Oracle Just Opened Its Cloud to Everyone Else's Chips

Oracle spent the last two years as the most Nvidia-dependent company in enterprise software. This week it decided to become the most chip-agnostic cloud on the market, and it did it in a single afternoon.

On October 27, at Oracle AI World in Las Vegas, the company announced four separate things that normally would each be a headline on their own: Google TPUs are coming to Oracle Cloud Infrastructure, OpenAI's custom Jalapeño inference chip is coming too, Cerebras hardware is coming by the first half of 2027, and Nvidia's Vera Rubin NVL72 racks are going generally available at multi-gigawatt scale. There was also a full suite of ChatGPT and Codex integrations baked into Fusion Applications and NetSuite.

Any one of those is a product announcement. Together they are a strategic repositioning, and the timing tells you why.

The part that actually matters

The TPU deal is the real story. Oracle says it will be the first hyperscaler outside Google to offer Google's tensor processing units, specifically the eighth-generation TPU 8t for training and TPU 8i for inference. Google's own release describes the 8t as delivering nearly three times the compute performance per pod over the previous generation. Both companies are careful to say availability begins in calendar year 2027 and ramps over a multi-year period.

That timeline matters, and so does the framing. TPUs have historically been internal-only hardware for Google. The first crack came in 2025, when OpenAI reportedly struck a cloud deal that involved TPU capacity. Now those chips are being offered as a product through a third-party cloud. AWS and Azure have not done this. Oracle got there first.

Why would Google help a competitor sell its silicon? Because TPU revenue is leverage, and because the customer Oracle is chasing is OpenAI. Follow the triangle for a second. OpenAI is Oracle's largest compute tenant and, through the Stargate buildout, effectively its largest creditor-by-contract. OpenAI also buys TPUs from Google directly. Now OpenAI can buy TPU capacity through Oracle, the same partner it owes billions in committed spend to. Every layer of that stack makes the suppliers more entangled rather than less.

The Jalapeño wildcard

OpenAI's chip effort finally has a name attached to public deployment plans. Jalapeño, unveiled in June with Broadcom, is OpenAI's first "intelligence processor," a reticle-sized ASIC built in a reported nine-month cycle and aimed at LLM inference. CNBC reports it will run on OCI alongside Cerebras. Oracle has not disclosed volume, pricing, or which customers get access first.

This is the announcement with the least detail and the most signal. A company that designs its own inference silicon and then places it in a partner's cloud is telling you it expects inference demand to be enormous and margin-sensitive. Training stays on Nvidia. Inference goes wherever the economics work.

Nvidia is not being abandoned

Worth being clear, because the "Oracle dumps Nvidia" take wrote itself within minutes of the keynote. Oracle's Mahesh Thiagarajan flatly denied a pivot, calling Nvidia chips the best and most fungible option available. The Nvidia announcement was arguably the largest of the day: Vera Rubin NVL72 and Vera CPUs are now generally available on OCI, with plans to scale OCI Superclusters to multi-gigawatt capacity using Vera Rubin racks. There is also a planned HGX Rubin NVL8 configuration with two 64-core Intel Xeon 6 processors, 4TB of DDR5, and 6.4 Tbps RDMA networking for training and fine-tuning workloads.

The proof point Oracle keeps repeating is Abilene, Texas, where OpenAI trained and served GPT-6 Astra on an OCI supercluster with more than 100,000 Nvidia Blackwell GPUs. OpenAI's Uday Ruddarraju said the company is bringing Vera into production on OCI. So the same company that just announced its own inference ASIC is also standing up the next generation of Nvidia hardware at scale. That is not a contradiction. It is a hedge, and hedges are what you buy when you are not sure which part of the market grows fastest.

The business apps angle

The software side of the announcement is less dramatic but probably more immediately monetizable. Oracle announced an AI Workspace that puts OpenAI models inside ChatGPT while connecting to Fusion Applications, NetSuite, the Oracle AI Database, and OCI, with governance and approval flows staying inside Oracle's stack. An App Composer lets teams build applications with Codex and have Oracle deploy and manage them. An AI Security Mesh applies OpenAI's cyber models to vulnerability discovery across OCI, other clouds, and on-premises environments. And an SMB bundle throws in $12,000 a year in cloud credits to get smaller companies started.

Read Oracle's own disclaimer carefully, though. The press release carries a future product disclaimer stating this is general product direction and not a commitment to deliver. No general availability dates were given for the app-side products. Treat the demos as intent, not shipping software.

The debt clock in the room

Here is the tension nobody on stage wanted to discuss. Oracle's remaining performance obligations, the contracted revenue it has booked but not yet recognized, grew from roughly $455 billion at the end of the first quarter of fiscal 2026 to $664 billion as of August 31, 2026. Analysts estimate a large share of that is OpenAI-related, though Oracle has never broken out the figure. The stock is down more than 30% year to date on financing concerns, and the company faces a December air-quality hearing over its Project Jupiter data center in New Mexico after a contractor delivered a force majeure notice.

Against that backdrop, "multi-gigawatt" commitments and 2027 accelerator ramps are not free. They are capital calls. Diversifying silicon suppliers reduces single-vendor risk and gives Oracle negotiating leverage on pricing, which is genuinely valuable. But it does not generate cash this quarter, and the market has started grading Oracle on cash.

What to watch

Three things will tell you whether this was a real inflection or a very good keynote. First, whether TPU capacity on OCI actually appears in 2027 on schedule, since multi-year ramps slip quietly. Second, whether Jalapeño shows up in a production inference service with published pricing, which is the difference between a partnership and a press release. Third, whether Oracle's investor day targets for fiscal 2030 come with a credible financing structure, because right now the company is promising compute it has not yet paid for to customers who have not yet paid it.

The strategic logic of an open cloud is sound. The hard part is that openness costs money before it makes any.