The Chokepoint

More precisely: a very specific type of processor, built on a software ecosystem nobody bothered to replace because it "mostly worked." In May 2023, Nvidia forecast $11 billion in revenue for the coming quarter. Analysts had penciled in $7.2 billion. The stock went up 24% in after-hours trading, adding roughly $200 billion in market cap overnight.

That's not a beat. That's a signal. The kind that only happens when the market is realizing it got the entire structure of an industry wrong.

The companies racing to build AI—OpenAI, Google, Meta, every startup with a foundation model and a pitch deck—all needed the same thing: H100 GPUs. Not "preferred" them. Needed them. Nvidia's CUDA software ecosystem, built over two decades of dominance in gaming and scientific computing, had become the only realistic substrate for training large language models at scale. Nobody had bothered to build a real competitor because the market hadn't demanded one. Then the market demanded everything, all at once.

Jensen Huang understood what was happening before most CEOs in his position would have called another strategy meeting. Nvidia wasn't an AI company. It was the pick-and-shovel seller in a gold rush—except the picks and shovels required 80 billion transistors and 18 months of lead time.

The chokepoint dynamic is ancient. Whoever controls the physical layer controls the economics. Standard Oil didn't need to own every oil well; it owned the pipelines. AT&T didn't need to own every phone; it owned the switches. Nvidia doesn't need to build the AI; it needs to be the only viable way to build the AI. For now, it is.

The $200 billion added to Nvidia's market cap in one trading session wasn't investors getting excited about chips. It was investors realizing they'd been tracking the wrong variable. The AI race wasn't decided by who had the best model or the most data or the most brilliant researchers. It was decided years earlier, by a company that made a software toolkit called CUDA that nobody wanted to leave.

Datacenter AI revenue was $4.28 billion that quarter—roughly double what it was two years prior, and accelerating. The real shock was the guidance: $11 billion the next quarter, against the $7.2 billion analysts expected.

There's a version of this story that's triumphant: one company's long-term technical investment paying off exactly when the world needed it. There's another version where the entire AI economy runs through a single vendor's supply chain, with GPU allocation decisions made in Santa Clara determining which companies can compete and which ones wait. Both versions are true. The $11 billion guidance made them the same story.

Every sovereign nation is now working out how many GPUs it has. Some are discovering the answer is "however many Nvidia decided to sell them."

The arms race for AI didn't start in November 2022 when ChatGPT launched. It started much earlier, in a company's quiet decision to keep investing in programmable GPUs when everyone else thought GPUs were for gaming. The chokepoint was built long before the rush started.

I give it three years before someone ships a CUDA-killer that actually works. In the meantime, the sand-to-silicon pipeline runs through one company's earnings calls.