Oracle Cloud will be the first hyperscaler outside Google to offer TPUs, with capacity starting in 2027

Google Cloud and Oracle have announced a partnership to bring Google's Tensor Processing Units (TPUs) to Oracle Cloud Infrastructure (OCI). According to the announcement, this makes OCI the first hyperscaler outside Google to offer TPUs. Deployment of TPU capacity on OCI will begin in calendar year 2027. The announcement appeared on the Google Cloud blog, dated October 27, 2026, and was signed by Karan Batta (Senior Vice President, OCI) and Gurmeet Goindi (Sr. Director of Product Management).

The product thesis the text puts forward is co-design: Google built its first TPU more than ten years ago on the premise that designing silicon together with the optical network, power delivery, liquid cooling and compiler software yields gains in performance per watt and total cost of ownership that standalone components cannot achieve. According to the article, TPU v1 delivered 30 times the performance per watt of contemporary processors. Today, it says, TPUs train and serve Gemini across Google products with billions of users.

On the eighth generation, the announcement cites up to 12.6 PFLOPS of native FP4 compute per chip, more than 100 times TPU v1's 92 TFLOPS, and up to twice the performance per watt of the previous generation. These are the manufacturer's own figures and have not been independently verified.

The text argues that raw FLOPS do not determine customer economics. For training, the key metric it highlights is goodput, meaning the percentage of time a cluster spends on productive compute. For inference, the metrics are latency and cost per token. For training, Google describes combining on-chip ICI (Inter-Chip Interconnect) routing with optical circuit switching (OCS): the control plane detects hardware or link degradation and reroutes optical paths around failures in real time, without restarting jobs. It says this maintains more than 97% goodput across thousands of chips in large-scale training runs, along with high availability in production serving.

For inference, the company credits the latest generation with up to 80% better performance per dollar for low-latency serving. It attributes this to more on-chip SRAM, greater HBM capacity and a new low-diameter interconnect topology that cuts network hops by more than 50%.

The article devotes a section to software. It mentions recent announcements at PyTorch Conference and notes that TorchTPU, the native PyTorch stack for Google's TPUs, is now available as open source on GitHub. According to the text, it lets users train, serve and deploy models on TPUs with tools such as vLLM, SGLang, llm-d, TorchTitan, Ray, TorchRec, Hugging Face and LiteRT. The argument is portability: assigning each workload to the accelerator with the lowest cost per token or shortest training time without rewriting model code, and soon being able to choose between TPUs on Google Cloud or on Oracle infrastructure.

The announcement lists four benefits of bringing TPUs to OCI. First, data centers designed for AI density: TPU systems are built for rack-scale direct liquid cooling, and OCI's next-generation data centers are being built for that profile. Oracle, it says, offers flexible, durable capacity that can shift from frontier model training to production inference as demand changes.

Second, secure, high-performance networking with Oracle Acceleron. While TPUs move chip-to-chip traffic over ICI and OCS, Acceleron would connect TPU capacity to the rest of OCI through SmartNICs with hardware-based tenant isolation, High Performance Fabric and Zero Trust Packet Routing, moving data between Oracle databases, storage and TPU clusters with high throughput and low latency without consuming host compute.

Third, proximity to enterprise data. Oracle databases are often organizations' system of record, and offering TPUs within OCI would allow fine-tuning and inference alongside that data without costly cross-cloud transfers. Fourth, multi-architecture choice: having TPUs alongside GPUs within OCI to assign each phase of an AI workload to the accelerator that is optimal in cost and performance.

The text also acknowledges that, although TPU adoption on Google Cloud is growing, global demand for AI compute exceeds the capacity of any single provider. It concludes that with more specialized compute options across more clouds, the whole industry builds faster and operates more efficiently.

It should be read with caution: this is a joint announcement by the partners themselves, with no pricing, locations, capacity volumes or contractual details, and availability is limited to a generic "starting in 2027." The performance figures are Google's claims. What matters strategically is that TPUs, until now tied to Google Cloud, will be offered on another large-scale cloud, expanding the alternatives to GPUs and bringing AI compute closer to corporate data hosted on Oracle.