AI is long, AGI is short

Silicon Valley is currently pushing a narrative that AGI is a scarce resource, and the company that builds the strongest model will capture most of future value. But the author sees this story as exaggerated hubris, and bets against that expectation.

Models are already commoditizing fast. Two years ago, GPT-4-level intelligence cost about $30 per million tokens, but now similar performance is available for under $1. DeepSeek, Kimi, and Qwen are also closing the gap at even lower cost.

Historically too, value has accumulated in the products and workflows built on top, rather than in the infrastructure itself.

• Railroads did not dominate the industrial economy; Standard Oil captured greater economic value.
• AWS built cloud infrastructure, but the emblematic winners were Stripe, Shopify, and Snowflake.
• AI will likely be the same: companies with customer relationships, proprietary data, and hard-to-replace workflows are more likely to win than the models themselves.

The author cites the case of Tellme Networks. This company had no proprietary speech recognition engine of its own, but using Nuance, the best-in-class engine at the time, it earned over $120 million by automating enterprise call centers. The conclusion is clear: the most important AI company has not yet been built, and founders obsessed with complex, seemingly trivial problems will define the next era.