Stay informed with free updates

Nvidia and OpenAI this week announced the framework of an agreement in which Nvidia will invest “up to $100bn” in OpenAI. Here are a few key questions you might have about the news.

Dunno.

Dunno.

Nvidia has said previously that AI capacity costs approximately $50bn to $60bn per GW, of which it takes $35bn to $40bn in exchange for chips, etc. That implies OpenAI has agreed to pay Nvidia between $350bn and $400bn. Is that value in the right sort of ballpark?

Dunno.

Dunno.

Is it all incremental new revenue for Nvidia? Are the values additional to OpenAI’s reported five-year commitment to spend $300bn on Oracle data centres, or is there some double-counting?

Dunno.

Dunno.

Nvidia has estimated its total available market at $3tn to $4tn by 2030. Was this size of deal already in company guidance, and was this a competitive tender or should we presume Nvidia’s outer-years market share holds steady at around 30 to 35 per cent?

Dunno.

Dunno.

The press release also says Nvidia will become the “preferred strategic compute and networking partner” of OpenAI. However, it’s a non-exclusive arrangement and OpenAI has always described Nvidia as a key partner. What does this new collaboration mean in practice?

Dunno.

Dunno.

Dunno.

Dunno.

Dunno.

What would such an arrangement say about Nvidia’s relationship with the incumbent hyperscalers, Google, Microsoft and Amazon?

Dunno.

Does this recent slew interlocking deals between Nvidia, OpenAI and Oracle, as well as Nvidia’s aggressive backing of data centre start-ups like Nscale, suggest that Nvidia is building up dependencies with the industry’s smaller and weaker players to bolster longer term sales forecasts?

Dunno.

Dunno.

OpenAI’s revenue this year will be about $20bn, its cash burn is expected to hit $115bn by 2029, and it has yet to give a target for break-even. Where’s it going to get the $300bn to $700bn, approximately, it would need to pay for the Oracle and Nvidia deals?

Dunno.

“If Nvidia accounts for the full $100bn as an equity investment, then in order for Nvidia to achieve a venture-acceptable 3x return, OpenAI will have to double revenue every year for the next six years, and turn profitable somewhere along the line,” says analyst Jay Goldberg at Seaport Research Partners. Is 3x a reasonable sort of target or should we assume that the deal will be part-structured as debt, which would lower the return threshold?

Dunno.

In its second-quarter results last month, Nvidia reported cash and equivalents of $53.7bn. The consensus has Nvidia’s cash pile growing to about $88bn by the year end. What percentage of the $100bn investment (if any) will be payment-in-kind and how will it change Nvidia’s cash trajectory from 2026?

Dunno.

Compute revenue in Nvidia’s most recent results missed expectations. There was a theory that corporate AI budgets have been moving towards single-purpose silicon, where Nvidia’s lead over competitors may be less defensible. Is this latest announcement an indication that fears of a slowdown in demand for Nvidia’s general-purpose chips are overblown?

Dunno.

Dunno.

Dunno.

Further reading:
— Oracle’s astonishing jam-tomorrow OpenAI trade (FTAV)