Wages Catch Up, Billionaires Move Out

The problem: there's no clean ratio to read

A price-to-income ratio is one of the most basic checks on whether housing is overvalued. People buy homes out of income, so prices can't outrun earnings forever. Building the ratio is harder than it sounds, and McBride names the obstacles:

• Lag. Most income data come out long after the year they cover. The 2025 numbers arrived in August and October 2026.
• Which income? Average total income is skewed upward by a small number of very high earners. Medians avoid that problem but come from surveys with their own quirks.
• Indexes, not prices. The most-watched national price measure, the S&P CoreLogic Case-Shiller National index, is a repeat-sales index. It tracks price changes on the same homes over time and is expressed relative to a base period, not in dollars.

The fix is to divide the price index by an income series. The resulting ratio has no natural unit, such as "five years of income." It is useful only compared with its own history: its median, its peaks and its troughs.

• $72,800.33 in 2025, up 4.2% from $69,846.57 in 2024
• The fifth consecutive year of strong wage gains
• Up almost 31% over the past five years

McBride takes the mid-year value of the Case-Shiller national index (available since 1976) and divides it by NAWI divided by 1,000. Scaling the wage index down only keeps the ratio readable and doesn't change its shape. For 2026, the wage index doesn't exist yet, so he assumes a 4% increase.

Result: by this measure, house prices are about 7.5% below the 2022 peak but "still well above the median level."

Result: prices are about 11% below the 2022 peak and, again, well above the median historical ratio.

The two series differ in what they capture. NAWI is wages per worker. Household income includes every earner in a home and non-wage income such as benefits and investment income. Since 2022 they have also grown at different rates, so they give somewhat different distances from the peak. The free portion of the post doesn't break that difference down. The point is that two independent income measures agree on the direction and the rough size of the gap.

Why McBride points to 1978–82, not 2006

McBride notes that five-year wage growth near 31% matches the late 1970s and early 1980s. He calls this "another reason" to compare the current housing cycle with 1978–82 rather than with the 2000s bubble and bust.

The comparison matters because the two episodes resolved affordability in opposite ways:

• 2006–2011: Prices were inflated by loose lending, and the correction came through prices. The Case-Shiller national index fell roughly a quarter in nominal terms, with widespread foreclosures.
• 1978–82: Inflation was high and mortgage rates climbed above 18% in 1981. Sales slumped, but nominal prices mostly stalled instead of collapsing. Fast-rising incomes, plus inflation, did the work of bringing the ratio down.

Today's setup resembles the second case. Prices rose sharply in 2020–22, mortgage rates then jumped, and wages have been rising quickly. That combination favors a long stall over a crash.

What "purgatory" looks like in numbers

McBride's conclusion is that prices are "too high based on income levels," and that he expects them "to spend 7 years in purgatory." In practice that means a prolonged period of flat or slow nominal price growth while incomes rise underneath them.

An illustrative calculation, not from the article: suppose nominal prices stay flat and incomes grow 4% a year. Each year the ratio falls by about 4%. After seven years incomes are up about 32% (1.04⁷ ≈ 1.32), so the ratio is about 24% lower. Inflation-adjusted prices fall a lot, while the sticker price of a home barely moves. That is how affordability can improve without a crash.

Caveats

• The 2026 points are estimates. The 4% wage and 4.9% median-income assumptions affect the latest reading. Slower income growth would make the ratio look worse.
• Averages skew. NAWI is an average, so gains concentrated among high earners can flatter it relative to what typical buyers earn.
• Ratios ignore financing. A price-to-income ratio leaves out mortgage rates, which drive the monthly payment buyers actually face. The same ratio is far less affordable at 7% than at 3%.
• Paywall. Part of the post, including chart commentary, is for paid subscribers. The free section doesn't explain how McBride arrived at seven years specifically, so treat that figure as his judgment, not the output of a model.

The open question is whether incomes keep rising at the 4–5% pace these estimates assume. If they do, the gap closes slowly through income growth. If wage growth slows sharply, the stall either lasts longer or prices have to give.

