Infrastructure

Compute Doubles Faster Than Its Price Falls

Epoch AI publishes at least three doubling times for AI compute. The one in every headline measures something else, and the curve that reaches your invoice is moving about three times slower.

InfrastructureBy Servola Tech Desk2026-08-016 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Compute Doubles Faster Than Its Price Falls

Key takeaways

AI computing capacity is doubling every 7 months while the compute you get per dollar doubles only every 2.2 years, so supply is compounding roughly three times faster than affordability and an abundance story upstream reaches your budget as a price story.
The nine month rate now in every headline comes from a paper about the performance of 57 leading supercomputers, submitted in April 2025. Epoch's own chip capacity series, published this January, says seven months. They measure different objects.
Applied to the reported 20 million chips, the three published rates give roughly 107, 186 or 354 million by the end of 2028. Same starting count, same window, a spread of more than three to one, all from constants published by one institution.
Five hyperscalers already hold 71 percent of global AI compute, and Microsoft has told investors that about 25 billion of its 190 billion dollar 2026 capital budget is component price inflation rather than additional hardware.

Three doubling times are in circulation

On 1 August, Epoch AI updated the database it keeps of the world's AI data centres. It now tracks 12.8 million H100 equivalents across 75 sites, assembled from high resolution satellite imagery, construction permits and company statements, and the organisation estimates that this represents about 27 percent of the AI compute delivered worldwide. Two days earlier the New York Times had built a long feature on Epoch's work, describing an incoming deluge of computing power: roughly 20 million AI chips in data centres today, doubling about every nine months, roughly 200 million by the end of 2028.

The buildout is real and that reporting is careful. The difficulty is the constant. Epoch publishes at least three different doubling times for AI compute, they measure different objects, and the one that has entered general circulation is not the one that describes a worldwide count of chips.

The nine month figure is about 57 machines

The nine months comes from Trends in AI Supercomputers, a paper by Konstantin Pilz, James Sanders, Robi Rahman and Lennart Heim submitted in April 2025. It builds a dataset of 500 AI supercomputers from 2019 to 2025 and finds that computational performance doubled every nine months, while hardware acquisition cost and power needs both doubled every year. Epoch's shorter data insight narrows the sample to 57 leading systems, those ranking in the top ten most powerful when first operational, and states plainly that these clusters account for roughly 10 to 20 percent by performance of all AI chips produced up to January 2025. The nine months decomposes into 1.6 times more chips per year multiplied by 1.6 times more performance per chip.

So the figure describes the peak throughput of the largest machines, measured to a cut off in early 2025 and published before the current build cycle began. Epoch's capacity series, published on 9 January 2026, answers a different question and gives a different number: global computing capacity growing 3.3 times per year since 2022, with a 90 percent interval of 2.7 to 4.1, which is a doubling every seven months. That series carries its own caveat in plain language. Its data tracks chip sales, not deployments, and it is expressed in H100 equivalents rather than raw chip counts. Neither series counts chips in service. Epoch does not publish one that does.

The same projection, run three ways

From August 2026 to the end of 2028 is about 29 months. Take the reported 20 million as the starting count and apply each published constant in turn. At a nine month doubling the multiple is 9.3, which lands at roughly 186 million and is where the reported figure of about 200 million, or ten times current levels, comes from. At the seven month rate the multiple is 17.7, or roughly 354 million. And there is a third path: if total capacity grows 3.3 times a year while each new chip is about 1.6 times faster than the one before, which is Epoch's own decomposition, then the number of chips grows around 2.1 times a year, a doubling of roughly twelve months, which gives a multiple of 5.3 and about 107 million.

That arithmetic is ours, applied to Epoch's published constants; it is not Epoch's projection and the institution has not endorsed it. But the spread is the finding. Three rates from one research group, applied to one starting number over one window, produce answers that differ by more than three to one. None of the rates is wrong. They answer different questions, and the difference between them is invisible by the time a number reaches a slide. Any capacity assumption carried into a three year budget quietly inherits whichever constant the source happened to pick up.

