Key points
- Oracle plans to offer Google's AI chips in 2027
- Cerebras hasn't disclosed the size of its Oracle deal
- Oracle says it will keep expanding its Nvidia systems
Oracle (ORCL) plans to offer AI chips from Google, Cerebras Systems (CBRS), and OpenAI on its cloud, while expanding its Nvidia (NVDA) systems to "multi-gigawatt scale." The companies announced the plans at Oracle AI World in Las Vegas on October 27.
Google's tensor processing units, or TPUs, are expected to arrive in 2027. The companies called Oracle Cloud Infrastructure "the first hyperscaler outside Google to offer Google TPUs." OpenAI also agreed to bring ChatGPT and Codex into Oracle's business software.
Cerebras shares closed up 10.1% on Tuesday, while Oracle rose 3.5%. Alphabet (GOOGL), Google's parent, gained 1.2%.
Google's TPUs arrive in 2027
Google designs TPUs to train AI models and run them once they're trained. Oracle and Google said the chips will become available on Oracle's cloud during 2027 and ramp up over several years.
Oracle will provide the power, cooling, and networking. The companies didn't disclose financial terms or how many TPUs Oracle will install.
Their release describes Google's eighth-generation chips: the TPU 8t for training and the TPU 8i for inference, the work of running a trained model to answer requests. It doesn't say which chips will go into Oracle's data centers.
The companies said in a joint blog post that customers could fine-tune and run AI models alongside their existing enterprise data "without costly cross-cloud transfers."
Google's chief technologist for AI and infrastructure, Amin Vahdat, said the arrangement "will give more organizations access to this technology."
Cerebras gets another cloud partner but won't give the size
Cerebras Inference on Oracle's cloud is planned to be available in calendar year 2027. Cerebras builds each processor from a whole silicon wafer and says putting compute and memory on one chip speeds up responses.
"Every step an agent makes requires inference, whether it's reasoning, checking its work, or taking action," said Andrew Feldman, co-founder and chief executive of Cerebras. "The faster the inference, the more work AI can do without making users wait."
A Cerebras spokesperson called the deal material but said the company won't disclose its scale because it is in a quiet period. Amazon Web Services announced its own plans with Cerebras seven months earlier.
Tuesday's gain left Cerebras at $119.31, still well below the $185 price of its May IPO. The stock fell below that offering price in early October.
Oracle says the new chips aren't a pivot from Nvidia
Nvidia's Vera Rubin NVL72 systems and Vera CPUs are now generally available on Oracle's cloud, the companies said. Oracle plans to grow its Vera Rubin installations to multiple gigawatts and to add a smaller eight-GPU Rubin server. The release gives no dollar figure for the expansion.
OpenAI has trained and served its GPT-6 Astra model on an Oracle supercluster in Abilene, Texas, the release says. "We are bringing Vera into production on OCI to support these workloads at scale," said Uday Ruddarraju, chief technology officer of compute at OpenAI.
In a CNBC interview at the conference, OCI president Mahesh Thiagarajan said the new chips don't mark a pivot away from Nvidia. He called Nvidia's chips "the best, most fungible," and said chip scarcity remains an issue for customers.
OpenAI's Jalapeño chip and ChatGPT come to Oracle
Oracle's cloud will also offer OpenAI's Jalapeño chip. OpenAI developed the inference chip with Broadcom (AVGO) and published its first benchmarks on August 25. Oracle's releases didn't say when it will be available.
On the software side, Oracle AI Workspace will let ChatGPT users work with Oracle's applications and data through plugins. A new tool, Oracle App Composer, lets employees build applications with Codex while Oracle runs them on its cloud. Oracle's small-business bundle includes ChatGPT Business and $12,000 a year in Oracle cloud credits.
"More businesses are choosing OpenAI as we make it easier to put AI to work through the tools and systems they already use," said Greg Brockman, OpenAI's president and co-founder.
Oracle shares are down more than 30% this year, partly on concerns about how the company will pay for its data center buildout. Oracle is now adding chips from three more providers, and none of Tuesday's releases said how it will pay for them or for the multi-gigawatt Nvidia expansion. On October 8, Oracle was in talks with Apollo and Goldman Sachs on chip financing.