What crowding out is supposed to look like

When a big new borrower arrives, the demand curve shifts out and the interest rate rises. That higher rate does two things:

• Some other projects no longer pay off at the higher rate. That lost investment is the crowding out.
• Households and firms save more, because saving now earns more, so the pool of funds gets bigger.

Because of the second effect, crowding out is "generically less than 1 to 1." Here is an illustration with made-up numbers: if tech firms borrow an extra $50 billion, the higher rate might draw out $20 billion of new saving. Other investment then falls by about $30 billion, not $50 billion.

Tabarrok adds that economists usually reserve "crowding out" for government borrowing. Private borrowers face a market test, so funds moving to them from lower-return uses is a gain, not a cost. He sets this point aside for the argument.

Problem one: the numbers don't add up

The chart that started the debate came from Steven Rattner, a financier and commentator. It showed data-center construction spending up $51 billion while other investment fell $120 billion. Under the model above, $51 billion of new spending can't crowd out more than $51 billion of other investment, let alone $120 billion. Something else has to be pushing other investment down.

Problem two: a different, much larger measure

Problem three: shares versus dollars

Shares of GDP can create a false picture. If tech investment's share goes up, other investment's share tends to go down. That can happen even when both are growing in dollars, as long as GDP itself is growing. Tabarrok redrew the same data in real (inflation-adjusted) dollars:

• From Q2 2023 to Q2 2026, real tech investment rose about 46%.
• Over the same period, real other investment rose about 3%.
• Other investment has weakened since its 2024 peak but is still above its starting level.

Problem four: interest rates moved at the wrong time

Today's real rates of roughly 2–3% are in line with historical averages. In his words, "if there are good projects they can be funded."

He offers several possible reasons for the earlier rise:

• recovery from unusually low pandemic-era rates
• federal government borrowing
• people saving less today because they expect AI to make them richer later

On the last point, he jokes about AI workers who have reportedly taken up smoking because they expect a future cure for lung cancer.

Real rates have also risen across advanced economies, even though most AI spending is in the US. Capital markets are global, and the world's pool of savings is much larger than America's. That makes it harder for US data-center borrowing alone to squeeze US investment.

His alternative explanation

What remains open

The post is short and leaves gaps. It doesn't publish data sources or definitions for its redrawn chart. It doesn't measure how much tariffs or expiring subsidies have actually cut investment. And it can't settle the counterfactual it raises itself. Crowding out can also happen through real resources rather than money: electricians, power-grid capacity, transformers and construction crews all have limited supply. A rate-based argument doesn't address those limits, and the post doesn't discuss them. Tabarrok's case is mainly that the evidence offered so far is weak. He doesn't claim to prove that no crowding out is happening.

Medicaid spends close to $1 trillion a year. About 65% comes from the federal government and the rest from the states. That is almost one in five dollars of all US health spending, a bit less than Medicare at about $1.2 trillion. Timothy Taylor's Conversable Economist blog summarizes a new analysis by James Hamlette and Karen Kopecky of the Federal Reserve Bank of Cleveland, "Who's Driving Medicaid's Enrollment Growth?" (October 13, 2026). Their answer to why the bill keeps rising is simple. Many more people are covered.

What Medicaid pays for

Medicaid is joint federal-state health insurance for eligible low-income people. Groups that can qualify include children, pregnant women, low-income elderly and disabled people, and adults in low-wage jobs. Federal rules require states to cover certain services:

• inpatient and outpatient hospital care
• physician services
• nursing home care
• mental health services
• preventive care for children

States may also choose to cover prescription drugs, dental care, and home- and community-based care. Because each state makes different choices on top of the federal minimum, "Medicaid" is really about 50 different programs.

Enrollment, not cost per person

Medicaid outlays per US resident were four times their 1990 level in 2024, citing the Congressional Research Service. (The post does not say whether that figure is adjusted for inflation.) The authors break that growth into two parts:

• Coverage: In 1990 about 10% of Americans were on Medicaid. The share is now approaching 30%.
• Cost per enrollee: Average spending per enrollee, adjusted for overall inflation, has barely moved since 1990.

