A correct conclusion resting on broken premises
That combination is the entire result, and it is more interesting than either half alone.
The thesis is simple enough. Semiconductor output can double in a year. Substations cannot. So hardware arrives before power does, and the excess must either sit in a warehouse losing value or displace working machines that have not finished depreciating. Both cost money. Watch the substation queues, not just the wafer starts.
The shape of that argument is sound, and the data mostly supports where it lands. What does not survive is the evidence offered for it. The headline revenue figure is not the company’s. The arithmetic that turns two growth rates into a price decline is wrong by a factor of two. The claim that fabs are fast and grids are slow was true two years ago and is not true now. And the prediction that compute prices must fall is contradicted by the price of compute, which has risen by a third since its trough.
More usefully: the strongest evidence for the thesis is a number the argument never cites, and two of its conclusions are refuted by taking its own mechanism seriously.
Thirteen claims, scored
Every falsifiable assertion, and the test that decides it. Rhetoric is not listed — only things that could be wrong.
Verdicts are against primary sources: SEC filings, earnings-call transcripts, EIA and LBNL publications, and ISO/RTO reporting. Where a figure could not be verified it is marked as such rather than assumed.
The choice is not between two things
The argument says excess hardware takes one of two paths: sit stranded, or displace the installed base. Both carry a penalty, so the operator picks their poison.
That is a claim about what is reachable, which is exactly the kind of claim a model checker settles. I wrote the pipeline as a state machine and let TLC explore every operator strategy — every possible schedule of shipping, warehousing, racking and cannibalising — to see whether any of them escapes.
None does. But the outcome space has three regions, not two, and which one you are in is decided by a threshold in three quantities: what ships, what is energized, and what already occupies the envelope.
Regime C is the case the argument misses, and it is the one its own premise implies. There, cannibalising does not cure stranding — it adds a write-down to it. You pay twice.
Nine TLC configurations, each declaring its expected outcome before it runs. Three are canaries that must fail: a model checker that cannot report a violation has proved nothing. Three are reachability probes that separate a real result from a vacuous one — without them, “no strategy absorbs the programme cleanly” would hold trivially in regime C because nothing can be absorbed at all. An independent Python classifier reproduces all three regimes from separate code.
Growth does not subtract
The argument opens on a gap: units doubling against revenue growing seventy percent. Thirty points of daylight, read as evidence that prices per chip are collapsing.
Growth rates compose by multiplication, not subtraction. Revenue is units times price, so the implied price change is a ratio.
The arithmetic is the smallest of four problems, though. The two quotes come from different speakers, three weeks apart, covering different periods — one about “next year”, the other about a fiscal year ending in January 2028. The product basket was never defined; the chief executive said “chips”, and the reporter who filed the story flagged in the same article that the company also sells CPUs, switch silicon, optical networking and laptop parts.
And the gap already had an explanation, given on the same call the revenue figure came from:
“Customers’ forecasts point to our growth doubling next year. However… we expect to grow approximately 70% as we are supply-constrained.”
Chief Financial Officer, Q2 FY2027 earnings call, 26 August 2026
Supply, not price. Meanwhile the direction is wrong too: disclosed revenue per gigawatt of capacity went from $18bn to $25bn to $40bn across three generations, gross margin is guided up on “executed price increases”, and customers were notified of rises above fifteen percent.
The widely repeated $673 billion revenue figure is not the company’s. It is a news outlet’s arithmetic — an analyst consensus multiplied by 1.70. The string “673” appears zero times in the 8-K, the 10-Q and the corrected transcript, and neither filing contains any statement of that year’s revenue at all.
Both sides are slow now
The argument rests on an asymmetry: silicon is fast, the grid is slow. That asymmetry has closed, and the people who build the fast side say so on the record.
“It takes two to three years to build a new fab. No shortcuts. And it takes another one to two years to ramp it up.”
Chief Executive, TSMC — said three times on one earnings call, April 2026
Advanced packaging capacity did roughly double in 2024 and again in 2025, off a small base and into pre-built shells. Forward growth decays to +68%, then +54%, then +30%. The doubling era is behind, not ahead.
On the grid side the argument’s 36–48 month transformer figure is directionally right but high, and it rests on thinner evidence than its confidence implies. The best current survey figure is 29 months for power transformers and 33 for generator step-up units. Every published number traces back to one root: vendor interviews conducted over nine days in June 2023, and a single proprietary survey whose method and sample size are unpublished. There is no independent government measurement of transformer lead times at all.
The price went the other way
The economic prediction is that new capacity floods in against linearly growing demand, so rental prices must fall toward the cost of electricity. Prices bottomed in December 2025 and have risen by a third since.
Corroborated by the seller. On its own earnings call the chip vendor disclosed that H100 rental had “risen 20% year to date, while A100 cloud pricing is up nearly 15%.” The A100 launched in 2020. On the argument’s own schedule it should have lost half its competitiveness twice over. It is appreciating.
Take the power-scarcity mechanism seriously and it predicts exactly this. If energized megawatts are the binding constraint, a machine that already occupies an energized slot is priced off the scarcity of the slot, not the speed of the chip. Which is why a six-year-old part appreciates while newer ones ship — and why the cannibalisation path loses its premise.
The floor is worth naming precisely. It is not electricity and it is not physics. Full-cost recovery across four-, five- and six-year book lives computes to $1.43–$1.82 per GPU-hour; the independently observed market floor across two rental indices was $1.70–$2.15. The intervals overlap. The floor is the depreciation schedule — which means the same market price is above water on a six-year book and underwater on a four-year one.
Straight lines fall faster than markets
The strongest claim in the argument is about accounting, and it is correct: cloud operators stretched server lives from three or four years out to five or six, deferring billions of expense, and pulling hardware early forces that deferral to come due.
