs on the government and the economy. A summary of the main components of SB 482 is given in Table 1.

Subscribe now

I estimate that the annual cost of SB 482 in increased Florida government outlays in the first year of activity 2027 would be $4.1 million, followed by increased annual costs of $3.3 million per year thereafter. This is four times the size of the current Appropriations Committee cost estimate of $987,062.

Much more significant are the estimated $6.6 billion in extra costs in 2027 on Florida businesses, and $3.1 billion in extra private business costs every year thereafter. Table 2 shows a summary of the components of this cost estimate, and Section “Cost Estimate Detail” describes the estimate in finer detail.

Cost Estimate Detail

Table 2 gives the detail of the cost estimate of the Florida AI Bill of Rights (SB 482). The following two subsections give finer detail on how these estimates were calculated for the two sections of bill components in Table 2.

Florida regulation of AI companies interaction with minors

Louisiana’s cost estimate of the protections for minors from social media in SB 162 estimated the new regulations and enforcement would require 4 attorneys, 1 forensic investigator, and 1 administrator, costing $774,346 in FY 2024. They note that this staffing estimate was informed by “conversations with the Utah Attorney General and Office of Consumer Protection’s request of an additional 20 full time employees to handle an anticipated workload of complaints and enforcement actions from enactment of the Utah Social Media Regulation Act.”4

Table 3 shows that Florida has 4.2 times the under-18 population as Louisiana. Assuming some economies of scale in implementing the regulation I use a conservative factor of 3.5 to represent the equivalent cost the Florida might have with the regulation of AI with minors portion of SB 482. This adds up to a cost of $2.7 million in the first year. I further assume that the Florida law would cost a little more than the Louisiana law given SB 482’s extra enforcement components.

SB 482 includes not only age verification and parental consent, but also AI reminders and pop-ups, required screening of harmful materials to minors, and significant penalties for noncompliance. If we assume a conservative estimate for these extra complexity costs of 15%, the factor of the Florida law from the Louisiana law goes from 3.5 to 4.0, bringing the total cost of implementing the components of SB 482 dealing with regulations of AI to minors to $3,116,743 in the first year 2027. We estimate that the startup costs to government to set up the systems for monitoring and enforcement will be 25% of the total cost, or $779,186.

The private costs of the proposed restrictions, mandates, and regulations in SB 482 are much bigger than the government costs. Florida has at least 200 AI and social media companies operating in the state.5 The analysis of the UK Online Safety Act estimated that the costs of the age verification would be between $0.5-$10.0 million for a business with 10 million users.6 I assume the average AI or social media company in Florida has 200,000 users (conservative estimate).7 This means that the average cost for age verification is $100,000 per company, or $20 million for all Florida AI and social media companies. I assume the initial cost is twice the cost of the ongoing costs.

To estimate the cost of the parent monitoring platforms in SB 482 on private businesses, we use a similar cost estimate from the UK Online Safety Bill estimates. The specific component estimated in those estimates is labeled “moderation tech.”8 They estimate a one-time build in the first year for a small company to cost between $0.5-$5.0 million. For this component, I make an even more conservative estimate of $1 million for the average small company, which equals $200 million for all the Florida companies. In this case, I estimate the ongoing cost to be one-fourth of the initial cost.

I estimate the cost of the systems to regularly alert minors that they are chatting with an AI that is not human at very small. This is based solely on the fact that this is a fairly inexpensive feature to engineer and maintain. If it cost $50,000 per company on average, that would be $10 million in the first year for all the Florida businesses followed by one-fourth the cost in the following years.

Content moderation seems to be the most expensive compliance component for businesses under SB 482. Clark (2025) reported that “Meta spends as much as $5 billion annually on safety and security measures.” Clutch.co is a company that outsources content moderation, and they charge $50-$99 per hour for human content moderation.9 Those prices are almost certainly coming down as much of the content moderation can be done by AI (ignore the irony of that given the cost estimate for SB 482). I estimate the cost to be below the low point of that range at $25 per hour. If we assume 500 moderation events per day at an average of 5 minutes of labor per event, that is 2,500 minutes per/day or 41.7 hours per day or 15,208 hours per year. At $25 per hour, that is $380,208 per company or $76.0 million for all Florida AI and social media companies. That number seems to be in the ballpark given Meta’s $5 billion price tag.

Florida personal data, image, and likeness restrictions

Wagener (2025) estimated the cost of the No Fakes Act to be $1.14 million annual compliance cost for a digital startup, plus a $500,000 fixed cost to verify the existing catalogue of digital assets. If we assume a much smaller cost per Florida company of $200,000 annual compliance and $100,000 fixed cost, the cost to Florida’s AI and social media companies would be $60 million in the first year, and $40 million every year thereafter.

For SB 482’s restrictions of tech companies from selling or sharing users’ personal data unless de-identified and protected, Castro, et al (2022), Rinehart (2022), and TDWI (2022) provide a good framework for estimating the cost to businesses. These studies focus on the costs of the California Consumer Privacy Act (CCPA). Castro, et al (2022) estimate the in-state costs in California to be $23 billion per year. With Florida’s GDP being 40% of California’s, a conservative estimate of the costs of personal data restrictions in Florida of one-fourth the cost of California is $5.75 billion per year in the first year and $2.875 billion every year thereafter.

Because the most recent cost estimate from the Professional Staff of the Florida Committee on Appropriations estimated minimal costs to the government from these activities, I estimate a minimal $100,000 per year of increased costs from restricting AI use of name, image, and likeness without consent and from restricting tech companies from selling or sharing users’ personal data unless de-identified and protected, respectively.

References

Anderson, Monica, Michelle Faverio, and Jeffrey Gottfried, “Teens, Social Media and Technology 2023,” Report, Pew Research Center (Dec. 11, 2023).

Castro, Daniel, Luke Dascoli, and Gillian Diebold, “The Looming Cost of a Patchwork of State Privacy Laws,” Information Technology & Innovation Foundation (Jan. 24, 2022).

Census, “Annual and Cumulative Estimates of Resident Population Change for the United States, Regions, States, District of Columbia, and Puerto Rico and Region and State Rankings, April 1, 2020 to July 1, 2025,” NST-EST2025-COMP, US Census Bureau (2025a).

Census, “Annual Estimates of the Resident Population for Selected Age Groups by Sex: April 1, 2020 to July 1, 2024,” SC-EST2024-AGESEX, US Census Bureau (2025b). Florida data and Louisiana data.

Clark, Adam, “Facebook Moderation Move Is a Mistake. Why Zuckerberg Shouldn’t Copy Musk With Meta and 5 Other Things to Know Today.” Barrons (Jan. 8, 2025).

Druilhet, Daniel, “Louisiana SB 162,” Fiscal Note, Legislative Fiscal Office, State of Louisiana (Jun. 1, 2023).

DSIT, “Impact Assessment (IA): The Online Safety Act,” Impact Assessment, Department for Science, Innovation, and Technology, United Kingdom (Oct. 23, 2024a).

DSIT, “Potential impact of the Online Safety Bill,” Research and Analysis, Department for Science, Innovation, and Technology, United Kingdom (Oct. 24, 2024b).

Martin, Justin D. and Stephen W. Song, “What 1500 Florida kids told us about digital media policies,” Tampa Bay Times (May 7, 2025).

Professional Staff of the Florida Committee on Appropriations, “SB 482: Bill Analysis and Fiscal Impact Statement,” Committee on Appropriations, Florida Senate (Feb. 17, 2026).

Readycontacts.com, “AI Companies in Florida,” Readycontacts.com (Jan. 28, 2026).

Rinehart, William, “What is the cost of privacy legislation? A Collection of Estimates,” The Center for Growth and Opportunity (Nov. 17, 2022).

TDWI, “State of CCPA Report Reveals Strain, Rising Costs as More Consumers Exercise Privacy Rights,” Research and Resources, Transforming Data with Intelligence (Mar. 9, 2022).

Utah Senate, SB 194, Sen. Michael K. McKell and Rep. Jordan D. Teuscher, State of Utah (Mar. 13, 2024).

Wagener, Trevor, “The Real Costs of the NO FAKES Act,” Computer & Communications Industry Association (Sep. 10, 2025).

Wess, Sydney, “Content Moderation Services Pricing Guide 2026,” Clutch.co (accessed Feb. 18, 2026).

Will AI be the accelerant of broad-based wealth concentration in the hands of a few? Economic theory and economic data tell us that the answer is a resounding… “probably not”.

Wealth inequality had a moment back in 2014 with the publication of Piketty’s landmark book, Capital in the Twenty-First Century. Now the distribution wealth is having another smaller moment corresponding to increased worries by some about the progress and proliferation of AI. The questions from a decade ago are the same ones we are asking now. Can wealth inequality grow indefinitely? And will AI be the accelerant of broad-based wealth concentration in the hands of a few? Economic theory and economic data tell us that the answer is a resounding… “probably not”.

It is actually a comforting credibility of economics that most of its answers are not an absolute yes or absolute no. Instead, economics balances the costs with the benefits and weighs multiple kinds of evidence to arrive at a probability or what is usually something between yes and no.1

In this article, I provide evidence from economic theory and economic historical data that extreme wealth concentration is unlikely to result from the progress of AI. I start with a summary and context of some recent comments from economists, AI practitioners, and media about their forecasts and justifications for how they think AI will influence the future of wealth inequality.

I then summarize a 2019 paper of mine that tests the central hypothesis of Thomas Piketty (2014) that wealth inequality can grow without bound. Our answer is that only in rare and unlikely cases can wealth inequality grow indefinitely.

I conclude by presenting two economic principles that are taught in every introductory economics course that each provide natural offsetting forces to potential imbalances caused by AI-induced wealth inequality: (i) prices adjust to counteract imbalances in supply and demand, and (ii) labor and capital are almost always complements.

Subscribe now

Current Debate

To close out the year in 2025, Philip Trammell and Dwarkesh Patel published a Substack article entitled, “Capital in the 22nd Century”.2 Their title makes clear that they are commenting on and updating the empirical and theoretical predictions of Thomas Piketty’s (2014) much cited, Capital in the Twenty-First Century.

Piketty’s main thesis is that wealth inequality can grow without bound when the rate of return on capital exceeds the economic growth rate, when some individuals save more than others, or when some individuals receive higher returns on investment than others.

When the rate of return on capital significantly exceeds the growth rate of the economy (as it did through much of history until the nineteenth century and as is likely to be the case again in the twenty-first century), then it logically follows that inherited wealth grows faster than output and income. People with inherited wealth need save only a portion of their income from capital to see that capital grows more quickly than the economy as a whole. Under such conditions, it is almost inevitable that inherited wealth will dominate wealth amassed from a lifetime’s labor by a wide margin, and the concentration of capital will attain extremely high levels—levels potentially incompatible with the meritocratic values and principles of social justice fundamental to modern democratic societies. (Piketty, 2014, p. 26)

Trammell and Patel start with a critique of Piketty, that his interpretations of the 20th century and the resulting predictions for the 21th century are too strong because he neglects the importance of labor’s complementarity with capital.3 But Trammell and Patel’s main thesis is that Piketty’s prediction of unbounded growth in wealth inequality is more likely in a society in which the productive process is dominated by capital—robotics run by AI.

The key assumption and prediction in this argument is whether labor will continue to be important in the production process when AI and robotics become more capable than humans for many tasks. It is true that if capital exemplified by AI and robotics push human labor out of the production process (labor and capital are gross substitutes), then wealth inequality can grow without bound—a result with clear theoretical support to which I will refer in the next section.

But I will provide evidence in this section and in the last section of this article that labor and capital have empirically been gross complements, especially in times of rapid advancements in capital technology. In other words, when capital’s capability advances, humans tend to find ways to make their productivity advance and catch up.

