China is building its own Nvidia – and US trade policy is forcing their hand!

Watch on YouTube ↗  |  March 31, 2025 at 00:18  |  5:14  |  All-In Podcast
Speakers
GavinSBaker — Portfolio Manager, Atreides Management
Jason Calacanis — Angel Investor / Founder, LAUNCH

Summary

In this All-In clip, host Jason Calacanis questions guest investor Gavin Baker about newly expanded US export controls covering companies in the Nvidia ecosystem. Baker argues the controls are extremely hard to police, since a Blackwell GPU is far more valuable per unit than the illegal drugs the US already fails to intercept, but that the friction still buys America an advantage at the cost of pushing China to build its own semiconductor ecosystem and innovate algorithmically, as DeepSeek showed. He puts the odds of China matching or beating Nvidia within five years at about zero, while conceding a ten-year horizon is open because Beijing plans in decades. The discussion then moves to AI agents, where Baker expects Anthropic's model context protocol to become a standard, sees very high ROI on AI capex and Blackwell if agents scale, and identifies available compute as the binding constraint.

  • New export-control additions target quantum computing and much of the Nvidia GPU ecosystem
  • Baker: policing GPU smuggling is harder than stopping drugs given value per unit
  • Chips are routed through Singapore, Vietnam and other intermediaries; cat-and-mouse until a grand bargain
  • Export controls create an immense incentive for China to build a domestic semiconductor ecosystem
  • DeepSeek cited as proof of Chinese algorithmic innovation under hardware constraints
  • Odds of China matching Nvidia in five years seen as near zero; ten years is open
  • Model context protocol from Anthropic, adopted by OpenAI, expected to become the agent standard
  • Compute is the rate-limiting factor for agents; OpenAI's $2,000-$20,000 monthly pricing shows scarcity
Ideas
GavinSBaker Portfolio Manager, Atreides Management 1:45
Export controls force China's own chip ecosystem.
US export controls on advanced GPUs put enough friction into the system to give America a theoretical advantage, but the cost is that they create an immense incentive for China to build its own semiconductor ecosystem and to innovate algorithmically, which DeepSeek already demonstrated. Necessity is the mother of innovation, and China is actively trying now. Baker puts the chance that China creates something competitive with or better than Nvidia within five years at roughly zero, but says that over ten years it is genuinely open, because the CCP plans in decades or centuries while America thinks in much shorter cycles. The domestic Chinese chip and AI hardware build-out is therefore a slow-burn theme to monitor rather than an immediate trade.
GavinSBaker Portfolio Manager, Atreides Management 4:16
Agent adoption makes Blackwell ROI high.
If AI agents actually materialize as a reality, helped by standards like the model context protocol that make integration easy, then the return on AI capex and specifically on Nvidia's Blackwell generation will be very high. Baker frames real agent adoption as the demand validation that justifies the current spending on Nvidia hardware, because agents running tasks all day are a continuous consumer of inference compute.
GavinSBaker Portfolio Manager, Atreides Management 4:38
Compute scarcity sustains AI chip demand.
Baker's second point on agents is that compute, not model capability, is the rate-limiting factor: in a world where everyone has agents doing things for them all day long, it will be a long time before there is enough compute in the ground for that to be a widespread reality. He cites OpenAI pricing its first agents at $2,000 to $20,000 a month as evidence of how scarce and expensive inference capacity is, which implies a long runway of demand for AI chips and the hardware supply chain behind them.
Up Next

This All-In Podcast video, published March 31, 2025, features GavinSBaker discussing KSTR, NVDA, SMH. 3 trade ideas extracted by AI with direction and confidence scoring.

Speakers: GavinSBaker  · Tickers: KSTR, NVDA, SMH