Google TPUs Leave Google Cloud: Oracle Deal and a $13.3B Inventory Build

On 27 October 2026, Oracle and Google announced that Oracle Cloud Infrastructure will offer Google's eighth-generation TPUs, the TPU 8t for training and TPU 8i for inference, making OCI the first major cloud outside Google to do so. Capacity arrives in calendar 2027. On 29 October, Alphabet's third-quarter 10-Q showed what selling TPUs as hardware looks like on a balance sheet: Google Cloud revenue of $30.7 billion, roughly double a year earlier, and inventory up from $2.4 billion to $13.3 billion, which the filing says consists primarily of hardware related to TPU systems for sale to enterprise customers, plus consumer devices such as Pixel. The TPU is now a product Google ships, not just a service it rents.

What happened

The Oracle release says OCI will provide "the power, cooling, and datacenter networking" to run Google TPUs inside Oracle's infrastructure, ramping over several years. Pricing was not disclosed. Mahesh Thiagarajan, president of OCI, said the deal "gives customers access to multiple leading AI compute architectures in one cloud". Google's Amin Vahdat called TPUs "a powerful, purpose-built option for demanding training and inference workloads".

On Alphabet's earnings call on 28 October, CFO Anat Ashkenazi said TPU system sales were a meaningful contributor to Cloud growth and that she expects them "to continue to scale in Q4, followed by a meaningful revenue ramp in 2027." Google Cloud CEO Thomas Kurian said Google has partnerships to deliver TPU systems with Oracle and with Crux AI, the TPU cloud joint venture with Blackstone, and that Google is placing TPUs in customer data centres and third-party facilities. He put the TPU customer base at nearly 100 in prepared remarks and over 100 significant customers in Q&A, naming Apple among those scaling pretraining on TPUs.

This builds on Anthropic's October 2025 commitment to use up to one million TPUs, a deal Anthropic described as worth tens of billions of dollars. A year later, the TPU business has moved from one anchor tenant to a distribution channel.

The numbers

Why it matters

For a decade the TPU was a reason to use Google Cloud. Now Google is selling TPU systems into other people's buildings, including a direct competitor's cloud. That is a strategic reversal: Google is choosing accelerator volume over cloud exclusivity, presumably because the alternative is watching Nvidia set the price for everyone, including Google.

It also changes Nvidia's competitive picture more than any single benchmark could. Custom silicon has always been limited by who could buy it. If a TPU can be procured from Google directly, rented from Oracle, or rented from a Blackstone-funded neocloud, it starts to look like a merchant accelerator with a captive software stack. Kurian's mention of TorchTPU, built with Meta to let PyTorch models run on TPUs, goes after the one thing that kept most teams away: JAX.

Oracle's angle is simpler. OCI also widened its other accelerator options at the same event, including plans to deploy Cerebras inference hardware next year, according to The Next Platform. Oracle wants to be the cloud that sells whichever chip a customer can get.

What sceptics say

Read the 10-Q's revenue recognition note closely. Google says some customer arrangements "include options requiring us to repurchase the hardware and embedded software at pre-set prices" over a multi-year term, and where a customer assigns TPU systems to another party but keeps that repurchase right, Google accounts for it as a financial guarantee. That means Google is carrying some of the residual-value risk on hardware it has "sold". If accelerator prices fall faster than expected, some of that risk comes back.

The related-party piece is another flag. $16.8 billion of backlog is tied to Crux AI, an unconsolidated equity-method joint venture in which Google is an investor. Selling chips to a venture you partly own and counting the contract in backlog is legal and disclosed, and it is precisely the pattern that circular-financing critics point to.

Then there is cash. Ashkenazi said the TPU inventory build weighed on operating cash flow, and the call put third-quarter free cash flow at negative $9.1 billion, which included a $5.2 billion European Commission fine. Alphabet also said it began using third-party compute capacity at the end of the quarter because it is supply constrained, which will pressure Cloud margins. And none of the Oracle capacity exists yet: it starts in 2027, with no published price and no disclosed volume.

What it means for engineering and DevOps teams

My take: for most teams this is a 2027 planning item, not a 2026 one, but it is worth starting the portability work now. TPUs reward models written for XLA. If your training or serving code is PyTorch with custom CUDA kernels, the TPU price-performance claims do not apply to you until you have ported and profiled it. A small proof of concept on Cloud TPU today tells you whether OCI TPUs in 2027 are a real option or a procurement fantasy.

Second, multi-accelerator is becoming the default at the big buyers, and it will trickle down. Your observability, scheduling and cost tooling should treat accelerator type as a first-class dimension rather than assuming "GPU". The FinOps for AI work on tagging spend by model and accelerator pays off here, and the inference serving choices you make now decide how hard a migration is later.

Third, watch the contract terms if you are a large buyer. Repurchase options, multi-year commitments and capacity in third-party facilities are now normal in this market. Ask who owns the hardware in year four, and what happens to your reserved capacity if the facility partner slips.

FAQ

Can you buy Google TPUs outside Google Cloud?

Increasingly, yes. Alphabet now records TPU system sales as revenue, Oracle will offer TPU 8t and 8i on OCI from calendar 2027, and the Google and Blackstone joint venture Crux AI plans to offer TPU capacity. Pricing for the third-party offerings has not been published.

What are TPU 8t and TPU 8i?

They are Google's eighth-generation TPUs: 8t for training and 8i for inference. Google claims up to 2.7x better training performance for 8t and 80% better inference price performance for 8i versus the previous generation; these are vendor figures, not independent benchmarks.

How big is Google's TPU business?

Alphabet does not break it out. It says TPU system sales were a meaningful contributor to $30.7 billion of Cloud revenue in Q3 2026, and its inventory, primarily TPU systems hardware plus consumer devices, rose to $13.3 billion.

Related: FinOps for AI in 2026: 98% Manage Token Spend, 3 in 4 Can't Prove the Value, GPUs Run at 5%, and FOCUS Gets a Token Column in 1.5, LLM Inference in 2026: Prefill Is a Different Machine Now, and the KV Cache Has a Disk Tier and GPU Clouds in 2026: H100s at $2.59 an Hour, $104B of CoreWeave Backlog on $35B of Debt, and a Six-Year Bet on Silicon.