Brian Albrecht had a nice response article to Trammell and Dwarkesh as well as some insightful interactions on social media.4 One of the most salient pieces of evidence from Albrecht’s article is his reference to U.S. Federal Trade Commission report from 2019 that gives a visual representation of a survey of estimates from the literature of the elasticity of substitution between capital and labor.

Note that the vast majority of empirical estimates of the elasticity of substitution between capital and labor are less than one. This means that for most of the estimates of this parameter in the 20th and 21st centuries, labor and capital have been found to be gross complements (a low elasticity of substitution). As Piketty, Trammell and Patel, and Albrecht all recognize, it is the counterfactual relationship of labor and capital being substitutes that is a necessary condition for wealth inequality to grow without bound.

Economic Theory: Natural Limits to Increasing Wealth Inequality

Scott Condie, Kerk Phillips, and I published a paper in 2019 entitled, “Natural Limits of Wealth Inequality and the Effectiveness of Tax Policy.” In this paper, we tried to answer how likely it is for the wealth distribution to grow without bound. We did this by studying how the wealth distribution evolves over time in an economic model that has all the key characteristics that you would expect to matter. We then tested whether the most common explanations for causes of growth in wealth inequality resulted in indefinite increases in wealth inequality in the model.

A summary of our findings is that all the standard hypothesized causes of wealth inequality do actually increase wealth inequality. But only in the rarest and least empirically plausible cases can wealth inequality grow without bound. Hence, my answer of “probably not” to the question of whether AI will cause wealth inequality to grow without bound.

We wrote this paper as a response to Piketty’s (2014) main thesis that I have cited above. And the results of our paper are relevant to the current question of how AI will influence the growth of wealth inequality.

The laboratory for our experiments was an economic model that had all the salient component for a question of how the distribution of wealth would change over time. The model was general equilibrium so that prices adjust to offset imbalances in supply and demand. We represented individuals in the model as overlapping dynasties who pass wealth from one generation to the next in order to capture the most important characteristics of how wealth is accumulated over time.

As shown across the columns of Table 1 we tested whether wealth inequality could increase indefinitely under three main scenarios, each of which is listed in the literature as the most likely causes of disparities in wealth: (i) individuals saving at different rates, (ii) individuals having different labor endowments or labor productivity, and (iii) individuals earning different rates of return on their savings.

We studied these three potential causes of wealth inequality under two different assumptions about the elasticity of substitution between capital and labor. The two rows of Table 1 show the results for the case in which capital and labor are gross complements (the most empirically plausible) and the results for the case in which capital and labor are gross substitutes.

Our results, summarized in Table 1, were that none of the drivers of wealth inequality resulted in perpetually increasing inequality in an economy in which capital and labor are gross complements. And under the much stronger assumption that capital and labor are gross substitutes, only differences in savings rates among individuals could generate ever increasing wealth inequality.

Two Basic Economic Principles

1. Prices adjust until supply equals demand

One of the first principles of economics is that supply equals demand. Put more carefully, prices adjust until supply equals demand. This is fundamentally a general equilibrium argument. That is, it requires a model of both the supply side and demand sides of an economy. Then prices are determined that set supply equal to demand. This differs from partial equilibrium analysis in which prices are taken as given and one studies either the supply or demand of an economy or individual separately.

This principle of prices adjusting to offset imbalances in supply and demand is the key intellectual framework for understanding the large costs imposed by governments implementing price controls, like minimum wage, rent control, drug price limits, tariffs, quotas, etc.

This mechanism of price equilibrating supply and demand also helps dispell extreme economic predictions. The end of the 1990s and early 2000s saw a swell of articles and papers suggesting that the earth was going to run out of petroleum fossil fuel. These arguments were collectively labeled “peak oil.”5 What these thought leaders did not account for was that oil prices would rise as supply became scarce, thereby rationing the consumption, and that substitute energy use applications like EVs and renewable generation would expand in response. The extreme predictions of peak oil were never realized. And the simple principle of prices adjusting to equilibrate supply and demand could have prevented these arguments from gaining credibility.

This principle also applies to discussions of AI and capital and labor. As AI applications replace the functions and positions of humans in businesses, a key step in the hierarchy progression is broken. The entry step to middle- and upper-management is diminished, and the price for those skills and those individuals will increase. Worker wages will rise to increase labor supply and decrease labor demand. A lot more work will be done on the effects of AI on the interaction between capital and labor.

2. Capital and labor have always been complements

With the exception of short episodes, primarily restricted to particular industries, capital and labor have always been complements. And the thesis of Trammell and Patel, which I pose as a sequence of questions, is the following. Will the progress of AI and robotics evenutally force human labor out of the productive process? And will AI be the accelerant of broad-based wealth concentration in the hands of a few?

If the answer to the first question is yes, then the answer to the second question is yes. If capital and labor are substitutes and AI pushes human labor out of the production process, then wealth inequality can expand indefinitely. The Condie, et al (2019) study further qualifies that this is only true if capital and labor are substitutes and if individuals have different savings rates (see row 2 of Table 1).

But I argue that this scenario is highly unlikely. Labor and capital have been complements over all of recorded history if you limit yourself to economy-wide estimates over time periods that are not too short.

We’ve already seen in Figure 1 from Raval (2019, Fig. 3) that the vast majority of estimates of the elasticity of substitution between capital and labor suggests that they are complements. One proposed counterexample is that capital and labor were substitutes in Britain in the Industrial Revolution, when capital broadly replaced labor. Another proposed counterexample is the US post-1980 automation and IT revolution. However, Acemoglu and Restrepo (2019) point out that this was task substitution not factor substitution in the production function sense. In lay terms, the Industrial Revolution and Internet revolution created sectoral shifts and changed the types of tasks labor performed, rather than replacing labor.

Bringing this principle back to artificial intelligence, this principle is the reason why I predict that AI will augment labor’s productivity rather than replace it. In every other technological advancement of the last 150 years—of which there have been many—increases in capital productivity were quickly followed by increases in labor productivity. With AI, I see the technology increasing my own productivity every day. I look forward to the day in the not too distant future when education is customized and administered by expert teachers managing AI agents to a larger number of students learning in the way that is optimized for them.

All evidence points to AI and robotics augmenting the value of our human labor rather than replacing it. The valid fear expressed by Trammell and Patel has also been a natural and consistent response of many over the last 150 years to every new technology.6 But history and economic theory tell us that human labor has the potential and likelihood to benefit greatly from AI progress.

References

Acemoglu, Daron and Pascual Restrepo, “Automation and New Tasks: How Technology Displaces and Reinstates Labor,” Journal of Economic Perspectives, 33:2, pp. 3–30 (2019).

Acemoglu, Daron and J. Robinson, “The Rise and Decline of General Laws of Capitalism,” Journal of Economic Perspectives (2015).

Albrecht, Brian, “Will AI Prove Piketty Right?,” Substack post (Jan. 9, 2026).

Appenzeller, Tim, “The End of Cheap Oil,” National Geographic (Jun. 2004).

Bakhtiari, A.M.S., “World Oil Production Capacity Model Suggests Output Peak by 2006–07,” Oil & Gas Journal (Apr. 26, 2004).

Bardi, Ugo, “Peak Oil, 20 Years Later: Failed Prediction or Useful Insight?” Energy Research & Social Science, (2019).

Campbell, Colin J., Jean H. Laherrère, “The End of Cheap Oil,” Scientific American, 278:3, pp. 78-83 (1998).

Carroll, C., J. Slacalek, K. Tokuoka, and M. White, “The Distribution of Wealth and the MPC”. Quantitative Economics (2017).

Condie, Scott S., Richard W. Evans, and Kerk L. Phillips, “Natural Limits of Wealth Inequality and the Effectiveness of Tax Policy,” Public Finance Review, 47:1, pp. 32-57 (Jan. 2019).

Deffeyes, Kenneth S., Hubbert’s Peak: The Impending World Oil Shortage, Princeton University Press (2001).

Goldin, C. and L. Katz, The Race Between Education and Technology (2008).

Hirsch, Robert L., Roger Bezdek, and Robert Wendling, “Peaking of World Oil Production: Impacts, Mitigation, and Risk Management,” Hirsch Report, Department of Energy (2005).

Jones, T.H., “A Critique of Hubbert’s Model for Peak Oil,” FACETS Journal (2018).

Kaplan, S., and J. Rauh, “Wall Street and Main Street: What Contributes to the Rise in the Highest Incomes?” Review of Financial Studies (2010).

Mankiw, N. G., “Yes, r > g. So What?” American Economic Review Papers & Proceedings (2015).

Piketty, Thomas, Capital in the Twenty-First Century, translated by Arthur Goldhammer, The Belknap Press of Harvard University Press, 2014.

Raval, Devesh, “Whats Wrong with Capital in the Twenty-First Century’s Model?” Federal Trade Commission (Apr. 13, 2019).

Rognlie, Matthew, “Deciphering the Fall and Rise in the Net Capital Share”. Brookings Papers on Economic Activity (2015).

Solow, Robert, “Thomas Piketty Is Right… But”. The New Republic (2014).

Trammell, Philip and Dwarkesh Patel, “Capital in the 22nd Century,” Substack post (Dec. 30, 2026).

I estimate the macroeconomic effects for the state of Florida of adding AI regulation beyond what currently exists at the federal level. I make the conservative assumption that Florida restrictive AI regulation could reduce business productivity by 1.0% and could reduce individual labor productivity by a range of 1.0% for the highest earners down to 0.1% for the lowest earners. My macroeconomic simulation shows that increased AI regulation would have the following effects over the 2026-2035 period:

• Reduce Florida GDP by an average of 1.9% or $38 billion per year.

• Reduce Florida employment by an average of 0.5% or 54,100 workers per year.

• Decrease the average wage by an average of 1.4% per year.

Subscribe now

2. Background

Big picture debate in the US. An active debate is ongoing in the US and abroad about how much artificial intelligence (AI) should be regulated. One extreme is that AI is dangerous and should be heavily regulated. Another side is that AI needs to be allowed to flourish unfettered by overly precautionary regulatory policy. In addition, states are wrestling with the federal government over what level of government should have jurisdiction over these questions.

OBBBA. As part of the One Big Beautiful Bill Act (H.R.1, OBBBA), the House of Representatives included a 10-year moratorium on AI regulation by the states.1 Rep. Jay Obernolte (R-CA) has argued that the moratorium is “necessary to avoid a ‘labyrinth of regulation’ with ‘50 different states going 50 different directions on the topic of AI regulation.’”2 There have been many other voices supporting the moratorium with these arguments.3

A number of Republican Senators and Representatives initially opposed the moratorium.4 Forty attorneys general signed a May 16 letter to Congress opposing the measure.5 All of the main arguments against the moratorium are based on the complaint that it violates the federalist principle that states should have the freedom to independently regulate issues within their borders. The primary specific issues cited are limits on any state regulation and child safety regulations.6 The Abundance Institute has drafted an AI moratorium fact sheet addressing each of the main arguments against the moratorium.7

Florida proposed regulations. In the last 12 months, Florida has proposed at least 5 different AI regulations in the state. These Florida proposals are summarized in Table 1 below. I am not opposing these Florida regulations, but I am highlighting the action being taken in one state and that this action is representative in quantity of the many other differing regulations currently occurring in other states.8

The argument is compelling that a patchwork of regulations across states will stifle AI in the United States. It is not just the worry that a particular state will enact its own AI regulatory policy that slows down productivity. It is more that a patchwork of 50 different regulatory policies will create its own compliance, adoption, and investment cost that might dwarf the costs of any one specific set of state regulations.

3. Results

In the analysis in this article, do not assume that AI regulation is bad. I only assume that patchwork regulation across states is inefficient and will create limits on the benefits that can be obtained by optimal development, implementation, and use of AI.

