Oracle expands cloud compute options across its global network, announcing multiyear agreements at its Oracle AI World conference in Las Vegas to add Google Tensor Processing Units, Cerebras wafer-scale processors, and SambaNova inference systems to Oracle Cloud Infrastructure alongside fresh hardware from Nvidia, Intel, and AMD.
Key points
- Oracle will offer Google Cloud TPUs and Cerebras wafer-scale inference systems on Oracle Cloud Infrastructure starting in 2027.
- The cloud provider is scaling Nvidia Vera Rubin NVL72 deployments to multi-gigawatt capacity alongside upcoming Intel Diamond Rapids and AMD Venice server processors.
- Oracle also deepened ties with OpenAI, adding Broadcom-designed Jalapeño accelerators and native enterprise ChatGPT tooling across its corporate software suite.
A multi-vendor silicon roster
Oracle Cloud Infrastructure, known as OCI, is broadening its hardware catalogue to include specialised accelerators from several competing chip designers. The announcements in Las Vegas mark a shift from traditional single-architecture procurement, giving enterprise customers access to different processor designs within the same cloud environment.
Instead of restricting developers to standard graphics processing units, Oracle is opening its platform to dedicated inference processors, wafer-scale systems, and custom silicon originally built for rival hyperscale clouds. The move addresses growing corporate demand for alternative hardware as AI workloads diversify between resource-heavy model training and cost-sensitive real-time execution.
By bringing different architectures into its data centres, Oracle wants to prevent corporate lock-in and keep enterprise clients on its networking fabric. Customers will be able to run distinct parts of their AI software pipelines across specialised chips while keeping their underlying datasets hosted in Oracle databases.
Google TPUs and Cerebras arrive on OCI
The most unusual addition to OCI is Google’s proprietary Tensor Processing Unit family. Under a new partnership, Oracle and Google will deploy TPUs in Oracle data centres beginning in 2027, with installed capacity scheduled to expand over several years.
The agreement allows Oracle customers to rent Google’s custom machine learning silicon directly through OCI to train and run models without opening separate Google Cloud accounts or moving data between rival cloud networks. It represents a rare licensing of Google’s internal silicon to a direct cloud competitor.
Oracle is pairing the Google partnership with an agreement with Cerebras Systems, also slated for deployment in 2027. Cerebras manufactures wafer-scale processors that place compute cores and memory across an entire silicon wafer rather than cutting it into individual dies. This layout provides high memory bandwidth, reducing latency when serving generative models.
The Cerebras integration is aimed specifically at inference workloads where rapid response times are mandatory. Oracle said the service will focus on businesses building autonomous software agents and production applications that trigger repeated, multi-step calls to large language models throughout the day.
Inference capacity will also expand through hardware from SambaNova Systems. Oracle confirmed it will deploy SambaNova SN50 systems to OCI, providing another non-GPU option tuned for high-throughput model execution.
Nvidia Vera Rubin and enterprise processors
Alongside alternative accelerators, Oracle is maintaining its heavy commitments to Nvidia. The company said it plans to expand its Nvidia Vera Rubin deployments to multi-gigawatt capacity across its data centre footprint.
Oracle now offers Vera Rubin NVL72 rack systems alongside Vera central processing units on OCI. The company designed these clusters to manage both large-scale model training and high-volume inference. Oracle noted that the accompanying CPUs carry critical intermediate tasks, such as parsing enterprise code, calling external APIs, and retrieving records from relational databases between neural network cycles.
Traditional host compute is receiving an overhaul as well. Oracle announced a multiyear collaboration with Intel to build OCI X14 compute shapes powered by Intel’s upcoming Diamond Rapids server processors. These chips will handle core enterprise enterprise resource planning software, database management, and hybrid AI tasks. Oracle has not yet published an exact launch date for the X14 instances.
On the AMD side, Oracle is readying OCI Compute E7 instances. These instances will run on AMD’s next-generation EPYC server processors, codenamed Venice, providing higher core densities for cloud-native software and memory-intensive databases.
OpenAI accelerators and Snowflake integration
Oracle also broadened its commercial relationship with OpenAI. The two companies introduced integrations connecting enterprise databases and enterprise applications directly with ChatGPT, supported by three new software products: Oracle AI Workspace, App Composer, and AI Security Mesh.
On the hardware side, Oracle plans to deploy OpenAI’s custom Jalapeño inference accelerators. Developed in partnership with Broadcom, the Jalapeño chips are purpose-built to execute OpenAI models efficiently at scale, giving Oracle early access to custom silicon designed by the artificial intelligence laboratory.
To support data pipelines feeding these models, Oracle announced native availability of Snowflake’s AI Data Cloud on OCI. Enterprise clients will be able to buy, provision, and operate Snowflake services natively through Oracle’s billing and control console, co-locating data queries with the newly announced accelerator clusters.
The rollouts will begin in phases, with immediate availability for selected Nvidia Vera systems, followed by server updates from Intel and AMD, before Google TPUs and Cerebras wafer hardware enter commercial deployment on OCI in 2027.
