To rival Nvidia, Google surrenders the moat

AMIN VAHDAT has spent a decade helping build what Google calls a "tech island" — an end-to-end AI compute stack in which TPU silicon, Google data centers, and Google's own orchestration software all depend on one another. In an interview with Bloomberg this week, he explained why Google has decided to abandon it.

Bloomberg's reporting was ostensibly about Google's likely split into separate training and inference TPUs, to be teased at Cloud Next in Las Vegas. Behind that technical announcement sits a more consequential one. Google is now letting Anthropic run TPUs in Anthropic's own data centers rather than Google's; it has enabled customers to use PyTorch and third-party schedulers rather than its own orchestration stack; and it has signed multibillion-dollar deals with Meta, Citadel Securities, and (in advanced talks) G42. Alongside an existing October commitment to supply Anthropic with up to one million TPUs and a separate 3.5-gigawatt Broadcom-built TPU pact starting in 2027, they amount to a strategic concession.

Drawbridge down

The trouble is that the island was, for a decade, the whole point. Jeff Dean started building what became the TPU in 2013 precisely because off-the-shelf silicon could not meet Google's own demands; the chip's appeal was that it was optimized for workloads only Google ran. Vertical integration let DeepMind researchers co-design hardware with model architecture, which is how Gemini 3, released in November and trained entirely on TPUs, reached benchmark parity with frontier GPU-trained models. It is also how Google built what SemiAnalysis reckons is the only hyperscaler ASIC program besides AWS Trainium with a serious shot at scale. The integration was the moat.