Nvidia started pouring money into CUDA in 2006 — a way to make graphics cards run general-purpose computation. For most of the decade that followed, it looked like a mistake. Analysts questioned it. It ate margin on every chip because the capability shipped whether or not anyone used it. The addressable market was academic researchers doing work nobody could monetize.
Then the deep learning era arrived and it turned out Nvidia had spent ten years building the only road into it.
The story gets told as vision. It’s more useful read as tolerance. Jensen Huang didn’t know AI was coming in that form. He believed parallel computation would matter and he was willing to be publicly wrong about it for longer than any quarterly-driven executive could survive. He runs the company on the stated premise that it’s always thirty days from going out of business — paranoia and conviction operating at the same time, which is a stranger combination than it sounds.
The transferable lesson isn’t to make big bets. Everyone says that. It’s to budget for the drought before you start, because the drought is what kills the bet, not the thesis being wrong.
Every compounding asset has a period where the effort is real and the evidence is absent. Content is the obvious one. The first ninety days of publishing produce almost nothing measurable — no traffic worth reporting, no inbound, no signal. That’s not the system failing. That’s the shape of the system. Search engines and answer engines need a body of work and a history before they’ll weight you, and there is no version where that arrives in week three.
The failure mode is predictable: someone starts, holds for six weeks, sees flat numbers, concludes it doesn’t work, and rebuilds the strategy. Then repeats that cycle three times over two years. Total elapsed effort would have been more than enough. It got spent restarting instead of continuing.
Same structure applies to the RWA thesis. Tokenizing real assets is going to look overhyped and underdelivered for a stretch, because the infrastructure is arriving before the demand and the regulatory picture keeps moving. Anyone who needs the market to validate them quarterly will exit at the bottom of that curve. The only way through is to decide the duration up front.
So make it explicit. Before starting anything that compounds, write down the honest answer to two questions: how long until this should show real signal, and what am I committing to do regardless of what the numbers say during that window.
Then treat that window as pre-paid. Not something to re-litigate every month.
The second half of Huang’s approach matters just as much. Conviction on the direction, paranoia on the execution. He wasn’t relaxed about CUDA — the company shipped, iterated, and stayed operationally terrified the whole time. That’s the distinction between conviction and stubbornness. Conviction holds the thesis and adjusts everything underneath it. Stubbornness holds the whole thing rigid and calls that discipline.
Being early looks identical to being wrong right up until it doesn’t. The only thing that separates them, from the inside, is whether you decided in advance how long you’d wait.