Illiquid Millionaires and a Trillion-Dollar Data Center Bet

Maggiulli, who writes the blog Of Dollars and Data, argues that the headline number hides a basic question: what counts as being a millionaire? Someone with a $1 million home and $10,000 in checking meets the classic test, yet may not feel rich or be able to spend much.

Four definitions, four answers

He measures the population four ways:

Two illustrative households, not taken from the article, show how these definitions separate.

Nationally, removing home equity and retirement accounts cuts the number of millionaire households by more than half. On liquid net worth, the share falls from 21.5% to 6.1% of households, which implies roughly 8 million households. Only about 28% of millionaires by the classic measure are also millionaires by the liquid one. Maggiulli says there is nothing wrong with this, but it shows how much of this wealth is tied up in illiquid assets.

The trend, and what housing did to it

The share of millionaire households has risen under all four definitions, particularly in the last six years. The absolute jump in households above $1 million during that period is the most striking change in the data going back to 1989, in Maggiulli's view.

The detail he finds most telling is 2007 to 2010. Over those years, the share of millionaires by net worth fell while the share by net worth excluding the home rose. He reads this as evidence that an inflated asset class, in this case U.S. housing in 2007, can temporarily distort the count. Households near the line can cross it on paper because of a house price, then drop back when that price falls.

Millionaires aren't one group

What people own shifts as wealth grows, with more in businesses and stocks and less in home equity and retirement accounts:

Age explains a lot

Only 2% of households under 35 are millionaires, and 13% of households aged 35–44. Among households 55 and older, the figure is over 30%. The median millionaire household is 63, so half are older than that.

Maggiulli's advice follows from this: if you benchmark yourself against millionaires, make sure you are around their age. People with ten or more extra years of saving and compounding aren't a fair comparison. His analogy is measuring a kindergartner's strength against a teenager's.

What the numbers don't settle

The article doesn't name a correct definition. Maggiulli presents the data and leaves readers to choose. Each measure has limits. Net worth treats a house you live in like cash, while the liquid measure ignores retirement savings that may be a retiree's main source of income. None of the four, as described, adjusts for taxes owed when money leaves a pre-tax retirement account, or for local living costs.

Two other gaps are worth keeping in mind. The article doesn't separate how much of the rise since 1998 comes from higher asset prices, more saving or an older population, although the age figures suggest demographics could matter. The global comparison also mixes units: the U.S. figures count households, while the UBS ranking counts adults, so a couple sharing $1 million isn't the same as one adult holding it.

Klein says the level is less striking than the speed: much of the rise has come just since the end of 2024.

Software breaks the pattern. Business investment in software has kept growing at its pre-ChatGPT pace, which Klein calls intriguing. On these numbers, the boom is showing up in hardware and buildings rather than in company software budgets. The article's public text does not explain why.

Six companies carry much of the load

Working backward gives a clearer picture (our arithmetic from Klein's rounded multiples):

The current pace equals a little over 7% of the group's combined market value each year.

The $1 trillion and 2.5% of GDP figures should not be compared directly. The first is six companies' reported spending, which can include projects outside the U.S., while the second is a national measure of business spending by category. The available text does not reconcile them.

The spending the data misses

The filing describes several pieces of that plan:

• Sites: long-term leases and partnerships with third parties for real estate, power and facilities, with specialist operators managing the buildings.
• Equipment: buying some compute and infrastructure gear directly and running it in dedicated facilities built to Anthropic's specification.
• Energy: arrangements involving dedicated power generation, and possibly developing, operating or controlling power stations or other energy infrastructure.

Anthropic says its commitments are already "substantial" and that it expects capital spending "to increase significantly".

This matters for reading the data. If the labs move from renting to buying, company-level tracking of the listed six will capture a smaller part of the boom until the labs publish comparable accounts.

Where the money lands

Klein writes that the spending has flowed mostly to a handful of manufacturers that capture much of the value in the equipment inside data centers and connecting their server racks: Broadcom, Micron, Nvidia, Samsung Electronics, SK Hynix and TSMC.

Three of those six are based outside the U.S.: Samsung and SK Hynix in South Korea, and TSMC in Taiwan. That points to a caveat of our own. GDP subtracts imports, so spending on imported equipment does not add to U.S. output one for one. The 2.5% figure therefore shows the scale of the investment, not its net contribution to growth. The available text does not estimate how much of the spending leaks abroad.