Google and Alphabet CEO, Sundar Pichai, has recently said that Google’s internal metrics estimate that the use of AI has increased the productivity of its employees by about 10 percent.9 This 10 percent productivity increase from AI at Google is certainly larger than we would expect the benefit to be for the United States as a whole, even if some workers are seeing an even higher productivity increase. But there are many current use cases that are extending broadly to US residents in general.10

In our simulation, we use Florida as our example state. Based on the assessments above, we assume that extra AI regulation from Florida would decrease business productivity by 1 percent. We further assume that worker productivity in Florida would decrease by 1.0 percent for the highest earners, phased down to 0.1 percent for the lowest earners. From the discussion above, we think this is a conservative estimate for the potential productivity loss from restrictive AI regulation. But we provide the results of a smaller decrease in productivity in Table 3 below for those who think the effect might be smaller.

I use the open-source OG-USA macroeconomic model to simulate the effects of increased AI regulation at the national level, then we project those percent changes onto Florida GDP, capital stock, employment, and wages. I discuss some of the weaknesses of this approach in Section “4. Caveats”, but I think this approach is an appropriate first pass at the effects of extra AI regulation in Florida. All the code and results for running these analyses are available in the GitHub repository https://github.com/openSourceEcon/FL-AIreg.

Table 2 shows the results of extra Florida restrictive AI regulatory policy on state GDP, capital stock, employment and wages, assuming a 1 percent decrease in business productivity (total factor productivity) and a varying individual workforce productivity decline of 1.0 percent for the highest earners down to 0.1 percent for the lowest earners. The average annual loss in Florida GDP over the period 2026-2035 is -$38.1 billion per year, and the average annual loss in employment over the same period is -54,100 workers per year.

In the case that one thinks that the decrease in Florida business productivity would be less than 1 percent, we ran an extra simulation of the model in which the extra Florida restrictive AI regulation decreased business productivity by 0.5 percent and decreased worker productivity by 0.5 percent for the highest earners and declining to 0.05 percent for the lowest earners. Table 3 below summarizes those results, which are close to half as big as the results in Table 2.

The average annual loss in Florida GDP over the period 2026-2035 for a lower assumed loss in productivity of 0.5 percent is -$19.1 billion per year, and the average annual loss in employment over the same period is -27,730 workers per year.

4. Caveats

In these analyses, we are imputing results from a model of the entire United States onto the state of Florida. There are several potential weaknesses and sources of bias and error from this approach, which include the following:

• US demographics to which the model is calibrated are different from the demographics of Florida. One example is that Florida has an older population than that of the United States as a whole.

• The macroeconomic model is calibrated to US taxes, spending, deficits, and debt trajectory. This is very different from the state of Florida, which has no state income tax and has a balanced budget each year.

• The macroeconomic model is calibrated to the US being a large open economy in which international capital flows into the country to purchase a percentage of government debt and to supply private capital to domestic producers. More realistically, the state of Florida should be modeled as a small open economy that takes the world interest rate as given. International capital flows to Florida are likely very different from those to the US as a whole.

Despite these potential sources of bias in error from using a national model and scaling its results down to a particular state, I think this is a good approximation of what the effect of extra AI restrictive regulatory policy in a state would be.

5. Conclusion

It is likely that AI regulation will be needed as the capacity of the tools grow quickly and our adoption of these tools becomes more widespread. But supporters of the AI moratorium argue that those regulations should be made in one place, at the federal level. Not only is there a race among US tech companies to build the best AI tools and increase market share, but there is also a race among the tech companies of various countries and regions. The main competitor to the US right now in AI is China.

In this article, I estimate the potential economic costs to the state of Florida by enacting its own restrictive AI regulation. These costs are generated by the restrictive regulation reducing Florida business and worker productivity. I found that a 1 percent decrease in business productivity coupled by a varied worker productivity reduction generates annual losses in Florida GDP of $38 billion and annual losses in employment of $54,100 workers.

These losses are significant and might be a lower bound on the losses that could come from every state having its own varied restrictions on AI. The moratorium on state AI regulation in the One Big Beautiful Bill Act (H.R. 1), which is now being debated in the Senate, is an important piece in the global competitiveness framework of US productivity and growth. Furthermore, it can help states, businesses, and individuals better take advantage of the emerging benefits from artificial intelligence.

Model Notes

I used the open source OG-USA macroeconomic model to simulate these results. The documentation for the OG-USA model is online at https://pslmodels.github.io/OG-USA, and the documentation of all the theory of the model is online from the OG-Core model repository at https://pslmodels.github.io/OG-Core. The code for the OG-USA model is available in the GitHub repository https://github.com/PSLmodels/OG-USA. The model can also be downloaded as a Python package (ogusa) through the Python Package Index website https://pypi.org and API. All the code and data to replicate the analyses in this article are in the GitHub repository https://github.com/openSourceEcon/FL-AIreg.

References

Abundance Institute, “Fact Check: Addressing Misconceptions about the AI Law Pause”, Abundance Institute, June 25, 2025.

Adragna, Anthony, “Senate parliamentarian green lights state AI law freeze in GOP megabill”, Politico, Jun. 22, 2025.

Barkley, Taylor, “Protecting Kids Requires Smart, National AI Policy—Not Fifty Shades of Red Tape”, Now + Next, Substack, May 13, 2025a.

Barkley, Taylor, “The State AI Regulatory Moratorium in the States”, Now + Next, Substack, Jun. 2, 2025b.

Barkley, Taylor and Sebastian Griffin, “Op-Ed: AI regulatory moratorium provides time to get the standard right”, Opinion, The Center Square, May 20, 2025.

Chow, Andrew R., “Why AI Regulation has Become a ‘States’ Rights’ Issue”, Time, June 25, 2025.

Frazier, Kevin and Adam Thierer, “1,000 AI Bills: Time for Congress to Get Serious About Preemption”, Lawfare, May 9, 2025.

Kolas, Logan and Nate Karren, “Governors Break Ranks Over AI Regulation Patchwork”, Now + Next, Substack, Jun. 4, 2025.

Maslej, Nestor, Loredana Fattorini, Raymond Perrault, Yolanda Gil, Vanessa Parli, Njenga Kariuki, Emily Capstick, Anka Reuel, Erik Brynjolfsson, John Etchemendy, Katrina Ligett, Terah Lyons, James Manyika, Juan Carlos Niebles, Yoav Shoham, Russell Wald, Tobi Walsh, Armin Hamrah, Lapo Santarlasci, Julia Betts Lotufo, Alexandra Rome, Andrew Shi, and Sukrut Oak, “Artificial Intelligence Index Report 2025”, Institute for Human-Centered Artificial Intelligence (HAI), Stanford University, April 2025.

Nguyen, Britney, “Big Tech’s AI spending spree is going strong. Here’s how big it could be this year”, Quartz, Mar. 5, 2025.

Oremus, Will and Andrea Jiménez, “Republican's’ bid to stop state AI laws heads for a crucial vote”, Washington Post, Jun. 26, 2025.

Endnotes

Summary

We find that restoring full R&D expense deductibility effective back to 2022 would:

• Reduce the marginal effective tax rate for the average U.S. business impacted by the change from 9.8 percent under current law to 6.2 percent in 2026 and to 7.4 percent in 2027 and beyond.

• Generate a $92 billion increase in GDP in 2026 (+0.29 percent) and an average increase of $58 billion per year over the 2027-2035 period (+0.17 percent).

• Generate an increase in the aggregate capital stock of 0.70 percent in 2026 as well as an average increase of 0.46 percent per year over the 2027-2035 period.

• Reduce federal tax revenue by $54.3 billion in 2026 and by an average of $28.0 billion per year for the 2027-2035 period.

Subscribe now

1. Introduction

The Tax Cuts and Jobs Act of 2017 enacted a provision that became effective in tax year 2022 that replaced the ability of firms to fully expense research and development costs in the year they were incurred with a less generous 5-year amortization schedule for domestic R&D and a 15-year amortization schedule for international R&D. The part of the tax code where this law is specified is Section 174.1 In this article, we estimate the effect on business investment incentives and on macroeconomic outcomes from restoring the Section 174 law to full deductibility of R&D expenses, beginning in 2026, with the ability to be effective back to 2022.2

In our microeconomic analysis of the effect on individual business investment incentives, we find that restoring full R&D expense deductibility backdated to 2022 would reduce the marginal effective tax rate for the average U.S. business impacted by the change from 9.8 percent under current law to 6.2 percent in 2026 and to 7.4 percent in 2027 and beyond. This reduction in the marginal effective tax rate on investment can be thought of as an increase in the incentives for businesses to invest in R&D. We estimate that this reform would reduce federal tax revenue by $54.3 billion in 2026 and by an average of $28.0 billion per year for the 2027-2035 period (see Table 1).

In line with our microsimulation results, our macroeconomic model predicts that the reform would generate an increase in the aggregate capital stock of 0.70 percent in 2026 as well as an average increase of 0.46 percent per year over the 2027-2035 period. The change would also generate a 0.29 percent increase in GDP in 2026 and an average increase of 0.17 percent per year over the 2027-2035 period. That is equivalent to a nearly $92 billion increase in GDP in 2026, followed by average annual increases in GDP of about $58 billion per year over the 2027-2035 period (see Table 2A and Figure 1). The larger increases in 2026 are due to the policy potentially increasing R&D investment in the last half of 2025. Table 2B and the dashed lines in Figure 1 show the effects of the policy without the backdating component (without the effect of 2025).

The macroeconomic effects of the reform, especially in 2026 are significant given the average annual growth in US GDP is 2.1 percent over the last 20 years. And our macroeconomic revenue results are consistent with the estimated annual cost of the policy from 2027 and beyond of decreased federal revenue of $28 billion per year.

2. Background on Section 174 and R&D deductibility

Guenther (2022) provides a concise summary of the history of Section 174 and the R&D deduction. Congress created Section 174 of the tax code in 1954 to clarify arguments between businesses and the IRS regarding what research expenses could be immediately deducted and what expenses had to be amortized over 5 or more years. The creation of Section 174 gave businesses two options: (i) to immediately deduct R&D expenses, (ii) to amortize them over at least 5 years. After the creation of Section 174, most businesses chose to immediately deduct their R&D expenses.

Bloom, et al (2019) begin and end their survey of policies that promote innovation with the intuitive statement that “market economies are likely to underprovide innovation” because of the public good nature of knowledge. This market failure suggests a potential role for government to incentivize innovation. In their final ranking of policy reforms that incentivize innovation, R&D tax credits rank as the highest quality and conclusiveness of evidence, highest net benefit, and in the fastest effectiveness time frame.

Cowx, et al (2024) is a recent study that looks at the changes in US R&D investment since the 2022 repeal of immediate full deductibility resulting from the TCJA. They find that this policy change resulted in the effective tax rate faced by businesses increasing by 11.9 percentage points or 62% on average. This is a similar percentage change to our estimated changes in marginal effective tax rates in Table 1. They also found a significant decrease of $12.2 billion in domestic R&D investment in the first year of the policy among research-intensive and borrowing constrained firms. Because of their restricted sample of firms, they characterize their results as underestimating the economy-wide effects.

The most recent OECD (2025) data on “R&D tax expenditure and direct government funding of BERD [business expenditure on R&D]” show that the US ranks 38th among OECD countries in its tax subsidy rates on R&D expenditure for large profitable firms. One reason for the US not ranking closer to the top is that the US private sector likely has more incentive to invest in R&D relative to other countries. However, this low ranking might also be an indicator that the US would benefit in its role as a global innovation leader by increasing its subsidization of R&D.

