Martin Casado: AI market is expanding, not consolidating — open and closed source will both win big

Summary

Martin Casado, general partner at Andreessen Horowitz, argues that the AI market is expanding fast enough that zero-sum thinking is the cardinal investor mistake. Open source and closed source will both win — not at each other's expense, but by occupying different territory as the overall market grows.

His read on the open/closed dynamic draws on software history. Closed solutions capture value first because you need revenue to change behavior at scale — Unix before Linux, Oracle before MySQL. The same pattern is playing out in AI: OpenAI dominated language, Anthropic carved out code, Midjourney led image, ElevenLabs audio. But first-mover closed dominance doesn't foreclose open-source value; it creates the conditions for it, by training user behavior and proving the market.

The strongest evidence for Casado's thesis is OpenAI itself. OpenAI was first to image generation with DALL-E and lost that market. First to coding with Copilot and lost that to Cursor. First to video with Sora and is not the leader there either. And yet it remains the dominant AI company on language. Every one of those verticals became a massive standalone market — too big for any single company to hold.

The Cursor case

Cursor's rise follows a specific sequence. Microsoft GitHub Copilot trained early user behavior around AI coding, but the underlying models weren't good enough to deliver on the promise. The reinforcement learning wave of the past year changed that. Cursor caught that wave with better models at the right moment, and is now developing its own smaller specialized models on top of the foundation layer — building compounding advantages rather than remaining purely model-dependent.

Enterprise revenue quality

On enterprise AI spending, Casado is direct: most of what's being called enterprise AI adoption is actually a bottoms-up prosumer phenomenon. Individuals inside companies are driving behavior; it isn't large procurement budgets moving. He draws the parallel to the internet, PC, and post-Blackberry smartphone cycles — all of which started with individual adoption before institutional budgets caught up.

The accounting murkiness in AI revenue reporting — confusion between GMV and take rate, run rates presented as if they were GAAP figures — is, in his view, less a sign of fraud and more a familiar artifact of hypergrowth cycles where the underlying demand is real but the industry hasn't yet standardized how to measure it. The signal investors should focus on is individual behavior change, not enterprise budget allocation, which he regards as a lagging and noisier indicator.