RSI Is Closer Than People Think, Per Tae Kim
Speakers
Tae Kim
— Senior writer, Barron's; author of The Nvidia Way
Summary
Tae Kim argues that AI trade fears are overblown and compute demand will accelerate. He sees Kimi and better open-weight models increasing, not reducing, compute demand, and says recursive self-improvement is closer than expected. He defends Nvidia's moat, supply position, and strategic investments while highlighting overwhelming demand for AI silicon and HBM.
- AI sentiment is negative, but Kimi's large model and 64-GPU inference require massive compute.
- RSI could arrive within 3-9 months and consume enormous compute.
- SK Hynix sees 5-6x demand versus supply; AMD raised its AI market forecast to $220B.
- Hyperscaler cloud revenue growth and GPU cloud margins justify AI capex.
- Nvidia's CUDA moat, co-design, supply lockups, and ability to double revenue are underappreciated.
- Nvidia data center lease and vendor financing fears are overstated; its strategic stakes may pay off.
- Meta capex fears are false; AI ROI shows up in ads and its open-weight push.
Ideas
Meta AI capex fears are overblown.
Meta's AI spending fears are based on a leaked out-of-context town hall comment, but Meta is actually going full force and is expected to raise capex dramatically. AI investment is already delivering ROI through its advertising business, so the market's pullback on Meta over AI spend is premature.
Kimi boosts compute demand, not reduces.
Kimi is not a small efficient model but a 2.8 trillion parameter model that requires large 64-GPU servers and runs well on Nvidia, so open-weight models will increase compute demand rather than cut it. He sees the same dynamic as DeepSeek's reasoning models, where more capable models create more uses; agentic AI is taking off and the next 6-9 months will surprise positively.
AMD, SK hynix demand overwhelming.
AI semiconductor demand is overwhelming: SK Hynix customers are asking for 5-6x what it can serve, and AMD's CEO raised the 2030 AI market forecast from $120B to $220B in three months. HBM memory is in shortage, and sober management would not raise TAMs this aggressively without visible demand.
Cloud growth justifies AI capex.
Hyperscaler GPU cloud businesses are not dot-com capex: inference GPU cloud can earn 60-80% margins, and cloud revenue is growing fast with Azure around 40%, Google Cloud around 80%, and Amazon high double digits. That growing cash flow justifies the AI buildout and can make data centers highly free cash flow generative in 12-18 months.
