Home โ€บ Blog โ€บ Google's TPU fleet
AI Factory Case Studies ยท Part 6 ยท Google's TPU fleet

The Fleet: How Google Built AI Factories a Decade Before Everyone Else

Every other case study in this series centres on one campus. Google's does not. It started designing its own AI chips in the mid-2010s, joined them into pods with optical switches that can be rewired in software, trained Gemini across several data centers, and now rents its chips to rivals. This case study is about a fleet, not a site: what a ten-year head start buys, and what it costs.

FL
FrontierAGI Team
โš ๏ธ How this was researched (as of 26 Oct 2026) Chip figures come from Google: its Ironwood announcement, Cloud TPU documentation, the 2023 TPU v4 paper and a 2026 paper by Google authors comparing five generations of TPU supercomputers. Spending comes from Alphabet's earnings as reported by DCD; emissions from Google's 2026 Environmental Report as reported by Axios and S&P Global; the Anthropic deals from CNBC and DCD. Google does not publish per-site megawatts or chip counts, so site sizes are Epoch AI-based estimates and marked as such. This site is written partly with Anthropic's Claude, which runs on Google TPUs among other chips. See Reading the numbers.

1. Case file

๐ŸŸข
Google's TPU fleetMany sites, including New Albany and Columbus (Ohio) and Pryor (Oklahoma)
OwnerGoogle (Alphabet)
Current chipIronwood (TPU v7)
Pod9,216 chips ยท 42.5 EFLOPS FP8
Per chip4,614 TFLOPS FP8 ยท 2,307 BF16
Pod memory~1.77 PB HBM
2026 capex$195โ€“205B guidance
Big outside userAnthropic (up to 1M TPUs)
Largest sites (est.)New Albany ~453 MW ยท Pryor ~368 MW

Google compares an Ironwood pod's 42.5 exaflops with the El Capitan supercomputer's 1.7 exaflops, but that compares FP8 with FP64, which is not like for like. Site estimates come from an Epoch AI-based ranking; Google does not publish them.

2. Timeline

Google's AI factory, 2015 โ†’ 2027

The longest timeline in the series. Solid dots are confirmed by Google or a partner, hollow dots come from the press, โš ๏ธ marks claims we could not confirm and dashed dots are planned. Click a dot for its source.

Click a milestone to see where it comes from.

3. The bet: build the whole stack, early

Google started deploying its first Tensor Processing Unit in its data centers in 2015 and revealed it at Google I/O in 2016. That first chip only ran trained models; from TPU v2 in 2017 onward, TPUs could train them, joined into pods. Google now designs every layer: the chip (with Broadcom as design partner, a deal recently extended to 2031), the optical interconnect, the data center network, the compiler and frameworks (XLA and JAX) and the models (Gemini).

๐Ÿ’ก Rainier, ten years earlierAmazon's Project Rainier (Part 4) bets that owning the chip pays off. Google made the same bet a decade earlier, which is why it now has seven generations of experience, a mature software stack and enough spare capacity to sell to other labs.

4. Seven generations

Peak compute per chip, by generation

BF16 TFLOPS per chip on a log scale, with the largest pod for each. Toggle to show pod size. Peak numbers are not the same as real training throughput.

Sources: Google 2026 five-generation paper (v2โ€“v5p, Ironwood BF16); Cloud TPU docs (v6e Trillium); Ironwood launch post (FP8). TPU v1 was an inference chip measured in INT8 operations, so it is not on the BF16 scale.

5. The pod and its optical switches

NVIDIA's largest building block, the NVL72, joins 72 GPUs in one rack with copper NVLink. Google's Ironwood pod joins 9,216 chips. Inside each block of chips, TPUs connect directly to their neighbours; between blocks, Google uses optical circuit switches (OCS): mirrors that redirect light from one fiber to another. Google's 2023 TPU v4 paper described how this lets the pod be rewired in software, so it can route around a failed block or carve one pod into slices of different shapes for different jobs, without anyone touching a cable.

Rewire the pod

An illustration with 64 blocks instead of the real number. Click blocks to fail them, then choose a job size: the switch picks healthy blocks and leaves failed ones out.

