Stocks for Thoughts · The Wage Bill

Books were never the market.

Big Tech signed a trillion dollars of rent. The market has priced twenty‑five trillion behind it. Only one bill is big enough for that — the one that pays you.

The accounting is dull and it matters, so: two minutes, plain English.

When you sign a lease, nothing happens. The liability only lands on your balance sheet the day the landlord hands over the keys — the day the thing is, in the rulebook's phrase, "available for use." Until then it lives in a footnote. Microsoft's $329.1 billion of leases start handing over keys somewhere between next year and 2033.

There's a second reason the gap looks bigger than it is. The trillion is undiscounted future cash, added up nominally across two decades. The $285 billion is today's money. Even if every one of those buildings opened tomorrow, the balance sheets would swell by a good deal less than a trillion. The gap is a calendar and a discount rate, not a lie.

Here is the part that is genuinely different, and it has nothing to do with the accounting.

Amy Hood, Microsoft's finance chief, gave the reassuring version out loud in July: "If the demand environment changes, you just slow down what is, in fact, the largest component and the driver of COGS." She's right about capital spending. You can stop buying chips. You can leave a hole in the ground.

On the same call, Microsoft told the market that from the new financial year it was stretching the assumed life of its data centres and office buildings from fifteen years to twenty‑five, and reclassifying more of its future data‑centre leases from finance leases to operating leases. Finance leases count as capital spending. Operating leases don't. Headline capital spending for the calendar year duly fell from the $190 billion guided in April to about $175 billion. Hood was explicit that the underlying investment expectation was unchanged; only the classification moved. The reported number got smaller. The ambition didn't.

I'm not calling that a trick. It's a defensible estimate, and the strongest evidence for that is Amazon, which went the other way last year — shortening the life of some of its servers because AI was ageing them faster, at a cost of about $1.0 billion of its own 2025 net income. Meta lengthened its estimate the same month, and picked up $2.6 billion. Two of the most sophisticated operators alive moved the same number in opposite directions in the same January, which tells you it's a real judgement call rather than a scheme.

But notice what the two moves do together. The flexible spending is the part shrinking in the headline. The inflexible spending is the part growing in the footnote.

Underneath the trillion sit three timelines, and none of them line up.

So: a twenty‑year obligation, holding six‑year equipment, sold against a customer book that turns over in a little over two. The building outlasts everything inside it, and every promise made to fill it, several times over. That isn't fraud. It's a duration bet, and somebody has to be right about the far end of it.

So who pays? Follow it forward.

Since ChatGPT launched, the S&P 500's total market value has gone from about $36 trillion to about $70 trillion. The best‑known decomposition of that — Lisa Shalett's at Morgan Stanley Wealth Management, from October last year — put roughly three‑quarters of the gain down to the Magnificent 7 and the AI infrastructure names alongside them. Nobody has published a refresh since, so call the AI bet $25 trillion of created market value, treat the second digit as decoration, and hold the whole figure loosely.

Twenty‑five trillion at a flat twenty times earnings corresponds to $1.25 trillion of annual net profit. That is a back‑of‑envelope multiple, not a valuation model — no growth, no discount rate, no reinvestment. It is meant to give the number a size, not a decimal place. For scale: every corporation in America, every industry, made $3.6 trillion after tax last year. The AI increment alone is priced to earn about a third of all the profit made by every company in the United States.

At a 30% net margin — software‑like, and generous for an industry this capital‑hungry — that profit needs about $4.2 trillion of revenue a year.

The entire world software market this year is $1.47 trillion. All software. Every seat, every country, every vendor.

So the price needs AI to earn nearly three times the whole software industry. That isn't a forecast of software sales. It's a size check: what the price needs, held up against the biggest thing it could plausibly be sold into. They aren't the same kind of number and I'm not pretending they are. But the size is the point, and it's where I stopped and went looking for the last time a market did this.

On 10 December 1999, Amazon closed at its high: about $37 billion. That year it sold $1.64 billion of goods — mostly books, by then also music, video and toys — and lost $720 million doing it.

Run exactly the same arithmetic. Thirty‑seven billion at twenty times needs $1.84 billion of net profit. At the margin real booksellers earned — Barnes & Noble kept a net 3.57 cents on the dollar — that means $51.6 billion of revenue.

The entire US bookstore market in 1999 was $14.17 billion. Every book Americans bought through every channel, textbooks and warehouse clubs included, came to $31.25 billion.

Barron's ran it on the cover that May under the headline Amazon.bomb, arguing the shares had a great deal further to fall. They did. Amazon closed at $5.97 on 28 September 2001, down 94%. Barron's was right on the price, to within about two years.

