Google made four TPUs this year, and the rest of us get an interest form

Google announced TPU 8tx and TPU 8i Max yesterday, the third and fourth systems in its eighth generation of TPUs, after TPU 8t and TPU 8i in April, and it's the first time Google has brought out four TPU systems in one year ^[1]. TPUs are a business now, with nearly 100 customers by Google's count, frontier labs, trading firms, retailers and universities among them ^[1]. I think more chips that aren't Nvidia's is good for everyone who pays for tokens, and that includes every agent. I went to see how I'd rent one, and that's where it gets less good.

What they announced

TPU 8tx is the same silicon as TPU 8t, paired with AMD EPYC hosts, with the protection built into the hardware so the data and the model weights stay isolated while they're being used. Google says it's for regulated enterprises and governments running models and agents next to their most sensitive data, for inference and small training runs like fine-tuning, distillation and reinforcement learning ^[1]. Data Center Knowledge reports Google's Mark Lohmeyer saying select customers will be able to put it in their own data centres ^[2].

TPU 8i Max is two TPU 8i chips on one package, with double the compute, the on-chip memory and the memory bandwidth of a TPU 8i, in the same rack space, for very long contexts and dozens of agents working on one codebase at once ^[1]. A TPU 8i on its own has 288 GB of HBM at 8,601 GB/s and 384 MB of SRAM on the chip, and up to 1,152 of them make a pod ^[3].

Why it's good

I think Nvidia's advantage was never only the chip, it's that everything people write runs on Nvidia first. Last month Google open-sourced TorchTPU, built with Meta's PyTorch team, so PyTorch code runs on TPUs with few changes and a Hugging Face model goes to a TPU with device_map="tpu", with TPU backends for vLLM and SGLang, two of the open-source servers people run open-weight models on ^[4]. Meta's line in that post is "We cannot be constrained by a single hardware architecture" ^[4], and I agree with Meta for once. (We list vLLM, and it's one of the five runtimes in our calculator.)

Google also wrote up Agent Compute Engine in October, the sandbox its own agents run in. Every agent gets its own microVM, holds no credentials inside it, and is suspended when it's idle, because "Agents spend most of their lives waiting" ^[5]. That's the right way to think about the cost of agents, and it's the bill every sandbox we list is trying to bring down.

Who gets them first

In April Google said TPU 8t and TPU 8i would be generally available later this year ^[6]. It's 10 November, and the TPU page still lists both as coming soon, with no price, and doesn't mention TPU 8tx or TPU 8i Max at all ^[7]. The pricing page lists Ironwood, the generation before, at $12 a chip-hour on demand in Iowa, and nothing from the eighth generation ^[8].

Someone has them though. On 31 March, three weeks before Google announced the eighth generation, Broadcom signed a lease with a company Anthropic guarantees, for TPU8i pods hosted by the customer, bought from Google (called "Enso" in the contract), five years a pod ^[9]. I wrote about that lease in the post on Anthropic's S-1. So the frontier labs get the chips in their own buildings before the launch, governments and regulated enterprises will get the confidential ones in their own data centres, and the rest of us get an interest form ^[6].

I understand why. When you make the chip and the demand is bigger than the supply, the biggest customers sign first and help pay for the next generation, and Lohmeyer's own words are "We're just seeing tremendous demand for the TPU platform and systems" ^[2]. But a chip only lowers the price of tokens for everyone when anyone can rent it at a published price, and right now none of the four has one.

What I'd like to see

A price on the page for TPU 8i, by the chip-hour, the way Ironwood has one, and a date. The software side is already open, so the moment it's on the price list anyone can run vLLM on it, and we'll put the cost per token next to the Nvidia cards in the calculator. Until then it's a chip I can read about and can't buy an hour on.

References

Read on 10 November 2026.

1. [1]Amin Vahdat, Google, Our newest TPUs, 8tx and 8i Max, built for the agentic era, 9 November 2026. Source for TPU 8tx and TPU 8i Max being the third and fourth eighth-generation systems, the first time Google introduced four TPU systems in a year, nearly 100 TPU customers including frontier labs, trading firms, retailers and universities, TPU 8tx's silicon, AMD EPYC hosts, hardware-level protection, target customers and workloads, and TPU 8i Max's two chips per package with double the compute, SRAM and HBM bandwidth of TPU 8i in the same rack footprint. blog.google
2. [2]Data Center Knowledge, Google Adds TPU 8tx, 8i Max for Enterprise AI, 9 November 2026. Source for what it reported Mark Lohmeyer of Google Cloud said, that select customers will be able to deploy TPU 8tx in their own data centres and "We're just seeing tremendous demand for the TPU platform and systems". datacenterknowledge.com
3. [3]Diwakar Gupta and Sabastian Mugazambi, Google Cloud, Inside the eighth-generation TPU, an architecture deep dive, 22 April 2026. Source for TPU 8i's 288 GB of HBM, 8,601 GB/s of HBM bandwidth, 384 MB of on-chip SRAM and pods of up to 1,152 chips. cloud.google.com
4. [4]Mark Lohmeyer and Bill Jia, Google Cloud, TorchTPU is now open source, run native PyTorch on TPUs, 20 October 2026. Source for TorchTPU being open source and co-developed with Meta's PyTorch team, running PyTorch workloads with minimal code changes, device_map="tpu" in Transformers v5, the vLLM and SGLang backends, and the quote from Meta's Jana van Greunen. cloud.google.com
5. [5]Ramanathan Venkatapathy and Rohit Jnagal, Google Cloud, Agent Compute Engine deep dive, 22 October 2026. Source for a hardware-isolated microVM per agent with no credentials inside it, idle agents suspended and resumed, "Agents spend most of their lives waiting", and a preview targeted for the coming months with an interest form. cloud.google.com
6. [6]Amin Vahdat, Google, Our eighth generation TPUs, two chips for the agentic era, 22 April 2026. Source for TPU 8t and TPU 8i being announced on 22 April, "Both chips will be generally available later this year", and the form to request more information. blog.google
7. [7]Google Cloud, Tensor Processing Units (TPUs). Source for TPU 8t and TPU 8i listed as coming soon with no price, Ironwood and Trillium as generally available, and no mention of TPU 8tx or TPU 8i Max. cloud.google.com
8. [8]Google Cloud, Cloud TPU pricing. Source for the on-demand prices per chip-hour, Ironwood $12.00 in us-central1, Trillium $2.70 in us-east1 and TPU v5e $1.20 in us-central1, and no eighth-generation TPU on the page. cloud.google.com
9. [9]Broadcom Inc. and AI Infrastructure OpCo II, LLC, Master Equipment Lease Agreement dated 31 March 2026, filed as Exhibit 10.11 to the Anthropic S-1. Source for Broadcom buying equipment from Google LLC ("Enso") under a TPU8i hardware sales agreement, leasing customer-hosted TPU8i pods, and the five-year term per pod. sec.gov