American AI dominance would be destabilizing; the real risks are commodity economics, funding timing and compute dependency.
Very problematic, he argues. Take the fantastical case: controlling AI makes your military better than anyone else's. China's game-theory optimal response is then simply to blow up TSMC. If the U.S. reaches meaningful superiority, particularly from a military and national security perspective, that is very dangerous for the world.
Not outside a conflict. Talk of fixing fabs, actuators and other precursors is magical thinking; sourcing from the U.S. while your competitor sources from China is such a disadvantage you simply won't do it. Apple was browbeaten into some iPhone production in India, but it is diversifying, not truly leaving China.
OpenAI and Anthropic are clearly on the frontier, who knows what is happening with Google, and Grok and Meta are chasing. The Chinese are very capable and very smart, distilling these models to stay about six to nine months behind. He calls that a pretty good equilibrium, generally favorable to the U.S.
No. Calling them free confuses R&D with marginal cost. You still have to run inference, like GLM or Kimmy, and Kimmy is very expensive to serve, with a significantly higher cost per answer. You are not paying the R&D that created the AI, but you definitely pay the inference.
See's is a tremendously high margin business, but the absolute profit is capped with no reinvestment runway. BNSF's margins are far worse, yet in one year it threw off more free cash than See's had over its entire lifetime. Google search may be See's Candies, and AI the BNSF.
He is not fully convinced. AI is clearly incredible at coding and obviously very good at math, but those are verifiable domains. The compelling evidence that this translates cleanly to unverifiable domains, or ones with very long verification loops, is missing. Chess and Go were knowable, bounded domains solved by scale.
Zero marginal costs mean you already have distribution; your problem is demand, which is why you pay for ads. In a world of abundance the hard problem is discovery, so the companies that solve discovery in their domain dominate it and get a virtuous feedback loop. Transaction costs are zero too.
Frustrating but not surprising. Silicon Valley relearns two things every ten years: consumers do not want to pay for software, and consumers do not care about being productive. This market should be supported by advertising. OpenAI replayed the Dropbox story at a hundred times scale, selling many subscriptions but not enough.
Because TSMC did not invest over the last few years, the shortages will be so acute that big tech companies foregoing enormous revenue and profits will go through the pain of getting Intel, and Samsung's logic, up to speed. Scarcity ultimately saved Intel; he expects a major partner announced this year.
Not really. The circular financing — a 25% backstop, equity in the neoclouds, guaranteeing to buy their compute through 2030 — exists so those entities get a lower cost of capital, which means NVIDIA assumed the risk. Taking on risk has a price, so holistically that is a price cut.
“To blow up TSMC.”
“I think we need to push forward because we can't go back.”
“Number one, consumers do not want to pay for software.”
“NVIDIA's position is, I think, definitely unnatural.”
“What Happens When the AI Boom Runs Out of Money.” https://www.youtube.com/watch?v=h-0NZ-oIjlk. Transcript summary by WhipScribe, https://whipscribe.com/transcript-pages/what-happens-when-the-ai-boom-runs-out-of-money.html.