Oracle AI World: Google TPUs, Nvidia Rubin and Cerebras land on Oracle Cloud

Oracle turned its AI World event in Las Vegas on October 27 into a roster change for its cloud. Google Tensor Processing Units, multi-gigawatt Nvidia Vera Rubin systems and Cerebras inference will all run on Oracle Cloud Infrastructure, giving customers accelerator choices that did not exist on the platform six months ago.

The Google deal is the structural one. For the first time anywhere, a hyperscaler outside Google will resell Google's custom silicon. Oracle's president of cloud infrastructure, Mahesh Thiagarajan, said in a statement that OCI will provide the power, cooling and datacenter networking to deploy Google TPUs at scale, with availability beginning in calendar year 2027 and ramping over a multi-year period.

Related reading: OpenAI Models Coming to Google Cloud: Frontier GPT-6 Line Lands in Model Garden.

The hardware in question is Google's eighth-generation Tensor Processing Unit family, which splits into two designs. The TPU 8t targets training and, according to Google, delivers nearly three times the compute performance per pod over the previous generation. The TPU 8i is built for inference and the iterative work of multi-step AI agents. Google frames both as accelerators designed to maximize processing per watt of electricity, which matters to hyperscalers that pay to power and cool every rack.

Google's press release dates the offering and Oracle confirms it. The word "first" is doing real work here: Oracle is not slotting TPUs into a niche reseller program but presenting them as a standard option in a compute portfolio it markets as broad. Thiagarajan put the pitch in terms of choice, saying in the announcement that customers gain "the flexibility to match their most demanding workloads with infrastructure designed for near-linear scalability."

Related reading: Nvidia DSX POD: Modular AI Data Center Cuts Deployment From 6 Months to 14 Days.

The Nvidia announcement runs in parallel and is larger in immediate scale. Nvidia and Oracle said October 27 they will expand OCI Superclusters with Vera Rubin NVL72 systems to multi-gigawatt scale, supporting frontier training, inference and agentic workloads. The two companies also made Vera Rubin NVL72 and Vera CPUs generally available on OCI, and outlined plans for an HGX Rubin NVL8 eight-GPU configuration built on two 64-core Intel Xeon 6 CPUs, 4 terabytes of DDR5 memory and 6.4 terabit-per-second networking. Nvidia's release notes that OpenAI has already trained and served GPT-6 Astra on an OCI Supercluster with more than 100,000 Nvidia Blackwell GPUs in Abilene.

Cerebras adds the speed story. The two companies said October 27 they plan to make Cerebras fast inference available on OCI in calendar year 2027, aimed at AI labs and developers building interactive, multi-step agents where every extra model call adds latency. Cerebras designs its Wafer-Scale Engine to keep compute and memory on a single wafer-scale processor, cutting the data movement that slows long reasoning chains. The market read the news as material to Cerebras: the stock closed at $119.31 on October 27, up 10.10 percent on volume of more than 14 million shares, after touching a 52-week low of $106.13 earlier in the month.

Related reading: AI Networking: NVIDIA, Cisco Target 12.8 Tb/s Capacity.

A Cerebras spokesperson told CNBC the Oracle deal is material but that the company will not disclose its scale because it is in a quiet period ahead of earnings.

OpenAI rounds out the day. Oracle said October 27 it is embedding ChatGPT and Codex into its applications so that databases can be made actionable in natural language, and it plans to make chips from several vendors, including OpenAI's Jalapeno accelerator and Cerebras hardware, available on OCI. OpenAI chief executive Sam Altman was scheduled to join Oracle co-chief executive Clay Magouyrk on stage for a keynote the same afternoon.

Related reading: HPE Books a $1.2 Billion AI Order From Vultr Without an Nvidia Rack, as the Buildout's Value Moves to the Network.

The pattern across the announcements is customer demand for alternatives in a market where chips remain scarce. Thiagarajan told CNBC that scarcity is an ongoing issue and that customers such as OpenAI benefit from having different accelerators to choose from. Pressed on whether this is a pivot away from Nvidia, he said no, describing Nvidia chips as "the best, most fungible" and the overall AI market as incredibly large.

That is the counterweight for investors. Oracle is clearly broadening its silicon menu, but it is adding most of these options over a 2027 runway rather than all at once. Cerebras scale stays undisclosed through its quiet period, TPU availability begins next year and ramps over years, and Nvidia remains the incumbent that Oracle says it is not leaving. The bull case is that a multi-architecture cloud captures more of a growing pie; the bear case is that "choice" is easier to announce than to profitably integrate, and that customers may rent multiple chip types without boosting any single vendor's economics as much as the day's stock moves imply.

What to watch next: the earnings call where Cerebras quantifies the Oracle relationship after its quiet period lifts, whether the Google TPU ramp in 2027 converts into announced capacity numbers rather than intent, and how much of Oracle's own capacity is tied to Nvidia versus the newcomers. The market is now pricing an Oracle that sells Google hardware and Nvidia hardware side by side. That is the real change from October 27, and it will show up in how carefully customers divide their workloads across accelerators.

This content is for informational purposes only and does not constitute financial advice. Past performance is not indicative of future results. Consult a qualified financial advisor before making investment decisions.