The decade of contempt

Nvidia committed to CUDA on November 8, 2006, and every fact below was knowable on that date.

• Academic researchers were already contorting Nvidia graphics chips into general purpose parallel computers, including the Brook project at Stanford.
• Nvidia hired Ian Buck in 2004 to productize general purpose GPU computing.
• Single thread CPU performance scaling was visibly decelerating, and the published semiconductor roadmaps showed the wall.
• The available thesis held that parallel compute would matter far beyond graphics, and that whoever owned the developer layer would own the market when it arrived.
• No market for general purpose GPU computing existed in 2006, at any size.
• Nvidia gaming chips were profitable enough to fund a large multi year software program out of internal cash flow.

Option one: stay a gaming GPU company. Nvidia could keep selling graphics chips at fat margins and spend nothing on compute. That path cost nothing and bought nothing.

Option two: license a compute layer as a side project. Nvidia could build the software once and price it to its demand at the time, which was approximately zero. That path capped the spending and capped the ecosystem with it.

Option three: build CUDA into every chip. Nvidia could raise cost across the entire product line for a market that did not exist. That path was the expensive one, and it was the only one that put a developer platform in every developer's hands.

Nvidia shipped CUDA with the G80 generation in late 2006 and released the public SDK in 2007, on every chip. Jensen Huang has said on the record that Nvidia was spending around $500 million a year against a zero billion dollar market.

The market then graded the bet, and the grade ran seven years. Nvidia gross margins suffered from 2008 through 2013.

Nvidia stock fell about 80% in the 2008 crash, and Nvidia market capitalization dipped into the single digit billions. Analysts asked Huang repeatedly, on the record, to cut the compute spending.

Jensen Huang did not cut it. The GTC keynotes from 2009 onward carried the same pitch every year: accelerated computing, with the first applications in oil and gas, molecular dynamics, and imaging.

The story never wobbled because the story was a technical thesis with early users attached.

A decade long option is holdable when three conditions hold at once, and Nvidia held all three from 2006.

1. The core business funds the bet, so nobody can call the loan. Bets funded by narrative force a founder to manage the story, because the story is the collateral. Bets funded by internal cash flow let a founder look wrong for as long as the thesis needs.
2. The bet compounds a moat during the wait. Every year of CUDA added developer ecosystem, the exact asset the thesis said would decide the market. Waiting accrued value instead of burning it.
3. Serious early users create real pull, however small. Applause fails this test. Usage by people with real problems and real budgets passes it.

Vision without those three conditions is burn, and the graveyard of visionary burns is deep. Conviction is the structure that keeps a company solvent, compounding, and honest until the world catches up.