On this page · 13 sections
- What Oracle and Google announced on 27 October 2026
- TPU 8t and TPU 8i: the specs that matter for inference
- Where TPUs stand today: Google Cloud availability and price
- Framework support: JAX, PyTorch, vLLM and SGLang on TPU
- When TPUs beat GPUs for inference
- When GPUs still win
- Worked cost example: the throughput a TPU must hit
- Lock-in and exit
- India, EU and US notes
- Decision table: pick the right accelerator
- Pilot checklist before you commit
- How eCorpIT can help
- References
Summary. On 27 October 2026 Oracle and Google announced that Google's Tensor Processing Units will be available on Oracle Cloud Infrastructure in calendar year 2027, ramping over several years, and the companies say OCI will be the first hyperscaler outside Google to offer them. OCI supplies the power, cooling and data centre networking. Neither company has named the TPU generation, regions, shapes or prices for OCI. Google's eighth-generation TPU 8t (training) and TPU 8i (inference) are still listed as coming soon, and Google claims 8i delivers up to 80% better inference performance per dollar than Ironwood. Today Ironwood costs $12.00 per chip-hour on demand in us-central1 and $5.40 on a 3-year commitment, while OCI lists B200 at $14 and H100 at $10 per GPU-hour. The software moved in October: TorchTPU went open source, and vLLM now reaches TPUs through two plugins, vllm-jax and vllm-torchtpu. Our verdict: TPUs suit large, steady, high-utilisation inference on supported models; GPUs still win for CUDA-bound code, small or spiky traffic, and anything needed before OCI's TPUs arrive.
About this article. For CTOs, ML platform leads and FinOps teams deciding where to run inference in 2027. On 7 November 2026 we checked every announcement, specification, price and software status against Oracle and Google press releases and blogs, Google Cloud's TPU documentation and pricing page, Oracle's price list and price API, and the vLLM project's GitHub repositories. Performance figures are the vendors' own; we have not benchmarked these chips. Prices change, this is not financial advice, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.
What Oracle and Google announced on 27 October 2026
The release was issued at Oracle AI World in Las Vegas and published on both Oracle's newsroom and Google Cloud's Press Corner. It says Google TPUs will be available through OCI, which the companies describe as the first time TPUs will be offered on a hyperscale cloud outside Google. OCI provides the power, cooling and data centre networking. OCI president Mahesh Thiagarajan framed it as offering several AI compute architectures in one cloud.
Google confirmed its side in a Google Cloud blog post the same day, co-written by Oracle's Karan Batta and Google's Gurmeet Goindi: the two will deploy TPU capacity on OCI starting in calendar 2027, close to customers' Oracle databases. Batta's OCI blog says the ramp reaches gigawatt scale over multiple years, and places TPUs beside NVIDIA Vera Rubin NVL72 systems (now generally available on OCI), SambaNova SN50 systems and Cerebras inference planned for 2027.
The table summarises what those four documents do and do not say.
| Question | Status on 7 November 2026 |
|---|---|
| TPUs coming to OCI | Confirmed by Oracle and Google |
| Earliest availability | Calendar year 2027, ramping over several years |
| Scale | Gigawatt scale over multiple years, per Oracle's blog |
| Who provides facilities | OCI: power, cooling and data centre networking |
| Generations described in the release | TPU 8t and TPU 8i, Google's eighth generation |
| Which generation launches first on OCI | Not announced |
| OCI regions, shapes, quotas | Not announced |
| OCI TPU pricing and billing unit | Not announced |
| Support split between Oracle and Google | Not announced |
Oracle's blog carries a future-product disclaimer: release timing and pricing may change at Oracle's sole discretion, and the post should not drive purchasing decisions. Treat 2027 as a target, not a delivery date.
The pattern is broader. On Amazon's third-quarter earnings call, Andy Jassy said Amazon intends to make Trainium available outside AWS late in 2027, and AWS already installs Trainium in customer data centres through AWS AI Factories. Our companion piece covers what on-premises buyers should plan for Trainium outside AWS.
TPU 8t and TPU 8i: the specs that matter for inference
Google's eighth-generation deep dive, published on 22 April 2026, splits the lifecycle: 8t for pre-training, 8i for sampling, serving and reasoning. The figures below come from that post.
