Google Cloud launches two new AI chips to compete with Nvidia

The detail worth sitting with is that Google still sells Nvidia hardware alongside its own TPUs. That's not a contradiction — it's a hedge that keeps enterprise customers from defecting to AWS or Azure — but it also means Google's TPU adoption numbers will be hard to read cleanly from the outside, since any TPU growth could be offset by continued Nvidia revenue within the same platform.

The broader silicon story has been building fast. Cerebras filed for IPO just days ago (covered here April 18), signaling that specialized AI hardware is now mature enough to attract public-market capital, not just venture bets. Google's announcement lands in that same window and reinforces the same thesis: the infrastructure layer is where serious competitive positioning is happening, a point MIT Technology Review made explicitly in 'Treating enterprise AI as an operating layer' (April 16). Meanwhile, Upscale AI raising at a $2B valuation after seven months also points to how much capital is chasing this layer. Google's move is the incumbent version of that same race.

Watch whether Google publishes third-party reproducible benchmarks for the new TPUs within the next two quarters. If independent results match the launch claims on training cost-per-token, that puts real pricing pressure on Nvidia's H100 contracts with hyperscalers. If they don't materialize, this announcement reads as positioning ahead of Cerebras's IPO roadshow rather than a substantive capability shift.