On the last day of November 2022, a research lab in San Francisco quietly put a chatbot on the web. There was no launch event, no advertising, no keynote. OpenAI called it a "research preview" and expected, by its own later admission, a modest response. Within five days, a million people had signed up. Within two months, ChatGPT had reached an estimated hundred million users, making it the fastest-growing consumer application in history to that point.
Almost nobody, in those first giddy weeks of screenshots and disbelief, was thinking about the hardware. But every clever answer ChatGPT gave, every poem and pun and block of Python, was a burst of arithmetic running on a rack of chips in a data center. And nearly all of those chips carried the same logo: a green, swirling eye. NVIDIA had spent a decade quietly building the machinery of artificial intelligence. Now the whole world wanted to use it at once.
The shovel-sellers strike gold
The chips doing the heavy lifting were the H100 — the Hopper-generation accelerators we met last time, purpose-built for the transformer models that power large language systems. Training a frontier model consumes thousands of them running in concert for weeks. But the surprise of 2023 was inference: once ChatGPT was live, every single query needed GPUs too, and there were suddenly hundreds of millions of queries a day. Demand did not so much grow as detonate.
What followed was one of the strangest supply crunches the tech industry had ever seen. The H100 became the most coveted object in Silicon Valley. Lead times stretched from weeks to the better part of a year. Startups listed their GPU count in pitch decks the way earlier founders had listed revenue. Reports put the price of a single H100 in the tens of thousands of dollars, and even at those prices you had to wait in line. Elon Musk quipped that the chips were, at that point, harder to get than drugs — a line that was funny precisely because everyone in the room knew it was true.
NVIDIA had spent thirty years learning a lesson from the gold rushes it had lived through — the crypto boom of Part 11 chief among them. When everyone is digging for gold, sell the shovels. In 2023 the entire technology industry was digging for the same gold at the same time, and NVIDIA was very nearly the only shovel store on earth.
The day the numbers broke belief
Wall Street felt the tremor before it understood the earthquake. On May 24, 2023, NVIDIA reported a solid-but-unremarkable quarter — and then delivered guidance that looked like a typo. The company told investors to expect about eleven billion dollars in revenue for the following quarter, against analyst expectations closer to seven. The next morning the stock leapt roughly 24 percent, adding on the order of 184 billion dollars of market value in a single trading day — one of the largest one-day gains any company had ever recorded.
Then the actual quarter arrived, and it was somehow bigger than the impossible guidance. On August 23, 2023, NVIDIA reported revenue of $13.51 billion for the quarter ended July 30 — up 101 percent from a year earlier and 88 percent from the previous three months. The engine underneath was the Data Center division, which booked $10.32 billion, up a staggering 171 percent year over year. Gross margin climbed past 70 percent. This was not a good quarter. This was a company changing shape in real time.
"A new computing era has begun," Jensen Huang said as the results landed. It was the kind of line CEOs deploy on ordinary Tuesdays, but for once the arithmetic backed it up. Just four quarters earlier, the same Data Center line had brought in under four billion dollars. In the span of a year, ChatGPT had roughly tripled it.
Everyone builds an ark
The scramble was not confined to startups. The giants of the cloud — Microsoft, Google, Amazon, Meta, Oracle — announced massive H100 build-outs almost in unison, each terrified of being the one platform that couldn't offer customers the GPUs they were begging for. Microsoft, fresh off its investment in OpenAI, wired ChatGPT into Bing and Office. Every one of those integrations was, at the silicon level, another order for NVIDIA.
It was the vindication of a bet placed years before anyone had a word for it. The CUDA gamble of 2006, the DGX box hand-delivered to OpenAI in 2016, the transformer engine baked into Hopper — each had looked, at the time, like an expensive act of faith. In 2023 they all paid off at once. NVIDIA didn't just sell into the AI boom; it had spent a decade building the boom's foundation, and now it owned the only road into it.
The money, by any historical standard, was absurd. And it was about to get more absurd still. Because a company growing this fast, throwing off this much cash, with this much of the future running through its chips, was accelerating toward a number no chipmaker had ever touched.
Next time — Part 18: "One Trillion Dollars." How a company that started in a Denny's booth became the first chipmaker to cross the twelve-zero milestone.
Comments
Post a Comment