3. Modeling full R&D expensing deductibility

We use two open source models to simulate and estimate the effects of reforming Section 174 to return to full R&D expensing deductibility effective back to 2022. The first is Cost of Capital Calculator, a business tax microsimulation model of business investment incentives.3 The second is OG-USA, a macroeconomic model of US fiscal policy.4

Neither model has an explicit R&D expensing parameter, but both have investment tax credit rate parameters for businesses. To simulate the effects of returning to full R&D expense deductibility, we have to calculate what increase in the investment tax credit rate would be equivalent to moving back to full R&D expensing deductibility.

The Joint Committee on Taxation (JCT, 2017) estimated that replacing R&D expenditures deduction with expensing amortization in 2022 would result in an increase in federal revenue of $102.0 billion between 2022-2025, or about $25.5 billion per year.5 Thus, our goal is to determine how much of an increase in the investment tax credit (a parameter we have in our models) would reduce federal revenues by $25.5 billion per year (the amount of revenue lost from restoring R&D deductibility, which we don’t have in our model).

In 2014, the Joint Committee on Taxation (JCT, 2014) estimated that the American Research and Competitiveness Act of 2014 (H.R. 4438), which raised the investment tax credit rate 6 percentage points (from 14 to 20 percent), would reduce federal revenue by $19.9 billion in 2023 and $21.3 billion in 2024 (which averaged out to $20.6 billion per year). This was a 43 percent increase in the investment tax credit rate. We use these two estimates (JCT, 2014 and JCT, 2017) to assume that a 53 percent increase in the investment tax credit rate is equivalent to the implicit tax cut incentives from returning to full R&D expensing deductibility.6

The statutory federal and state investment tax credit rates vary by industry and asset types. Furthermore, a limited set of investment expenses qualify for the credit. In our baseline parameterization for both Cost of Capital Calculator and OG-USA we use the investment tax credit rate under current law of 1.5 percent for machines, 2.5 percent for buildings, and 1.0 percent for intangibles, representing an average overall rate of 1.5 percent. So our reform simulation is to model the change to full R&D expensing deductibility from 2027 and beyond as a 53 percent increase in the investment tax credit to 2.295 percent for machines, 3.825 percent for buildings, and 1.53 percent for intangibles, representing an average overall rate of 2.295 percent.

In our analysis, the year 2026 is different because it includes the increased investment tax credit on tax year 2025 R&D, which would affect business investment decisions for half the year if the legislation were passed in July 2025. We assume that 50 percent of tax-year 2025 business filers will either invest more in 2025 or push those investments forward to 2026. Either way, we are accounting for the effect of the policy on 2025 in the year 2026 in our analysis. As such, our calculation for the percent increase in the investment tax credit in 2026 is 79.5 percent, which equals the 53 percent increase in 2026 plus 50% of the 53% increase in 2025 (accounting for behavioral responses in the last half of 2025).

So our reform simulation is to model the change to full R&D expensing deductibility for 2026, which includes the behavioral response from 2025, as a 79.5 percent increase in the investment tax credit to 2.692 percent for machines, 4.488 percent for buildings, and 1.795 percent for intangibles, representing an average overall rate of 2.992 percent.

We estimate the cost of the policy in terms of revenue by using the inflation-adjusted average annual cost of reinstating full R&D deductibility of $28.0 billion per year from JCT (2017) (see the second row of Table 1). For a policy with backdating the deductibility for the years 2022-2025.

We assume the JCT (2017) cost estimate average for the years 2022-2025 of $25.5 billion per year. Then we inflate that amount from 2023/2024 dollars to 2027 dollars using the most recent CPI inflation forecast from the CBO (2025) March 2025 Long Term Budget Outlook to get a revenue loss estimate from the policy of -$28.0 billion per year. We then use the equation below that assumes a 50 percent decaying adherence rate to calculate the cumulative effect of the backdating effectiveness of the 2022-2025 years on the cost estimate for 2026 of -$54.3 billion.

4. Business Investment Incentives: Cost of Capital Calculator

The open source Cost of Capital Calculator model simulates the effect of tax policy on the investment incentives of corporate and noncorporate businesses.7 Table 1 shows the effect of reinstating full R&D expensing in Section 174 with backdating to 2022 on business incentives to invest in equipment and machinery through a measure called the marginal effective tax rate (METR). The METR measures the total burden of the tax system on new, equity-financed investments by corporate businesses. Under current law, the METR on new investments is about 9.8 percent. This is much lower than the federal corporate income tax rate of 21 percent because corporations are able to utilize a number of provisions—such as bonus depreciation and state investment tax credits—that can reduce effective tax rates relative to the statutory rate.

When we change Section 174 to allow for full expensing of R&D in the current year effective back to 2022, the METR decreases to 6.2 percent in 2026 (due to the addition of a half-year of increased R&D investment in 2025) and to 7.4 percent in 2027 and beyond. We estimate that the decrease in federal tax revenue from this effective business tax cut would be $54.3 billion in 2026 and $28.0 billion in 2027 and each following year. This decrease in METR from 9.8 percent to 6.2 percent in 2026 then to 7.4 percent in 2027 and beyond shows the increase in pro-growth business investment incentives (a tax cut) from allowing full R&D expensing deductibility.

5. Macroeconomic Effects: OG-USA

OG-USA is an open source, large-scale, general equilibrium overlapping generations macroeconomic model with rich representations of fiscal policy including both individual and business taxation, as well as government spending, deficits, and debt. The model also has detailed population demographic dynamics and international capital flows.8 In each of these simulations, we are assuming that the reform is permanent rather than the 5-year expiration that is in the current version of the bill passed by the House.

Tables 2A and 2B show the percent change from the reform, with and without backdating to 2022, in GDP, aggregate capital stock, aggregate employment, and average wage over the 10-year budget window 2026-2035. Figure 1 plots those macroeconomic time series from 2026 to 2035. Our simulation results show that the policy change will result in an increase in the growth rate of GDP equal to 0.29 percent in 2026 followed by an average increase of 0.17 percent per year from 2027 onward. Our simulations also show that the annual increase in the aggregate capital stock would go up by 0.70 percent in 2026 followed by an average increase of 0.46 percent per year from 2027 onward. This is equivalent to a nearly $92 billion increase in GDP in 2026, followed by average annual increases in GDP of about $58 billion per year in the 2027-2035 period. Total employment would be mostly unaffected by the policy change. And average wages rise with percentages similar to that of GDP.

Figure 1. Percent change in macroeconomic variables from Section 174 reform allowing full R&D expensing deductibility, backdated to 2022 versus no backdating: 2025-2035

Figure 2 shows the percent change in government debt and total tax revenue from the reform. The decrease of 1.0 percent in 2026 followed by a decrease of just over 0.7 percent per year from 2027 and beyond is consistent with the JCT (2017) estimate of the increased revenue from removing the full and immediate expensing deductibility of R&D being $28 billion per year in federal revenue.

Figure 2. Percent change in fiscal variables from Section 174 reform allowing full R&D expensing deductibility, backdated to 2022 versus no backdating: 2025-2035

These relatively small macroeconomic changes are the reason for the small changes in government debt as a percent of GDP, shown in Figure 3. This plot shows the time series of debt-to-GDP in both the baseline and reform. The reform barely influences debt-to-GDP, with the reform values (with and without backdating) in 2035 being 130 percent compared to the baseline value in 2035 of 129 percent.

Figure 3. Change in government debt-to-GDP from Section 174 reform allowing full R&D expensing deductibility, backdated to 2022 versus no backdating: 2026-2035

As a final analysis of the macroeconomic output, Figure 4 shows the average percent change in household savings over the first 10 years of the simulation (2026-2035) by lifetime income group. This is a distributional analysis of the reform results. In Figure 1, we see that the biggest macroeconomic result from this reform is an increase in the aggregate capital stock of 1.10 percent in 2026 with an average increase of 0.46 percent per year from 2027 and beyond. Figure 4 shows that this increased aggregate capital stock is coming primarily from households in the top one percent of lifetime income earners. Everyone in the bottom 99 percent of income earners has roughly no change in savings behavior.

Figure 4. Average annual percent change in household savings by lifetime income group from Section 174 reform allowing full R&D expensing deductibility, backdated to 2022: 2026-2035

6. Conclusion

We find that reinstituting full deductibility of R&D expenses with backdated effectiveness to 2022 would have a significant positive effect on investment and GDP in 2026 with moderate increases in 2027 and beyond.

Although the reform does not significantly change the amount of labor supplied by US workers, the average wage does increase by percentages similar to those of GDP.

References

Bloom, Nicholas, John Van Reenen, and Heidi Williams, “A Toolkit of Policies to Promote Innovation”, Journal of Economic Perspectives, 33 (3): 163–84 (Summer 2019).

CBO, “The Long-Term Budget Outlook: 2025 to 2055”, Congressional Budget Office (March 27, 2025).

Cowx, Mary, Rebecca Lester, and Michelle Nessa, “The Consequences of Limiting the Tax Deductibility of R&D”, Working Paper No. 4192, Stanford Business (July 2024).

Guenther, Gary, “Tax Treatment of Research Expenses: Current Law and Policy Issues”, CRS Insight, Congressional Research Service (Dec. 19, 2022).

JCT, “Description of H.R. 4438, the ‘American Research and Competitiveness Act of 2014’”, JCX-38-14, Joint Committee on Taxation (April 29, 2014).

JCT, “Estimated Budget Effects of the Conference Agreement for H.R. 1, The ‘Tax Cuts and Jobs Act’: Fiscal Years 2018-2027”, JCX-67-17, Joint Committee on Taxation (December 18, 2017).

JCT, “Estimated Revenue Effects of Provisions to Provide for Reconciliation of the Fiscal Year 2025 Budget: Fiscal Years 2025-2034”, JCX-22-2 R, Joint Committee on Taxation (May 13, 2025).

OECD, “R&D tax expenditure and direct government funding of BERD”, Organization for Economic Cooperation and Development, (last updated May 19, 2025, accessed May 31, 2025).

What is amazing about this most recent jobs report is that it makes the post-COVID US labor market (Jan. 2021 to Dec. 2024) tied for the second longest streak of consecutive positive monthly employment gains—48 straight months—in the last 86 years.

In a time when inflation pressures, high interest rates, and fiscal worries cloud the economic horizon, the US labor market is a bright spot of strength, resilience, and consistency.

Subscribe now

The December 2024 US employment report came out last Friday, January 10, 2025.1 It showed a continued monthly increase in US nonfarm payroll employment of +256,000 jobs in December 2024, which is pretty close to the average monthly job growth over the last two years (+218,542 jobs per month). But what is amazing about this most recent jobs report is that it makes the post-COVID US labor market (Jan. 2021 to Dec. 2024) tied for the second longest streak of consecutive positive monthly employment gains in the last 86 years—more specifically since 1939.2

Figure 1 below shows the time series of total US nonfarm employment (seasonally adjusted) from 1919 to 2024. This figure shows the total number of employed workers, not the change. The total number of workers in the December 2024 report was 159.5 million. This compares to 90.8 million workers in January 1980 and 29.9 million in January 1939. You can see that periods of declining employment are associated with recessions. You can also see that the long-run trend is positive employment growth—average +0.14% per month and average +1.75% per year over the 1919 to 2024 period.

Figure 1. US total nonfarm payroll employment, monthly, seasonally adjusted, 1919-2024

If you look at all the monthly changes in employment—current month total employment minus the previous month total employment—some are positive (employment gains) and some are negative (employment losses). The average monthly percent change in employment over the entire time series is a monthly employment gain of 0.14% (fourteen hundredths of a percent), which would have been an increase of +226,374 jobs in December 2024—pretty close to the actual monthly increase of +256,000.

Define a positive employment streak as the number of consecutive months with an increase in employment relative to the previous month. There have been 83 such positive streaks since January 1939. The smallest of these streaks is just one month, representing a positive employment change month that is both preceded and followed by negative employment change (22 of these). The largest of these positive employment streaks is 113 months from October 2010 to February 2020 (see Figures 2 and 3 and Table 1).

