Oracle AI Chips Expand: Google TPUs, Nvidia Rubin and Cerebras Explained

October 28, 2026 WIS Editorial

Oracle AI chips became a major technology story on October 27, 2026, when the company announced an expanded Nvidia Vera Rubin deployment alongside plans to offer Google’s Tensor Processing Units (TPUs), Cerebras inference hardware and SambaNova systems. Unveiled at Oracle AI World in Las Vegas, the agreements suggest that the future of cloud artificial intelligence may depend on several processor designs rather than a single dominant supplier.

Oracle is selling flexibility: customers train advanced models, serve millions of responses and run software agents that repeatedly call other tools. Those jobs do not necessarily need identical hardware. Crucially, the announcements cover different timelines. Some Nvidia products are available now, but Google TPU and Cerebras services are planned for 2027. Reuters independently reported the developments and raised questions about financing.

Why It Spiked

Three announcements converged on one day. Nvidia and Oracle outlined a plan to expand Vera Rubin superclusters to multi-gigawatt scale. Google and Oracle said TPUs would be offered on Oracle Cloud Infrastructure (OCI), the first hyperscale cloud outside Google’s own to offer them. And Oracle described future inference options from Cerebras and SambaNova, which compete for work after an AI model has been trained.

The story resonates because companies are questioning the cost of AI infrastructure. Nvidia remains crucial to large training systems, yet response speed, energy efficiency, software compatibility and cost per answer increasingly matter to customers. More suppliers could improve bargaining power and hardware choice. It does not follow that every new chip will outperform Nvidia or that a proposed service is already operating.

Oracle AI Chips: What Was Announced on October 27

At the Las Vegas conference, Nvidia said Vera Rubin NVL72 systems and Vera CPUs were generally available on OCI. The partners also outlined future HGX Rubin NVL8 configurations and much larger planned clusters. A multi-gigawatt target describes intended power-scale capacity; it does not specify a completed GPU count, signed contract value or construction timetable.

Oracle separately announced that Google TPUs would be offered in calendar 2027, with deployment expanding over multiple years. Its infrastructure overview named planned SambaNova SN50 and Cerebras wafer-scale inference services. Reuters also reported that Oracle intends to use its own Acceleron networking in some Nvidia-powered installations. The infrastructure competition therefore includes the connections between processors, not merely the processors themselves.

Why Nvidia Remains Central

Vera Rubin NVL72 links many GPUs in a rack-scale design intended for demanding model training and inference. Nvidia’s announcement emphasized that AI agents also perform CPU-intensive tasks, such as running code, retrieving data and invoking tools between model calls. Oracle says customers can scale CPU and GPU resources separately, potentially reducing expensive accelerator idle time.

That is a plausible engineering benefit, but it is not a guarantee of savings for every workload. Electricity, cooling, software, utilization and the price of capacity affect total cost. Nvidia and Oracle did not announce the monetary value of their multi-gigawatt expansion. It would be misleading to multiply a supplier’s illustrative revenue-per-gigawatt estimate into a supposed signed deal.

Nor is Oracle abandoning Nvidia. The new arrangements make more sense as diversification around an existing anchor supplier. Training very large models remains a different problem from serving small, frequent responses at low latency.

Google TPUs Offer Another Cloud Route

Google’s TPUs are purpose-built machine-learning accelerators. Oracle said its offering would include Google’s eighth-generation TPU 8t for training and TPU 8i for inference. The partners argue that customers could choose hardware based on throughput, power consumption and cost. These are vendor performance claims, and actual comparisons depend on software, models, batch sizes and operating conditions.

Making TPUs available through OCI could appeal to organizations already using Oracle databases or enterprise systems that want to test Google’s silicon without moving their entire cloud environment. It also offers Google another distribution channel. But the joint announcement explicitly targets 2027 availability and a multiyear ramp. Pricing, geographic availability and independent benchmarks remain unknown.

Cerebras and SambaNova Target Inference

Training produces a model; inference produces the answers people actually see. A complex AI agent may make repeated model calls while checking data or executing tasks. That makes response latency a commercial issue, not just a laboratory benchmark. Oracle’s Cerebras partnership announcement says it plans to offer wafer-scale inference in 2027, using a processor design intended to reduce data movement.

Oracle also identified SambaNova SN50 systems among its planned options. Reuters quoted Cerebras chief executive Andrew Feldman describing the agreement as an opportunity to reach more large customers. His statement expresses commercial expectations, not a verified count of deployments. Oracle has not released a consistent independent price-performance comparison across Cerebras, SambaNova, Google and Nvidia.

Buyers should evaluate reliability, security, software support and portability as well as speed. A faster chip is not necessarily cheaper if migrating applications requires extensive engineering work.

What the Expansion Means for AI Spending

Financing may prove as important as processor selection. Reuters reported that Oracle maintained its fiscal 2027 capital expenditure target of roughly $90 billion to $95 billion, while leaving the funding arrangements for these chip purchases unspecified. The company has said customers are increasingly paying for chips, but that does not mean every planned deployment has secured financing.

Building AI capacity also requires electricity connections, land, buildings, cooling and networking. Announced spending is not revenue, and capacity targets are not completed installations. Investors will ultimately need evidence of utilization, durable customer demand and margins that cover operating and financing costs.

The issue resembles the questions raised in WhyItSpiked’s analysis of Meta’s proposed $95 billion AI financing plan. The companies have different businesses, yet both illustrate why the market is scrutinizing how AI infrastructure gets paid for rather than treating every expansion as guaranteed profit.

Chronology: What Is Live and What Comes Later

Before October 27: Oracle already offered Nvidia-based cloud infrastructure and had outlined substantial capital spending. Its AI business was expanding before this week’s announcements.

October 27, 2026: At Oracle AI World in Las Vegas, the company announced the expanded Nvidia relationship, Google TPU agreement and new inference options. Reuters published independent reporting that day.

Calendar 2027: Oracle expects Google TPU and Cerebras inference availability, with the Google rollout growing over several years. Those dates are forward-looking, not evidence that all systems are deployed or that every customer can order them today.

What to Watch Next

Watch for exact launch dates, cloud regions, pricing and independent benchmarks for Google, Cerebras and SambaNova offerings. Monitor whether Oracle provides firm financing details and reports customers using the new hardware at scale. The balance between training and inference demand will determine which processor designs win business.

For now, the supported conclusion is narrower than either an Nvidia-collapse narrative or a guaranteed Oracle success story: OCI has announced a broader computing menu, with meaningful parts still scheduled for the future. Its commercial payoff remains to be demonstrated.