| Feature | TPU 8t | TPU 8i |
|---|---|---|
| Primary workload | Large-scale pre-training | Sampling, serving and reasoning |
| HBM capacity per chip | 216 GB | 288 GB |
| HBM bandwidth per chip | 6,528 GB/s | 8,601 GB/s |
| On-chip SRAM | 128 MB | 384 MB |
| Peak FP4 compute per chip | 12.6 PFLOPS | 10.1 PFLOPS |
| Network topology | 3D torus, 9,600 chips per superpod | Boardfly, up to 1,024 active chips per pod, maximum 7 hops |
| Special hardware | SparseCore for embeddings, LLM Decoder Engine | Collectives Acceleration Engine |
| Host CPU | Arm-based Axion | Arm-based Axion |
| Google's claim vs Ironwood | Up to 2.7x training performance per dollar | Up to 80% better inference performance per dollar |
Three 8i choices matter for inference. Three times the previous generation's on-chip SRAM keeps a larger KV cache on silicon, which Google says cuts idle time in long-context decoding. The Collectives Acceleration Engine cuts on-chip collective latency by 5x. Boardfly reduces the worst-case path across a 1,024-chip pod from 16 hops on a 3D torus to 7, which Google says gives up to 50% lower latency for communication-heavy work such as Mixture-of-Experts routing. Google also claims up to 2x better performance per watt than Ironwood for both chips.
These are Google's figures, and Google says the 80% gain applies most at low-latency targets for large MoE models. A dense 8-billion-parameter chatbot is not that workload.
Where TPUs stand today: Google Cloud availability and price
Until OCI capacity arrives, Google Cloud is where you rent TPUs. Google's TPU page lists Ironwood (seventh generation, 9,216 liquid-cooled chips per pod, 42.5 exaflops) and Trillium (sixth generation) as generally available, and TPU 8t and 8i as coming soon. Every Ironwood VM has four chips, so four chips is the smallest unit you can rent.
Cloud TPU is priced per chip-hour, though the console bills in VM-hours. Selected list prices on 7 November 2026, in USD per chip-hour:
| TPU version | Region | On demand | DWS Flex-start | 1-year commitment | 3-year commitment |
|---|---|---|---|---|---|
| Ironwood | us-central1 (Iowa) | $12.00 | $6.00 | $8.40 | $5.40 |
| Ironwood | europe-west2 (London) | $13.20 | $6.00 | $9.24 | $5.94 |
| Trillium | us-east1 (South Carolina) | $2.70 | $1.35 | $1.89 | $1.22 |
| Trillium | europe-west4 (Amsterdam) | $2.97 | $1.35 | $2.08 | $1.34 |
| TPU v5e | us-central1 (Iowa) | $1.20 | $0.60 | $0.84 | $0.54 |
Google prices Ironwood's DWS Calendar mode at $8.40 per chip-hour in both regions, and Spot prices are dynamic and can change up to once every 30 days. The 3-year Ironwood rate is 55% below on demand ($5.40 ÷ $12.00 = 0.45), so commitment drives TPU economics.
OCI prices GPUs per GPU-hour, and a server's hourly price is that rate times its GPU count. On 7 November 2026 Oracle's price API (last updated 2 November) gave pay-as-you-go rates of $14 per GPU-hour for B200, $15 for B300, $16 for GB200, $18 for GB300, $10 for H100 and H200, $6 for AMD MI300X and $8.60 for MI355X. A TPU chip and a GPU are not equivalent units, so compare cost per million tokens at your latency target, not price per accelerator-hour. OCI has not said how it will bill TPUs. For a chip-level comparison, see our Ironwood TPU7x vs B200 inference cost analysis.
Framework support: JAX, PyTorch, vLLM and SGLang on TPU
The software picture changed in October 2026.
JAX. JAX is Google's native TPU framework, compiled through XLA. Google says code that runs on Ironwood scales to the eighth generation, with XLA handling Boardfly routing and CAE synchronisation. Pallas lets you write TPU kernels in Python, and MaxText (pre-training) and Tunix (post-training) are Google's JAX reference libraries.
PyTorch through TorchTPU. TorchTPU registers TPUs through PyTorch's PrivateUse1 interface, so tensors are ordinary PyTorch tensors. It offers three eager modes, uses torch.compile with XLA as the backend, and supports DDP, FSDPv2 and DTensor. Google called native PyTorch support a preview in April. On 20 October 2026, at PyTorch Conference 2026, it released TorchTPU as open source on GitHub, co-developed with Meta's PyTorch team. Its documentation now says TorchTPU replaces PyTorch/XLA and describes dynamic shape bounds in torch.compile to reduce recompilation. In April Google had named recompilation from changing sequence lengths and batch sizes as a main open problem. Inference shapes vary constantly, so measure compile and recompile time on your own traffic.
vLLM. Google's inference documentation, updated 3 November 2026, says Cloud TPU works with vLLM through two plugins: vllm-jax for JAX models and vllm-torchtpu for PyTorch models.
- vLLM JAX is the renamed tpu-inference project (Apache-2.0 licence), now moving to a new repository. It recommends v7x (Ironwood), v5e and v6e (Trillium). Its release support matrix, last updated 22 October 2026, marks unit, correctness and performance tests as passing for Gemma 4, Llama 3.1 8B, Llama 3.3 70B, Qwen3-30B-A3B, Qwen3-32B, Qwen3-Coder-480B-A35B and Qwen3.5-397B-A17B. DeepSeek-R1, Kimi-K2.6 and gpt-oss-120b pass unit and correctness tests with performance untested, and DeepSeek-V3.2, GLM-5 and Llama 4 Maverick multimodal are untested.
