SemiAnalysis CEO Dylan Patel: Google Has the Most Global AI Compute but Models Lag, Citing Internal Resource Competition
Dylan Patel, CEO of semiconductor and AI infrastructure research firm SemiAnalysis, argued on a podcast that although Google holds approximately 25% of global AI compute, its cloud division's incentive to sell compute to competitors rather than allocate it to DeepMind forms the backdrop for Gemini 4 Argon's third-party scores trailing Claude Opus 5.5.
ImportanceMaterialEvidenceE2 unreplicatedWrite-upQuick
Dylan Patel, CEO of semiconductor and AI infrastructure research firm SemiAnalysis, offered an explanation on the Big Technology Podcast that launched on October 9: Google possesses the most AI compute globally, yet its model scores lag behind competitors. The root cause, he argued, is not insufficient total compute but internal competition for resources.
AI research firm Epoch AI estimated that as of the end of 2025, Google held approximately 25% of global AI compute, making it the largest single holder. However, third-party scores from AI model evaluation firm Artificial Analysis showed that Gemini 4 Argon, Google's highest-tier model released on September 30, scored 53, tying with OpenAI's GPT-6 Astra but still below Anthropic's Claude Opus 5.5 at 58.
Patel's central argument is an organizational incentive distortion: there is resource competition between Google Cloud head Thomas Kurian and DeepMind co-founder Demis Hassabis. Selling compute to external customers (including competitor Anthropic) generates revenue more directly than allocating it to DeepMind for internal use, so the cloud division has an incentive to expand external sales while the proportion of compute allocated to DeepMind actually declines. Patel stated that Google's share of its own compute used for model training and inference over the past six months was lower than two to three years ago.
Patel further predicted that comparing only DeepMind, Anthropic, and OpenAI, DeepMind would have the least compute by the end of 2026; in 2027, OpenAI and Anthropic would each add 5 to 10 GW of data center capacity, and if Google adds only a few GW for its own use, the gap would widen further. This prediction has not been independently verified, and model performance depends on multiple factors including training methods, data quality, and compute allocation; Patel's explanation focuses on the single dimension of compute allocation.