KM4BIKM4BI

RSI Is Closer Than People Think, Says Tae Kim

Tae Kim argues AI compute demand remains underestimated as agents, possible RSI and HBM shortages offset current market fears.

Tae Kim frames the recent pullback in AI and semiconductor stocks less as a broken thesis than as a fear cycle: profit taking after a sharp rally, geopolitical noise, and alarming headlines around Meta, Kimi, ASML, OpenAI and Nvidia. His core point is that markets may still be underestimating the amount of compute that stronger models, agents and a possible RSI acceleration could consume.

Reading the market mood

Kim says sentiment has turned very negative even though demand signals remain tight. Meta is his example: one internal quote about agentic AI fed a story about weaker returns and lower capex, but subsequent signals suggested Meta was still leaning into heavier AI spending.

Efficiency does not mean less compute

The conversation compares Kimi with the DeepSeek moment. A more capable or cheaper model can create more use cases rather than a compute glut. Kim stresses that Kimi is not a tiny model: at 2.8 trillion parameters, it still requires substantial infrastructure to serve.

Agents, RSI and enterprise adoption

The strongest claim is about the next six to nine months. Kim argues agentic AI is taking off now, and RSI may be closer than many investors believe. If models begin using more compute to improve themselves, infrastructure demand could accelerate again.

Nvidia’s moat beyond CUDA

Kim’s Nvidia case is not just about CUDA. He points to system integration, networking/CPU/GPU co-design, balance-sheet strength and the ability to lock up scarce components such as optics, TSMC wafers and HBM memory.

What to watch

Headlines about vendor financing, large OpenAI-related commitments and power bottlenecks can move markets quickly. Kim’s advice is to wait for the economics of the actual deals before concluding that AI infrastructure demand has cracked.

Source