Data

The US is Building a Lot More Data Centers Than Five Years Ago, But We Are Still Building More Warehouses

Jeremy Horpedahl1 Comment

Data centers seem to be popping up everywhere. And based on the value of current construction, the US is indeed building a lot more data centers than we were in 2020 or 2021, about four times as much data center construction (inflation adjusted).

But… did you know that we build a lot more good-old manufacturing than data centers? Almost four times as much in recent months. And that’s even after a decline in manufacturing construction over the past year and a half.

The US also builds about the same amount of warehouses and chemical plants as we do data centers. Data centers may exceed those two categories in a few years, but for now they are pretty similar.

Keep in mind that manufacturing and chemical facilities also use a lot of electricity and water, and have plenty of local negative externalities! Warehouses probably have a lot less resource consumption and external effects, but it’s not zero either.

Are data centers popping up everywhere? Well, people are certainly noticing them. But so are lots of other types of buildings, which rarely register more than a peep from concerned citizens and local media, unless there is some clear and obvious external effect.

Fuel Costs Are Way Up, But It’s Still Pretty Affordable to Fill Up Your Tank (relative to wages)

Jeremy HorpedahlLeave a comment

Two months ago I wrote about gasoline prices and tried to give the current prices some historical context. Gas prices have, of course, only continued to increase since then. Here’s a chart I created to give a bit more context, using an idea from Ryan Radia: how much does it cost to drive a car 250 miles? Since fuel efficiency has increased over time, we might be understating how much it costs to drive today relative to the past. And of course, to give the “cost” proper context I have stated in terms of hours worked at the average wage (note: the final data point is from April 2026, as we don’t have wage data for May yet):

In April 2026 it took about 1.4 hours of work at the average wage ($32.23) to purchase enough gasoline to drive 250 miles (10.7 gallons) at the average fuel efficiency (23.4 miles per gallon). That average fuel efficiency figure is from 2024, the latest available, so it could be a bit higher today. Maybe it’s a little easier than 1.4 hours of work to buy it, but even if fuel efficiency had crept up to 25 mpg (that would be a big increase in 2 years, historically speaking), it would still be 1.3 hours of work.

1.4 hours of work is certainly a big jump from earlier in 2026, but you’ll notice it is still on the low end in this chart, and well below the peak we saw in June 2022 of just over 2 hours of work to buy 250 miles worth of gasoline.

But 23.4 miles per gallon is pretty low, as this is includes lots of trucks and SUVs with pretty bad fuel efficiency. What if we looked at some more fuel efficient vehicles?

Here’s a few I checked on (all for 2026 models, with gas and electricity at current national averages):

  • Toyota Camry: 0.71 hours of work
  • Chrysler Pacifica Hybrid: 0.61 hours on electric, 1.18 hours on gasoline
  • Tesla Model Y: 0.37 hours of work

It will probably not surprise you that the all-electric Tesla Model Y is cheaper than the average car to operate at current prices, but you may not have realized that it is almost four times cheaper. But the Toyota Camry, with all models operating as hybrids now, also comes in pretty good at about half the cost of the average vehicle to operate (and the Camry is a very affordable car to purchase). The Chrysler Pacifica hybrid minivan does pretty well too, though even operating only on electricity (30 miles at a time), it’s only slightly more fuel efficient than the Camry.

arXiv will ban authors who submit papers with LLM mistakes

Joy Buchanan1 Comment

In the world of academic preprints, arXiv has long been the go-to platform for researchers to share work quickly. But with the explosion of generative AI tools, the repository is drawing a line in the sand.

On May 14, 2026, arXiv moderator Thomas Dietterich announced a clarified enforcement policy. If a submission contains incontrovertible evidence that authors didn’t properly check LLM-generated content, all listed authors face serious consequences.