Figure 2. US employment streaks: consecutive positive monthly gains and cumulative employment gains, 1939-2024

Figure 3. US employment streaks: consecutive positive monthly gains and average monthly employment gains, 1939-2024

In summary, the US labor market since the end of COVID has been growing at rate that is close to the long-run average rate. But its consistency has been historic. The current streak of 4 years (48 months) of consecutive monthly employment gains is currently tied for second longest. But it will likely stand alone in second place when the January 2025 employment report comes out on Friday, February 7, 2025. We have a long way to go to reach the nearly 10-year employment gain streak of the 2010-to-2020 period, and it is unlikely we will get there. But the average monthly employment gains in our current streak (+354,542) are higher than the other four streaks of more than 40 consecutive months. In a time when inflation pressures, high interest rates, and fiscal worries cloud the economic horizon, the US labor market is a bright spot of strength, resilience, and consistency.

A note on open source resources and replicating these data and images

All the source data and corresponding data visualizations from this article can be obtained and replicated by running the code locally on your personal computer by using the resources in the USempl_Streaks-2025-01 (https://github.com/OpenSourceEcon/USempl-Streaks-2025-01) GitHub repository. Instructions are found in the Jupyter notebook USemplStreaks_2025_01.ipynb in the repository. One advantage of creating the images on your local machine is that the images created with the Python code have more dynamic functionality, including mouse hover-over full data description, muting and unmuting of particular series, zoom, and selection.

The presence and frequency of ad hominem arguments in a policy debate is an indicator of a weak argument.

We don’t want to restrict the speech of others, but we have the right to choose what information we consume. And the aggregate effect of those choices has the potential change the calculus of polarization.

This is not an article about open source tax policy, but it is about openness and transparency in policy debate. I am revisiting an article I wrote nearly 15 years ago on June 30, 2009 on the original Econosseur blog. My article was entitled “On Economic Debate—the Ad Hominem Index”.1 Given that we are entering the heavy season of the 2024 general US Presidential election, this article seems especially relevant.

My old article starts off by highlighting a characteristic of the Nobel Prize winning (1976) economist, Milton Friedman (1912-2006), as described by his longtime coauthor and collaborator, Anna Schwartz (1915-2012), in her June 2009 review of a biography of Milton Friedman.2

“For example, Friedman's style as a debater reveals an aspect of his personality. He was always courteous to his opponents in a debate, never attacked ad hominem. He concentrated on weaknesses of the opponent's arguments and invariably emerged as the victor in the debate.”

Subscribe now

The term, “ad hominem”, is latin for “to the man” or “to the person”. An ad hominem argument is a well known logical fallacy in which a case is made using “an attack on the opponent’s character rather than the opponent’s argument.”3

My point in the original blog post was to highlight my admiration of policy debates and arguments that avoid ad hominem arguments. I then looked at an example of a long series of exchanges between two popular economist/bloggers of the time, Greg Mankiw and Paul Krugman (Nobel Prize in economics, 2008). I noted a general lack of ad hominem attacks in the arguments of one author in contrast to the other author’s near ubiquitous use of ad hominem attacks.

My thesis of the old article was that the presence and frequency of ad hominem arguments in a policy debate is an indicator of a weak argument. The ad hominem index was a conceptual proxy heat measure of the weakness of a debater’s position. Unfortunately, the logical fallacy of ad hominem arguments are believed to be particularly effective, even if the literature on this is inconclusive.4

One of my favorite characteristics of open source policy modeling is that the openness and transparency of the models renders them fundamentally apolitical. That is, a user might not agree with my previous policy positions, my training, or the organizations for which I have worked. But because all of my modeling assumptions are out in the open, an interested policy proponent has more of an incentive and requirement to comment on the assumptions of my model rather than any personal characteristics of mine. This is a much more productive scientific and democratic form of debate and collaboration.

What if our two presidential candidates used no ad hominem attacks? What if we never heard the phrases “crooked Joe Biden” or “loser Trump” in campaign events? The ad hominem attacks distract from the substance and nuance of the policy stances and characteristics the candidates support. Ad hominem attacks divert voter attention away from the policies.

Negative political attack ads and arguments are rightfully protected under the freedom of speech component of the First Amendment of the US Constitution, as long as they do not contain patently false statements or libel.5 If freedom requires minimal restriction on ad hominem arguments and negative attack ads, what can be done? I propose three approaches to sift through distracting ad hominem attacks to be able to better get to the core substance of policy arguments.

1. Media should explicitly calculate an ad hominem index for campaign speeches, political advertisements, articles, and political debates. I don’t know what the mathematical formula for this index should be. But it could include quantification of the number of unrelated personal attacks per substantial topic or policy area or the number of personal attacks per word. This index value could become viewed as a sufficient statistic for the soundness of an advocate’s arguments (low value ad hominem index value being associated with good arguments). Imagine if CNN and Fox News both posted independently calculated ad hominem index values for political candidates.

2. Big tech companies should offer AI (artificial intelligence) ad hominem filters for both written and video speech. These filters could be trained on previous year’s debates, ads, and articles with ad hominem arguments labeled by experts. As such, an AI ad hominem filter would depend on how it was trained. Because of this variance, it would be important for a competition of filters to exist. Versions of these AI filters already exist in OpenAI’s ChatGPT model.6

3. Voters must choose to exclude the ad hominem attacks of their candidates. This is the hardest requirement. The truth is that we humans love our biases to be confirmed by outside sources.7 This is why we tend to more happily tolerate ad hominem attacks from our preferred candidates and are more likely to condemn ad hominem attacks from our less preferred candidates. We must choose to exclude the ad hominem attacks of all candidates. This is a hard sell. But we need to opt in on the AI ad hominem filters (which don’t explicitly exist yet) and assess the ad hominem index values (which don’t exist yet) on the speeches, ads, articles, and debates we consume. At the very least, we can be individually more discerning as to which arguments are valid.

I am tired of polarization, demonizing of opponents, and negative campaigning on all sides of the political spectrum. I long for a day when candidates debate the costs and benefits and winners and losers of different policy options rather than attack their opponent’s person. How nice it would be for policy advocates to self impose logical consistency (the lack of logical fallacies) in their arguments.8

If our political system and candidates are currently in a bad equilibrium in which all sides have the incentive to use ad hominem arguments and negative attack ads, we the voters might be able to change their incentives by opting in to metrics and filters that restrict our access to these fallacious approaches. We don’t want to restrict the speech of others, but we have the right to choose what information we consume. And the aggregate effect of those choices has the potential change the calculus of polarization.

UPDATE (June 20): Hattip to Bob Ewing of Ewing School (those who know, know). He sent me this investment pitch video on WeFunder.com for a new group called AllSides Technologies, Inc. Among other things, AllSides says their BiasChecker AI product will be released soon. Their description is the following:

“The AllSides Bias Checker provides instant bias ratings and analysis for any news article. Since it is based on the AllSides patented bias system, it reflects the average judgment of all Americans, not just a panel of insiders or the musings of a biased AI system.”

• Broad benefits for Nevada residents

• Low loss of state revenue

• Maintain Nevada tax revenue from tourism

Subscribe now

1. Nevada fiscal landscape

Nevada is a state that has seen its total state tax collections and sales tax collections higher than its 15-year trend for the last three years (since Q2 of 2021, see Figure 1). And the state’s total reserves and Rainy Day fund balances are still at historical highs (see Figure 2). The state legislature has an opportunity next year to cut taxes in a way that maintains fiscal balance and provides the most benefit to Nevada residents. Taxes that could be cut include Nevada business taxes, property taxes, and sales taxes.

Figure 1: Nevada tax revenue collection as percent of trend: 2008Q3 to 2023Q2

Figure 2. Nevada Rainy Day fund and total reserves as a percentage of general fund expenditures: 2000-2023

Nevada is one of 9 states with no state tax on individual employment income. This is one of the reasons why the Nevada total state and local sales tax burden ranks 14th highest among US states. You can also see in Table 1 that 4 of the top 15 highest sales tax states are ones with no state individual income tax.

Without a state employment income tax, Nevada relies heavily on its state and local sales and use taxes and on its taxes on tourism. Figure 3 below shows the percentages of total state tax revenues of each broad category of tax revenues for Nevada fiscal year 2022-2023. State and local sales and use taxes represented 70.5% of total state tax revenue.

Table 2 below shows the detail of Nevada state revenue subcategories for fiscal year 2022-2023. In addition to tourism affecting the sales tax, many of the tax revenues from tourism are also in the “Other taxes” category, which represented in total 5.3% of Nevada state tax revenues in FY 2022-2023. The largest tourism tax revenue streams were from the lodging tax ($256 million), short term car lease fee ($97 million), live entertainment tax ($80 million), intoxicating beverage tax ($51 million), and transportation connection tax ($40 million).

2. Nevada sales tax landscape

We present evidence here that points to the Nevada sales tax as the most beneficial area to cut taxes for residents of the state. Other alternative reforms include cutting the state’s business taxes, insurance and real estate taxes, mining taxes, and tourism taxes.

Nevada has a minimum sales tax rate of 6.85%, with optional additional local sales and use tax rate of up to 1.525% extra (a maximum total sales and use tax rate of 8.375%). Table 3 breaks down the minimum sales tax rate of 6.85% into its components. The first two categories represent the 4.6% state sales tax rate, with the third row 0.5% going to county government and fourth row 1.75% going to city government. The fifth category is the optional local sales and use tax rate, which ranges between 0% and 1.525%.

The distribution of total state and local sales and use tax rates across Nevada’s counties is show in Figure 4. The most rural and least populated counties have the minimum sales and use tax rate of 6.85% (Esmeralda, Eureka, Humboldt, and Mineral), while the most populated and most visited for tourism Clark County (includes Las Vegas, Henderson, and Mesquite) has the maximum sales tax rate of 8.375%. Washoe County (includes Reno and Sparks) has the second highest population in the state, the second most tourism, and the second highest sales tax rate of 8.265%.

The Nevada sales and use tax covers most retail goods, vehicles, and restaurants (prepared food). Goods and services that are not covered by the state sales tax include farm machinery and equipment, groceries (unprepared food), and most professional services.2

3. What are the most efficient taxes to cut on consumption goods?

We want to address two measures of efficiency in this section. The first is distributional equity, which can loosely be defined as an efficiency measure. We want to broadly avoid reforms that benefit or make worse off one group of the Nevada population more that other groups. For example, a regressive tax makes low-income residents worse off than high-income residents. Conversely, a progressive tax penalizes high-income earners more than low-income earners. We want to avoid both extremes.

The second measure of efficiency can be referred to as economic distortion, deadweight loss, or behavioral effects. These inefficiencies are related to how much people change their consumption decisions when the price of the good changes. The economic term for this response behavior is price elasticity, which is the percent your consumption of a good decreases in response to a one-percent increase in the price of that good.

We start with distributional equity. Figure 5 shows the share of total expenditure for eight broad consumption categories for households with different income levels. These data come from a national sample of consumers.3

The primary pattern that emerges from Figure 5 is that most of the consumption categories represent fairly constant shares of total expenditure for households across the income spectrum. High income households spend the same percentage of their total consumption dollars on these goods as low income households. This is particularly true of the “Other, taxed” category, which includes most of the Nevada consumption goods that fall under the state sales tax. This category also represents the biggest share of expenditure for all incomes.

The exceptions to this constant expenditure share across income pattern are “Owned dwellings” and “Rented dwellings”, both influenced by the property tax (highlighted in red and orange, respectively). For higher income households, a larger share of consumption expenditure is spent on owned dwellings and a lower share is spent on rented dwellings. As such, a cut in property taxes would benefit high income home owners and low income renters most.4

In terms of distributional equity, a cut in the Nevada sales tax rate would have the broadest base across Nevada residents and would influence high- and low-income households similarly.

