Google Unveils TPU 8tx and 8i Max: One Chip for Data Encrypted in Use, One for Huge Contexts
Google didn't unveil the fastest chip in the world. It unveiled two TPUs that solve two problems that used to force you out of the public cloud: one encrypts data while it is being processed, the other doubles memory so a million-token context actually fits.
What Google Announced on November 9
On November 9, 2026, Google expanded its eighth-generation TPU lineup with the TPU 8tx and the TPU 8i Max. Neither competes for the biggest number of operations per second: both target use cases that had no good answer in the public cloud, protecting sensitive data during processing and holding huge context windows without running out of memory.
Four Systems in the Eighth Generation
The family now has four members: the 8t for large-scale training, the 8i for low-latency reasoning, the 8tx for confidential computing and the 8i Max for massive contexts and rack density. Worth sorting out early, because much of the coverage mixes up the 8i with the 8i Max and the 8t with the 8tx, and these are four different systems.
They Aren't New Chips From Scratch
The 8tx is a modified version of the 8t; the 8i Max is two low-latency 8i chips in a single multi-chip package. The base silicon already existed: what is new is the packaging and what gets added on top.
Read also
TPU 8tx: The Data That Cannot Leave the Chip
It adds hardware-level confidential computing and is paired with fifth-generation AMD EPYC processors, codenamed Turin, on custom hosts, with Boardfly networking. Google claims up to 50% lower latency than the 8t with almost no performance penalty: that is a vendor figure, not an independent measurement.
What Confidential Computing Is, in One Sentence
An enclave is a region of memory that neither the operating system nor the hypervisor can read: the owner of the server cannot see what is inside while it is being used.
At Rest, in Transit and in Use
Encrypting a disk is easy because the data is not being used while it is encrypted. Encrypting the data a TPU needs right now forces the chip to decrypt it internally and to keep the key out of everything else. That encryption in use is what changes the math for banking, healthcare and government: they no longer have to choose between a shared accelerator and keeping data at home. Google also says the protection covers model weights, so the provider cannot read the model it runs either.
TPU 8i Max: Long Context Is a Memory Problem
Every token in the window takes space in the attention cache, and the memory required grows with the context. So the answer was not doubling raw compute but doubling memory: two 8i chips in a multi-chip package, with double the compute, SRAM and HBM bandwidth of a single 8i. SRAM is the fast memory next to the chip; HBM is the high-capacity memory feeding it.
Why Bandwidth Rules, Not FLOPs
As the context grows, the accelerator spends more time waiting for data than doing math. For long-context inference the practical limit is memory. The 8i Max also targets the highest compute density per rack in the family, and in a data center space and power are the real constraint.
Who Each One Is For
The 8tx is aimed at organizations with strict data rules, such as banking, healthcare and government. The 8i Max fits wherever a model has to read contracts, case files or entire repositories without chunking. Large-scale training remains the 8t's job.
What We Still Don't Know
There is no announced general availability, no pricing, no regions and no quotas; some of the press places availability in 2027. A portfolio announcement is not a product you can buy tomorrow, so this post promises no improvement to your API bill: with no published price there is nothing to pass through. When numbers show up, what matters is cost per token on a real workload and latency at your context size, not a theoretical peak.
Why This Is News and Not Just Another Press Release
The bottleneck moved twice in the same announcement: from FLOPs to memory, and from memory to trust. Optical connectivity was already setting the industry's pace, as Marvell's purchase of Lumentum showed, and now trust is setting the pace of contracts. It is the same question running through this year's silicon race: not who owns more accelerators, but who can prove what they did with your data.
Conclusion
The 8tx answers a compliance question and the 8i Max answers a memory question. Both are product announcements, not availability announcements: the fine print is still when, where and at what price. If you work with sensitive data or enormous contexts, these names will show up in architecture conversations next year.
We keep covering AI compute with numbers and with what is still unknown. If you care about the practical side, keep reading the development guides and technology news on this blog.