This matters because health spending is usually described as a price problem: medical inflation, expensive new treatments. For Medicaid as a whole, the cost of covering each person has roughly kept pace with prices in general. The growth comes from how many people the program covers.

Who drove the growth, decade by decade

Hamlette and Kopecky find that the groups behind enrollment growth changed over time:

Enrollees and costs don't match up

Average cost per enrollee hides big differences between groups. Elderly and disabled people make up 17.3% of Medicaid enrollees but just over half of spending. They need nursing homes, long-term care and treatment for chronic conditions.

A rough calculation (ours, using the post's figures) shows the gap:

• Elderly and disabled: about 51% of spending divided by 17.3% of enrollees gives a ratio near 3.
• Everyone else: about 49% of spending divided by 82.7% of enrollees gives a ratio near 0.6.

So an elderly or disabled enrollee costs roughly five times as much as the average for everyone else.

Taylor applies this to current policy. The main proposals to hold down Medicaid costs would require recipients to work. Congress wrote work requirements for expansion adults into a 2025 budget law, set to take effect around the end of 2026.

Taylor argues the savings are likely to be limited, for two reasons:

1. Non-disabled, non-elderly adults, the group these rules target, account for only about 15.8% of Medicaid spending.
2. Many of them already work. Since 2014, working adults have driven most enrollment growth.

An illustration (our numbers, not the article's): suppose work requirements cut spending on this group by 10%. That would save about 1.6% of total Medicaid spending, around $16 billion on a roughly $1 trillion program. The big cost centers, elderly and disabled enrollees, would be untouched.

Where Taylor would look instead

He points to Mark Shepard and Jacob Wallace's article "Understanding Medicaid Managed Care: The Procured Competition Model," in the Spring 2026 Journal of Economic Perspectives. Taylor is the journal's managing editor, which readers should keep in mind. In the model they describe:

• The state runs a procurement to choose which insurers may take part.
• Those insurers then compete to enroll beneficiaries.
• The state controls who gets in, how the insurers compete, and how enrollees choose among them.

Shepard and Wallace say states set up these programs in very different ways, and there is little evidence on which designs work. They call this a promising area for research.

What's left open

The post is short and leaves several gaps:

• It does not say what the remaining third or so of spending goes to. Children are the likely answer, but that is an inference.
• It does not reconcile the fourfold rise in per-resident spending with the flat cost per enrollee. Partly this is a matter of nominal versus inflation-adjusted dollars, which the post doesn't spell out.
• It offers no estimate of what better managed-care design might save. The research it cites says the evidence doesn't exist yet.

The firm conclusion is narrower: Medicaid's cost is mostly about how many people it covers. Within that, a small group of elderly and disabled enrollees accounts for most of the spending.

The news

Tyler Cowen, who flagged the paper on Marginal Revolution, calls the issue "a good litmus test for who still thinks like an economist." The post itself is just the paper's abstract plus that line. The detail below comes from the abstract alone.

Why scoring a wealth tax is hard

A simple revenue estimate multiplies the tax rate by the base. Here the base is the total net worth of California's billionaires. At 5%, a $100 billion yield implies about $2 trillion of taxable wealth. That arithmetic assumes the base stays put, and three things can shrink it:

• Migration. Wealthy people can change their residence before the tax applies. A billionaire who leaves takes their whole fortune out of the base.
• Litigation. Wealth taxes on worldwide assets invite legal challenges. If part of the tax is struck down or tied up in court, some of the expected money never comes in.
• Spillovers to other taxes. A departed billionaire stops paying California income tax every year after leaving, not just once.

The third point drives the paper's main result. The wealth tax is collected once. A lost income tax payer is lost every year.

How the authors estimated it

Jaros and Rauh used three main sources:

1. The Forbes Billionaire List, to measure the size of the base.
2. News reports on billionaires' California homes, to judge who is really a resident.
3. Public announcements of billionaires leaving the state.

