It’s the business of commodities. The magic is in tiny incremental improvements and distribution. DeepSeek forces us to question if AI—possibly intelligence—is a commodity.
No, but it's good enough to replace some office jobs. Which forces us to ask, to what degree is intelligence--unique intelligence--required for useful production? (We can ask the same about physical strength.)
I find it interesting that so much discussion about “LLM’s can do some of our work” is centred around “are they intelligence” and not what I see as the precursor question of “are we doing a lot of bullshit work?”
My partner is in law, along with several friends and the amount of completely _useless_ work and ceremony they’re forced to do is insane. It’s a literal waste of their talent and time. We could probably net most of the claimed AI gains by taking a serious look at pointless workloads and come out ahead due to not needing the energy and capital expenditure.
Surely that would be amazing for NVDA? If the only 'hard' part of making AI is making/buying/smuggling the hardware then nvidia should expect to capture most of the value.
No. Before Deepseek R1, Nvidia was charging $100 for a $20 shovel in the gold rush. Now, every Fortune 100 can build an O1-level model with currently existing (and soon to be online) infra. Healthy demand for H100 and Blackwell will remain, but paying $100 for a $20 shovel is unlikely.
Nvidia will definitely stay profitable for now though, as long as Deepseek’s breakthroughs are not further improved upon. But if others find additional compression gains, Nvidia won’t recapture its old premium. Its stock hinged on 80% margins and 75% annual growth, Deepseek broke that premise.
There still isn't a serious alternative for chips for AI training. Until competition catches up or models become so efficient they can be trained on gaming cards Nvidia will still be able to command the same margins.
Growth might take a short-term dip, but may well be picked up by induced demand. Being able to train your own models "cheaply" will cause a lot more companies and departments want to train their own models on their own data, and cause them to retrain more frequently.
The time of being able to sell H100 clusters for inference might be coming to an end though.
It’s the business of commodities. The magic is in tiny incremental improvements and distribution. DeepSeek forces us to question if AI—possibly intelligence—is a commodity.