SemiAnalysis Determines DeepMind is No Longer a Leading AI Laboratory
SemiAnalysis's latest report directly concludes that Google's DeepMind is no longer a leading AI laboratory, with the probability of regaining SOTA in the future dropping to zero.
The report points out that DeepMind is simultaneously losing top talent and computing power. Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals have recently left Google to start their own ventures. Previously, Noam Shazeer, co-lead of Gemini, moved to OpenAI, and Nobel laureate John Jumper joined Anthropic. A significant amount of computing power has also been sold to competitors, with estimates suggesting that over 20% of TPU shipments will be sold directly to Anthropic between Q3 2026 and Q4 2027. The issue is attributed to Google's bureaucratic, slow, and strategically conservative organizational culture, drawing parallels to IBM and Intel.
This represents a strong external critique of Google's AI strategic positioning, shifting focus to the prioritization of talent and computing power allocation. Google Cloud benefits from external TPU sales and growth expectations, while DeepMind and Gemini are under pressure due to declining assessments of their competitive edge.
Source: Public Information
ABAB AI Insight
SemiAnalysis interprets the recent leadership changes and core talent departures as a signal that DeepMind is shifting from a "full sprint at the frontier" to a more commercial path. The large-scale sale of TPUs is seen as an indication that internal computing power priorities are yielding to cloud business monetization.
In terms of capital strategy, Google is opting for short-term cloud revenue and system sales rather than maximizing internal model R&D investment. The motivation is to balance shareholder returns with long-term bets on AI, accepting the trade-off of relatively decreased model competitiveness.
This follows a historical trend of large tech companies shifting from research-oriented to profit centers, currently in a clear phase of prioritizing cloud business over frontier models.
Essentially, this represents capital concentration: computing resources are shifting from internal frontier laboratories to external high-paying clients, mechanically converting scarce hardware into predictable cloud and hardware revenue.
ABAB News · Cognitive Law
The collective departure of top talent is often the strongest signal of a change in laboratory priorities.
Selling computing power to competitors equates to actively lowering one's own ceiling.
When a company begins to think like IBM, the riskiest competition is no longer the top priority.