What the available text leaves open

The public version stops just as the section on suppliers begins, under the heading "When Unlimited Demand Meets Finite Supply". The rest is behind a "Read more" cutoff. We can't report what Klein concludes about suppliers, or what he means by the "changing shape" in his title.

The available text also doesn't address whether the investment will earn an adequate return, how it is being financed, or how long it can continue. It establishes the size, speed and main buyers of the boom, but not whether the spending will pay off.

Tipping is large and growing, but there has been little evidence on what it does to workers and firms. A new working paper by economists Dylan Balla-Elliott and Isaac Norwich tries to fill that gap. Tyler Cowen highlighted its abstract on his blog, Marginal Revolution, and noted that Norwich is on this year's academic job market from the University of Chicago. Only the abstract is available here, so the details of the method and the precision of the estimates aren't covered below.

The problem

Tips behave like a commission

The first finding is a simple regularity: tips are a nearly constant share of revenue. So a worker's hourly pay rises and falls with the firm's sales, much as a salesperson's commission does.

Here is an illustration, using invented numbers. A server handles $50 of sales an hour, and customers tip 20%, or $10. If the server handles $60 an hour instead, tips rise to $12. The firm hasn't changed anything, yet serving more customers has earned the worker another $2 an hour.

The clawback

Firms that begin collecting tips don't let workers keep the whole gain. They claw back 23 to 33 cents of each tip dollar through lower base wages.

Continuing the illustration, suppose a shop pays a flat $18 an hour and then adds tipping that brings in $10 an hour. The clawback implies base pay falls by $2.30 to $3.30, to between $14.70 and $15.70. Total pay rises to between $24.70 and $25.70. The worker is still ahead by $6.70 to $7.70, not the full $10.

This means a tipped job's lower wage isn't straightforward evidence that workers are worse off, and it also means that a tip isn't purely a transfer to the worker.

Testing whether tips change effort

Four results follow from those shocks:

• Revenue: the same kind of shock raises revenue nearly twice as much at tipping firms as at non-tipping firms.
• Base wages and prices: neither responds. So price increases can't explain the larger revenue gain, and the extra pay reaches workers through tips, not raises.
• Hours: they rise by more at tipping firms.
• Output per hour: it also rises by more at tipping firms.

Their account runs as follows. Each worker becomes more productive, so the firm finds it worthwhile to expand, and it also recovers about a third of the tips through lower wages.

Why every firm doesn't tip

If tipping helps both sides, every firm might be expected to adopt it. The authors say the gain is smaller where customers tip less, which would explain why adoption stops short of universal. A weak tipping norm provides a weaker incentive for effort and less money to share.

The size of the gains

The headline welfare numbers come from a model, a stylized version of the economy tuned to match the demand-shock estimates, rather than being directly measured. In firms where customers are prone to tip, the model implies:

• Job value up about half: the abstract doesn't define this, but in labor economics it usually means how much a job is worth to a worker compared with the alternatives.
• Firm profit up about 30%.

The authors suggest these gains "may" explain why tipping has recently grown in many industries. That is a hedged claim, not a demonstrated cause.

What to keep in mind

The abstract leaves several questions open:

• Customers: the stated gains are for workers and firms. The abstract reports no estimate of what customers gain or lose, though they pay the tips. It also doesn't address tipping fatigue or social pressure.
• Scope: the large gains apply to firms where customers are prone to tip, so they don't extend to any business that adds a tip screen. Firms that chose to adopt tipping may also differ from those that didn't.
• Who shares in tips: the abstract doesn't break results out by role, such as servers versus kitchen staff, or by whether tips are pooled.
• Evidence and review: the sample's representativeness and the design behind the clawback estimate aren't described. Cowen's post also offers no view of his own on the paper and doesn't say whether it has been peer-reviewed.

The takeaway that holds up best is the distinction between the pieces. The clawback range and the demand-shock responses are empirical estimates. The size of the benefit to workers and firms rests on the model built on top of them.

Its headline result: across 330 standard lock-ups, the typical stock did not fall around the expiry date.

Why the date looks dangerous

The fear rests on simple supply and demand. Founders, employees and early investors may hold large stakes they have been unable to sell. When the restriction ends, those shares can reach the market all at once, and more supply should push the price down.

There is also historical support for some caution. Outside this study, a well-known paper by Field and Hanka (2001), covering IPOs from 1988 to 1997, found an average three-day abnormal return of about -1.5% around expiry, with volume up about 40%. That is an average from a different era, so it isn't directly comparable with the new median figure.

