COMPUTE ATLAS Supercomputer supply-chain graph
1153 systems 482 sites Sourced data

Blog/2026-10-07

· Compute Atlas

Anthropic and Google TPUs: stated vs modelled

Google and Anthropic say Anthropic can use up to one million Ironwood TPUs and well over 1 GW in 2026. No site, chip count or power figure has been disclosed, and the campus-level numbers in circulation are Epoch AI estimates.

In October 2025 Google Cloud and Anthropic each announced that Anthropic would gain access to up to one million of Google’s seventh-generation TPUs, with well over a gigawatt of capacity arriving in 2026. That is the whole of the primary-source record on the one-million figure. Neither company has said where the chips sit, how many are installed, how much power they draw, or what Anthropic pays. The only campus-level figures for Google’s TPU fleet come from Epoch AI, which models them from satellite imagery and permits and does not claim to know which customer uses which building. This article separates the three kinds of evidence (stated, reported, estimated) and shows how far apart they are. Our entry for the commitment is google-tpu-anthropic-2025, and it deliberately carries no installed-chip count.

How the commitment came about

Google Cloud’s press release of 23 October 2025 describes access to up to one million TPUs, “tens of billions of dollars” of value and well over a gigawatt of capacity coming online in 2026. Anthropic’s own announcement of the same day repeats the figures and places them inside a three-platform strategy: Google TPUs, Amazon Trainium (Amazon remains the primary training partner, via Project Rainier) and NVIDIA GPUs. The two documents agree with each other, which is why we mark the commitment verified. They agree on a ceiling, though. “Up to one million” is a limit on what Anthropic may draw, not a delivery count, and “tens of billions” is a range rather than a price.

The commitment has since been stacked on top of. On 6 April 2026 Anthropic announced a further agreement with Google and Broadcom for “multiple gigawatts” of next-generation TPU capacity that it expects to come online starting in 2027, with the vast majority sited in the United States. Broadcom’s securities filing, as relayed by The Register, puts the Anthropic portion accessed through Broadcom at about 3.5 gigawatts from 2027 and says Anthropic’s consumption depends on its continued commercial success. That qualifier matters for anyone reading the headline figure as a purchase: the filing describes capacity Anthropic can access, conditional on its business.

On 24 April 2026 TechCrunch reported, citing Bloomberg and Anthropic, that Google plans to invest up to $40 billion in Anthropic: $10 billion at once at a $350 billion valuation and $30 billion more if performance targets are met. The same article says Google Cloud will supply a further 5 gigawatts over five years. We have not found a Google or Anthropic primary page for the 5 GW figure, so we treat it as reported, not stated. We also did not find a primary page on the term, price or structure of the underlying TPU contract, and nothing we fetched says whether Anthropic buys chips, rents capacity by the hour or commits to minimum spend.

What the hardware is

The seventh-generation part is called Ironwood, and Google’s documentation names it TPU7x. According to Google’s TPU7x documentation, each chip peaks at 4,614 TFLOPs in FP8 and 2,307 TFLOPs in BF16, with 192 GB of high-bandwidth memory at about 7.4 TB/s, built as two chiplets. The page lists a pod footprint of 9,216 chips in the text (its table shows a different figure, which we read as a typo), and Google’s Ironwood launch post of 6 November 2025 gives the same 9,216-chip superpod with 1.77 petabytes of shared HBM and optical circuit switching that reroutes around failed links. The post quotes Anthropic’s head of compute on Ironwood’s inference and training gains, and shows liquid-cooled racks with third-generation cooling distribution units. Google claims 10 times the peak performance of TPU v5p and more than 4 times the per-chip performance of TPU v6e, which are Google’s own comparisons.

One million chips is therefore roughly 108 pods of 9,216 (our arithmetic). Google documents the pod as a product configuration and does not say where any pod is installed, which is why our TPU7x pod entry is a documented configuration and not an installation. No per-chip power figure for Ironwood appears in the pages we fetched, so we cannot convert one million chips into megawatts from primary data.