The second measure of efficiency of a tax is how much economic distortion it imposes. When you create or increase a tax, you increase the price of the good. This price increase makes consumers choose to buy less off that good. How much less they buy in response to a price change is called the price elasticity of demand.5 For this reason, the price elasticity of demand is a good proxy statistic for the degree of distortion a tax imposes. In other words, if you impose a tax on a highly elastic good (like “Leisure” goods shown in Table 4), you’ll create the biggest drop in demand for that good.6

Looking back at Table 2, many of Nevada’s entertainment industry taxes are in the “Other taxes category”. The ideal Nevada tax cut would maintain its taxes on tourism, while mostly cutting taxes on Nevada residents. Most of the categories of retail goods that fall under the Nevada sales tax would also be categorized under the “Non-durables” consumption goods category in Table 4. This category has a relatively high elasticity compared to the other categories and, therefore, fits the criterion of a category that would benefit from a tax cut. In other words, the estimated -0.816 elasticity on non-durables in Table 4 suggests that a cut in the Nevada sales tax would significantly benefit Nevada consumers because they would increase their demand for those goods.

4. What would be the effect of a Nevada sales tax reduction?

In this section, we estimate the size of the loss in state tax revenue from a 0.6 percentage point reduction in the Nevada state sales tax. We assume that the change would take place in 2025, although any policy change would more likely take place in FY 2025-2026. We assume a growth rate in sales tax receipts of 4% for the three fiscal years between the most recent data FY 2022-2023 and FY 2025-2026. This is a conservative growth rate assumption compared to the 21.32% and 6.73% growth rates between the prior two fiscal years, respectively.7

We present the estimated revenue loss in Table 5 both with and without the behavioral response as estimated by the -0.816 elasticity in Table 2 due to the lower price from the reduced sales tax.

In the static case in which we assume no change in demand from the sales tax reduction, we estimate that the 0.6 percentage point reduction in the Nevada sales tax would reduce state revenue by $254 million in FY 2025-2026, a 13% reduction.8 Given the historically high reserves in the state, this reduction is within the realm of fiscally feasible policies. But the actual cost in revenue loss is likely to be much lower.

The last row of Table 5 shows our estimate of revenue loss from the 0.6 percentage point reduction in the Nevada state sales tax if we incorporate the estimated price elasticity on nondurable goods from Table 4 of -0.816. This means that for every 1% decrease in the price of nondurable consumption, consumers will on average increase consumption of those goods by 0.816%. The 0.6 percentage point decrease in the sales tax represents an average 7.3% decrease in the price of those goods across the state. The estimated price elasticity means that quantity demanded would increase by about 6.0% in response to the average 7.3% price decrease. We estimate that the revenue loss with the estimated increase in demand from the policy change would only be $153 million, a 7.9% decrease in revenue.9

5. Conclusion

Nevada’s current fiscal health provides an opportunity for the state to implement some pro-growth policies that make its residents more likely to stay, make prospective residents more likely to move to the state, and encourages businesses to locate and grow in the state.

Cutting the state sales tax will have the broadest effect on Nevada residents and will affect them similarly across the income spectrum. Also, cutting the state sales tax will leave Nevada tourism revenues relatively untouched.

The strongest case for cutting the state sales tax is that the decrease in sales tax rate will be partially offset by an associated increase in demand by Nevada consumers. This increase in demand would represent a tangible increase in standard of living for residents across the income spectrum.

The Minnesota Timberwolves beat the Denver Nuggets last Monday night 106-80 in Game 2 of the NBA Western Conference Semifinals. The Timberwolves’ defense holding the Nuggets to 80 points represented the minimum points that Minnesota allowed this season, both in the regular season and in the playoffs.1 In this article, we discuss another Minnesota minimum about which there is less consensus.2

Minnesota is one of only four states that currently have a state individual alternative minimum tax or state AMT (see Figure 1). Iowa had a state AMT up until the beginning of 2023, when it was repealed.3 The stated goal of the Minnesota AMT is “to ensure that anyone who benefits from certain tax advantages pays a minimum amount of tax”, focused primarily on high income tax filers.4 As we detail below, the AMT is much more complicated than a minimum tax liability level or percentage that a filer must pay. The Minnesota individual AMT really creates two separate and independent ways of calculating state income tax liability.

In this article, we show how the Minnesota state AMT affects many different types of household tax filers in the state. We run our simulations using the FiscalSim-US open source model of US federal and state individual income taxes and benefits. Our experiment is to simulate the standard (non-AMT) current law Minnesota individual income tax for a number of different types of state tax filers and show how much extra tax liability the state AMT adds to these filers. Because all of our modeling infrastructure is open source, we provide all of our code and source data in a way that is easy to replicate and customize.5

Figure 1. Four states with state alternative minimum tax (state AMT) as of May 2024

We find that Minnesota’s AMT does guarantee that middle-to-middle-high income taxpayers pay an increased minimum amount of tax by limiting certain deductions. However, this introduces significant complexity to the tax code. Additionally, Minnesota’s AMT only applies to a small portion of tax payers, and raises a relatively small amount of money. Even high income households will often not pay significantly higher taxes because the AMT tax is more dependent on the amount of deductions and dependent exemptions filed than on employment income.

However, the Minnesota AMT does impose a significant extra tax penalty for each additional child in Minnesota households for a wide range of middle-to-high incomes. And the state AMT extra tax can also be significant on households with large expenses that are deductible in the standard state tax liability calculation but are not allowed in the AMT calculation.

This article is organized in the following sections:

• 1. Introduction

  • Figure 1. Four states with state alternative minimum tax (state AMT) as of May 2024

• 2. Details of Minnesota’s Alternative Minimum Tax

• 3. Effects of Minnesota AMT on Minnesota tax filers

  • 3.1. Scenario 1: MN tax filers, varying employment income and number of children

    • Figure 2. Minnesota AMT minus standard tax liability for married filers by employment income and number of children

    • Figure 3. Minnesota AMT minus standard tax liability for single filers by employment income and number of children

  • 3.2. Scenario 2: Effect of increasing AMT-non-allowed deductions

    • Figure 4. Minnesota AMT minus standard tax liability for married filers by AMT-non-allowable deductions and number of children

    • Figure 5. Minnesota AMT minus standard tax liability for single filers by AMT-non-allowable deductions and number of children

• 4. Conclusion

• Footnotes

2. Details of Minnesota’s Alternative Minimum Tax

The first federal alternative minimum tax was passed in 1969 as part of the Tax Reform Act.6 The minimum tax was passed as a response to high income households that were paying little to no income tax because of large subtractions from their taxable income or deductions. Tax filers would add a percentage of certain deductions and subtractions over a certain limit. In 1977, Minnesota enacted its first state alternative minimum tax that had a similar structure to the federal AMT.7

In 1982, the U.S. Congress passed the Tax Equity and Fiscal Responsibility Act, which changed the federal alternative minimum tax structure from its simple add-on original form to a requirement for some filers to calculate their taxes two different ways and pay the greater of the net liabilities.8 Minnesota passed state tax reforms in 1987, that similarly amended the Minnesota AMT.9 Instead of adding a portion of deductions to the final tax liability, the current minimum tax uses an alternative tax system with fewer deductions, subtractions and exemptions and different tax rates. Tax payers must compute their tax liability under both the normal tax system and the AMT system and pay the higher of the two.

There are  some notable differences between the standard federal tax system and the federal AMT system. The standard deduction for the federal AMT is much higher. For a household that is married, filing jointly, the standard deduction for the AMT is $82,150 versus $25,800 under the standard non-AMT system. This ensures that the federal AMT will not affect lower income households. Only some itemized deductions are allowed under the federal AMT. For example, intangible drilling costs can not be subtracted from the federal AMT liability while charitable contributions and medical expense deductions still can. Lastly, after taxable income is calculated under the limited set of deductions and subtractions allowed in the AMT, it is taxed at 26% and 28% for income over $206,100, instead of using the 7 tax brackets under the normal tax code.10

Minnesota’s state AMT closely follows the federal AMT with some differences. Minnesota’s AMT does not allow dependent exemptions. This creates an effective penalty for children under the AMT. In addition, home mortgage interest is not deductible under the Minnesota AMT. Lastly, the Minnesota AMT uses a 6.75% flat tax. This is 6.75% marginal tax rate on all income under the Minnesota AMT structure has is a lower rate than the all Minnesota’s progressive marginal tax rates for any income over $41,050 for married filing jointly (6.80%, 7.85%, and 9.85%) in the Minnesota standard system. For this reason, many higher income taxpayers will often pay more under the normal tax system with the higher rates. The Minnesota AMT only applies to state tax filers for whom the AMT-excluded deductions are a big part of their tax filing.

Minnesota’s state AMT does not affect most tax filers in the state nor does it raise a lot of revenue. A 2017 white paper from the state House of Representatives forecast that Minnesota’s AMT would raise $23.6 million in the 2017 tax year from around 9,500 taxpayers.11 That is about 0.2 percent of Minnesota tax revenue from about 0.3 percent of registered Minnesota taxpayers.

3. Effects of Minnesota AMT on Minnesota tax filers

In this section, we look at the effect of income and deductions not allowed under the AMT and how they impact total tax liability. To do this, we model two separate scenarios of the difference between the tax liability computed under the Minnesota AMT and the standard Minnesota tax liability calculation. We refer to this difference as the “extra MN AMT above the standard tax liability”. Where the difference is zero, the calculated Minnesota AMT amount is less than the standard calculation tax liability. Where this difference is positive, the household pays the Minnesota AMT calculated tax liability as an extra amount above the standard calculated tax liability.12

3.1. Scenario 1: MN tax filers, varying employment income and number of children

In our first scenario, we look at the effect of number of dependent children on the extra MN AMT above the standard tax liability for a married couple filing jointly. As discussed in section 2, the Minnesota AMT calculation does not allow for the state dependent exemption of up to $4,800 for each child.13 In order to show the effects of the AMT, we model the net tax liability of a married household with $100,000 of AMT-non-allowed deductions (e.g. gambling-loss deduction or home mortgage interest deduction).

We use the FiscalSim-US open source microsimulation for federal and state income tax to show how different levels of before-tax employment income and number of children affect the extra Minnesota AMT above the standard tax liability. In Figures 2 and 3, we show the extra MN AMT above the standard tax liability for household employment incomes between $0 and $600,000, for married (Figure 2) and single (Figure 3) filers, and for number of child dependents between 0 and 6.

Figure 2. Minnesota AMT minus standard tax liability for married filers by employment income and number of children

Figure 3. Minnesota AMT minus standard tax liability for single filers by employment income and number of children

The first characteristic to note about the Minnesota AMT is that it works exactly as it was designed for all tax filers, adding extra tax liability for married filers with employment income between $82,150 and $543,000 and extra tax liability for single filers with employment income between $61,610 and $500,000. In this case of $100,000 of AMT-non-allowed deductions, the increased standard deduction under the state AMT calculation of $82,150 for married tax filers and $61,610 for single filers ensures that there is no extra state AMT liability for households with employment income under these respective standard deduction amounts.

At incomes for married and single filers above $543,000 and $500,000 in Figures 2 and 3, respectively, the standard tax liability calculation becomes greater than the state AMT calculation. This is because the standard calculation has a progressive schedule of marginal tax rates that exceed the flat 6.75% tax rate on taxable income in the state AMT calculation.

Figures 2 and 3 clearly show the tax penalty on each additional child in a household resulting from the Minnesota AMT. The difference between each line in figures 2 and 3 in the positive extra tax range is close to $350, which is approximately equal to the 6.75% flat AMT tax rate multiplied by the state dependent exemption of $4,800 per child, which is not allowed in the state AMT calculation.

