Hacker News Reader

Comment by besterman23 | Hacker News Reader

This fits, if NVidia fully believed in the LLMs/AGI future, why would they waste their super valuable GPUs by selling them? Why not just make your own LLM with the advantage of being one of the few major compute providers?

View on HN

There's some kind of skill needed in an organization to make those llms. Openai and anthropic were able too. The other llm creators, not really. There's no guarantee that Nvidia would be another leader and not end up like meta or xai's attempt

Deepseek claims to be making $38/gpu/hr on deepseek 4.1 flash, and I would say that’s probably marginally close to the frontier, so they don’t have to be as good as OpenAI/Anthropic. Even at half that revenue, NVidia makes back the cost of a b200 cluster in roughly half a year. (19*8*24*5*52 = 948,480)

It’s estimated NVidia sold 4.5-5.2 million gpus in 2025 with total revenue numbers at about $130b.

With this terrible math in mind, they could produce half of that and make substantially more money, basically an infinite money glitch.

That Deepseek number doesn’t account for any costs, though. That’s the biggest barrier to the infinite money glitch - it relies on zero cost LLM creation and operation.

Sure, but the dumb math winds up around $70b more a year in revenue per year with half their gpu output. Like it is actually kind of an absurd number if it’s true. They’re leaving essentially $30-40k on the board with each gpu they sell.

Unless it costs more than that figure to make a near frontier model and serve it, I can’t see how it wouldn’t be a slam dunk after a two or three year period.

They would even have extensive leg up in providing the most efficient inference on their own hardware.

This is all sort of in jest, but my point is, if the gold is real and all it takes is the shovels and money to hire people to use them, then it’s a big question as to why the shovel maker isn’t capitalizing.

Yeah, but I also don't think it's that deep. E.g., by the same line of thinking, TSMC should be doing what Nvidia / Apple / AMD are doing, and ASML what TSMC is doing.

NVidia is good at making the shovels, and sticks to that.

It’s not, but for fun I think that depends on if you see the models as just another technology or as a “country of geniuses in a datacenter”. The former means NVidia believes LLMs are fundamentally limited, and if that’s the case OpenAi and Anthropic are basically screwed as their core products become commodities.

Nvidia already does that, their models just aren't competitive.

Google is doing it successfully to the point where they only let you rent their shovels and never let you buy them.

What’s interesting about Google is, they don’t seem to care to be the best LLM either, and merely maintain comparability to the leading LLMs as they see fit. I think they realize their best position is to serve an LLM as a complement to their search function, and if people want a more LLM heavy experience they can have that as an alternative.

Reply on news.ycombinator.com