Where the capacity might physically be

Epoch AI’s AI data centers dataset is the only public attempt to place Google’s TPUs. Its methodology page says IT power is derived from cooling equipment visible in satellite imagery, permits and company statements, and compute is usually calculated from IT power using the efficiency of the chips likely deployed. Epoch states roughly 80 percent confidence that power estimates fall within a factor of 1.4, with cost within 1.6 and timing within six months, and says the dataset covers an estimated 43 percent of global deployed AI compute (90 percent interval 23 to 81 percent).

We track four Google campuses from that dataset. Every figure below is Epoch’s estimate as of early October 2026, not a Google disclosure.

CampusEpoch chip estimateEpoch IT powerGoogle TTM PUE (Q2 2026)
Pryor North, OK23.5k v5p, 77.9k v5e368 MW1.11 (Mayes County)
Columbus, OH193.9k v5e, 28.2k v5p, 204.6k v6e, 38.9k v7303 MW1.06
New Albany, OH247.0k v5e, 66.6k v5p, 262.7k v6e333 MW1.07
Council Bluffs East, IA84.2k v5e, 125.2k v6e, 54.3k v5p, 41.9k v7237 MW1.11 (Council Bluffs)

The individual Epoch pages for Pryor North, Columbus, New Albany and Council Bluffs East give the building-by-building timelines behind these totals. Columbus, for example, shows its first building operational in April 2024 and a fourth in May 2026, with a fifth projected for July 2027. New Albany reaches its fourth building in October 2025 and 333 MW. Council Bluffs East reaches 237 MW with a third building on 7 July 2026.

Three points limit what these tables show. First, none of the four Epoch pages mentions Anthropic. The Pryor page lists the user as Google DeepMind and marks that “speculative.” These are Google’s campuses, filled mostly with older generations (v5e, v5p, v6e) that Google runs for its own services and Cloud customers. Nothing in our sources ties any of them to the Anthropic commitment. Second, Ironwood is a small share of what Epoch sees. Summing the two campuses with v7 chips gives about 80,800 (38.9k plus 41.9k, our sum), against a ceiling of one million, and Epoch’s own coverage is partial. A reader should not conclude that Anthropic has 80,800 Ironwood chips; the figure is only the amount of v7 that Epoch believes sits at two of the campuses it tracks. Third, the four campuses together add up to about 1,241 MW of estimated IT power and 1.45 million chips of all generations (our sums). That is close to the “well over a gigawatt” headline, and the closeness is a coincidence of scale: the headline refers to capacity for Anthropic coming online during 2026, and the Epoch total is a whole-fleet stock of mixed-vintage chips, much of it built before the deal was announced.

Power and cooling

Google publishes a power usage effectiveness (PUE) table by campus on its efficiency page. PUE is the ratio of total facility energy to IT energy, and Google’s version is a trailing twelve-month figure measured during stable operations that, per the page, includes overhead such as cooling, power distribution losses, substations, transformers and water treatment. For the second quarter of 2026 the fleet figure is 1.10 for the quarter and 1.09 on a trailing-year basis. The campus values in the table above are the trailing-twelve-month numbers: Mayes County 1.11, Columbus 1.06, New Albany 1.07 and Council Bluffs 1.11. The page gives one line for Council Bluffs, Iowa, and does not say which campus it covers, so we show it against the East campus only as a proxy and do not record it on our system entry.

Two cautions apply. A campus PUE covers every load on the campus, not the Anthropic share, and it says nothing about water. Epoch’s pages count 42 to 70 evaporative cooling towers per campus (64 at Pryor, 70 at New Albany, 42 at Council Bluffs East), which implies substantial water use even with low PUE, but Epoch gives no water figure and neither does Google on the page we fetched. A low PUE from evaporative cooling trades electricity for water. Ironwood itself, per Google, is liquid cooled, so the racks added for it use a different thermal path from the air-cooled generation preceding it. How a given campus mixes the two is not disclosed.