Figure 2 shows that a married Minnesota taxpayer with $309,000 of employment income, $100,000 of AMT-non-allowed deductions (e.g., gambling losses, mortgage interest) and 7 children would have an extra AMT tax liability of $6,099. Figure 3 shows that a single Minnesota taxpayer with $260,000 of employment income, $100,000 of AMT-non-allowed deductions and seven children would have an extra AMT tax liability of $7,467.

3.2. Scenario 2: Effect of increasing AMT-non-allowed deductions

The other big difference between the state AMT tax liability calculation and the standard Minnesota calculation is the expense and loss deductions that are allowed under the standard calculation but are not allowed under the Minnesota AMT, such as gambling losses, home mortgage interest deductions, state or local taxes, and state or local taxes (income, sales, or property).14 In this second scenario, we fix the employment income of a married tax filer (Figure 4) and a single tax filer (Figure 5) at $200,000. And we show how the extra Minnestota AMT above the standard tax liability changes as the amount of AMT-non-allowed deductions increases.

Figure 4. Minnesota AMT minus standard tax liability for married filers by AMT-non-allowable deductions and number of children

Figure 5. Minnesota AMT minus standard tax liability for single filers by AMT-non-allowable deductions and number of children

We only plot the values of the MN AMT non-allowed deductions on the x-axis up to $200,000 in Figures 4 and 5 because employment income is fixed for all these filers at $200,000, and the standard tax liability does not increase for deduction amounts above employment income. This is seen as the flat line for filers are the top end of the AMT-non-allowed deduction values. The extra child penalty of the state AMT described in the previous section is also evident in Figures 4 and 5.

As the size of AMT-non-allowed deductions increases going left-to-right on the x-axis, the extra MN AMT above the standard tax liability calculation increases. And the extra tax liability increases at the rate of $675 dollars for every $10,000 extra of AMT-non-allowed deductions—the 6.75% flat tax rate of the MN AMT.

The left side of Figure 5 for single Minnesota tax filers is interesting. The extra MN AMT liability above the standard tax liability calculation is positive for zero AMT non-allowed deductions for filers with 5, 6, and 7 children. This means that for single filers with 0-to-4 children, the higher marginal tax rates on $200,000 of employment income in the standard tax liability calculation are greater than the AMT liability that does not allow for the dependent exemption. However, at 5-to-7 children, the lack of dependent credit makes the AMT tax liability calculation greater than the standard calculation.

4. Conclusion

The Minnesota individual alternative minimum tax has existed in some form for the last 47 years. Its stated purpose is to guarantee that high income tax filers are not able to avoid paying taxes by deducting certain expenses, like gambling losses and mortgage interest. However, the Research Department of the state House of Representatives estimated in 2017 that Minnesota’s AMT would raise $23.6 million in that tax year from around 9,500 taxpayers–about 0.2 percent of Minnesota tax revenue from about 0.3 percent of registered Minnesota taxpayers.15

In this article, we used the FiscalSim-US open source microsimulation model of federal and state individual income tax and benefit policy to simulate Minnesota tax liability for different types of Minnesota tax filers. We showed that the Minnesota AMT imposes a significant extra tax penalty on each additional child in households with a wide range of middle-to-high incomes. Furthermore, the state AMT can also impose significant extra tax liability on household with large deductions that are not allowed under the state AMT, such as gambling losses, mortgage interest, and state and local taxes.

Given the degree of extra complexity introduced by the Minnesota AMT, the small number of households to which it applies, the small amount of revenue it raises, and the associated child tax penalty, it is not likely that the Minnesota AMT would pass a cost-benefit test. The MN AMT places costs and disincentives in places that would be hard to justify given the narrowness of its target demographics and the small amount of revenue it raises. If the MN AMT could not simply be removed, it is at least likely that more efficient ways exist of taxing middle-to-high income Minnesota households.

I want to highlight and celebrate the most important “streak”er to the United States economy: the US labor market.

In the spirit of John Cena at this year’s Oscars—whose streaker bit was a reference to a different Oscars streaker from 50 years ago1—I want to highlight and celebrate the most important “streak”er to the United States economy: the US labor market. Although the job market has shown some recent weakness,2 its current streak of now 40 consecutive months of positive job growth is representative of a broader trend.

Despite having a few other 20-year periods of labor market performance similar to the last 20 years in terms of total job growth and average job growth, there is one dimension on which the last 2 decades are unmatched. No previous 20-year period has had more streaks of consecutive positive employment months, nor has any previous period’s streaks had more total jobs added than those of our most recent two decades.3

Figure 1 plots each of the 83 positive consecutive US nonfarm employment streaks since January 1939, also showing the cumulative employment total over each streak. The orange streaks are the shortest and smallest streaks in terms of both consecutive months (< 40) and cumulative employment gains (< 10 million), respectively, and are not labeled in the legend. The remaining 7 streaks have either consecutive positive employment months greater than or equal to 40 (longest) or cumulative employment gains of more than 10 million (largest) and are labeled in the legend, with the most recent streaks being dark purple and the older streaks fading to green then yellow.

Figure 1. US employment streaks: consecutive positive monthly gains and cumulative employment gains, 1939 to 2024

Table 1 shows the statistics for those 7 longest or biggest streaks. Obviously, positive employment streaks that last the longest will tend to have the largest total employment gain, as is the case with the 2010-2020 streak. For this reason, I also calculate the average monthly employment gain for each streak.

I have highlighted in yellow the rows in Table1 for the four most recent large streaks, starting in 2003, 2010, 2020, and 2021, respectively. It is interesting to note that the 7-month COVID recovery streak from May to November 2020 added nearly the same number of total jobs (12.3 million) as the positive streak of nearly 4 years from July 1975 to March 1979 (13.0 million jobs).

Figure 2 shows a scatterplot for all the 83 consecutive positive employment streaks in the US data with months in streak on the x-axis and average monthly employment gain across the streak on the y-axis. Not only is our current streak of 40 consecutive positive months (January 2021 to present April 2024) ranked 5th highest among the 83 total consecutive streaks, but it is also ranked 5th highest in terms of average monthly employment gain during the streak.

Figure 2. US employment streaks: consecutive positive monthly gains and average monthly employment gains, 1939 to 2024

Resilient employment in the face of headwinds

Last week’s US GDP report came in lower than expected at 1.6% annual growth, inflation has been significantly higher than its 2% target level since March 2021, and the most recent nonfarm job openings data released two days ago was at a 3-year low. But the US labor market has been surprisingly resilient in the face of these headwinds. Figure 3 shows the time series of monthly US nonfarm employment from July 1919 to the most recent month.4 The vertical gray bars designate periods of recession.

Figure 3. US nonfarm employment, monthly, seasonally adjusted: 1919 to 2024

When you look at the Figure 3, the most prominent characteristics are the three biggest drops in employment at the beginning of the Great Depression (Aug. 1929 to Mar. 1933), the Great Recession (Dec. 2007 to Jun. 2009), and the COVID-19 pandemic recession (Feb. to Apr. 2020), and the otherwise steady upward trend of growth in US jobs.

The average time between recessions since 1945 has been 5 years and 5 months.5 The time since the end of our last recession has been 4 years. The US will certainly experience another recession in coming years. And there is always a chance that an economic slowdown arrives sooner than later. But I want to highlight and celebrate the broad US labor market, which has been the steady foundation to which our recent economic resilience has been anchored.

Figure 4 gives a cleaner comparison of employment losses and subsequent recoveries at the beginning of each of the last 15 recessions, from the Great Depression to the most recent COVID-19 pandemic recession. Figure 4 is a normalized peak plot in which the peak employment level at the beginning of each recession is normalized to a value of 1.0, then each previous value and each subsequent employment level is shown as a percentage of that peak value.

Figure 4. US nonfarm employment, normalized peak plot, comparison of last 15 recessions, 1929 to 2024

Figure 4 shows how historically unprecedented were the employment losses at the outset of the most recent COVID recession (bold black line). But we also start to see that the employment recovery since the COVID recession has been extraordinarily steady. The steady upward climb of the most recent positive employment streak (Jan. 2021 to Apr. 2024) has been historically strong, as we observed in Figures 1 and 2 and in Table 1.

And the longest positive employment streak of 113 months from Oct. 2010 to Feb. 2020 (nearly 10 years) came on the heals of the Great Recession of 2007-2008, one the biggest economic shocks since the Great Depression. It is interesting to note from Figure 4 that the labor market after the Great Recession (bold pink line for recession Dec. 2007 to Jun. 2009) took 76 months (more than 6 years) to get back to its pre-recession levels.

I close this article with one final table of what the US labor market has done in the last 20 years (and 7 months) since the beginning of the September 2003 consecutive positive employment streak. The US economy has added 28.0 million jobs since September 2003, an increase of 21.5 percent. Table 2 decomposes those job gains into the job gains in the most important broad US industry categories.

The total employment in the industries in Table 2 add up to total US nonfarm employment. While it is true that the broad services industries have grown the most, it is not true that blue-collar jobs have declined the most. Among blue-collar industries, manufacturing has lost the most jobs with a 9.7% decline over the last two decades. But other blue-collar industries, such as Mining and Logging and Construction increased by 12.5% and 21.2%, respectively.

The biggest winner over the last 20 years was Transportation and Warehousing, with an increase of 57.3%. Similar to Wholesale Trade (also up 11.4%), Transportation and Warehousing is largely an intermediate goods industry. It often involves trade between producers. The fact that this was the highest growth industry in the US over the last 20 years suggests that the main engine of growth in the US economy is more about the relentless march forward of our broad production sectors, and not just a shift from manufacturing to services industries.

Epilogue

This US labor market strength seems to come from deeper cultural and institutional advantages than any one political party or ideology can claim. The four major positive employment streaks from the last two decades cut squarely across four presidential administrations that included two Republican Presidents and two Democratic Presidents.

In some cases, recovering from big negative economic shocks has required long, steady, and reliable creation of new jobs. But in all cases, the performance of the US labor market demonstrates a dynamism and resilience that is the envy of the world.

A note on open source resources and replicating these data and images

All the source data and corresponding data visualizations from this article can be obtained and replicated by running the code locally on your personal computer by using the resources in the USempl_Streaks-2024-05 (https://github.com/OpenSourceEcon/USempl-Streaks-2024-05) GitHub repository. Instructions are found in the Jupyter notebook USemplStreaks.ipynb in the repository. One advantage of creating the images on your local machine is that the images created with the Python code have more dynamic functionality, including mouse hover-over full data description, muting and unmuting of particular series, zoom, and selection.

1. About me

2. Why Econosseur?

3. What this space will be

1. About me

I am Rick Evans, and I am an economist. Sounds like the beginning of a 12-step meeting. My professional website is here, a formal bio is here, and my CV is here. I am a macroeconomist who focuses on public finance questions using open source policy models. I love code and computation, math and theory, and real world policy problems. But I want to give a less formal, more focused intro to who I am by covering how I got here, what I am doing now, and where I am going. And I want to move through those topics in reverse order.

1.1. Where I am going

My goal for the medium run and long run is to help researchers, analysts, and policymakers in many countries and many local governments create, maintain, and use open source models for making their policies and related analyses more open, transparent, replicable, and collaborative. I also hope to be an active participant in the creation of those models and the analyses using them.

1.2. What I am doing now

I’ve spent the last two years at the Center for Growth and Opportunity (CGO) at Utah State University building an open source microsimulation model (FiscalSim-US) of US federal and state individual tax and benefit policies. That work has now transitioned over to my new position as Senior Economist at the Abundance Institute in Salt Lake City.

I published a number of state tax policy papers at the CGO in 2023, all using open source policy models.

