SemiAnalysis declares DeepMind is no longer a frontier lab as Google bets on cloud over AI research

DeepMind Is No Longer a Frontier Lab, SemiAnalysis Declares—Google Has Chosen Cloud Over AI Research

SemiAnalysis issued a stark assessment of DeepMind's competitive position: the lab is effectively out of the frontier race due to sustained talent departures from its reinforcement learning teams and what the research firm views as misallocated compute resources. The declaration follows the recent exits of senior researchers including Jeff Dean, Noam Shazir, Sanjay Kutcher, and Oriel Soria—departures SemiAnalysis characterizes as signals that top researchers no longer see DeepMind as the place to be.

The core argument is structural: researchers cluster around other researchers. As talent density falls, recruitment becomes harder. The hemorrhaging at DeepMind reflects not leadership failures in execution but a strategic choice by Google—one that subordinates frontier AI development to cloud infrastructure revenue.

Google's compute allocation tells the story. Google is selling more than 20% of its total TPU shipments from Q3 2026 through Q4 2027 directly to Anthropic on long-term contracts. That compute, once sold, is not available for internal research. Meanwhile, SemiAnalysis estimates Google is allocating only roughly 15% of its overall compute to DeepMind, with the remainder going to GCP customers. By contrast, labs like OpenAI allocate compute roughly 50-50 between training and inference. Google has inverted that calculus in favor of serving external customers.

The financial logic is real. SemiAnalysis projects GCP system-sales margins in the low 30% range—slightly below core cloud margins but still extremely profitable. The firm estimates more than $250 billion in additional TPU bookings could be added to GCP's remaining performance obligation in coming quarters as hyperscalers continue ramping. From a pure business perspective, Google is choosing the more certain, higher-margin revenue stream over the riskier and more resource-intensive bet on frontier model development.

The paradox some observers raised is incomplete. Sebastian Malabai argued that SemiAnalysis cannot simultaneously claim Google is exiting the AI frontier while saying GCP revenues will accelerate. A strong revenue base, Malabai reasoned, should strengthen Google's position over time. But that logic misses the dynamic at play. The race for frontier AI is fundamentally about compute allocation right now. OpenAI doesn't sell 20% of its chips to competitors on multiyear contracts; it keeps chips for training. The moment Google sold that capacity away, it made a bet that the margin on cloud services today matters more than leadership in model capability tomorrow.

The counterargument—that Google could sit out this cycle, build more capacity, then redirect compute back to internal research in 2027 or 2028—runs into a harder problem. By then, the talent will be gone. Demis Hassabis reportedly sought to work with investors fully committed to AI as a fundamentally distinct technology, not one competing against cloud infrastructure or other bets. Google's allocation signals the opposite priority. Hassabis has since become a major backer of Anthropic, helping assemble financing and data-center commitments despite Anthropic's limited access to traditional debt markets.

Google's 15% ownership stake in Anthropic comes with no voting rights, no board seat, and no board observation rights—a purely financial position. That structural constraint reflects Anthropic's determination to stay independent and control its own compute allocation decisions, a luxury DeepMind no longer has.