Google TPUs head to a hyperscaler for the first time: Oracle Cloud, starting 2027
Announced Oct 27: Google TPUs land on OCI in calendar 2027, the first hyperscaler outside Google to offer them; ORCL rose about 5% that day.
For the first time, Google will let another hyperscaler run its TPU chips. On October 27, at the Oracle AI World conference in Las Vegas, Oracle announced that Google Tensor Processing Units will be available through Oracle Cloud Infrastructure (OCI) starting in calendar year 2027 and ramping over multiple years. Per Oracle’s press release, OCI supplies the power, cooling and data center networking; Google supplies the silicon that until now has trained Gemini inside Google’s own fleet. The market voted the same day: ORCL rose as much as about 5% intraday (24/7 Wall St. logged it up 5% in afternoon trading; the Chinese outlet Cailianshe, carried by Sina Finance, counted “up about 4%” at its deadline — same move, different cutoffs).
Key points
- First outside Google: both companies confirmed OCI as the first hyperscaler outside Google to offer TPUs, with supply starting in calendar 2027 and a multi-year ramp — a deployment start date, not a live service.
- Two chip flavors: the eighth-generation TPU line splits into TPU 8t for training (nearly 3x compute per pod over the previous generation) and TPU 8i for inference; Google’s blog cites 12.6 PFLOPS of FP4 compute per chip and over 97% goodput across thousands of chips.
- More on the same day: Oracle also expanded its Nvidia deal to multi-gigawatt Vera Rubin superclusters and added Cerebras and SambaNova inference chips; the stock move and the balance sheet landed together (quarterly capex of $28.5B, RPO of $664B).
Background
For a decade, TPUs have been Google’s in-house advantage: designed, built and used to train Google’s own models, and rented out only through Google Cloud. Any company that wanted TPU compute had to buy it from Google. This deal turns the chip into something that can be installed in a rival’s data centers — Oracle provides the power, cooling and networking, Google provides the chips.
The tie between cloud buyers and sellers has been tightening across the industry. Anthropic’s seven-year, $11.6B cloud contract with Akamai came with warrants, and Nvidia, AMD and Broadcom signaled interest in the first $5B of Meta’s equity raise of up to $95B. With TPUs on OCI, the competition between hyperscalers stops being about who has the most GPUs and becomes about who has the longer chip menu.
The chip menu
OCI’s TPU offering is built on Google’s eighth-generation TPUs, with Oracle’s press release and Google’s Cloud blog supplying complementary numbers:
| Item | Detail | Source |
|---|---|---|
| TPU 8t (training) | Nearly 3x compute per pod over the previous generation; officially frames frontier-model development cycles dropping “from months to weeks” | Oracle press release |
| TPU 8i (inference) | Aimed at inference, agentic and long-context workloads | Oracle press release |
| Per-chip compute | 12.6 PFLOPS of FP4 compute, up to 2x the performance per watt of the previous generation | Google Cloud blog |
| Training stability | Over 97% goodput sustained across thousands of chips, via on-die ICI routing with optical circuit switching that reroutes around faults | Google Cloud blog |
| Inference economics | A new topology cuts network hops by more than half; up to 80% better performance per dollar for low-latency serving | Google Cloud blog |
The software side is being prepared too: Google’s TorchTPU, its native PyTorch stack, is now open source on GitHub, with vLLM, SGLang, Ray and Hugging Face on the supported-tools list. Porting a PyTorch training job or a vLLM deployment to OCI’s TPUs therefore doesn’t start from zero — but “supported” and “runs well” are different claims, and only real 2027 workloads will test the gap.
The rest of the day’s deals
TPUs were one course of a longer menu served at Oracle AI World. Also announced: an expanded Nvidia partnership, with Vera Rubin superclusters to reach multi-gigawatt scale and Vera Rubin systems plus Vera CPUs now generally available on OCI — co-CEO Clay Magouyrk said they are “performing better than expected across hardware quality, manufacturing yield and performance.” Per Reuters, Oracle plans to offer chips from Intel-backed SambaNova Systems and Cerebras; separately, Oracle and Cerebras announced that Cerebras’s service is scheduled to arrive on OCI in 2027. Alongside these sits the earlier-announced AMD Helios rack of 72 MI455X GPUs. Per CTOL Digital Solutions’ tally, Oracle has delivered 850 MW of capacity and more than 300,000 GPUs since the end of May.
While the menu grows, Oracle is sliding its own parts into the stack: per Reuters, even Nvidia-based systems will skip some of Nvidia’s networking chips in favor of Oracle’s in-house Acceleron — and the Google blog itself touts Acceleron SmartNICs and Zero Trust Packet Routing among OCI’s selling points. (Chinese outlet Cailianshe, carried by Sina Finance, reports the same swap.) The chips can be anyone’s; Oracle wants the network entry point to be its own.
The money behind it
Who pays for the wider menu? CTOL Digital Solutions ran the numbers from the latest 10-Q: quarterly capex of $28.5 billion, 3.4x a year earlier, of which $11.4 billion came from customer prepayments — customers funded 40% of the build. Cloud and software margin fell from 60 cents to 39 cents kept per revenue dollar. RPO stands at $664 billion (up about $209 billion year over year), but only about 13% — roughly $86 billion — converts to revenue within 12 months. Mizuho framed the October 28 investor day as the key window for management to address concerns about high capital spending and the path to positive free cash flow.
One line captures that analysis: the chip menu is wider, but how much of each dollar Oracle keeps is the question that hasn’t changed. The stock bought the menu on Tuesday; the investor day has to answer for the bill.
Our take
For teams buying compute in the cloud, the direct signal is that training compute now has a credible option beyond Nvidia, and for the first time taking it doesn’t mean moving your whole workload into Google Cloud. TPUs and GPUs coexisting in one cloud makes proximity-to-data the pitch: Oracle’s argument is that your enterprise data already sits in Oracle databases, so fine-tuning and inference can happen next to it without cross-cloud transfers.
Two reservations. First, supply starts in 2027 — no OCI region, SKU or pricing for TPUs has been announced, and a five-year plan is only as good as what actually ships. Second, the TPU software ecosystem is catching up on open-source stacks, but CUDA’s gravity is real; for many teams the migration cost could eat the paper gains in performance per dollar. A wider menu is a fact. Which dish is actually cheaper waits for the day the menu has prices.