SemiAnalysis founder Dylan Patel and technical staff member Jordan Nanos argue that the widely-cited AI infrastructure spending numbers are still too conservative. Speaking on the Big Technology Podcast, both contend that compute demand continues to outpace supply despite hard constraints on construction, financing, and equipment. Their core claim: AI value creation extends far beyond Claude Code subscriptions and per-token pricing into chip design, drug discovery, and broader economic sectors that most analysts are not yet modeling.
The optimism is conditional. Patel expects AI to accelerate the transfer of economic value from labor to capital. Nanos observed that GPU cloud providers can run profitable businesses while delivering poor service, a market structure that signals demand so strong it masks execution failure. The conversation also surfaces specific security vulnerabilities the two encountered while producing SemiAnalysis's ClusterMax ratings of GPU cloud providers, a detail that makes the full transcript worth reading on its own.
The interview covers Nvidia's financing backstops, Anthropic's business model, Oracle's execution problems, Google DeepMind's compute position, and why frontier AI labs may choose to withhold their best models from the market entirely. The infrastructure payback math is the thread running through all of it. Read this to understand what the bull case actually requires, not just what it claims.
[READ ORIGINAL →]