The decade of contempt

1. Academic researchers are already contorting Nvidia's graphics chips into general-purpose parallel computers: Stanford's Brook project, and Ian Buck, hired in 2004 to productize the idea. Single-thread CPU scaling is visibly decelerating; anyone reading the semiconductor roadmaps can see the wall. The thesis: parallel compute will matter far beyond graphics, and whoever owns the developer layer will own the market when it arrives. What doesn't exist: the market. Any of it.

Stay a gaming-GPU company with fat margins. License a compute layer as a side project, priced to its current demand of approximately zero. Or build CUDA into every chip, raising costs across the entire product line, for a market that does not exist.

CUDA ships with the G80 generation in late 2006, SDK public in 2007, on every chip. Huang's own later formulation, on the record: they were spending around $500M a year against a $0 billion market.

Then the market graded the bet, and the grade ran seven years. Gross margins suffered through 2008 to 2013. The stock fell about 80% in the 2008 crash. Market cap dipped into single-digit billions. Analysts on record, repeatedly, asked him to cut the compute spending. He did not. The GTC keynotes from 2009 onward show the same pitch every single year: accelerated computing, applications first in oil and gas, molecular dynamics, imaging. The story never wobbled, because the story was a technical thesis with early users attached, not a narrative that needed managing.

A decade-long option is holdable only when three things are true, and Huang had all three:

1. The core business funds it, so nobody can call the loan. This is the sixth conviction marker, and it's the load-bearing one. Bets funded by narrative (the next round, the patron's continued belief) force you to manage the story instead of the truth, because the story is the collateral. Bets funded by your own cash flow let you be wrong-looking for as long as the thesis needs.
2. The bet compounds a moat while you wait. Every year of CUDA built developer ecosystem, which is exactly the asset the thesis said would decide the market. Waiting wasn't idle; it was accruing.
3. There's real pull from serious early users, however small. Not applause. Usage, from people with real problems and real budgets.

Vision without those three is just burn, and the graveyard of visionary burns is deep. That's the whole file: conviction isn't the feeling of being right while others doubt. It's the structure that lets you stay solvent, compounding, and honest until the world catches up.