By Mari Hansen When Nvidia’s stock closed at $208.27 on a Friday in late April, a figure briefly surpassed $5 trillion once more. The business had previously encountered this ceiling in October, slipped away from it during a tense winter, and is now fighting its way back. It’s difficult to avoid feeling that the figure’s ability to shock has diminished as you watch this play out. A few years ago, $5 trillion wasn’t associated with a company that produces chips that most people will never see, but rather with a national economy. The market hardly blinks as it scrolls by on a Friday afternoon.
Nvidia wasn’t even a major factor in the rally that got it there. Of all the companies, Intel—the one that was essentially written off in the AI debate—posted earnings that exceeded forecasts, and its stock shot up 24%, the highest single day since 1987. AMD increased by 14%. Qualcomm increased by 11%. Like a tide coming in, the entire semiconductor floor suddenly rose. Prior to earnings from the hyperscalers—Microsoft, Google, Meta, and Amazon—who collectively invested about $400 billion in AI infrastructure last year and don’t appear to have any intention of stopping, there is a feeling that investors weren’t so much picking winners as they were purchasing the entire concept of chips once more.
And yet. A distinct mood is remembered by anyone who kept an eye on things through December. The headlines about “AI fatigue” began to write themselves, the stock had fallen about 8% from its October highs, and fund managers were discreetly locking in gains. Perhaps that was just year-end rebalancing, the standard financial operations. However, it’s also possible that it was the market’s first sincere inquiry in a long time: what precisely are you paying for?
The price isn’t the intriguing change. It has to do with the nature of the demand. For two years, the narrative revolved around training—the massive, labor-intensive process of creating sizable language models, where Nvidia’s hegemony was virtually unassailable. Inference and the daily grind of actually running those models for millions of people are now the main topics of discussion.
Inference is more sensitive to cost. It is more vulnerable to competition, including AMD’s less expensive options and the custom silicon that businesses like Google are covertly producing internally to wean themselves off of third-party chips. A significant Nvidia client, Alphabet, has already revealed its own chips that will compete with Nvidia later this year. The experience of a supplier witnessing its largest purchasers begin constructing the product themselves is almost familiar.
During his keynote address at the GTC conference in San Jose back in March, Jensen Huang exuded the confidence of a man whose order book extends beyond $500 billion until 2026. The fundamentals are truly remarkable: gross margins are still close to 75%, and data center revenue alone reached $51.2 billion in a single quarter. These are not the figures of a troubled business.
However, there is a certain feeling associated with maturity, and this is beginning to feel that way. Spreadsheets and return-on-investment calculations are replacing desperate, FOMO-driven purchasing. The NVLink interconnect has drawn criticism from regulators, who openly question whether Nvidia’s ecosystem is a wall or a moat. Selling the Rubin architecture without cannibalizing Blackwell will be a challenging task, as it is already sampling.
Whether $5 trillion is a ceiling or merely a step is still up for debate. Before the skeptics stopped talking, Tesla endured years of skepticism. Nvidia might follow suit. However, it appears that the easy part, when everyone just believed, is coming to an end. It will take more time to earn what comes next.