• Jason DeBacker and Richard W. Evans, “What Business Tax Reforms Would Be Best for Tennessee?” Research in Focus, Center for Growth and Opportunity (Jan. 12, 2023). Uses open source Cost-of-Capital-Calculator model.

• Evans, Richard W., “What are the Effects of Utah Income Tax Rate Cuts?” Research in Focus, Center for Growth and Opportunity (Mar. 22, 2023). Uses open source FiscalSim-US model.

• Evans, Richard W., “What are the Effects of Missouri Corporate Income Tax Rate Cuts?” Research in Focus, Center for Growth and Opportunity (Mar. 22, 2023). Uses open source Cost-of-Capital-Calculator model.

And I have a large number of new federal and state policy pieces using our open source models that I have been waiting to publish under the banner of the new Abundance Institute. Many of those articles will be published through this Substack.

In addition to my current work with microsimulation modeling of federal and state, business and individual, tax and benefit policy, I am also one of the core maintainers of a number of country-specific open source macroeconomic models (see models of USA, South Africa, India, and Malaysia as well as the core engine of those models). I have been working on the US model since 2011, and the other country calibrations came about through work with the European Commission, World Bank, and United Nations.

I am also currently involved in a very cool project with Raiany Romanni of the Amaranth Foundation and my long-time collaborator, Jason DeBacker, at the University of South Carolina estimating the value of research and development in anti-aging drugs and biotechnology. We are doing this work using the OG-USA open source macroeconomic model with its rich demographics—including mortality, fertility, and immigration rates by age and time period—and calibrated lifetime productivity profiles. We are answering questions like, “What would be the long-run value of medical breakthrough’s that allow us to have abundant organ replacement options?” I will write more on this topic later this year.

1.3. How did I get here?

I spent the first 12 years of my post-graduate career in academia with (i) 8 years as an Assistant Professor in the Economics Department at Brigham Young University (BYU) in Provo, Utah and (ii) 4 years as Senior Lecturer and Associate Director (for my final year) of the Masters in Computational Social Science (MACSS) program at the University of Chicago. My teaching at BYU focused on advanced undergraduate economics, mathematics, and computation. And my teaching in the MACSS program at Chicago focused on macroeconomics, data science, and machine learning for graduate student economists.

Among many great experiences during those first 12 years of university work, there are two main activities that stand out to me and were most formative for what I am doing now and what I plan to do with the rest of my career. And both of these threads started at about the same the same time back in 2009-2012.

1.3.1. Student computational economics boot camps and mentored research

My first job after receiving my PhD in economics from the University of Texas at Austin in May 2008 was as an Assistant Professor at Brigham Young University. BYU’s Economics Department had only undergraduate students, so all of our research assistants and teaching assistants were advanced undergrads. But BYU also had a long history of being one of the top departments in the country for getting its students into top 20 economics PhD programs.

In the summer of 2009, I met an innovative Associate Professor in the BYU Math Department, Jeff Humpherys. He saw that I was a macroeconomist who loved math, theory, and computation. And Jeff needed someone to teach a dynamic programming readings course to his students enrolled in an applied math sequence that he was putting together. I don’t remember exactly why I chose to teach an extra readings course that was in addition to my economics department teaching load, but I know that I have always loved teaching dynamic programming. And I am fairly certain that I discerned something innovative and exciting in what Jeff was doing.

To make a long story short, Humpherys created a multidisciplinary applied math program that involved a summer boot camp of core material for students, followed by more specialized courses, and mentored research. I became a co-PI on the first National Science Foundation grant ($599,918 over 5 years) that Jeff led for building a new applied math curriculum, which would later become the basis for the BYU Applied Computational Mathematics (ACME) degree program.

The first big lesson and skill that I learned from working with Jeff Humpherys was how to create a powerful, difficult, and focused curriculum for students that had the primary aim of turning them into co-researchers, coauthors, collaborators. This is a different training model approach than is commonly observed in either academic or private sector settings. The difference with the training model I learned from Jeff Humpherys is that your goal is to turn your students into collaborators.

As Jeff’s outside-funded student mentoring program in the BYU math department turned into the ACME degree program, his funder asked if I wanted to create an economics focused student mentoring program and boot camp. In 2012, the BYU Macroeconomics and Computational Laboratory (MCL) was born, founded by me and Kerk Phillips, my great friend, colleague, and coauthor. During my remaining time at BYU from 2012 to 2016, we had three cohorts of students (about 15 students a year, 46 total) come through the BYU MCL program.

I left BYU in the summer of 2016 for a job as Senior Lecturer in a new program at the University of Chicago called the Masters in Computational Social Science (MACSS). I got to help create and build the curriculum for that program, which included rigorous computational, data science, and machine learning training. I tailored tracks for economics-focused students, but we also had students focusing in sociology, psychology, and political science. While at Chicago, I also had a three-year appointment as a Fellow in the Becker Friedman Institute (BFI). It was through BFI that I received a new grant in 2016 from the Charles Koch Foundation to expand and extend the work I had previously done in the BYU MCL and create a new organization at Chicago that gave world class computational economics training to the best PhD-bound students from around the world.

The Open Source Economics Laboratory (OSE Lab) was born in January 2017, and the first OSE Lab bootcamp was held Summer 2017. We had about 25 students come to the program every summer for three summers (2017, 2018, 2019). All the curriculum was taught be superstar guest lecturers, all the curriculum was open source (see 2017, 2018, and 2019), and all the students were assigned mentored research projects for the academic year after the boot camp.

This type of training with a focus on creating collaborators is the foundation of all of my current outreach, marketing, and training work—from meeting with state and federal legislators and researchers using the FiscalSim-US microsimulation model to giving trainings to tax policy officials in developing countries. I want more collaborators.

1.3.2. Python and open source development and workflow

It was back in 2009 that my friend in the BYU Math Department, Jeff Humpherys, also convinced me to make the investment to leave my old graduate school software platforms of MATLAB and STATA and replace them with Python. It took me a couple of years to fully commit, but that has ended up being one of the pivotal positive decisions of my career. Jeff was an early adopter of Python. But he saw that its open source availability, large network of users, and fast development in areas of scientific computing and high performance computing made it ideal for both training students and for collaborating on research.

By the time we started the BYU Macroeconomics and Computational Laboratory in 2012, Python was the computational software of all of our modules, training, and research. This was obviously the direction of the newly starting Applied Computational Mathematics Emphasis degree program at BYU. I still use their open source, open access Python labs in my trainings. And by 2016, when we established the Open Source Economics Laboratory at the University of Chicago, our Python-based boot camp curriculum was still at the frontier of a broader trend in economics.

What we couldn’t have predicted back in 2012, 2016, and 2019 was that Python’s dominance among machine learning software packages would explode in importance due to the emergence of generative AI. Now Python is not only the most widely used software in economics, high performance computing, and machine learning, but the funding behind its network of developers has exponentially expanded.

Python’s open source availability makes it accessible and affordable for any user with a browser, whether a student at a university, a policy analyst at a state office, or a government official in a developing country.

Another benefit of using open source Python as the software foundation for computational modeling is its compatibility with open source collaborative workflow. This form of collaboration is the only way I have every seen to effectively scale collaboration on a single project beyond two or three participants.

You probably remember in college or high school doing a group project in which one of the members of the group did most of the work (maybe that was you) and other members of the work did almost none of the work. The open source workflow overcomes many of the coordination failures that led to your high school project group being so dysfunctional.

The modern open source workflow was pioneered by the development of the Linux kernel (the core version of the Linux operating system). If you look at the contributors page of the GitHub repository for the Linux kernel, you’ll see it has over 15,000 individuals who have contributed to the source code. That is a huge group project. The attribution, heirarchical permissions, testing, and approvals, have given open source collaborative workflows the ability to unlock effective and scalable collaboration among large groups of participants.

Although still in its infancy, I think the open source collaborative workflow is increasingly used among researchers. This approach is not only valuable for coding projects, but also for more general applications such as writing a paper and crafing legislation. My shift away from proprietary software back in 2019 to open source Python set me on a course toward open source computation, collaborative workflows, and policy modeling.

In 2012, I was introduced to Matt Jensen who was at the American Enterprise Institute at the time and was starting an Open Source Policy Center (OSPC). Kerk Phillips and I were the first macroeconomists to join this program—followed soon after by Jason DeBacker—starting the to build what are now the open source OG-Core and OG-USA macroeconomic models.

It was through the creation of these code bases and working with developers and engineers from Continuum Analytics (now Anaconda.com) that we learned coding best practices and open source workflows. I found that the open source workflows were not only the most effective and efficient way to scale up the number of collaborators on a coding project, but also on research projects. Now I have many collaborators, from students at Universities (we even had a high school student contribute to one of our models), to researchers at large nongovernmental institutions, to state legislative and research staff, to senior maintainers in the Policy Simulation Library network.

2. Why Econosseur?

So I’ve decided to give Substack a try as my primary outlet for publishing my work. My goal is to write short and long articles here on the subjects of economic policy, open source modeling, dynamic data visualizations, economic theory, or any other topic that is professionally relevant.

I chose the name Econosseur for this Substack and called the title of this post “[Re]Introducing Econosseur” because that was the name of the blog that I started with Kerk Phillips and later Jason DeBacker back in October 2008 during the Great Recession. Our hope when we named the blog Econosseur back in 2008 was that the reader would naturally think of an economics connoisseur, one with “informed and discriminating taste” and “with expert knowledge or training” in economics (see American Heritage Dictionary, 5th edition).

The last post on that blog was in August 2015. The econosseur.com website is no longer under our ownership, but you can see the content by using the Wayback Machine’s internet archive of the site. Economics blogs that were going strong back when we started in 2008 and are still going include Greg Mankiw’s (2006), Alex Tabarrok’s and Tyler Cowen’s Marginal Revolution (2003), and James Hamilton’s and Menzie Chinn’s EconBrowser (2005).

The part of the old Econosseur blog that generated the most traffic was the Economic Jokes page. This was really just a side thing where we posted funny jokes we had heard. But I guess it ended up being the top result when anyone searched for economic jokes.

There are two old Econosseur articles of mine that I still remember today as being either good, insightful, meaningful, or particularly timely. The first was written on October 17, 2008, and was entitled “International Bailout Arms Race”. This article documented proposed a game theoretic lens through which to interpret the responses of all major developed countries to follow the US with big bailout packages. And similar to the old cold war arms race between the US and USSR, the US had more capacity to absorb the fiscal strain of the bailouts, thereby putting more strain on the countries who had to follow the US response. I noted the irony that a global recession caused by US mortgage-backed securities actually strengthened US global leadership and made the dollar the safe-haven asset among international currencies.

The other old Econosseur article of mine that stands out to me was written on June 30, 2009, and was entitled, “On Economic Debate—the Ad Hominem Index”. I will probably repost that article here, because I think it is especially salient in today’s more polarized environment. I was responding to the generally conservative versus liberal debates of the time between Greg Mankiw and Paul Krugman. I noted that Paul Krugman’s arguments often included ad hominem attacks, whereas Mankiw’s were focused on logical arguments. I compared Mankiw’s approach to that of Milton Friedman, who was said to have been courteous to his opponents in a debate. I also detailed why ad hominem attacks are likely an indicator of the weakness of someone’s argument. Mankiw had a nice little response post to my article, entitled, “This makes my day”.

3. What this space will be

I hope to post short and long articles on this space about economic policy, open source modeling, technical novelties, dynamic data visualizations, and any other professionally relevant commentary I choose. I am happy to interact in the comments of the articles, on X (@RickEcon), otherwise you can get my email from my website.

If you want these articles emailed to your inbox, please subscribe to this Substack by filling in your address in the “Subscribe” field below and clicking the button. My Substack is free to everyone, and no subscription is required.