What counts as “Incontrovertible Evidence”? The policy targets clear signs of unchecked AI output, including:

  • Hallucinated or fake references
  • Meta-comments left by the model (e.g., “Here is a 200-word summary; would you like me to make any changes?” or placeholder instructions like “fill in the real numbers from your experiments”)
  • Other obvious errors, plagiarized text, biased content, or misleading claims generated by AI

arXiv’s Code of Conduct already holds every author fully responsible for the entire paper’s contents.

The Penalty

  • One-year ban from submitting new papers to arXiv.
  • After the ban, future submissions must first be accepted at a reputable peer-reviewed venue before arXiv will host them.

At first researchers discussing the policy online seemed happy about the one-year ban, but when I pointed out that it is essentially a ban for life to use it at a pre-print venue, some people became nervous.

Why now? arXiv has been overwhelmed by low-effort “AI slop.” These papers are marked by fabricated citations and shallow summaries. This erodes trust in the entire preprint ecosystem.

In response to the complaints (someone like me would be worried that I’ll somehow let an error slip through and then be banned for life from posting working papers), Scientific Director Steinn Sigurðsson shared:

on the whole @arxiv flap about hallucinated references etc

you don’t see the stuff we reject… some of it is really really egregious

the decision to impose additional consequences is largely to throttle that stuff so n00bs and bad actors don’t trash us trying repeatedly

This is the problem that we face with every internet forum. A few bad actors ruin it for good people.

In 2022 I wrote Content moderation strategy

Elon Musk buying Twitter is the big news this week. He wants to enhance free speech on the site and, according to him, make it more open and fun. Some fans are hoping that he will make the content moderation and ban policy more transparent. Maybe that’s possible. 

If no one can be banned, then bad actors will bring the whole platform down. Inevitably, good people get caught in the net, and it’s devastating to be locked out of a platform where your peers are sharing.

However, if you want to be taken seriously by tech folk then ask for a system that is possible. A substantially better experience might be incompatible with the site being free to users.

Part of the problem that I don’t hear people talking about is that a free platform is not easily compatible with good customer service.

For some not-fake work and citations: Buchanan et al. (2024) provided early clear evidence that a mark of LLM-written work is fake citations. And, Buchanan and Hickman (2024) show that certain framings can prompt people to be more suspicious of AI-generated writing, such that they are pushed toward doing a fact-check before believing all claims.

Buchanan, Joy, and William Hickman. “Do people trust humans more than ChatGPT?.” Journal of Behavioral and Experimental Economics 112 (2024): 102239.

Buchanan, Joy, Stephen Hill, and Olga Shapoval. “ChatGPT hallucinates non-existent citations: Evidence from economics.” The American Economist 69.1 (2024): 80-87.

How do Income Tax Brackets Work?

Zachary Bartsch1 Comment

I was listening to an episode of The Deduction, a podcast by the Tax Foundation. As if that first sentence isn’t evident enough, I was reminded of how confusing taxes are – period. Even experts disagree and see grey areas. As I was listening, I thought “man, they need a graph”. So, here we are.

Income Tax Vocabulary

The money that you are paid by your employer is your gross income. Not all of it is taxable. You can deduct money from your gross income to get your taxable income. Most people subtract the ‘standard deduction’ from their gross income, which is how I’ll proceed in this post. Since the standard deduction for 2026 is $16,100 for a single earner, that means that your taxable income is $16,100 less than your gross income. By following a formula, one can calculate the amount of money that they must pay the government. These payments can be all at once, throughout the year, or even directly from your paycheck. The total that’s due to the government by April 15 is called the total tax liability. Finally, the money that the government doesn’t take, and that you get to keep, is called your net income. It’s your income net of taxes.

If you’ve had a job, then you are probably most familiar with your gross income, what your employer pays you, and your net income, what you get to take home. The steps in between might include some hand-waving.

Marginal Tax Rates

One of the most confusing pieces of the income tax code is marginal income taxes. Below are the brackets for 2026.

