Google really would rather not rent its brains from Nvidia. On November 10, TechCrunch reported that the company is expanding its eighth-generation TPU family — first presented in April — with two new specialized chips, each aimed at a different pressure point in the AI buildout.

The 8tx: training with the curtains closed

The TPU 8tx is the stranger and more interesting of the pair. Designed together with AMD and built on the foundation of the existing 8t, it is meant to train and run models confidentially, with the security built into the hardware itself rather than bolted on in software.

That “confidential” label is doing real work. As TechCrunch describes it, the 8tx can reason and train at the same time: it can distill and train smaller models on the fly without shuffling data between devices. For organizations sitting on sensitive data — hospitals, banks, governments, anyone whose lawyers outnumber their engineers — the pitch is that the data stays sealed while the model learns from it. The chip doesn’t just compute; it keeps secrets while it computes.

The 8i Max: two chips, one package, twice the headroom

The second arrival is more straightforward brute force. The 8i Max takes two of Google’s 8i inference chips and fuses them into a single package, delivering double the memory, double the power and double the bandwidth. Inference — the part where a trained model actually answers your prompt — is where AI companies burn money at scale, serving billions of requests. A package that doubles the room to work in is a direct attack on the cost of every chatbot answer.

Why everyone is building their own silicon

Google is not acting alone, and TechCrunch’s context makes the pattern plain: Microsoft, Amazon, OpenAI and Anthropic are all developing their own chips as well, each trying to loosen Nvidia’s grip on the industry’s most expensive dependency. Custom silicon is slow, hard and eye-wateringly capital-intensive — and the entire industry has decided it’s worth it anyway. That tells you exactly how much margin and leverage currently sit in one supplier’s hands.

For Google specifically, the TPU line is the quiet backbone of its whole AI strategy: the same hardware family trains and serves the Gemini models that are rapidly being woven into Search, Android and Workspace. Every generation that narrows the cost or capability gap with Nvidia hardware strengthens everything built on top of it.

What does any of this mean if you never plan to buy a datacenter chip — which is to say, almost everyone reading this? Compute is the raw material of every AI feature on your phone and in your browser. When the biggest AI companies manufacture more of it themselves, and aim it at cheaper inference and confidential workloads, the downstream effects are the ones you actually feel: faster AI features, more of them bundled into products you already pay for, and AI tools that can finally be trusted with the sensitive stuff.

The Bit Take

Google’s new lineup splits the AI future in two: one chip that thinks out loud at double volume, and one that whispers. The 8tx’s “this stays between us” mode might matter more than raw speed — the next AI land grab isn’t for your prompts, it’s for the data you were too nervous to upload.

Sources