Google is splitting its homemade AI silicon into specialized parts, and Wall Street is starting to put a very large number on what that could be worth.
At the Hot Chips 2026 conference, Google presented its eighth-generation TPU family to a technical audience, according to ServeTheHome. The notable detail is that there isn't one chip anymore — there are two. Google described the TPU 8t, built for training, and the TPU 8i, built for inference. ServeTheHome also notes that Google remains one of the only hyperscale cloud companies that develops its own training hardware, rather than buying it.
The training-versus-inference split matters because those are genuinely different jobs. Training is the expensive, one-time-ish work of teaching a model, which rewards raw computational muscle. Inference is what happens every time someone actually uses the model — a chatbot reply, a search summary — and it happens constantly, so cost and efficiency per query dominate. A single chip designed for both is a compromise. Two chips are not.
Investors appear to be paying attention. Morgan Stanley sees a $200 billion opportunity in Google's custom AI chips, according to a report from finance.biggo.com.
The move also fits a broader shift in the chip supply chain. A weekly industry roundup from Digitimes grouped Google's diversification alongside a fall in TSMC's share price and SK Hynix deepening its co-design work on HBM memory — signs that the pieces of an AI system are increasingly being designed together rather than bought off the shelf.
Why it matters: if the largest buyers of AI computing keep building their own specialized silicon, the economics of the entire AI industry — and the dominance of the companies currently selling those chips — start to look less settled than they did.