The income tax side

An illustrative example (these are not the paper's numbers). Suppose departing billionaires had paid $1.5 billion a year in income tax. Treated as a permanent loss and discounted at 4%, that stream is worth about $1.5B ÷ 0.04 = $37.5 billion today. That alone is more than the wealth tax's one-time $30 billion. A lower discount rate or larger departures widen the gap. This is the logic behind the paper's conclusion that lost income tax "more than offsets" the one-time collection.

How confident is the result?

The answer turns on two uncertain inputs: how much the wealth tax actually raises, and what discount rate to use. So the authors ran 100,000 simulations across plausible combinations of the two. In 85% of them, the net present value to the state is negative. By the authors' own reckoning, the tax still breaks even or better in about one case in seven.

What the abstract leaves open

• What the $38.9 billion counts. The abstract does not say whether that figure includes only billionaires who have already left, or also departures the authors expect later.
• Permanence. The result treats lost income tax as permanent. Some people come back, and fortunes change over time, so the length of the loss matters a lot.
• Who disagrees. The $100 billion estimate from Galle et al. is the main counterpoint. The post doesn't present their side or any response to this paper.
• Status. This is a working paper, so it has not been peer reviewed. Its departure data comes from public announcements, which may miss quiet moves or count moves that are partly for show.

The general lesson goes beyond California. A tax that people can avoid by moving should be judged on all the revenue that leaves with them, not on the size of the base before anyone reacts.

The claim

Europe is less of a surprise. Klement notes that its biggest data centre hubs, the UK and Spain, have paired data centre growth with more spending on renewables and batteries. The US was the place where gas was expected to win.

Why gas was the expected winner

A data centre runs around the clock, and it wants power that is always there. Gas plants can be switched up or down on demand. Solar and wind can't, since their output depends on the weather and the time of day. Policy also seemed to favour gas. Klement mentions the "best efforts of the current US government". That is a reference to the administration's support for fossil fuels and its rollback of support for clean energy, which included cuts to wind and solar tax credits in 2025.

What the study did

The authors did three things.

What they found

Klement concludes that a data centre arriving in your area raises local power demand, and that this demand "is more likely than not" to be met by renewables and batteries rather than gas. His explanation is simple: "money talks." New renewable capacity is cheaper to build than new gas plants, so developers build it whatever the political mood.

He ends with a jab: "talk to anyone in Texas, and you wouldn't know any of it." Texas is the US leader in both wind and solar, yet it is politically identified with oil and gas.

What the article doesn't tell you

The post is short. It gives no actual figures: no megawatts, no growth rates, no gap between data centre and non-data-centre areas. It only describes the direction shown in the charts. Several questions are worth keeping in mind. These are the editor's caveats, not Klement's:

• Capacity is not energy. The charts appear to count capacity, meaning a plant's maximum rated output. A solar farm typically runs at about a quarter of its rated output over a year. A gas plant can run far more of the time. So 1 gigawatt of new solar delivers much less electricity than 1 gigawatt of gas. "Most new capacity is renewable" does not mean "most new electricity is renewable."
• Correlation and site selection. Data centres went to places with cheap power. In Texas and Nevada, that cheap power is largely wind and solar. Some of the link may run from cheap renewables to data centres, not the other way round. The article doesn't say how the study handles this.
• Timing. New gas turbines face long manufacturing waits, so gas plants planned for data centres in 2024–25 may not show up in completed capacity until later. A study ending in 2025 could understate gas.
• What "area" means. The article doesn't say how regions are defined. That choice affects how cleanly the two groups can be compared.

Why it matters for investors

If the finding holds up, the economics are pulling the AI power buildout towards solar, wind and batteries, despite the politics. That matters for anyone weighing utilities, power developers or energy-transition funds against a plain "AI means gas" story. Still, the evidence here is three charts and a one-paragraph summary of a study. It is better read as a challenge to the conventional wisdom than as proof that the conventional wisdom is wrong.