Europe counts chips, Epoch counts equivalents

There is a European illustration of this unit problem sitting in the open right now. On 30 July the EuroHPC Joint Undertaking opened its tender for AI gigafactories, up to 10 billion euros of public money against a target of at least 20 billion more from private investors. Four smaller facilities must each hold at least 75,000 AI chips and can draw up to 500 million euros; three larger ones must each hold at least 100,000 chips and can draw up to 1 billion. Bids close on 12 November and awards are expected in early 2027.

Now set those thresholds beside Epoch's measurement of the current record holder, a single facility at about 1.1 million H100 equivalents. The two figures are not comparable, and that is exactly the point. The European tender is specified in chips. Epoch's benchmark is denominated in H100 equivalents, a normalised unit in which one modern accelerator counts for several. A programme written in chips and a benchmark written in equivalents will diverge further every year that per chip performance improves, which is the same effect that separates the seven month curve from the nine month one. When you are sizing capacity, the unit is not a footnote.

The curve that reaches your invoice

Epoch also publishes the number that actually decides your bill, and it attracts almost no attention. Computing performance per dollar is improving at 1.37 times a year, a doubling every 2.2 years, which is about 26 months. Set that against capacity doubling every seven. Supply is compounding somewhere between three and four times faster than price performance. Abundance upstream does not arrive downstream as cheapness. It arrives as a great deal more capacity at a unit price that is falling slowly, and the gap between those two curves is where a compute budget goes wrong.

The buyers' own disclosures describe the same gap from the other side. Amazon has guided to about 200 billion dollars of capital expenditure this year, Microsoft to roughly 190 billion, Alphabet to between 175 and 185 billion, and Meta to as much as 145 billion, citing higher component costs among the reasons. Microsoft told investors that roughly 25 billion dollars of its 190 is component price inflation rather than additional hardware. When Alphabet raised its 2026 capital budget alongside second quarter results, its shares fell 7 percent. Epoch puts five hyperscalers at 71 percent of global AI compute. The deluge is being funded by a small number of buyers who are themselves watching unit costs rise, and it lands inside their own fleets first.

Frequently asked questions

Is the nine month doubling figure wrong?

No. It is correct for what it measures, which is the peak computational performance of leading AI supercomputers in a dataset built to an early 2025 cut off. The error is in transferring it to a different quantity. Applying a cluster performance growth rate to a worldwide count of chips assumes that chip counts and cluster throughput grow at the same speed, and Epoch's own decomposition of that figure says they do not.

How many AI chips are actually in service worldwide?

Nobody publishes a verified count. Epoch's capacity series is built from quarterly chip sales in H100 equivalents and explicitly notes that sales are not deployments. Its data centre database, updated on 1 August 2026, holds 12.8 million H100 equivalents across 75 sites and estimates that as roughly 27 percent of delivered compute, with an accuracy target of 80 percent within a factor of 1.5. Figures of about 20 million chips are reasonable estimates, not measurements.

What should we actually plan against?

The cost curve, not the capacity curve. Compute performance per dollar doubles roughly every 26 months, and that is the rate at which a fixed budget buys more work. Capacity doubling every seven months tells you what exists in the world, most of which sits inside five companies' fleets. Model your own three year spend on the 1.37 times per year improvement and treat anything better as upside.

The right question to put to any vendor quoting a compute abundance number is which series it came from and what that series counts. Capacity in H100 equivalents, cluster performance, chips in service and compute per dollar are four different curves with doubling times running from seven months to twenty-six, and they are routinely quoted as though they were one. The chips are genuinely coming. Budget against the slowest of those curves, because that is the one with your name on it.

Epoch AIAI computedata centresNvidiahyperscaler capexEuroHPCcapacity planning

Cadence And Synopsys Tuned Their Tools For Nvidia

Nvidia says its own Vera CPU runs Cadence and Synopsys verification up to 1.5x faster. The number is Nvidia's, and the design farm question is now open.

5 min read

Infrastructure

AMD Handed Half of Inference to a Rival Chip

On July 23 AMD and Cerebras built one inference pipeline from two rival chips: AMD Helios handles the prompt, the Cerebras wafer writes the answer. The 5x efficiency figure is a model, not a benchmark. Here is what disaggregated inference changes for your compute budget.

3 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.