TPU 8tx and 8i Max: Google adjusts its AI chips for inference
Google adds TPU 8tx and TPU 8i Max to its eighth-generation accelerator lineup. Both systems target sensitive uses and heavy memory loads, in a strategy more open to external customers.
On November 9, 2026, Google unveiled two new systems for its eighth-generation AI accelerator lineup: TPU 8tx and TPU 8i Max. The first targets confidential computing, while the second addresses use cases that saturate memory and data center racks. Together, they demonstrate that the chip battle is no longer just about transistor count, but also about how hardware is assembled around existing silicon.
Deux systèmes, deux problèmes très concrets
TPU 8tx is aimed at businesses and government agencies handling sensitive data. Google explains that the system keeps data and model weights protected and isolated during execution, whereas traditional encryption often only covers storage and transport. In other words, information does not become “naked” during computation, meeting a very specific need in regulated sectors. The company also states that TPU 8tx relies on the low-latency Boardfly network and can reduce latency by up to twice compared to TPU 8t, without notable performance loss.
TPU 8i Max tackles the other end of the problem: models and agents working on very large contexts, such as entire codebases or hours of video. Google chose a simple-to-explain but harder-to-manufacture approach: two TPU 8i chips in a single package. As a result, the manufacturer claims a doubling of compute, onboard SRAM memory, and HBM bandwidth, all while staying within the same rack form factor. The gain is measured as much in performance as in compute density per square meter. Data center operators like that. Their energy bills, less so.
Google pousse ses TPU vers des clients extérieurs
These announcements are not just technical. They also confirm Google's shift in stance, no longer reserving its TPUs for internal use only. The company now counts nearly 100 external clients, spread across cutting-edge labs, trading firms, distributors, and universities. The TPU lineup covers both Google Cloud infrastructure and deployments in clients’ own data centers.
This openness is based on a modular logic. Google does not present TPU 8tx and TPU 8i Max as new chip generations, but as new configurations around the same silicon. The compute core remains identical; what changes are the packaging, memory, network, hosts, and security. The company says it can deliver this type of system in a few months, whereas a new chip takes years to design and manufacture. It’s less spectacular than launching a brand-new GPU, but faster to industrialize.
Le vrai terrain de jeu devient l’inférence
The move aligns with a broader market trend: the center of gravity is shifting towards inference, that is, the daily execution of models rather than their initial training. This shift favors specialized accelerators capable of optimizing latency, memory, or cost per request depending on the use case. This is precisely the niche where Google is trying to position its TPUs against Nvidia, as well as against accelerators designed by Amazon, Microsoft, and OpenAI.
Google is also trying to reduce software friction. TorchTPU, released as open source, aims to enable running PyTorch code on TPU with limited changes. Inference frameworks like vLLM and SGLang are also supported. In practice, the company seeks to make TPUs a more general platform, less reserved for teams willing to rewrite their software stack. Adoption often hinges on this more than on presentation slides.
Un marché qui se segmente à vue d’œil
The two new products mainly show that Google now segments its offering around very specific use cases. TPU 8tx addresses confidentiality. TPU 8i Max addresses memory shortages and workload density. In both cases, the solution lies more in the system than in the chip alone. For clients, this can reduce the number of clusters needed and simplify operations. For Google, it reinforces the idea that its TPUs are no longer just an in-house tool but a full-fledged commercial platform.
One question remains about pace. If Google can truly multiply variants without waiting for a new silicon generation, the company gains time over its competitors. But it also ties itself more closely to the expectations of external clients, who demand very concrete guarantees of security, compatibility, and availability. The catalog grows, and so does the promise. The next chapter will be played out in data centers, not slogans.
By the numbers
- 2 new systems — TPU 8tx and TPU 8i Max, announced on November 9, 2026.
- 4 TPU systems in one year — first time Google reaches this pace, in 2026.
- Nearly 100 external clients — figure provided by Google for its TPUs.
- Up to 2x lower latency — TPU 8tx compared to TPU 8t, according to Google.
- 288 GB of HBM — base capacity of TPU 8i mentioned by Google.