It is correct, it is quantified in filings — about $43.8 billion of operating income across five companies in one year, eleven percent of what they reported — and it has already happened.
“…changing the useful lives of a subset of our servers and networking equipment, effective January 1, 2025, from six years to five years… In 2024, we also determined, primarily in the fourth quarter, to retire early certain of our servers and networking equipment. We recorded approximately $920 million of accelerated depreciation… due to an increased pace of technology development, particularly in the area of artificial intelligence and machine learning.”
Amazon, FY2024 Form 10-K, Note 1
That is the mechanism, in the filing, in the company’s own words, two years before the argument was written. And then the conclusion inverts, for a reason that is pure arithmetic: book value falls in a straight line, resale value decays exponentially, and the two curves cross.
Two more things sit in the same filings. The $920 million charge is 0.53% of the gross fleet — a rounding error. And on the same day, the company extended heavy equipment from ten years to thirteen, worth about $0.9 billion to operating income, almost exactly cancelling the server hit in a different segment.
On 1 January 2025, on comparable hardware, citing the same industry conditions, one company shortened server life from six years to five and another extended it from five to five and a half. Book life is a management judgment, not a measurement of how long hardware lasts.
Caveats that travel with the resale curve: the H100 sample is ten executed sales over ninety days and the A100 fourteen; only the V100, at 122, is a real sample. The index’s cost basis is system-allocated rather than chip list price — re-basing roughly doubles the residuals, which strengthens rather than weakens the conclusion. And the same part number prints 48.4% in one form factor and 79.4% in another at identical age.
A queue is an option book
Interconnection queues are quoted as though they were demand. They are not. Filing is cheap and priority is scarce, so the same megawatt is counted in several places at once, and most of it never arrives.
One utility states both numbers in a single filing: 47 gigawatts of signed data-centre contracts, and its own forecast of 16.6 gigawatts of load. A realisation ratio of 0.35.
Little’s Law reconciles two statistics that look contradictory. With 53 GW completing and about 750 GW withdrawing against a 2,061 GW queue, mean residence across all departures is 2.85 years, while the measured median for projects that actually finish is 5.08. Both are right: withdrawals leave early and pull the mean down, so completers wait roughly 1.8× the average departure. Two independent statistics agreeing is the point of running the check.
How much hardware is actually coming
The grid debate is conducted in gigawatts and the chip debate in dollars, so the two are never checked against each other. Both convert.
Set that against what the debate actually quotes — a 230 gigawatt state queue, 47 gigawatts of contracts at one utility — and the mismatch is not evidence of a shortage. It is evidence that an option book is being read as an order book.
The argument is about the wrong variable
Having reconciled the demand side, the question becomes which remaining uncertainty actually decides the answer. That is not the same question as which one people argue about.
Measuring the interconnection queue more precisely cannot settle this, because the queue is not where the variance lives. One of the two dominant factors can at least be bounded from physics, and doing so produces the strongest support the thesis gets anywhere in this audit.
An enterprise rack draws eight to ten kilowatts. A current AI rack draws 132 to 142. Converting a rack slot therefore recovers between five and eight percent of what the new hardware needs; the rest has to come from somewhere new. The headline “ten to thirteen percent annual fleet turnover” is not a reuse pool, because conventional server energy is flat — those slots get refilled, not released.
New hardware cannot hide in old power. It is the best argument for the thesis, and the argument never makes it. Bounded at 0–23%, against the 10–45% a casual reading would allow.
There is a third limit on top of that, which this model did not originally carry. A hall’s feeders, busway and local substation are sized for the density the hall was built at, so freeing 140 kW across fourteen cabinets does not put 140 kW at one cabinet — and the distribution gear in between is its own multi-month procurement. It binds the same way the density argument does, which makes 0–23% a ceiling that is not approached rather than a range. No published figure exists for how much retrofit distribution permits; it returns the same confirmed negative as decommissioned megawatts, and so becomes a third unsourced quantity on top of the two above.
What the author said back
This audit was posted as a comment on the original article. Dean Lee replied within the hour, conceded both sourcing errors, and returned two things that changed this page.
“When energized power is the binding constraint, the economic rent migrates from the chip fabricator to the entity holding the energized interconnection queue. The 30 percent rebound in compute lease rates is simply the shadow price of megawatts asserting itself over unenergized silicon.”
Dean Lee, in reply, 18 September 2026
That is a better sentence than any in this report, and it exposes an error in it. C13 — “the return distribution splits” — was scored partly here, on the grounds that the market is not currently inside the loss band. Wrong test. C13 is a claim about dispersion, between holders of energized interconnection and holders of unenergized hardware, not about whether capital is being returned on average. The rent migration is observable, and the rising rental price is evidence for it.
So the same datum cuts both ways, because the two claims describe different moments of one distribution: it falsifies C12, that the average must fall, and confirms C13, that the variance widens. The scoreboard above reflects the correction.
The strongest correction to this audit came from the author of the thing being audited. Which is roughly the best outcome a piece of work like this can have.
So: does the grid bind?
Two operators in the United States publish actual meter readings for data-centre energization. Both are delivering roughly one gigawatt a year. National models imply seventeen.
Probably yes — and modestly. Not a collapse. About a fifth of shipments, in the median case, landing ahead of the megawatts to run them.
The honest caveat is that the single number carrying this verdict is the one with the least evidence behind it: the scaling from two published meter readings to a national figure. No source provides it. It is my assumption, and if it is wrong, this conclusion moves.
What nobody knows
Listed because the sensitivity analysis says these dominate, and because an audit that hides its gaps is an advertisement.