Two hundred eighty billion dollars divided by 2.2 million chips is about $127,000 per accelerator.
No H100 costs that. No B200 costs that. Nothing Nvidia has ever shipped costs that.
That number comes from the Guardian's investigation into Microsoft's internal documents, which landed Monday and reported roughly 2.2 million AI chips installed across Microsoft's global fleet after about $280 billion in spending since 2022 — including more than $41 billion in the past quarter alone.
Do the division and you're left with a mystery that isn't actually a mystery. The chips are a minority of the spend. According to the Guardian's investigation, the rest went into land, buildings, substations, cooling plant, fiber, backup power, and networking — the slow physical layer every chip has to sit inside before it does a single calculation.
Which is the tell. This industry has spent three years counting chips because chips have names and launch dates. But it reportedly has been selling gigawatts — capacity — that it hasn't energized.
Last week in this space, we traced the leverage lane: the SPVs, the lease vehicles, Goldman's $1.5 trillion in combined data-center lease commitments with a full $1 trillion still "uncommenced." That was a single-company, balance-sheet story.
This week the question widens. It's no longer just how the buildout is financed. It's whether the capacity being financed exists at all — and the ECB just put a formal bubble flag on the answer.
Start with the arithmetic problem, because it forces the right question. The Guardian's leak gives us two more data points that sharpen it. An internal Microsoft target reportedly called for 1.8 million chips installed by the end of 2024 — meaning the fleet reportedly grew by only about 400,000 in the eighteen months since. And Blackwell is thin: fewer than 500,000 accelerators installed against roughly a million anticipated.
Microsoft's response was that the calculations are inaccurate and draw wrong conclusions from incorrect assumptions. Note what it did not say: which figures are wrong. Disputing a number without correcting it is not the same as the number being wrong.
So set the dispute aside and look at where the money physically went. The binding constraint on this buildout was never silicon. It was civil engineering.
A fab responds to demand in quarters. A substation, an interconnection agreement, a commissioned building shell with power and cooling actually turned on — those respond in years, and they answer to regulators and utilities, not to purchase orders.
That inverts three years of conventional wisdom. The story was always that accelerators are scarce and everything else is a formality. The operational reality is closer to the opposite: chips sit in inventory waiting for a warm shell to plug into. Satya Nadella has said as much on the record — the biggest issue isn't a compute glut, it's power.
The binding constraint on this buildout was never silicon. It was civil engineering. Chips sit in inventory waiting for a warm shell to plug into.
Here's where it gets dangerous for anyone holding these stocks. "Capacity" is not one number. It's five.
Announced capacity. Contracted capacity. Built capacity. Energized capacity. Commissioned IT capacity. Those figures can differ by a factor of several, and press releases almost never specify which one they mean.
Now watch what happens when you hold Microsoft's public claims against its regulatory filings. Microsoft says it added roughly five gigawatts of capacity in two years, and a 2024 internal presentation reportedly implied five gigawatts already installed — which would put the total near ten gigawatts today.
Its sustainability reporting tells a different story. According to the Guardian, those filings point to something closer to 1.2 gigawatts of AI capacity in 2024. Not five. Not ten.
Professor Shaolei Ren, quoted in the piece, argues the sustainability filings are the more credible source — because they are "prepared for regulators rather than for a keynote." If that's right, the public capacity claims are overstated by a factor of four or five.
That is the fork. Either "capacity" is loose vocabulary and everyone privately knows which number they mean — or the number being broadcast to markets is four to five times the number filed with regulators.
Hold that fork in mind and look at what Dell'Oro did on August 18.
The firm nearly doubled its 2030 data-center capex outlook to more than $3 trillion. The stated driver, from VP of Research Baron Fung: "higher hyperscale capex guidance, increased projections for global data center power capacity, and higher commodity costs."
Read that first clause again. Higher hyperscale capex guidance. The input to a $3 trillion forecast is the number the hyperscalers announced — the same category of self-reported figure the Guardian's leak just put in tension with regulatory filings.
Dell'Oro isn't auditing installed substations. It's extrapolating guidance. And its own caveat gives the game away: growth "will depend on the sustainability of investment, power availability, and supply chain conditions." Power availability is the civil-engineering bottleneck, listed as a footnote to a number built on the assumption it gets solved.
This is where the institutional bubble flags stop being mood and start being about measurement.
On August 17, ECB economists wrote that "economic research on past technological revolutions points to a worrisome conclusion: a correction of current stock market valuations is likely." Not possible. Likely.
Their frame is the one that matters here: valuations can correct sharply even if the technology succeeds spectacularly. Cisco was right about the internet. It still took eighteen years to recover its 2000 high.
The backdrop is a Shiller CAPE around 41 — the second-highest reading in modern history, surpassed only by the dot-com peak of 44-45. And roughly €440 billion in euro-area exposure to US tech means, in the ECB's words, a US correction "would not remain a US problem."
Then Cramer, on Thursday, reportedly from the construction site of Micron's fab in Boise: "There's an incredibly jarring gulf between stock prices and reality." He was talking macro — oil, the 30-year yield above 5.33%, Walmart down 9%. But the phrase describes the capacity problem exactly.
A gulf between what's said and what's real is only a mood until you can name the metric behind it. Now you can.
The metric is capacity. A CAPE-41 valuation, second-highest ever, rests on capex forecasts, and those forecasts rest on hyperscaler capacity guidance — a self-reported number that at least one company's regulatory filings appear to contradict by four or five times.
No one has audited the input. Dell'Oro extrapolated it. The ECB is flagging what happens when the extrapolation unwinds.
There's one adjacency worth naming and then leaving alone. The same civil-engineering bottleneck that strands chips in inventory is precisely what makes the lease-and-SPV financing we covered last week look circular — you're financing energized gigawatts against shells that answer to utilities, not quarters. That's a different lane. But it's the same crack, seen from the other side.
For 2026, the framing to carry is this. The consensus risk everyone repeats is chip scarcity. The Guardian's leak suggests the opposite: chips are arriving faster than the substations to power them, and "capacity" is being counted in press releases before it's energized in the field.
If you own these stocks, the number to distrust isn't the P/E. It's the gigawatt. Ask which of the five it is — and whether it shows up in a keynote or a regulatory filing. When those two disagree by a factor of four, the market is financing the keynote.