AI Value Capture
The article analyzes a shift in AI value capture away from model providers toward infrastructure, platforms, and application layers, arguing that commoditization of models and high capital requirements are driving consolidation and changing where profits accrue in the AI stack.
Background
Semianalysis is an influential tech-industry research newsletter focused on AI hardware, infrastructure, and economics, run by analysts Dylan Patel and Jeremie Eliahou-Shields. The "AI value capture" debate asks which layer of the AI stack—chipmakers (NVIDIA), cloud providers (Microsoft, Google, Amazon), model developers (OpenAI, Anthropic), or application companies—will retain the most profit as the industry matures. This piece argues that advantage is shifting from hardware to the model layer as inference (running AI responses) overtakes training in compute demand, benefiting companies with strong distribution, proprietary data, and 'flywheel' effects from user feedback. Key context: NVIDIA currently captures the majority of AI profit via its GPUs; model providers have been spending heavily on training without clear returns; and many investors have assumed the "commoditization" of models would push value to apps, similar to earlier tech waves like the internet or mobile.