What the new numbers show

So more shares changed hands, but the typical price didn't suffer. Oliveira's explanation is that the date is printed in the prospectus months ahead, so anyone expecting selling can trade before it. Put another way, a known event may already be reflected in the price by the time it arrives. A rise in volume also doesn't by itself mean a price drop, because every seller needs a buyer.

The terms differ more than the calendar suggests

Six months remains the norm: 392 of the 440 prospectuses, or 89%, lock up insiders for 180 days or six months. The differences lie elsewhere.

Early release. Among the 43 IPOs with the most trading in their first month, 14 let some locked shares sell early. Among the 320 with the least trading, only 3 did. That is about a third versus about 1%. For a heavily traded new stock, then, the headline expiry date may not be the first day insiders can sell. The post doesn't describe the conditions for early sales.

What to do with this

Oliveira's lesson is to read the terms rather than watch the calendar. Two questions matter: who is actually locked up, and can any of them sell early? Both are answered in the prospectus.

The data doesn't show that the date is safe for any individual stock. Nearly half of them fell, and a median of 0.7% says nothing about how large the worst drops were. It is better read as evidence against a blanket rule of selling before expiry.

What the summary leaves open

The post is a brief summary of a longer linked study, and it omits several things a careful reader would want:

• The spread of outcomes: it gives no average return or range around the median.
• The longer term: a three-day window can't show whether prices drifted in the weeks before or after expiry. If traders sell ahead of the date, as Oliveira suggests, any weakness would fall outside that window.
• The sample: it doesn't explain how the 330 "standard" lock-ups were chosen from the 392 six-month deals. The 43 and 320 trading groups also leave 77 IPOs unmentioned.
• The period: the sample covers less than three years, a short span for judging a market pattern.
• Practical gains: a typical edge of 0.7% over three days isn't evidence of a profitable trade, and the post doesn't test one.

The part of his post available to non-subscribers is short. It covers the Apartment List data and only the headline of a second data source. The rest is paywalled, so this write-up cannot assess his full evidence.

How small the change is

The news is the direction, not the size. As a rough calculation, a 0.1% rise on $1,384 is about $1.40 a month, or roughly $17 a year. That amounts to flat rents. If overall consumer prices are rising faster than that, rents are still falling in inflation-adjusted terms, even at the new positive reading.

"Bottomed" therefore means the decline has stopped. It does not by itself imply a strong rebound, and the visible portion of the post makes no forecast of how quickly rents might rise.

What caused the slump

The mechanism is straightforward. Each finished building adds empty units all at once, and its owner needs to fill them to cover costs. Owners compete for a limited pool of tenants by holding down or cutting rents, and that pressure spreads to nearby older buildings.

Rents can also stay soft after construction peaks. Apartment List says new deliveries, meaning finished units reaching the market, have slowed considerably since 2024 but remain fairly robust by historic standards. Even at what it calls the tail end of the boom, the market had still been struggling to absorb the new inventory, that is, to fill it with tenants.

The data point that disagrees

The disagreement isn't necessarily a contradiction. Realtor.com's rent report has generally focused on listings in the largest metro areas, while Apartment List publishes a national measure, and the two use different samples and methods. Indexes built differently can cross zero at different times, especially when one sits at just 0.1%. A small move in either series could change the sign. The visible text doesn't reconcile the two, and it's fair to treat "likely" in McBride's headline as doing real work.

Why this matters beyond renters

That is general background rather than a claim made in the visible post, but it explains why economists watch this data. Falling or flat asking rents have been a reason to expect slower shelter inflation. If asking rents begin rising, that downward pressure would eventually weaken. At 0.1% growth, it would not yet add much upward pressure of its own.

For renters, the practical change would be gradual. A market with falling vacancies generally offers less negotiating room over time. For landlords and apartment investors, higher occupancy would help revenue before any large rent increases arrive. Neither the size nor the timing of these effects is established by the available data.

What remains open

Three questions can't be answered from the excerpt:

• Durability. One reading just above zero is a thin base, and deliveries are still robust by historic standards.
• Demand. The visible text explains the turn through supply and occupancy, with nothing on household formation, jobs or immigration, any of which could change how quickly units fill.
• Geography. A national median can hide large differences between metro areas, and the available text doesn't break the figure down.