The financing backstops

A second layer of capacity for Anthropic is being built by neoclouds and former bitcoin miners, with Google’s balance sheet behind the leases. This is where the record is most concrete and also where the chip type is least clear.

Fluidstack, an AI cloud operator, is the lessee at three sites. At Cipher Mining’s Barber Lake in Texas, a Cipher exhibit filed with the SEC describes a 10-year lease for 168 MW of critical IT load (244 MW gross) worth about $3 billion, with Google backstopping $1.4 billion of Fluidstack’s obligations and receiving warrants that equal roughly 5.4 percent pro forma ownership of Cipher. At TeraWulf’s Abernathy campus in Texas, a TeraWulf release of 28 October 2025 describes a 25-year, roughly $9.5 billion commitment for 168 MW with a $1.3 billion Google backstop and, unlike Cipher, states that no TeraWulf equity or warrants were issued. TeraWulf’s second-quarter 2026 results say it agreed to sell its 50.1 percent stake in that joint venture for about $530 million, and that completing a Lake Mariner building made $600 million of Google credit support for Fluidstack’s lease there effective. At Hut 8’s River Bend in Louisiana, Hut 8’s release describes a 15-year, $7.0 billion lease for 245 MW with a Google financial backstop for the full base term and a first data hall due in the second quarter of 2027.

The link from Fluidstack to Anthropic is documented in two places. Anthropic’s 12 November 2025 post announces a $50 billion investment in US data centers built with Fluidstack in Texas and New York, with sites coming online through 2026, and names no chips. Hut 8’s December 2025 release describes Anthropic collaborating with Hut 8 while Fluidstack operates the clusters, with expansion options that could take the arrangement to as much as 2,295 MW of IT capacity. Separately, TeraWulf’s 6 July 2026 filing shows Anthropic signing a direct 20-year lease for about 401 MW at Hawesville, Kentucky, worth about $19 billion over the initial term, with initial capacity in the second half of 2027 and the rest by early 2028 (our entry: terawulf-justified-anthropic). The filing names no guarantor and no hardware.

The structure is worth stating plainly. Google guarantees part of the lease payments of a company (Fluidstack) that builds clusters for a customer (Anthropic) that Google also supplies with TPUs and has invested in. Google’s guarantee lowers the borrowing cost for the site developer. It does not by itself say the racks hold TPUs; in all the pages above the accelerators are unnamed.

What we could not confirm

We could not confirm any of the following from a fetched primary source: the site or sites of the one million TPUs; how many Ironwood chips are installed; the power draw of that capacity; the price or contract form; whether Fluidstack’s Texas, Louisiana or New York sites run TPUs, Nvidia GPUs or Trainium; and the 5 GW April 2026 figure beyond press reporting. Broadcom’s filing itself was blocked to us, so the 3.5 GW and “commercial success” language rests on The Register’s account. Google’s TPU7x page has an internal inconsistency on pod size. We also found secondary articles claiming other terms for the Google investment, and we left them out because we could not read them from a page that supports them.

What to watch

Three disclosures would change the picture. First, Epoch revisions: if v7 counts at Columbus and Council Bluffs rise sharply, or a new campus appears with mostly v7, that is the first observable sign of Ironwood scale. Second, the first Fluidstack data halls (Barber Lake delivery was targeted for September 2026, River Bend’s first hall for Q2 2027): whether the operators name the accelerator. Third, Google’s own cloud and Alphabet filings, which would be the place for a customer commitment figure. Until then, the honest statement is a ceiling of one million chips, a stated 2026 start, and a fleet that Epoch estimates in the low gigawatts with Ironwood still a minority of it.

Sources

deep-diveai-clustersacceleratorsprocurement