Oracle is not currently looking at developing its own custom chips for artificial intelligence (AI).
During an Oracle AI roundtable this week attended by DCD, Jason Rees, SVP of technology and cloud engineering EMEA at Oracle, confirmed the company's AI strategy.
When asked if Oracle was exploring the possibility of developing its own custom chips, following in the footsteps of Google, Microsoft, and Amazon Web Services (AWS), Rees said: "No, not at the moment. It's very much a case of we're working with Nvidia, with AMD, and all the other chip manufacturers. Our partnership is with those companies."
Google notably, as well as using Nvidia chips, creates its own 'Tensor Processing Units' (TPUs) for AI workloads. The cloud giant released its sixth generation TPU dubbed Trillium in May 2024.
AWS has its Trainium and Inferetia chip offerings; the latest generation of the former is expected to consume 1kW of power.
Microsoft has also developed its own Arm-based CPU and AI accelerator.
Oracle has offered access to Nvidia H100s since September 2023, and in October launched a Nvidia DGX Cloud supercomputer via the Oracle Cloud marketplace. Earlier this year, it was confirmed that Oracle would also be offering Nvidia's Blackwell GPUs.
Rees expanded on Oracle's relationship with Nvidia, stating: "We have a relationship with Nvidia, and what that means is the way we have architected Oracle Cloud Infrastructure (OCI) means that the network fabric of our cloud really allows us to get the massive scale required. Looking at training models, for instance, you've got the ability to cluster different regions together.
"We've got a concept for superclusters that can scale up to 32,000 GPUs - Nvidia H100s, H200s, Nvidia Grasshopper. These are all clustered together on our infrastructure."
Oracle also offers Arm-based chips to its customers, and has close relationship with Ampere, developer of CPUs aimed specifically at the data center market based on the Arm architecture. It made a $40 million investment in Ampere in 2019. In 2023, Reuters reported that Oracle spent more than $100m on Ampere chips.
Oracle customers can also access instances based on AMD Epyc Genoa processors.
The company has been a big beneficiary of the AI boom, renting GPUs to wide a range of customers, notably Microsoft. In its latest earnings call, the company revealed that in Q4 alone it had signed more than 30 AI sales contracts, totaling more than $12.5bn.
The company also revealed that it was building a "very large data center" for OpenAI, which would have the newest Nvidia chips and interconnect. OpenAI confirmed in June 2024 that it would be using OCI to train its large language model.
Earlier this month, Oracle took a hit when Elon Musk's xAI withdrew from talks which could have seen the latter invest $10 billion in using Oracle Cloud for AI workloads. On the decision to withdraw, Musk wrote on X: "xAI contracted for 24k H100s from Oracle and Grok 2 trained on those. Grok 2 is going through finetuning and bug fixes. Probably ready to release next month. xAI is building the 100k H100 system itself for fastest time to completion. Aiming to begin training later this month. It will be the most powerful training cluster in the world by a large margin.
"The reason we decided to do the 100k H100 and next major system internally was that our fundamental competitiveness depends on being faster than any other AI company. This is the only way to catch up."