- vllm-torchtpu, also Apache-2.0, runs vLLM's standard PyTorch model definitions through TorchTPU behind the usual CLI and OpenAI-compatible server. Ironwood is recommended and tested in CI, Trillium is supported without CI, and v5e and v5p are experimental. The first start compiles TPU graphs and can take several minutes.
Neither project lists TPU 8i yet. SGLang has matching backends, sglang-jax and sglang-torchtpu.
Google also notes that many models hardcode attention head dimensions of 64, while current TPUs reach peak matrix efficiency at 128 or 256. A model that runs correctly may still leave performance unused until adapted.
When TPUs beat GPUs for inference
On the vendors' published designs and prices, TPUs are the stronger candidate when most of these hold:
- Large MoE or long-context models at steady load. TPU 8i's SRAM, HBM, CAE and Boardfly target this, and Google frames its 80% claim around large MoE models.
- High, predictable utilisation. Ironwood drops from $12.00 to $5.40 per chip-hour on a 3-year term, which pays only if traffic keeps chips busy.
- Your model is in a tested TPU path. The vLLM JAX passing list, your own JAX code, or a PyTorch model validated on vllm-torchtpu. vLLM keeps your API identical to a GPU deployment.
- You need a large single domain. Ironwood pods reach 9,216 chips, and TPU 8i pods connect up to 1,024 active chips with at most seven hops.
- You want a second source of accelerators. From 2027 OCI adds a non-Google place to rent TPUs.
When GPUs still win
GPUs remain the safer choice when any of these apply:
- Your stack depends on CUDA. Custom CUDA kernels, TensorRT-LLM engines and CUDA-only attention or quantisation libraries do not port to TPUs; they need rewriting in Pallas or replacing.
- Traffic is small or spiky. The smallest Ironwood unit is a four-chip VM, $48 an hour on demand in Iowa (4 × $12.00). OCI's B200 and H100 shapes are eight-GPU servers ($112 and $80 an hour), but GPUs also come in single-GPU sizes, such as OCI's one-GPU A10 VM at $2 an hour. Idle capacity erases any per-token advantage.
- Request shapes vary widely and you serve through PyTorch. Check how the October TorchTPU release handles recompilation on your traffic before you commit.
- Your model is not yet tested on TPU. Budget engineering time before assuming parity.
- You need capacity now. TPU 8t and 8i are coming soon and OCI's TPUs start in 2027, while OCI already offers NVIDIA Vera Rubin NVL72 in general availability.
- You need on-premises or multi-cloud placement beyond Google and Oracle. GPUs run in every major cloud, specialist GPU clouds and private data centres.
The comparison draws on the announcements, Google's TPU and pricing pages, and Oracle's price list.
| Factor | TPU on OCI (from 2027) | TPU on Google Cloud (today) | NVIDIA GPU on OCI (today) |
|---|---|---|---|
| Availability | Calendar 2027, multi-year ramp | Ironwood and Trillium GA; 8t and 8i coming soon | Generally available, including Vera Rubin NVL72 |
| Generations | Not announced for OCI | Ironwood, Trillium, v5p, v5e and older | B200, GB200, GB300, B300, H200, H100 and older |
| Pricing basis | Not announced | Per chip-hour: on demand, DWS, 1-year, 3-year | Per GPU-hour list price |
| Example list price | Not announced | Ironwood $12.00 per chip-hour on demand | B200 $14, H100 $10 per GPU-hour |
| Serving software | Not stated for OCI | JAX, TorchTPU, vLLM JAX, vllm-torchtpu, SGLang | CUDA, TensorRT-LLM, vLLM, SGLang |
| Main lock-in | XLA plus OCI | XLA plus Google Cloud | CUDA plus OCI |
Worked cost example: the throughput a TPU must hit
Without our own benchmarks, the useful arithmetic runs backwards: given a list price, how many output tokens per second must one chip sustain to hit a target cost per million tokens?
Required tokens per second = hourly price ÷ target cost per million tokens × 1,000,000 ÷ 3,600.
| Ironwood pricing option (us-central1) | Price per chip-hour | Tokens/s per chip for $1.00 per million | Tokens/s per chip for $0.50 per million |
|---|---|---|---|
| On demand | $12.00 | 3,333 | 6,667 |
| 1-year commitment | $8.40 | 2,333 | 4,667 |
| DWS Flex-start | $6.00 | 1,667 | 3,333 |
| 3-year commitment | $5.40 | 1,500 | 3,000 |
Worked line: $12.00 ÷ $1.00 × 1,000,000 ÷ 3,600 = 3,333 tokens per second per chip. Beat that on your own model at your latency target and Ironwood on demand serves below $1.00 per million output tokens.
