Three stories about the same company landed on one day — and together they tell you where AI's money is, and isn't, going.

This morning, OpenAI showed off a custom chip, Jalapeño, built to wean itself off Nvidia's hardware and the margins that come with it. Around the same time, Nvidia's own researchers demonstrated that the next jump in capability may live in software: an agent "harness" that took an ordinary model from about 30% to a perfect score on a hard reasoning benchmark's public set — no new silicon required. Two signals, both pointing away from the idea that Nvidia's chips are the whole story.

Then, after the US close, Nvidia reported its quarter.

The number

$96.2 billion in revenue for the three months to late July — up 106% on a year ago, and comfortably past both its own guidance and Wall Street's expectations. Data-centre revenue alone was $89 billion, up 117%. Non-GAAP earnings came in at $2.22 a share against expectations nearer $2.08. Gross margin held at 75%. And the guidance for next quarter is $108 billion.

This is a company selling everything it can make, at the price it wants, to customers who cannot get enough. CEO Jensen Huang's line caught the mood: "AI has reached its inflection point… Now, compute is revenue."

What the three stories say together

Here is the tension worth sitting with.

Everything about the direction of the industry says Nvidia's position should be eroding. Its biggest customers — OpenAI, Google, Amazon, Meta — are all building their own chips. The frontier is visibly shifting toward the software wrapped around models, where Nvidia sells nothing. Custom silicon and clever harnesses are meant to be the escape route from the "Nvidia tax."

And yet the magnitude says the opposite. While everyone plots their escape, they are paying Nvidia more than ever — 106% more than a year ago. OpenAI's Jalapeño won't ship in volume until 2027. Google's and Amazon's in-house chips have existed for years, and Nvidia's data-centre revenue still more than doubled. The alternatives are real; they are also, for now, a rounding error against the demand.

The bet underneath the beat

The figure that should give pause isn't the revenue — it's the commitment. Nvidia's forward supply and purchase commitments jumped from around $119 billion to $279 billion in a single quarter, much of it locking in memory well ahead of need. That is the sound of a company wagering enormous sums that today's demand is not a bubble.

If it is right, "compute is revenue" is the defining equation of this cycle, and everyone racing to build their own chips is racing to catch a train that is still accelerating. If it is wrong — if the custom silicon lands at scale, or the harnesses make each GPU do more, or the AI-capex boom simply cools — those commitments become a very large number to have written down.

For one more quarter, at least, the tax everyone is trying to leave just went up.