Two Thousand Dollars

OpenAI’s Astra model solved ten open problems in mathematics last week. Not textbook questions. Problems that had resisted human mathematicians for decades. A Fields Medal winner said one proof was journal-ready. Total compute cost: roughly two thousand dollars.

Same week, BCG published updated numbers on enterprise AI adoption. Sixty percent of companies report minimal or no value from their AI investments. Not disappointing returns. Nothing.

Two thousand dollars did what tenured experts couldn’t. Millions bought conference badges and empty dashboards.

The gap isn’t capability. BCG found that 70 percent of AI value comes from reorganizing work around the tool — not from the model, not from the vendor, not from the data pipeline. The companies getting three-times cost reduction and 2.7-times return on capital aren’t running better prompts. They rebuilt how decisions get made, how tasks get routed, how teams learn from output.

Everyone else bought the same software and left the org chart untouched. That’s a $2,000 problem dressed up as a million-dollar one.

The part nobody wants to hear: every month without that reorganization, the companies that already moved add another month of compounded learning. The distance doesn’t shrink. It accelerates. And no model upgrade fixes that.