Google's first confidential TPU is still Google's own accelerator. What is different is the processor sitting next to it. For the host CPU, Google did not use its Arm-based Axion chip. It picked AMD EPYC. That one choice says a lot about where trust in AI hardware comes from.
What Google announced
Google has added two systems to its eighth-generation TPU family. TPU 8t (training) and TPU 8i (inference) were announced in April. The new ones are TPU 8tx and TPU 8i Max, described in a Google Cloud blog post, according to TechRadar and Tom's Hardware.
- TPU 8tx is built on the same silicon as TPU 8t but is aimed at secure, confidential inference and small-scale training. It uses 5th Gen AMD EPYC host CPUs instead of Google's Arm-based Axion hosts, and it takes the lower-latency Boardfly network from the 8i. Google claims up to 2x lower latency than TPU 8t with almost no performance penalty. Tom's Hardware notes it is not clear what measure Google uses for that claim.
- TPU 8i Max puts two TPU 8i chips on one package for long-context agent work. Tom's Hardware derives roughly 20.2 FP4 petaflops from the doubling. Google has not published that figure directly in what was reported.
Tom's Hardware reports that Google has not shared specs, prices or ship dates for either new system, and lists availability as "coming soon".
Why a confidential TPU matters
Normally, data has to be unlocked in memory while a chip processes it. Confidential computing tries to close that gap with hardware isolation, to keep the machine's operator from reading prompts, model weights or intermediate data. TechRadar reports that AMD's Secure Encrypted Virtualization (SEV) works with the TPU to extend the protected environment to the accelerators. Each accelerator's identity and security configuration is checked before it gets direct access to protected memory. AMD says prompts, weights and intermediate data stay encrypted in host memory and in transit.
The target customers are regulated enterprises and governments that want to fine-tune, distill and train agents on private data without moving it between clusters, according to the same reports.
The twist: AMD hosts, not Axion
Google has not said why it passed over Axion. TechRadar points to a hint in Google's own documentation: the supported Confidential VM configurations run on AMD SEV and Intel TDX machines, with no Axion-based instances, and TPUs cannot currently be attached to Confidential VMs at all. Its reading is that Google's audited confidential stack is x86-only for now. That is an interpretation by the reporter, not a Google statement.
AMD's release adds that Google security experts have reviewed AMD's security processor firmware, and that SEV already underpins Google's Confidential VMs, Confidential GKE and Confidential Space.
One detail for chip watchers: TechRadar, citing ServeTheHome, says the host is a 5th Gen "Turin" part that matches the TPU 8t's PCIe Gen5 x16 interface, so a newer Gen6 host would add no accelerator bandwidth.
What this means
- Trust is a stack, not a chip. A confidential claim is only as credible as the years spent attacking and auditing the hardware and firmware under it. Google reached for the stack that already has that history.
- TPUs are leaving Google's own buildings. TechRadar reports that select customers will be able to deploy 8tx in their own data centers, and that Oracle recently said it will offer TPUs on its cloud. Hardware attestation is how a model owner can place weights on machines someone else runs.
- Treat the claims as unproven for now. TechRadar says key security details remain unpublished, so the claims deserve scrutiny once customers can test them. Tom's Hardware adds that Google's best third-party performance numbers so far are for the previous-generation Ironwood TPU, and there is no eighth-generation result to compare yet.
What is still unknown
- Prices, specs for the 8tx (peak compute and memory figures are listed as not stated) and ship dates.
- How the security design holds up under independent testing.
- Whether Google will bring a confidential stack to its own Axion CPUs later. Nothing in the reports says so.
TechRadar also notes that an earlier market rumor of an AMD and Google TPU project in a future generation is not confirmed by this announcement, since the EPYC chips sit in separate sockets.
Sources: TechRadar, "Google turned to AMD Epyc, not its own Axion chips, to secure its first confidential TPU" (published November 15, 2026, techradar.com/pro/google-turned-to-amd-epyc-not-its-own-axion-chips-to-secure-its-first-confidential-tpu); Tom's Hardware, "Google adds two new specialized chips for AI agents and security to 8th-gen TPU lineup" (published November 12, 2026, tomshardware.com). Read November 16, 2026. This is a summary and analysis of those reports. We did not test the hardware.