NVIDIA expands on OCI—and a rival accelerator is invited in

On October 27, 2026, Oracle and NVIDIA announced general availability of Vera Rubin NVL72 systems and Vera CPUs on Oracle Cloud Infrastructure. They outlined a future multi-gigawatt expansion and planned HGX Rubin NVL8 systems. General availability today and longer-term capacity plans are different claims.

The same day, Oracle and Google said TPU capacity would become available through OCI starting in calendar 2027. OCI would become the first hyperscale cloud outside Google to offer TPUs. Google's eighth generation includes TPU 8t for training and TPU 8i for inference.

Oracle also plans to offer inference systems from Cerebras and SambaNova. It is building options around different computing workloads, not declaring which semiconductor supplier will ultimately win. None of these future plans should be mistaken for operating capacity or booked sales.

The relationship began in the opposite direction in 2024

On September 9, 2024, Oracle and Google announced general availability of Oracle Database@Google Cloud. Oracle database services operated on OCI infrastructure deployed inside Google Cloud data centers. This was not a sale of Oracle's database business or a wholesale transfer of its customer data to Google.

Customers could use Oracle Exadata and Autonomous Database services alongside Google Cloud applications, analytics and AI with simpler connectivity and procurement. The infrastructure and service ownership remained distinct despite sharing a data-center location.

The arrangement now points in both directions. First, Oracle's database stack entered Google's cloud facilities; next, Google's specialized compute is scheduled to enter Oracle's cloud facilities in 2027. Direct cloud competitors are distributing portions of one another's critical technology because their enterprise customers want both.

Why would Google sell its TPU option through a rival cloud?

Google argues global AI compute demand exceeds the footprint of any single provider. Providing TPU capacity through OCI can expand access to customers whose data, procurement agreements and applications already sit on Oracle's infrastructure.

Oracle gains an accelerator choice without requiring customers to relocate their complete enterprise systems solely to obtain Google silicon. Google may gain another channel for TPU technology without needing every user to migrate to Google Cloud.

The actual revenue split, pricing and operating economics of the partnership have not been disclosed. It does not follow that Google Cloud loses revenue or OCI automatically gains profit. The demonstrated structural change is that silicon supplier and customer-facing infrastructure platform need not be the same company.

Training, inference and agents do not spend compute in the same way

Frontier model training depends on clusters performing calculations and exchanging data at scale. Production inference is sensitive to throughput, latency and cost per token. Agents additionally run code, access data, test outputs and execute tools between model calls.

Oracle and NVIDIA emphasize Vera CPU performance for these CPU-intensive portions of agent execution while accelerators handle model computation. The bottleneck in one application may sit in GPU utilization, another in CPU tool execution, and another in memory or networking.

There is no universal rule that NVIDIA should perform all training or Google all inference. Model architecture, software frameworks, available capacity and unit economics determine the outcome. Giving customers several architectures is a way to preserve choices as workloads change.

Oracle may be competing for the place where enterprise work runs

Consider a company that holds customers, orders and inventory in Oracle databases. Running AI elsewhere can add cross-cloud transfer cost, data-access governance, latency, billing and security review. Bringing TPU and GPU services near that data in OCI may reduce some of those frictions.

Oracle and Google explicitly position the TPU announcement around proximity to Oracle databases and storage. Oracle does not have to manufacture the semiconductor that is optimal for every job if it can connect customers' preferred compute to the applications and data they already operate.

That is a strategic interpretation, not proof of cloud lock-in. Customers retain alternatives, and the economics depend on prices, actual integration quality and contracts. Nonetheless, the commercial battle for the operational environment differs from the engineering contest over which AI chip is fastest.

Oracle Acceleron connects layers; it does not erase NVIDIA networking

Serving several accelerator architectures requires networking and storage that can sustain data movement. Oracle Acceleron combines fabric, SmartNICs, isolation and routing technologies; Oracle also announced new RDMA, distributed communications and failover capabilities on October 27.

But Oracle's own networking layer does not displace every NVIDIA component. Oracle–NVIDIA superclusters combine NVLink, Spectrum-X Ethernet, Spectrum-6 switches and BlueField-4 DPUs with OCI's infrastructure. These technologies operate at different points in the stack.

