Nvidia's record $150 billion buyback shows where AI compute money ends up

A product team's monthly inference bill goes to a cloud provider or a model lab, and part of that money moves on to the chipmaker ^[8]. According to CNET, Nvidia gets paid whenever anyone builds or upgrades AI capacity, before the buyer has shown it can cover the cost ^[8]. "AI startups are burning through investor cash and handing it straight to Nvidia to buy chips, even though the startups themselves aren't making a profit," said Robin Wigglesworth, a Financial Times reporter ^[7].

Nvidia's own figures show what happens to that money next. The roughly $26 billion it returned to shareholders last quarter was about 27% of the quarter's revenue ^[1]. Growth of 106% means the same quarter a year earlier brought in about $46.7 billion ^[2]. Monday's $150 billion also comes on top of roughly $85 billion in repurchase room the board had already approved ^[3]. CNET calls Nvidia the only company in the AI ecosystem making real monetary gains ^[2].

The figure in the announcement is bigger than the figure being spent. Ed Zitron, who hosts the podcast Better Offline, called the buyback "an attempt to calm very nervous investors around AI" ^[9]. He also pointed out that board approval is not a legal agreement, and Nvidia is not committed to buying any shares ^[10]. CNET notes the company has made big promises before, including a $100 billion data-center deal with OpenAI that never materialized ^[11].

CNET's report does not include compute prices or a forecast for them. The case for planning on flat costs comes from where the cash sits. The supplier with the margin is putting its cash into its own shares, not into more AI startups, CNET notes ^[12]. The hyperscalers renting out its chips are borrowing ^[6], and the labs still cannot show investors standalone revenue ^[5]. Borrowers repay debt out of what their customers pay them. I'd plan on flat prices with moderate confidence, and I'd give low confidence to any roadmap that needs them to fall.

The test is to re-run the budget once. Price each compute-heavy feature at today's unit cost, held flat through fiscal 2028, the end of Nvidia's buyback window ^[1]. Then sort each feature two ways: whether it pays for itself at that price, and whether the price is locked by contract or can move when the vendor changes it. Features that pay at a locked price ship. Those that pay at a floating price also ship, with someone assigned to read the vendor's repricing notices. A feature that loses money at a locked price loses it for the whole contract term, so renegotiate it or cut it. The last box holds features that only pay if prices fall; those are bets on falling prices, and they get re-scoped. The cost of this approach is under-building. If compute does get cheaper, the team gives up a year of features it could have afforded, and that gap is easier to close than a product priced below cost.