Key Takeaways

  • Google possesses massive TPU capacity, yet internal teams like DeepMind have less dedicated R&D compute than OpenAI and Anthropic.
  • Google Cloud CEO Thomas Kurian optimizes for GCP revenue and operating margins by selling capacity to external buyers rather than reserving it for internal research.
  • Google contracts like the SpaceX deal include mutual 90-day cancellation clauses, introducing volatility into compute allocations.
  • Data gravity is fading fast; AI tooling makes migrating workloads across clouds easier, allowing funded researchers who leave Google to take their work elsewhere.

The Misalignment Inside Google's Silicon Empire

DeepMind researchers are feeling the pinch. At the AI Engineer Summit, swyx spoke with staff who were openly frustrated: “look our leadership keeps selling the compute that we want to other people.” Google built a massive hardware footprint with custom TPUs, but owning silicon does not guarantee that your frontier lab gets to train on it.

Dylan Patel explains that the market assumption of Google winning simply due to raw hardware volume has cracked. “The two labs have more compute now because they've reached that scale,” Patel notes, comparing OpenAI and Anthropic to DeepMind. “So this whole like, oh, Google has the most compute, they'll win is is actually not really the case anymore.” While OpenAI and Anthropic channel every available dollar and chip directly into frontier training runs, Google splits its hardware across commercial cloud clients and enterprise products.

Kurian's Margin Win Is Hassabis's Loss

The root cause is structural incentive misalignment between Google Cloud Platform (GCP) and Google DeepMind. Under GCP leadership, cloud executives receive rewards for hitting top-line revenue, operating income, and margin targets. Selling TPU hours to outside startups and enterprise partners drives quarterly cloud earnings.

Patel points out the flaw in this corporate division: “I mean I think it's a fantastic move if you look at just GCP as its own company, right? Thomas Kurian continues to pump his revenue, his operating margin, his operating income, but for Google as a whole, it's really dumb. Like you should not incentivize your best researchers leaving.”

When Google funds external teams or lets researchers walk out the door to start new labs, it often sells them compute. But those departing founders do not stay locked in forever. The old assumption that storage and database lock-in would prevent departures no longer holds. As Patel notes, “Data gravity and all this stuff that people used to talk about with clouds in the pre-AI era is a lot less relevant in the AI era because AI can just help you migrate everything.”

The Risk of Fragile Compute Contracts

Selling compute externally also introduces operational fragility through aggressive contract terms. Jordan points to Google's commercial agreements, such as its deal with SpaceX: “SpaceX's contract with Google, for example, has mutual 90-day cancellation rights. So SpaceX's research starts to work, they can pull the compute back. Google stops making money off of that compute on some inference endpoint, they pull the compute back.”

These short-fuse agreements create a mercenary market where hardware capacity shifts based on immediate inference margins rather than multi-year research horizons. For frontier AI development, where pre-training clusters require sustained, uninterrupted access for months, running on fickle cloud agreements undermines research consistency.

What to Do With This

Audit your compute supply agreements this week and calculate your true switching cost. If you rely on single-cloud arrangements or assume vendor lock-in protects your proprietary pipelines, build automated data migration scripts using modern code-generation tools to retain cloud optionality across providers.