Google's TPUs step out of the building

Three things are true at once. The hardware has, for the first time, closed the per-chip gap with Nvidia. The economics genuinely favour TPUs inside Google's own stack — and in the one good independent test, on inference. And the external business changed fundamentally in 2026: TPUs are now sold as systems into customer data centres, financed off Google's balance sheet, and — from 2027 — offered on a rival hyperscaler's cloud. What has not changed: Nvidia still takes an estimated 70–86% of AI-accelerator revenue, and the TPU story leans heavily on one very large, very leveraged customer.

Section 01The hardware: eight generations in

Google has been building TPUs since 2015. The latest full generation, Ironwood, was unveiled April 9, 2025, announced for availability in November 2025, and reached formal general availability on March 31, 2026 by Google's own release notes — a two-step rollout that explains why sources give different "GA" dates. An eighth generation (8t for training, 8i for inference) was announced April 22, 2026 and was still listed as "coming soon" in October 2026.

Who makes them: TSMC fabricates the chips Reported — Google does not name its foundry. Broadcom has co-designed TPUs with Google for roughly a decade and, in an April 2026 SEC filing, disclosed a long-term agreement to develop and supply future TPUs through 2031 Confirmed — filing. MediaTek is reported to be designing a cost-optimised inference variant for generation 8 Reported. Ironwood's list price, from Google's own price list (verified October 29, 2026): $12.00 per chip-hour on demand, $5.40 with a three-year commitment — while negotiated large-scale rates are far lower (SemiAnalysis estimates Anthropic pays ~$1.60 per chip-hour) Estimate.

Section 02TPU vs Nvidia: where each one actually wins

On training cost, there is no neutral scoreboard. Google does not submit TPUs to MLPerf, so no independent body has measured TPU against GPU on training. What exists is modelling: SemiAnalysis (Nov 2025) put Ironwood's all-in cost of ownership about 44% below an equivalent Nvidia GB200 server — for Google's own internal use Estimate. That number strips out Nvidia's margin and assumes Google's power, networking and utilisation; it does not automatically transfer to a third party's building. Google's cloud chief has said its AI servers pay for themselves in under two years, and those running Google's own silicon in roughly half that Reported.

On inference, there is one good independent test. SemiAnalysis's InferenceX preview (Sep 2026) ran the same large model on all three chips, with modelled hourly costs:

Ironwood ~18% cheaper than B200 and ~34% cheaper than B300 on this workload. The caveat rides along: under heavy load its time-to-first-token was 5.41 seconds against B300's 2.40 — slower to start answering — and Ironwood has no FP4 precision, where Blackwell and Rubin do their cheapest work. Nvidia's Rubin platform claims 5× Blackwell inference performance at 10× lower token cost; those are vendor claims, not measurements.

Where Nvidia still wins, on evidence: the software ecosystem (twenty years of CUDA kernels and tooling against a PyTorch bridge that only went native and open-source on October 20, 2026, when Google released TorchTPU, co-developed with Meta); latency-sensitive and FP4 inference; flexibility — a GPU runs anything and resells into a deep second-hand market, a TPU runs matrix maths and, until this year, only in Google's buildings; and roadmap cadence, with Rubin shipping while generation-8 TPUs are still "coming soon." Nvidia's public reply to the Meta reports (Nov 25, 2025) was that it is "the only platform that runs every AI model and does it everywhere computing is done."

Section 03The external business: confirmed here, reported there

Until 2026 the answer to "can I buy a TPU?" was no. Alphabet began selling TPU systems — whole assembled systems, hardware plus software, not bare chips — with first deliveries to customer data centres in Q2 2026 Confirmed. The deals, kept in their evidence tiers:

Anthropic — the anchor customer Confirmed

October 2025: Anthropic expanded to up to 1 million TPUs, "well over" 1 gigawatt coming online in 2026, worth "tens of billions of dollars" (the exact value was not disclosed; a ~$52 billion figure and a 400,000-bought / 600,000-rented split are SemiAnalysis modelling Estimate). Broadcom confirmed on its December 2025 call that Anthropic was its mystery $10 billion rack customer, with a further $11 billion order after. April 2026 brought a further multi-gigawatt agreement from 2027 — Broadcom's filing says ~3.5 GW; Anthropic later described 5 GW; that discrepancy is unresolved in the sources. And Anthropic's IPO filing (Oct 26, 2026) commits at least $114.5 billion to Google through July 2033, plus an estimated $125–150 billion of further TPU capacity in 2028–2031 and $161.2 billion of Broadcom equipment leases — by one tally, ~$400 billion of its lined-up compute is TPU-based, more than to Amazon, Microsoft or its Nvidia channels.

