Oracle Widens OCI Accelerator Menu: Google TPUs, Cerebras, SambaNova

Google's Tensor Processing Units will run on Oracle Cloud Infrastructure next year — the first time Google's AI chips have appeared in any hyperscaler's cloud other than Google's own. Oracle announced the addition at its AI World 2026 show in Las Vegas this week, alongside Cerebras and SambaNova hardware, as it scrambles to keep pace with Microsoft, Google and Amazon Web Services.

The hardware news arrives against a striking financial backdrop. Oracle's revenue rose from $57.4 billion to $67.4 billion between fiscal 2025 and 2026, and its OCI business hit $7.4 billion in the first quarter of fiscal 2027, a 121 percent year-over-year jump. Remaining performance obligations — contracted sales not yet booked as revenue — reached $664 billion, up $209 billion over the previous year.

The spending side is equally large. Capital expenditures hit $55.7 billion in fiscal 2026 and could exceed $90 billion this fiscal year, all while public resistance to datacenter buildouts grows. Oracle is not alone in the rush, though analysts frequently cite it as the cautionary example.

What new silicon is Oracle offering?

The accelerator expansion gives enterprise and HPC customers more hardware options for inference-heavy AI workloads arriving in OCI during 2027:

• Google TPUs, aimed at large language model training, post-training, and inference for reasoning and agentic AI. Google released its eighth-generation TPU in April, splitting the line into the TPU 8t for training and TPU 8i for inference.
• Cerebras CS-4, a rack-scale accelerator introduced in August, built on three Wafer Scale Engine 3 Turbo chips. Cerebras claims up to 30 times the tokens-per-second-per-user speed of GPUs.
• SambaNova SN50, fifth-generation Reconfigurable Dataflow Units released in February. They are designed to work alongside GPUs, with GPUs handling prefill tasks and RDUs handling decode.

Nathan Thomas, senior vice president and product manager for OCI, framed the strategy as deliberate breadth. "We want to enable that, so we partner very deeply with who we view are best-in-breed in markets," he told The Next Platform before the show. "We don't want to put ourselves in a position where we're dictating any of that stuff or trying to cut off any options for those customers."

Oracle works deeply with both Nvidia and AMD on chipsets, he noted. That is roughly the opposite of lock-in.

What is the Deep Trust AI Architecture?

On Monday, Oracle unveiled Deep Trust AI Architecture, which will be embedded in its AI Database, Autonomous AI Database and APEX AI Application Generator. The control layer is designed to protect enterprise data, enforce business rules, ensure accurate AI answers and govern how AI interacts with corporate environments — placed where the data lives rather than inside each application.

It dovetails with Oracle's Maximum Availability Architecture, which guides enterprises toward systems that survive failures.

"Deep Trust AI Architecture applies the same principle to AI: build trust into the architecture from the start, instead of trying to bolt it on later," wrote Juan Loaiza, executive vice president of Oracle database technologies, and Maria Colgan, vice president of product management. "Trust should be built into the architecture."

How does OpenAI fit in?

Oracle is also expanding its OpenAI partnership, linking the vendor's models, ChatGPT and the Codex developer tool with Oracle applications, data and cloud infrastructure. ChatGPT plug-ins in Oracle AI Workspace will connect OpenAI tools to Fusion applications, NetSuite, AI Database and OCI, making them accessible to agents and ChatGPT users alike.

In the Fusion portfolio, Oracle unveiled 25 new agentic applications — spanning HR, ERP, supply chain and customer management — for tasks such as scheduling deliveries, service evaluations and payroll. Fusion Companion lets users work with Fusion Applications through ChatGPT and other tools, automatically selecting the agents, workflows and approvals needed to complete a request. A public Git repository, the Fusion Setup and Configuration Registry, turns Oracle's implementation best practices into reusable skills for setup agents.

Co-CEO Mike Sicilia offered the philosophy behind it in his keynote. "Instead of navigating applications and managing processes, you simply tell the system what you'd like to accomplish," he said. "Because it understands your business, it can figure out how to get the work done."

The bet, in short: enterprises will run AI on whatever silicon suits them, provided the trust layer and the data stay in Oracle's cloud. Whether a $90 billion capex bill supports that wager remains the industry's favorite open question.