For the same reason shovels were a better business than mining for gold.

The foundational model providers are hemorrhaging VC cash, with no certain path to profitability.

And for such a profitable business as Nvidia, the opportunity cost is massive. Any dollar burned trying to compete with them is a dollar Nvidia can’t invest in its cash cow business that’s currently generating whopping returns with little uncertainty.

When they sell GPUs, they get money immediately which they can spend to make more GPUs

If they hoarded GPUs, they'd then need to figure out how to make "winning AI" produce money - which nobody has done yet. In the meantime, they'd need additional outside money to keep running operations.

They already face plenty of risk (eg long payoff times for capital investments in fabs etc), why would they take on more?

Selling the hardware is very profitable in the here and now. Making a new foundational model may not be (none of them are profitable so far). If I have the choice between making a lot of real money right now and gambling that I might make a lot of money later, I'll take the now.

Because they are not better at OpenAI/Anthropic/Google at training foundational models and then building real products around those models.

Nvidia sucks at consumer software.

They do have Nemotron models that they release for free to the public but those are most likely done to dogfood their teams.

Nvidia is making a profit. Their AI customers are not.

you forget that Nvidia is winning far more victories than any of the AI companies. they get to absorb all the benefits of the AI bubble with none of the risks