And completely wrong about the business. Amazon's revenue last year was $717 billion — more than fifty times the entire 1999 US bookstore market. E‑commerce went from 0.5% of American retail to 16.9%. The bulls' thesis wasn't too optimistic. It was too small.

Here's the part that gets left out when this is told as a bull story. A buyer at that December close waited nine years and ten months — on price, and Amazon has never paid a dividend, so that is the whole return — just to get back to even. Being right about the world and wrong about the price cost a decade.

Books were never the market. Retail was. Which is precisely what today's bulls say about software.

If AI isn't selling software, what is it selling? Every serious version of the bull case gives the same answer, and it isn't coy about it: labour.

Jensen Huang says it plainly. Human intelligence is somewhere between 55 and 65% of world GDP — "let's call it $50 trillion" — and AI might take ten trillion of it. Satya Nadella attaches no number but argues the business logic migrates out of software and into the agent tier, which is the same claim with the price removed. Worth saying: Huang's version is augmentation, not replacement. He has been consistent and public that the job‑destruction reading is the wrong one.

So take the frame seriously and size it properly.

American wages and salaries run at $13.4 trillion a year. Add the benefits and payroll taxes employers actually write cheques for and it's $16.2 trillion — about $102,000 a year for each of the 159 million people on a payroll. Apply the ILO's global labour income share — 52.7% — to world output and the whole planet's wage bill lands somewhere near $62 trillion.

Against that denominator, the $4.2 trillion the price requires is 6.7% of every wage paid on Earth. Which is — and I did not expect this when I started — less than half of what Nvidia's chief executive says out loud.

The three rungs below today's price are the ones with something behind them: money actually collected, the rate at which inventors have historically been paid, and the entire world software market. Not one of them clears nine trillion. Today's price is past all three, which is another way of saying the market has already stopped valuing this as a product and started valuing it as a claim on payroll.

It is a middling claim, though, not a wild one — uncomfortable, but well short of absurd. That is the honest position of the price, and it is not where either camp wants it to be.

Now the punches, because they're good ones.

William Nordhaus went and measured the thing everyone assumes. Across the whole post‑war American economy, 1948 to 2001, the people who invented things captured about 2.2% of the value their inventions created. The other 97.8% leaked away — to buyers, to workers, to consumers through lower prices, to competitors who copied it. Innovation has been, historically, a fabulous deal for everyone except the innovator's shareholders.

Now impose that rate on today's denominator. Take 2.2% of a $62 trillion wage bill and you get about $1.4 trillion of annual revenue — which, at the same 30% margin and the same 20× multiple as everything else here, supports roughly $8 trillion of value. The market has priced twenty‑five.

Be clear about what that sum is and isn't. Nordhaus measured a share of total social surplus, not a share of payroll; this is not his calculation, it's his historical capture rate borrowed and pointed at a different denominator. It is a deliberately hostile benchmark rather than a finding. But it's the only long‑run empirical anchor anybody has for the question that actually matters — how much of the value a technology creates its owners get to keep — and it points down.

Then there's diffusion, and this is the number I keep coming back to. Ramp watches the card and bill‑pay spending of seventy thousand American businesses. In June, the median business in that panel spent $10.66 per employee per month on AI. That's $128 a year, set against a wage of about a hundred thousand.

And Ramp's panel is not the American economy — it is corporate‑card customers, which skews young, digital and eager. Fifty‑five per cent of them bought something from an AI vendor in June, against the twenty‑one per cent of all US firms the Census finds. So this is the median of an enthusiastic sample. It spends $128.

For the price to work, that figure has to be roughly $6,900.

And now the counterpunch, because it's the best line the bulls have. In the top tenth of American firms, the typical one already spends $516 per employee per month. Annualise it: $6,194. That is ninety per cent of the number the whole edifice needs, being paid today, out of ordinary operating budgets, with no drama at all. So the bet isn't that anyone will ever pay it. The bet is that the median firm ends up where the top decile already is. Whether that takes five years or thirty is the entire trade — and that duration simply isn't observable from here, by me or by anyone selling you either side of it.

Two more things sit awkwardly. Around one in five American firms report using AI at all. The Census Bureau's fortnightly survey put it between 17 and 20 per cent across the first half of this year, with 20 to 23 per cent expecting to within six months. Rising steadily, which is the bull's point; still a long way from the four‑in‑five of the workforce the exposure studies talk about, which is the bear's. And one of the clearest measured labour effects anyone has found so far — Brynjolfsson's study of several million workers a month of payroll data — is a 6% employment decline for twenty‑two to twenty‑five‑year‑olds in the most exposed jobs, against a 6 to 9% rise for older workers in those same jobs. No visible effect on pay, at any age.

That last detail is the one that should bother a bull most. The adjustment is running through hires that never happen, not through wages that fall. A hire that never happens saves real money. It is also extremely hard for a vendor to send an invoice for.