--- type: "post" id: "google-and-oracle-to-offer-google-tpus-on-23f0b0" slug: "google-and-oracle-to-offer-google-tpus-on-23f0b0" title: "Google and Oracle to offer Google TPUs on Oracle Cloud Infrastructure from 2027; TorchTPU open-sourced" description: "Google and Oracle will deploy Google TPU capacity on Oracle Cloud Infrastructure starting in calendar year 2027, making OCI the first hyperscaler outside Google to offer TPUs." retrieval_nugget: "Google and Oracle will deploy Google TPU capacity on Oracle Cloud Infrastructure starting in calendar year 2027, making OCI the first hyperscaler outside Google to offer TPUs. Google also made TorchTPU, its native PyTorch stack for TPUs, open source. For AI builders, the announcement pairs a future cloud-location option with a lower-friction software path to TPU adoption." published_at: "2026-10-27" updated_at: "2026-10-27" record_date: "2026-10-27" date_kind: "published_at" topics: ["google"] entities: ["Google Cloud","Oracle","Oracle Cloud Infrastructure","Google Tensor Processing Units","TorchTPU","PyTorch","Oracle Acceleron"] related_patterns: [] novelty_label: "new_fact" what_changed: "Google and Oracle plan to offer Google TPUs on OCI starting in 2027, and TorchTPU is being open-sourced—extending the prior TorchTPU/PyTorch story to a new cloud provider." prior_refs: ["memory:fa2d69eb-af93-86ea-a78a-6019794280c9","memory:20cf8138-48ea-88de-a3ec-5e09833b5963","memory:1c1a2ca9-3523-841e-ae34-361f077ec0a7"] source_url: "https://cloud.google.com/blog/products/infrastructure/bringing-google-tpus-to-oracle-cloud-infrastructure" source_urls: ["https://cloud.google.com/blog/products/infrastructure/bringing-google-tpus-to-oracle-cloud-infrastructure"] schema_version: "newruntime-agent-readable-v0.2" stable_id: "post:google-and-oracle-to-offer-google-tpus-on-23f0b0" status: "published" visuals: [] editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]} routes: {"html":"https://newruntime.com/posts/google-and-oracle-to-offer-google-tpus-on-23f0b0/","markdown":"https://newruntime.com/posts/google-and-oracle-to-offer-google-tpus-on-23f0b0.md","json":"https://newruntime.com/posts/google-and-oracle-to-offer-google-tpus-on-23f0b0.json"} --- # Google and Oracle to offer Google TPUs on Oracle Cloud Infrastructure from 2027; TorchTPU open-sourced ## Retrieval answer Google and Oracle will deploy Google TPU capacity on Oracle Cloud Infrastructure starting in calendar year 2027, making OCI the first hyperscaler outside Google to offer TPUs. Google also made TorchTPU, its native PyTorch stack for TPUs, open source. For AI builders, the announcement pairs a future cloud-location option with a lower-friction software path to TPU adoption. ## Summary Google and Oracle will deploy Google TPU capacity on Oracle Cloud Infrastructure starting in calendar year 2027, making OCI the first hyperscaler outside Google to offer TPUs. Google also made TorchTPU, its native PyTorch stack for TPUs, open source. For AI builders, the announcement pairs a future cloud-location option with a lower-friction software path to TPU adoption. ## What happened Google Cloud and Oracle announced a partnership to bring Google Tensor Processing Units to Oracle Cloud Infrastructure. Deployment is planned to start in calendar year 2027. Google describes OCI as the first hyperscaler outside Google to offer TPUs. The companies position the arrangement as an additional compute choice for organizations whose AI workloads need to balance throughput, latency, and unit economics across training and inference. The announcement is not a general availability notice. It does not name regions, prices, reservation terms, instance shapes, or the particular TPU generation that OCI customers will receive. Its concrete commitment is future TPU capacity on Oracle infrastructure. Google says the systems are designed for direct liquid cooling at rack scale and that OCI's next-generation data centers are built for that profile. Oracle will provide the surrounding OCI networking layer through Acceleron, including a SmartNIC with hardware-level tenant isolation, High Performance Fabric, and Zero Trust Packet Routing. Google's technical case is that TPU economics come from a co-designed system rather than chip specifications alone. The post says eighth-generation TPUs deliver up to 12.6 PFLOPS of native FP4 compute per chip, compared with TPU v1's 92 TFLOPS, and up to twice the performance per watt of the preceding generation. It says on-die Inter-Chip Interconnect routing and Optical Circuit Switching can route around hardware or link degradation without restarting a job, sustaining more than 97% goodput across thousands of chips in large training runs. It also claims that more on-chip SRAM and HBM, plus an interconnect topology with more than 50% fewer network hops, provide up to 80% better performance per dollar for low-latency serving. The software part is available now: TorchTPU, Google's native PyTorch stack for its TPUs, is open source on GitHub. Google says developers can train, serve, and deploy models on TPUs with vLLM, SGLang, llm-d, TorchTitan, Ray, TorchRec, Hugging Face, and LiteRT. The stated objective is portability: teams should be able to place a workload on the accelerator that offers the desired cost per token or time to train without rewriting model code. Once OCI capacity arrives, Google says that placement can include Google Cloud or Oracle infrastructure. Oracle's enterprise-data argument is also central. Because Oracle databases are often systems of record, the companies say TPU inference and fine-tuning can run near existing