How to read this page. Each section starts Simple, then goes to a Deep dive: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where the maker has not published a figure -- a price, a die size, a factory address -- this page says so rather than estimate it.
1.At a glance
SimpleStart here
TPU 8t is the training half of Google’s eighth TPU generation, announced 22 April 2026 at Google Cloud Next alongside an inference-focused sibling, TPU 8i 1. Google says both chips “will be generally available later this year” 1, and as of this check its own product page still lists them as “Coming Soon” 2 — so this page describes a chip Google has specified in full, not one already in customers’ hands.
Deep diveThe technical detail
Designed in partnership with Google DeepMind 1. Splitting a generation between training and inference parts is not new for Google: TPU v1 was an inference-only chip (see its page on this site), and v5e and v5p ship as two separately-priced chips 2. What Google does claim as a first for this generation is narrower — “for the first time, both chips run on Google’s own Axion ARM-based CPU host” 1.
2.Launch and history
SimpleStart here
Google split its eighth generation because training and inference have started pulling in different directions: training needs raw, brute-force compute, while serving AI agents at scale needs memory bandwidth and low latency above all else — one chip design was starting to compromise on both.
Deep diveThe technical detail
Google’s own stated rationale: rising inference demand “as frontier AI models are deployed in production and at scale,” where “interactions between agents at scale magnify even small inefficiencies” 1. TPU 8t is the side of that split built for the first half of the problem: training the models in the first place.
3.What’s inside it
SimpleStart here
216GB of HBM memory and 128MB of on-chip fast memory per chip, built for raw training throughput at superpod scale 3.
Deep diveThe technical detail
Google’s technical deep-dive gives per-chip figures separately from the pod-level marketing numbers: 216 GB of HBM capacity, 6,528 GB/s of HBM bandwidth, 128 MB of on-chip SRAM (“Vmem”), and 12.6 petaflops of peak FP4 compute per chip 3. Google has not published a process node or foundry for TPU 8t in any official source found; this page does not guess one.
4.Full spec table
SimpleStart here
A full TPU 8t superpod scales to 9,600 chips, delivering 121 exaflops of compute and 2 petabytes of shared memory 1.
Deep diveThe technical detail
| Spec | Google TPU 8t |
|---|
| Peak compute (fp4, per chip) | 12.6 PFLOPs |
|---|
| Memory (per chip) | 216 GB HBM, 6,528 GB/s bandwidth |
|---|
| On-chip memory (per chip) | 128 MB SRAM (“Vmem”) |
|---|
| Superpod scale | Up to 9,600 chips |
|---|
| Superpod compute | 121 exaflops |
|---|
| Superpod memory | 2 petabytes shared HBM |
|---|
| Reliability | >97% “goodput” (Google’s own uptime/efficiency metric) |
|---|
Per-chip figures per Google’s technical deep-dive 3; pod-level figures per Google’s launch blog 1. These are measured at different levels (single chip vs. full superpod) and should not be read as directly comparable to each other. Google’s product page separately states TPU 8t delivers roughly 3 times the compute per pod of the previous generation 2.
5.Where it’s made
SimpleStart here
Google has not disclosed who manufactures TPU 8t or on what process node. No official source confirms this yet.
Deep diveThe technical detail
Broadcom’s April 2026 SEC filing describes a long-term agreement to develop and supply custom TPUs for Google’s “future generations,” running through up to 2031, without naming TPU 8t specifically or disclosing a process node 4. Press reports mentioning MediaTek in connection with this generation are not corroborated by any official Google or Broadcom statement this page could confirm, and are not repeated here as fact.
6.Which systems use it
SimpleStart here
Google names one early customer voice in its launch announcement, Citadel Securities, though the surrounding context does not make clear whether that quote applies to TPU 8t specifically or to the eighth generation broadly 1.
Deep diveThe technical detail
As of this check, TPU 8t remains listed as “Coming Soon” on Google Cloud’s own product page, with no region or zone availability published 2.
7.Official pricing
SimpleStart here
No price has been published. TPU 8t remains listed as “Coming Soon” on Google Cloud’s own product page 2.
Deep diveThe technical detail
Google Cloud’s pricing page, checked September 2026, lists no TPU 8t entry at all. This page reports that plainly rather than estimating a rate from earlier generations.
9.What came before, what came next
SimpleStart here
Came before: Ironwood (TPU7x). Comes after: not yet named — Google’s own product page lists TPU 8t and TPU 8i as its newest generation, with no ninth generation announced 2.
Deep diveThe technical detail
TPU 8t succeeds Ironwood as the training-focused half of Google’s eighth generation, with TPU 8i as its inference-focused sibling, covered on its own page. As of this check, Google has not announced anything beyond the eighth generation.
10.Hidden in plain sight
SimpleStart here
Google announced this generation’s two chips together, and named each for the job it is meant to do — a training chip and an inference chip, specified side by side on the same day 1.
Deep diveThe technical detail
Announcing a matched training/inference pair at once is the clearest statement Google has made that the two workloads have diverged enough to justify separate silicon 1. It is a change of emphasis rather than a first: Google built an inference-specific TPU before (v1 itself was inference-only, covered on its own page on this site), and v5e and v5p already sat side by side as two separately-priced chips in one generation 2. What is new here is that both halves were named, specified and positioned as a deliberate pair from the outset.