Google TPUs likewise use inter-chip ICI and optical circuit switching (OCS) for their own cluster communication while Oracle Acceleron connects TPU capacity with OCI databases, storage and front-end networks. Chip-to-chip networking and the cloud platform's data-connectivity layer should not be conflated.

Cerebras and SambaNova make the portfolio more heterogeneous

Cerebras announced plans to supply its wafer-scale inference through OCI in calendar 2027. SambaNova announced a collaboration to deploy SN50 reconfigurable dataflow-based inference infrastructure in OCI data centers; the release did not establish an identical 2027 launch commitment.

In multi-step agent workflows, repeated inference calls accumulate latency. Enterprises may therefore evaluate architectures by token throughput, responsiveness and total operating costs, not only training peak performance.

However, heterogeneous chips are not automatically interchangeable. CUDA, TPU software, Cerebras and SambaNova have different tooling and optimization considerations. Google points to TorchTPU and an expanding open software stack, but every production model still needs application-specific compatibility and cost validation.

Choice expands at a moment of heavy capital expenditure

Supporting more hardware types introduces power, cooling, network, procurement, security and utilization challenges. Multi-vendor supply can create flexibility but may also increase engineering complexity and idle-capacity risk.

At its October 28 Investor Day, Oracle forecast peak capital spending in fiscal 2027–2028, positive free cash flow during the second half of fiscal 2028 and at least $30 billion annual free cash flow in fiscal 2029. These are management projections, not delivered results.

Oracle also outlined a limit of $20 billion in additional capital-market funding over fiscal 2027–2030, excluding refinancing, and cited vendor financing, customer-provided hardware and prepayments as ways to reduce upfront spending. If customer workloads do not fill the new systems, multi-architecture capacity could become a financial burden.

Will customers change accelerators without changing clouds?

Oracle is attempting to sell a stable infrastructure operating relationship even as the customer's optimal accelerator changes. The 2024 database-in-Google partnership and 2027 TPU-in-Oracle plan show that competing platforms can also serve as one another's distribution channels.

Execution may require specialists in accelerator runtimes, distributed networking, heterogeneous scheduling, data security and workload economics. These are capability requirements inferred from the architecture, not a company announcement of additional hiring.

The measurable tests will come after product launches: whether TPUs and Cerebras arrive when planned, whether customers adopt them, whether integration lowers costs and whether spending turns into free cash flow. The winning AI chip vendor and the winning enterprise AI operating platform do not necessarily have to be the same company.

BANSEOG VIEW

BANSEOG VIEW | When customers switch chips, where does their work remain?

Oracle's October 27 announcements do not mean it is abandoning NVIDIA: its Vera Rubin business is expanding. In parallel it wants to offer Google's TPUs beginning in 2027, Cerebras inference and a separate SambaNova option.

In 2024 OCI-based Oracle database services became available within Google Cloud facilities. In 2027 Google's TPU capacity is planned on OCI. Moving a technology closer to an existing enterprise customer may be easier than asking that customer to migrate every system to another cloud.

Neither product announcements nor the breadth of chip choices establish higher customer retention or profitability. Integration quality, real availability, workloads and infrastructure spending will determine whether Oracle's multi-architecture platform strategy works.

SOURCES

Primary sources and references

Published Nov 1 2026. Oct 27 official Oracle/NVIDIA: Vera Rubin NVL72 and Vera CPUs OCI generally available, multi-gigawatt expansion planned, HGX Rubin NVL8 planned. Oracle/Google: eighth generation TPU 8t training and 8i inference planned on OCI starting calendar 2027; not yet GA. Sep 9 2024 Oracle Database@Google Cloud GA: Oracle DB on OCI infrastructure physically inside Google Cloud data centers; not complete DB ownership transfer. Cerebras inference OCI calendar 2027 plan, SambaNova SN50 separate collaboration without confirmed same launch year. Oracle Acceleron cooperates with NVIDIA NVLink/Spectrum-X and Google's ICI/OCS at different layers. Oct 28 Oracle Investor Day: positive FCF targeted second half FY28 and $30B+ FY29, cumulative additional financing limited to $20B FY27–30 excluding refinancing, all forward-looking. Pricing/revenue split, new hardware utilization and personnel changes unverified. Strategic interpretation by Banseog Search.