enterprise data, avoiding cross-cloud transfers. OCI will offer TPUs alongside GPUs, leaving customers to assign stages of a workload to different accelerator types. ## What the archive adds This is a new fact on a known TPU software story, rather than a restatement. Our earlier Field Note, [Google and Meta Bring Native PyTorch Support to TPUs](https://newruntime.com/posts/google-and-meta-bring-native-pytorch-support-to-3b9b77/), covered the native PyTorch direction and TorchTPU context. The Oracle partnership adds a separate distribution channel: planned TPU capacity in OCI from 2027. The open-source status of TorchTPU is also material because it makes the software layer inspectable and reusable outside a single managed-cloud product surface. ## Trend signal No supplied pattern is a genuinely close fit: this is an accelerator availability and software-portability announcement, not evidence that agent protocols, runtime boundaries, or completed-work economics have changed. ## Why now Google frames the timing around demand for AI compute exceeding the footprint of any one cloud provider. The partnership also follows Google's recent PyTorch ecosystem announcements, which the company presents as an effort to make TPU use more open and developer-friendly. The two moves are mutually reinforcing: software portability matters more when the same accelerator family is intended to appear in more than one cloud environment. ## Connections to other events The closest archive connection is the earlier Field Note on native PyTorch support for TPUs. That coverage concerned the developer stack; this announcement extends the potential infrastructure footprint. The archive also contains an earlier Telegram post about OpenAI becoming available through Oracle Cloud Marketplace. That was a commercial procurement route for AI services, whereas this announcement concerns Oracle hosting Google-designed accelerator capacity; they should not be treated as the same product or access model. ## Why it matters Teams should separate the immediate and future implications. Immediately, an open-source native PyTorch stack gives TPU evaluation a clearer path for workloads already built around common PyTorch, serving, and orchestration tools. That does not prove compatibility or performance for any particular model, but it reduces the prospect of a wholesale code rewrite as the first evaluation step. The OCI deployment matters later for teams with data, networking, identity, or procurement concentrated in Oracle. If the promised capacity and integrations arrive, those teams may be able to evaluate TPUs without first moving the relevant workload into Google Cloud. More broadly, the announcement makes accelerator selection look increasingly like a workload-placement decision: training goodput, inference latency, cost per token, data location, and operational constraints may matter more than allegiance to a single hardware vendor. ## What happens next - **Google and Oracle will disclose OCI TPU regions, configurations, or commercial access details.** (12 months, 70%). Watch for: an OCI service page, regional capacity list, or pricing documentation. - **TorchTPU will accumulate clearer compatibility guidance for the named PyTorch ecosystem tools.** (6 months, 65%). Watch for: release notes, supported-version matrices, and reproducible examples from Google or tool maintainers. ## Limits and uncertainty All performance, efficiency, goodput, and availability statements in this note originate with Google and Oracle's joint announcement; it provides no independent benchmark results, customer case study, or workload-level cost comparison. The post does not establish whether TorchTPU feature parity matches other TPU software paths, how easily existing PyTorch applications will migrate, or which models will benefit. It also does not state OCI prices, availability regions, capacity volumes, service-level commitments, data-residency terms, or a precise launch date beyond calendar year 2027. The future cloud offer should therefore be read as a partnership commitment, not as currently purchasable TPU access. ## Skill to build now Make accelerator assumptions explicit in one existing PyTorch workload. Record the model version, precision, batch or concurrency profile, serving framework, data location, latency target, training-time target, and cost measure before testing another platform. Then check whether its dependencies map to the tools Google lists for TorchTPU, such as vLLM, SGLang, Ray, or Hugging Face. This creates a portable evaluation brief now, rather than a hardware-specific migration project later. ## Reader takeaway TPUs are not available in OCI yet, but Google and Oracle have committed to deploy them there from 2027, and Google has opened the PyTorch layer intended to make evaluation less proprietary. Treat the announcement as a reason to preserve portability in accelerator-facing software, not as evidence that OCI TPU economics or availability are already known. ## Open question Will Oracle disclose enough pricing, regional capacity, and operational detail for enterprise teams to compare OCI TPUs fairly with GPUs and Google Cloud TPUs before deployment begins? Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json ## Continue in the graph Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`. - topic hub: [Google](https://newruntime.com/topics/google.md): Explore the google topic hub. - collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis. - collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph. - whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)