Oracle — TPUs on a rival's cloud Confirmed

October 27, 2026, by joint press release: Google's TPUs will be offered through Oracle Cloud Infrastructure — the first time on a hyperscale cloud outside Google — built on generation 8, available in calendar 2027, ramping over multiple years. No pricing or regions disclosed. Oracle simultaneously expanded its Nvidia Rubin capacity to multi-gigawatt scale: it is adding TPUs, not replacing Nvidia.

Crux AI — the financed neocloud Confirmed

A Google–Blackstone venture, launched September 9, 2026 with $5 billion of Blackstone equity: Google supplies TPUs and software, first 500 MW in 2027. Reuters reported a $22 billion loan from ten banks to buy the chips, secured by the chips' value and customer contracts — lenders are, in effect, treating TPUs as bankable collateral for the first time.

OpenAI — a reported plan to buy Reported

The Information reported on October 27, 2026, citing a person familiar with the matter, that OpenAI plans to buy Google TPUs; Reuters relayed it the same day; CNBC separately confirmed the plan by its own sourcing. No price, volume, generation or timing has been disclosed. OpenAI declined to comment; neither company has announced anything. This is a credible, multiply-reported plan — it is not a signed deal, and it should never be quoted as one. Context: OpenAI has rented TPUs through Google Cloud since mid-2025 reporting, and remains one of Nvidia's largest buyers.

Meta — a reported rental, a confirmed collaboration Reported / Confirmed

The Information reported in February 2026 that Meta had signed a multi-year, multibillion-dollar deal to rent Google TPUs; Meta and Google declined to comment, and as of October 2026 neither has confirmed it. What is confirmed is the engineering: on October 20, 2026 Google open-sourced TorchTPU, co-developed with Meta's PyTorch team, whose engineering director said it "enables us to deploy our PyTorch workloads to Google TPUs." Deep collaboration, confirmed. The rental contract's size: not confirmed.

Everyone else

Apple trained its Apple Foundation Models on rented Cloud TPUs (8,192 v4 chips for the server model) — from Apple's own 2024 paper Confirmed. Safe Superintelligence said in April 2025 it would run research on Cloud TPUs Reported. Google has named Lightricks, Essential AI, Thinking Machines Lab, Hudson River Trading, Boston Dynamics and Citadel Securities as TPU customers. On its Q3 2026 call, Google said over 100 significant customers now use TPU systems. One correction worth making plainly: no credible evidence links xAI to Google TPUs — the traffic runs the other way (Anthropic rents Nvidia capacity at xAI's Colossus site).

Section 04The money

Alphabet reported its third quarter of 2026 on October 28 — and the premise behind the TPU story checks out in the release itself: Google Cloud revenue more than doubled, and "sales of TPU systems" were called a meaningful contributor to growth. Alphabet does not break out a TPU revenue line; anyone quoting you a precise TPU revenue figure is quoting an analyst model (Section below).

The cash cost is visible now; the revenue mostly is not. Alphabet's CFO says the significant majority of revenue from existing TPU agreements lands in 2027; inventory — mostly TPU systems built for outside customers ahead of that revenue — rose from $2.4 billion at the end of 2025 to $9.991 billion by June 2026. Full-year 2026 spending guidance stands at $195–205 billion, more than double 2025's ~$91 billion, partly funded in Q3 by $28.7 billion of new bonds and ~$12.4 billion of stock sales, with no buybacks.