Utilisation changes everything. Those figures assume the chip is busy every hour you pay for. At 40% utilisation the effective on-demand price per busy hour is $12.00 ÷ 0.4 = $30.00, and the 3-year rate becomes $5.40 ÷ 0.4 = $13.50. A committed TPU that sits idle half the day can cost more per token than on-demand capacity you release when traffic drops.
Monthly scale. An 8-chip Ironwood slice (two VMs) over a 720-hour month uses 8 × 720 = 5,760 chip-hours: $69,120 on demand or $31,104 on a 3-year commitment, $38,016 apart. One eight-GPU OCI server at list price for the same month costs 8 × 720 × $14 = $80,640 with B200 or $57,600 with H100. The units differ; what matters is the tokens each delivers.
What TPU 8i's claim would mean. If Google's up-to-80% gain held for your model, the same tokens would cost 1 ÷ 1.8 = 0.56 of Ironwood's cost, about 44% less. Treat that as a ceiling until 8i prices are published and you can test.
The GPU side. At OCI list prices, a B200 at $14 per GPU-hour must sustain 14 ÷ 1.00 × 1,000,000 ÷ 3,600 = 3,889 output tokens per second for $1.00 per million, and an H100 at $10 must sustain 2,778. Our B200 vs H100 cost-per-token guide covers measuring that throughput.
Lock-in and exit
TPUs on OCI weaken the old argument that choosing TPUs means choosing Google Cloud, but the compiler dependency stays: TPU code runs through XLA, as GPU code runs through CUDA. Neither company has said a TPU workload on Google Cloud will move to OCI unchanged, so do not assume parity until OCI publishes images and documentation.
The practical hedge is an engine that runs on both. vLLM runs on GPUs and, through vllm-jax or vllm-torchtpu, on TPUs behind the same OpenAI-compatible server. Keep weights in standard formats, keep accelerator-specific code out of application logic, and keep a GPU deployment of each critical model warm until TPU results are proven. Our guide to autoscaling AI inference with HPA and KEDA applies on either accelerator.
EU buyers should note that the EU Data Act bars cloud switching charges from 12 January 2027. Our EU Data Act switching-fees briefing covers what that removes; read OCI's TPU terms against it once published.
India, EU and US notes
India. Neither company has named regions for TPUs on OCI. Google's TPU price list and its regions and zones page, updated 26 October 2026, show no India region for any TPU version. TPUs rented today run abroad, so where inference touches Indian personal data, DPDP obligations make region and processor terms part of the decision. Our India GPU cloud rental pricing guide covers in-country GPU options.
EU and UK. Ironwood is priced at $13.20 per chip-hour on demand in London against $12.00 in Iowa, and Trillium at $2.97 in Amsterdam against $2.70 in South Carolina. But Google's regions page lists public Ironwood zones only in us-central1, and warns that the price list can show regions without generally available public zones, so check before budgeting for London. Trillium has a public zone in europe-west4.
US. Ironwood's public zones are all in us-central1 (Iowa). Plan reservations: Google's OCI post says global AI compute demand exceeds any single cloud's footprint.
Decision table: pick the right accelerator
| Pick this | If |
|---|---|
| Ironwood on Google Cloud now | You serve a TPU-tested model at high, steady utilisation from us-central1 and can commit for 1 to 3 years |
| Wait for TPU 8i (Google Cloud or OCI) | Your main workload is large-MoE or long-context serving and launch can slip into 2027 |
| TPUs on OCI from 2027 | You already run on OCI and can wait for regions and prices |
| NVIDIA GPUs on OCI or elsewhere | You depend on CUDA kernels or TensorRT-LLM, need capacity now, or need on-premises placement |
| Both, behind vLLM | You want failover and negotiating power across TPU and GPU |
Pilot checklist before you commit
- List every model you serve with parameter count, MoE structure, context length, precision and custom kernels.
- Check each against the vLLM JAX support matrix, or test it on vllm-torchtpu.
- Benchmark on Ironwood or Trillium with production-like prompts; record throughput at your p95 latency target, compile time and recompilation frequency.
- Calculate cost per million output tokens at realistic utilisation, not peak.
- Fix the head-dimension and shape issues Google flags before judging performance.
- Ask Oracle in writing for OCI TPU regions, shapes, billing unit, support model and software images.
- Re-run the benchmark on OCI's actual TPU offering before signing a commitment.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. We benchmark your models on TPUs and GPUs, build vLLM serving that runs on both, and model cost per million tokens at real utilisation before you commit, through our AI and machine learning services. Talk to us at /contact-us/.
Last updated: 7 November 2026.