6. Training across many data centers

Microsoft's Fairwater (Part 3) made headlines in 2025 for linking two sites into one AI superfactory. Google had already done it: its December 2023 Gemini technical report says Gemini Ultra was trained on TPU pods across multiple data centers, which it described as a step up in scale from its earlier flagship model. That is why Google's AI factory is better understood as a fleet of campuses joined by Google's own network than as any single site.

Related reading: AI factory networking and custom silicon vs NVIDIA.

7. Renting chips to rivals

For years TPUs mainly served Google's own products and a few cloud customers. That changed in October 2025, when Google and Anthropic announced that Anthropic would get access to up to one million TPUs, bringing well over a gigawatt of capacity in 2026, in a deal worth tens of billions of dollars. In 2026 Anthropic signed a further agreement with Broadcom and Google for about 3.5 GW of TPUs from 2027, and reports in April 2026 described Google's investment in Anthropic as locking in 5 GW of TPU capacity โš ๏ธ. The Wall Street Journal reported Google funding a $3.2 billion data center in western New York for Anthropic to lease โš ๏ธ.

Who uses whose chips

A simplified map of the labs in this series and the chips they train on, as reported. Tap a lab.

8. Power and emissions

Google spreads its AI capacity across many campuses, buying power from utilities and contracting clean energy, rather than building one giant gas-powered site. But the scale is showing. Alphabet's capital spending was $91.4 billion in 2025, and its 2026 guidance rose from $175โ€“185 billion in February to $195โ€“205 billion in July.

Spending and footprint

Google's 2026 Environmental Report, covering 2025, showed electricity use up 37% and total greenhouse gas emissions up 18%, its largest annual rise, driven mainly by manufacturing chips and servers. Operational emissions fell 2% thanks to clean energy contracts, including over 12 GW of new deals, but supply-chain emissions rose 25%, the share of carbon-free electricity on an hourly basis stayed roughly flat, and water use rose 34% to 10.9 billion gallons. Google set a goal in 2020 to run on carbon-free energy every hour of every day by 2030.

9. Six bets compared

ColossusAbileneFairwaterRainierMetaGoogle
The betSpeedOutside capitalNetwork of sitesOwn siliconEvery building typeWhole stack, a decade early
ChipsNVIDIANVIDIANVIDIATrainium2NVIDIATPU v7 Ironwood
Unit of scale1 campus1 campus2 linked sites1 campus + 1Linked buildings + mega-campusA fleet of campuses
PowerOwn turbinesGas microgridUtilityUtility GenCoOwn gas / utility gasUtilities + clean energy contracts
Main issuePermitsTax breaksUtility gasWellsStranded plantsRising emissions and water use
Ten-year head startSeven TPU generations
Optical switchesPods rewired in software
Chips as a productRivals rent TPUs at gigawatt scale
Footprint growsEmissions up 18% in 2025
VerdictGoogle is the only company in this series whose AI factory predates the AI boom. Owning every layer for a decade lets it build pods larger than anyone else's, train across data centers routinely and now sell its chips to a direct competitor. The cost is showing up in its own climate numbers. Next, the neoclouds: companies like CoreWeave that built AI factories as a rental business, financed with debt.

10. Reading the numbers

From Google
Per-chip and pod figures; TPU first deployed 2015, revealed 2016; OCS described in the 2023 TPU v4 paper; Gemini Ultra trained across multiple data centers (Gemini report, Dec 2023); environmental report figures.
From Alphabet earnings
2025 capex $91.4B; 2026 guidance $175โ€“185B (Feb) โ†’ $195โ€“205B (Jul), via DCD.
Deals
Up to 1M TPUs and well over 1 GW in 2026 (Google and Anthropic, Oct 2025); ~3.5 GW from 2027 with Broadcom (DCD). The 5 GW figure and $3.2B New York data center are press reports โš ๏ธ.
Estimates
Site megawatts come from an Epoch AI-based ranking, not Google.
Illustrations
The pod rewiring demo uses 64 blocks and a simple allocation rule; the "who uses whose chips" map is simplified.