Google has added two new chips to its eighth-generation tensor processing unit (TPU) family. TPU 8tx is built for confidential AI training and inference. TPU 8i Max packs twice the memory and compute of Google’s current inference chip into a single package. According to TechCrunch AI, Google announced both on Monday as part of a wider effort to rely less on outside chipmakers for its cloud business.
TPUs are Google’s in-house processors for AI workloads. The eighth generation arrived in April, focused on more inference capacity and support for big training jobs. These two variants push that lineup into more specialized uses.
What Google Just Shipped
- TPU 8tx: a chip that keeps secrets. Google designed TPU 8tx with AMD and based it on the existing TPU 8t. Its main job is confidential training and inference. In its blog post, Google said the chip “builds security directly into the hardware layer, keeping sensitive data and model weights protected and isolated even while in active use.” That last part matters most. Encrypting data in storage or in transit is standard practice. Protecting it while the chip is actually computing on it is the harder problem.
- Training and reasoning on the same chip. TPU 8tx also comes with a custom inference network, so it can run reasoning alongside training. Google says that lets customers distill and train small models on the fly. Distillation means using a large model to teach a smaller, cheaper one. Doing both jobs on one chip removes a handoff step that usually happens across separate hardware.
- Your data never has to move. Google’s main use case for TPU 8tx is a company that wants to train a model on its own proprietary data without moving that data between devices. Every transfer is another chance for exposure. Banks, healthcare providers and anyone holding sensitive IP have good reason to want fewer of them.
- TPU 8i Max: two inference chips in one package. TPU 8i Max essentially fuses two of April’s inference-focused TPU 8i chips into a single package. That doubles memory capacity, processing power and bandwidth. In practice, models can work with a larger range of information during each query. That helps with long documents, big codebases and agent workflows that build up a lot of context.
- Part of an industry-wide break from Nvidia. Google isn’t alone here. TechCrunch AI notes that Microsoft, Amazon, OpenAI and Anthropic are all building their own AI chips and software to lean less on Nvidia. Google has been at it longer than most, but every hyperscaler now wants more control over its hardware costs and supply.
Why This Matters
What stands out is that Google is specializing instead of just scaling up. Earlier TPU generations mostly chased raw performance. This launch splits the lineup by job: one chip for regulated, privacy-sensitive work and one for memory-hungry inference.
The confidential computing angle is the more interesting bet. Many enterprises still hold back from cloud AI because they don’t trust where their data goes or who can see their model weights. Hardware-level isolation speaks directly to that worry. If it works as described, it could open up workloads that have stayed on-premises so far.
The AMD partnership is notable too. It shows Google is willing to co-design chips with an established player instead of building everything alone. It’s also a quiet win for AMD, which has spent years trying to grab a bigger share of AI compute.
The timing adds some drama. TechCrunch AI points out that the launch lands at a shaky moment for Nvidia, after CEO Jensen Huang admitted over the weekend to an affair with someone alleged to be a Chinese spy. That has nothing to do with the chips themselves, but rivals rarely mind launching into a competitor’s bad news cycle.
What We Don’t Know Yet
The announcement leaves some big gaps:
- No pricing. Google didn’t say what either chip will cost on Google Cloud.
- No availability dates. There’s no timeline for when customers can actually rent them.
- No benchmarks. Google hasn’t published performance numbers against Nvidia’s GPUs or its own earlier TPUs.
- Unverified security claims. Hardware-level isolation is only as good as its implementation, and independent security reviews will matter.
What Comes Next
Expect Google Cloud to pitch TPU 8tx hard to finance, healthcare and government customers, since compliance requirements in those sectors have kept AI adoption slow. TPU 8i Max will likely go to teams running large-context models at scale. The bigger question is whether custom chips like these can actually loosen Nvidia’s hold, or whether they just give hyperscalers more leverage when they negotiate with it. Full details are in TechCrunch AI’s original report.