Marginal Tax rates work like this: Every dollar that you earn faces a tax rate. If your taxable income would be below zero, then you pay zero in taxes. But if your taxable income is $5k, then it gets taxed at a rate of 10%. That part should be pretty straightforward. But what if your taxable income is $15k? According to the table, you face a tax rate of 10% for dollars earned up to $12,400. That would be a tax liability of $1,240. But the remainder of your $15k in taxable income exists in the next tax bracket. That portion of your taxable income faces a tax rate of 12%. Sticking with the example, $2,600 is in the 12% tax bracket, so the tax liability for that portion of your taxable income is $312 (=$2.6k*0.12). Therefore, your total tax liability would be the sum of your tax liabilities across all applicable tax brackets: $1,552 (=$1,240+$312).

There are some features of marginal tax rates that are worth mentioning. Since the tax rates on the lower taxable income brackets don’t change, earning more gross income never reduces your net income unless the tax rate exceeds 100% (which it doesn’t here). So, when someone says that their taxable income is in the 35% tax rate bracket, they probably just mean that their last dollar earned is there. They’re only paying 35% on the taxable income that’s above $256,225. They’re not paying 35% of all earned dollars to the Internal Revenue Service (IRS).

Below is a graph that details the different marginal tax rates with shaded areas. The blue line is the average tax rate. It’s calculated by dividing the tax liability by the gross income. Even though one might earn an income that’s greater than $257k where the marginal tax rate is 35% or greater, the average tax rate remains lower, topping out at about 30% in this figure. The average tax rate is lower than an earner’s top marginal tax rate because the income in those lower brackets never disappears or get taxed at a higher rate.

Continue reading →

Most Published Research Findings Are Directionally Correct

James BaileyLeave a comment

As a new quick rule of thumb inspired by the Nature papers, you could do worse than “cut estimated effect sizes in half”. If a published paper says that a college degree raises wages 100%, then chances are the degree really does raise wages, but more like 40–50%. In 2005, John Ioannidis said that “most published research findings are false”. By 2026, we seem to have improved to “most published research findings are exaggerated.”

That’s the conclusion of my piece out today at Econlog: “Is Economics Finally Becoming Trustworthy?“

There’s plenty of both good and bad news for economics and the social sciences in both my piece and the Nature special issue it describes. It’s kind of like the Our World in Data motto:

In short, our attempt to replicate hundreds of papers showed that published social science results shouldn’t be trusted precisely today, but they seem to be getting more reliable over time, and they are much more reliable than chance. Economics and political science look the best, though we are still very far from perfect:

You can read the full piece here.

Gerrymandering Doesn’t Give an Obvious Edge to Either Party in the US House

Jeremy HorpedahlLeave a comment

Congressional districts must be redrawn after each US Census. In fact, that is one of the main functions the Census: to determine how many seats of the US House of Representatives that each state is allotted. A related function is to give states information about the distribution of the population in their state. Even if a state doesn’t gain or lose seats after a Census, the population in their state may have grown, shrank, or simply moved around within the state. If each Congressional district is to represent roughly the same number of people, district boundaries will still need to be redrawn even absent a change in the state’s total share of the US House seats.

That much is clear. However, given that historically and still largely today Congressional districts are drawn by state legislatures, there is a temptation and a real possibility that the party in power of a state legislature will draw boundaries in a way that benefits that party. There is nothing illegal about doing this as far as the federal Constitution is concerned (that I am aware of), but it does seem a bit unsporting. But I guess much of politics might be deemed “unsporting.”

Nonetheless, sometimes the shape of districts is so obviously weird and not representing an cohesive group of citizens or communities that it gets the derisive term “Gerrymander,” which derives from a historical example of a very odd looking district. But even if a district doesn’t look weird, it may still give one party an advantage that some deem unfair, such as by diluting one party’s supporters into multiple districts so they get no seats, or alternatively cramming all the supporters into one district so they have a very lopsided victory in just one district, rather than controlling multiple districts. This practice is known as “partisan Gerrymandering,” and it will be my focus in this post today (there are other forms, such as racial Gerrymandering, which are also important but are beyond the scope of this post).