What the models say Estimate — quoted as a range of sell-side models, not as consensus fact, because the two named firms disagree by $36.5 billion on 2027 alone and may not even be counting the same thing: TD Cowen models $9.4 billion of TPU sales in the second half of 2026 and $62.5 billion in 2027 (at roughly a 20% operating margin on outside sales, against ~35–36% for Cloud overall); Wells Fargo models ~$9 billion in Q3 2026 alone and $99 billion in 2027.

The partner's economics. Broadcom's AI semiconductor revenue (all custom silicon plus AI networking, not TPU-only) reached $16.7 billion in its fiscal Q3 2026, up 221%, with management giving "line of sight" to roughly $115 billion of AI revenue in fiscal 2027. And the financing machine is now real: on October 26, 2026, Bank of America, Blackstone, Citi and Morgan Stanley committed $63 billion to a vehicle that will buy Broadcom accelerators and lease them to Anthropic, deliveries from 2027 Confirmed — with Broadcom itself providing residual-value support on the senior portion, a backstop that exists precisely because there is not yet a liquid resale market for these chips. (Reuters reports the first-lien portion at ~$43.5 billion, about 69% Reported. The $63 billion is financing capacity, not revenue — do not let anyone add it to a sales line.)

Section 05The strategic picture — and the bear case

Does this threaten Nvidia in numbers yet? At the margin of new commitments, yes; in current revenue, no. There is no audited market-share series, but 2026 estimates cluster at Nvidia holding ~70–86% of AI-accelerator revenue, AMD ~5–7%, and all custom chips (TPU, Trainium and others) ~12–18%. Nvidia's data-centre revenue was $89.0 billion in the single quarter ended July 26, 2026 — still growing at more than double. Where TPUs do show up: the largest disclosed forward spending ledger in the industry (Anthropic's filing) allocates more to TPU capacity than to all Nvidia channels combined; Broadcom's AI trajectory is substantially a TPU story; and Alphabet's Cloud growth and $561 billion backlog are explicitly TPU-assisted.

What would have to be true for TPUs to become a genuine third platform: TorchTPU must reach JAX-level performance in ordinary production use, not just in launch benchmarks; gigawatt-scale outside deliveries (Anthropic 2027, Oracle 2027, Crux's 500 MW) must land on time, generation 8 included; a second and third Anthropic-scale customer must buy, not only rent — OpenAI and Meta are the candidates, both unconfirmed; lenders must keep financing the chips without a vendor backstop as large as Broadcom's; and generation 8's claimed gains (2.8× training price-performance, 80% better inference performance per dollar) must arrive before Rubin's claimed reset of the baseline.

The strongest bear case, fairly stated:

• The advantage may not travel. The 44% cost gap and the sub-one-year payback are Google-stack numbers — Google's power, networking, software and utilisation. Hand a bare system to a third-party operator and, as one September 2026 critique put it, "the margin of superiority compresses." Outside sales also earn ~20% operating margin against Cloud's ~36%: Google trades its best business for a merchant-chip business against an incumbent running ~75% gross margins.
• Single-vendor, single-roadmap supply. A TPU buyer is locked to Google for silicon, compiler and interconnect, with Broadcom (and now MediaTek) underneath and TSMC underneath that. No second source, no spot market — lenders price the chips off customer contracts, not resale value. If a generation slips, there is no fallback inside the platform.
• Software is still a tax. Every non-JAX customer pays a porting cost; TorchTPU is weeks old; latency-sensitive work and FP4 still favour Nvidia today; most AI engineering hiring and tooling assumes CUDA.
• Google's incentives are conflicted. Google competes with its TPU customers in models, is short of compute for its own products, and can rationally prioritise itself in a crunch. Customers hedge accordingly — Anthropic runs four silicon platforms, Meta builds its own chips, OpenAI is building one too. TPU is everyone's second source, which caps its share of each wallet.
• Concentration and leverage. The outside-TPU story is disproportionately Anthropic — a company whose filing shows an $80.7 billion first-half accounting net loss and over $600 billion of compute commitments against $16.3 billion of half-year revenue, financed through SPVs, payment-in-kind notes and vendor guarantees. Broadcom's own filing language makes its 3.5 GW conditional on Anthropic's "continued commercial success." If frontier-lab economics wobble, TPU revenue, Broadcom's backlog and Alphabet's $10 billion chip inventory wobble together.