Surely this practice occurs. Some states have tried to avoid it the problem of Gerrymandering by using non-partisan commissions, though this is a minority of states (less than a dozen), and when push-comes-to-shove they don’t actually seem that committed to the idea (both California and Virginia have essentially abandoned these commissions in 2025-26 to attempt to, once again, gain a partisan advantage). But lately a particular question has come up: does partisan Gerrymandering benefit one major party more?

In total for the US House, whatever Gerrymandering at the state level that is happening seems to roughly wash out in national representation: in the 2024 election, Republicans received about 51.7% of the two-party share of votes totaled over all House elections, and Republicans have about 50.6% of the seats in the House. Perhaps you could say that the GOP effectively loses 5 seats to what they “should” have in a truly proportional sense, but this ignores many factors, some of which I will discuss below. But even so, the GOP has a slim majority in the House and they won a slim total of national House votes. It’s about right.

But that “washing out” at the national level ignores some very large disparities at the state level. In some states, one party has all the House seats, even though they got nowhere near 100% of the House vote. Many of these are states with 1 or 2 House seats, which are less interesting because either there is no possibility of Gerrymandering (1 seat) or there is no obviously “fair” division, but it is not only those small states. For examples, Massachusetts gives all 9 seats to the Democrats, even though Republicans received 31.5% of the two-party vote share. Do Republicans deserve 3 of the seats? Is the fact that they don’t have 1/3 of the seats evidence of Gerrymandering? Conversely, in Oklahoma Republicans hold all 5 seats, even though Democrats got 30% of the vote. Should Democrats get a seat or two in Oklahoma?

(Note: for all vote data, I have queried Google Gemini Pro. I found multiple errors along the way, but I am fairly confident the numbers are all correct now. Please let me know if you spot any errors).

Neither Massachusetts nor Oklahoma’s Congressional representation is an obvious case of Gerrymandering on its face. It’s possible that 1/3 opposition party support in both states is perfectly even distributed across the state, such that it would not be possible to draw any “fair” districts that give the opposition roughly 1/3 of the seats. But it could be the result of Gerrymandering, or at least an indication we should look deeper. We can tally up all of the differences across states in the following chart:

Chart 1

Continue reading →

GDP Forecasts for the First Quarter of 2026

Jeremy Horpedahl1 Comment

Forecast models, betting markets, and surveys of experts all drastically overstated the actual growth of GDP in the last quarter of 2025. They were off in the initial release, which was just 1.4 percent, but this was even further revised down to 0.5 percent. All four of the sources I track were forecasting over well over 2 percent, with some over 3 percent.

Does that mean we shouldn’t trust the forecasts? Perhaps, but last quarter was largely pulled down by government spending cuts, which the models completed missed. You can see this very clearly in the Atlanta Fed GDPNow model. Perhaps they shouldn’t have been surprised by this drop in government spending, but that is where the major error was.

So what do these forecasts think about the first quarter data for 2026, which comes out tomorrow? The two best predictors historically, GDPNow (Atlanta Fed) and Kalshi, are pretty far apart on this one, over a percentage point difference, with GDPNow being the only forecast under 2 percent:

Price Level: Noise vs Signal

Zachary Bartsch3 Comments

My university recently hosted a guest speaker. Among their content, they included some nominal macroeconomic values from pre-2020, back in the era when inflation was very low. That roughly includes the years 2012-2019. Truly, inflation stayed below 2% through February of 2021, but I think that we can all agree that the economy was different in a few ways beginning in 2020.

I asked the speaker why not express the nominal values in real terms. They were emphatic that the low rates of inflation at the time implied that the signal-to-noise ratio was too low. Therefore, the ‘real’ inflation adjusted values would not be more precise because excessive noise would be introduced into the series during a period when not much deflating was necessary in the first place.

My answer to this is a firm ‘maybe’. It makes sense and it’s plausible (Jeremy has written about error and revisions in the past). We can think about the noise in price indices in a few ways.

1) It may be information is incomplete and becomes more complete as time passes. This sort of noise only exists in the short-run and is resolved as more information becomes available later in time. Revisions tend to happen each month for prior months, as well as each year for prior years. There are also big revisions after methodological, consumption weight, and data source changes.

2) Another type of noise is due to incomplete information that is never resolved. After all, the government statisticians can’t see literally all of the transactions. Those unobserved transactions will never make it into the official inflation measures and we’ll never get a perfect picture.  

3) Methodological artifacts may also include known biases. This type of noise doesn’t get corrected except after major changes to the series. If those changes never happen, then we just sort of live with imprecision. Luckily, so long as the bias is consistent, then percent change in the price indices will approximate the underlying true levels. However, if there are non-random biases in the percent change, then it can cause some trouble.

One way to get an idea for the amount of noise in the data is to observe the magnitude of revisions. Of course, this only helps us with the first type of noise above that eventually gets resolved with more information. It’s much harder to get a handle on the imprecision that is not identifiable. The Philadelphia Federal Reserve Bank provides an easy-to-use database that puts all of the archival and revised numbers for many macro series in a single place: the Real-Time Data Set (RTDS). It includes every historical PCE price index value for each publication month. Let’s limit our sample to the 21st century.

Continue reading →

Are Americans Thriving Under Trump? No, According to the Cost of Thriving Index

Jeremy HorpedahlLeave a comment

The Cost of Thriving Index from Oren Cass’s American Compass is an attempt to calculate how well US families are doing financially, but without using traditional inflation adjustments to income. Instead, Cass and crew have chosen 5 categories of goods and services, and tracked those over time relative to median earnings for men ages 25 and older (in the baseline model — it can also be applied to different categories of workers).

Scott Winship and I wrote a detailed critique of the COTI, which I summarized in a previous blog post. Our critique comes from several angles, including correcting several major errors in COTI, as well as arguing that standard inflation adjustments to median income are superior to this new approach.

Based on our critique, I don’t think COTI is a very good measure of how well US families are doing financially. But the COT Index still has many fans. And Cass seems to think Trump is in large part pursuing many policies that should help out US workers and families, such as Trump’s tariff policies. Thus, it will be useful to see if Trump’s policies are leading to American workers “thriving” in the first year of Trump’s presidency.

Unfortunately, even using Cass’s preferred approach, Americans don’t appear to be thriving under Trump.

Continue reading →

The United States Has A Progressive Tax System

Jeremy Horpedahl5 Comments

For Tax Day 2026, here are some estimates of how progressive the US tax system is, drawing primarily from published academic work. While there is disagreement about exactly how progressive the tax system is (and should be), these papers all agree that as income rises, average tax rates rise. These estimates attempt to include, as best as possible, all federal, state, and local taxes, and to take account of tax incidence.

From Auten and Splinter in the Journal of Political Economy:

Piketty, Saez, and Zucman in the Quarterly Journal of Economics (Figure IX):

And here is a chart that I created, which comes from the appendix data for PSZ (2018), which is roughly comparable to the Auten-Splinter chart above. Note that it isn’t perfectly comparable: the income groups on the x-axis aren’t exactly the same, and the latest year in PSZ is 2014 rather than 2019 (they do have estimates for later years in updates to the work, but I am trying to stick with the published academic work). But they are roughly comparable:

Auerbach, Kotlikoff, and Koehler in the Journal of Political Economy take the additional step of computing lifetime average tax rates, rather than for a single year, showing the US tax system is even more progressive when considered this way. Note: they also include the value of transfers, which makes these results not directly comparable to the papers above:

Finally, here are two estimates from think tanks that work on tax policy. Even though the Tax Foundation is considered more right-leaning and ITEP is considered more left-leaning, both agree that the overall US tax code is progressive.