Tae Kim: sentiment is negative but AI compute demand is unstoppable — RSI may be closer than markets think

Tae Kim: AI compute demand is unstoppable — RSI may be closer than markets think

Tae Kim argues the current selloff in semiconductor and AI infrastructure stocks is a sentiment story, not a fundamentals one. Negative headlines have outpaced the underlying demand signal, and the market is making the same mistake it made during the DeepSeek panic a year ago.

The FUD cycle

The pressure on chip stocks traces to a few converging anxieties: the resumption of U.S. bombing in Iran (chip names fell sharply within two to three weeks of that development), a Reuters headline mischaracterizing a Zuckerberg internal town hall comment as a signal Meta would cut AI CapEx, and then a cluster of stories — a Wall Street Journal piece on vendor financing, the CMXT IPO in China, and an Information article on ASML — that hit the market in quick succession. Kim's read is that each of these, examined individually, either reversed quickly or was badly overstated.

The Reuters/Meta example is illustrative. One quote from an internal town hall leaked out of context, headlines ran with it for days, and then Reuters itself published a follow-up reporting that Meta is planning to raise CapEx dramatically this year and next. The panic had no durable basis.

On the Financial Times piece reporting NVIDIA is backstopping a $50 billion data center lease in Texas, Kim says the actual structure is a 15-year lease commitment with a renewal option — roughly $1–2 billion a year against a revenue run rate he estimates at $320 billion, heading toward $400–500 billion next year. He calls it a rounding error.

“Sam is on the record at the YC event that the next six months are going to be much more dramatically better for AI than the last two years. RSI, I think, is a lot closer than people think — both frontier labs are winking that it's going to happen very soon, and that's going to soak up an unbelievable amount of compute. SK Hynix executives said their customers are asking five to six times more than they're able to serve.”

The demand signal

The strongest piece of evidence Kim cites is AMD CEO Lisa Su raising her AI GPU total addressable market estimate from $120 billion to $220 billion in the space of three months. Kim's point is that serious CEOs of non-meme businesses running publicly traded companies for decades don't expand a TAM by that magnitude on a hunch.

SK Hynix executives said during their IPO roadshow that customers are requesting five to six times more HBM memory than they can supply, with plans to double capacity over five years. Kim reads Jensen Huang's statement that the chip industry has enough supply to double revenue every year as a forward signal the market hasn't priced in — NVIDIA's published revenue estimates for next year are well below that implied trajectory.

The Kimi K2 scare follows the DeepSeek template. The model has 2.8 trillion parameters, requires a 64-GPU server to run optimally, and its servers were overwhelmed on day one. Kim argues this is the opposite of a compute efficiency story — it's a demand accelerant. More capable models find more uses. Reasoning model adoption drove a compute surge last year; agentic AI is doing the same now.

RSI timeline

Kim says recursive self-improvement is closer than markets think. Both Anthropic (which published a blog post on the topic) and OpenAI have signaled it publicly, and Sam Altman stated at a YC event over the weekend that the next six months will see more AI capability progress than the previous two years combined. Kim describes AI researchers consistently liking his RSI-related posts, which he takes as implicit confirmation from people close to the frontier.

If RSI arrives in the next three to nine months, the compute absorption would be on top of an agentic AI ramp that is itself still early. The enterprise adoption gap remains enormous — one cited estimate suggests usage would 100x if every company adopted AI to the degree of the most advanced users. Kim puts the combined ARR of OpenAI and Anthropic at roughly $120 billion, against a $6 trillion global IT and knowledge management market, and sees no structural reason that figure can't reach $200–400 billion within a couple of years.

NVIDIA's structural position

Kim argues NVIDIA's moat is less about CUDA software lock-in than about three reinforcing advantages: the co-design of networking, CPU, and GPU into a coherent system; a balance sheet large enough to prepay and lock up constrained components (optical parts, TSMC wafers, HBM memory); and a growing portfolio of equity stakes in companies across the supply chain, from CoreWeave to optical component makers Lumentum and Coherent.

The vendor financing concern around NVIDIA's reported talks with OpenAI and SoftBank — reportedly up to $250 billion — Kim treats as premature. Neither NVIDIA nor OpenAI is commenting, and he argues that if hyperscale GPU cloud carries 60–80% inference margins (a Morgan Stanley estimate he cites), the financing structure looks like a supply-chain investment, not a subsidy. Jensen Huang has a track record: his CoreWeave stake is the most visible example.

Andy Jassy's shareholder letter framing sits underneath all of this. Amazon is committing $200 billion not on a hunch, Jassy wrote, but because the demand is visible and the free cash flow case is legible over a medium-to-long horizon. Azure is growing 40%, Google Cloud 80%, and Amazon cloud in high double digits. Kim's argument is that critics applying a static revenue model to a business compounding at those rates will keep drawing the wrong conclusion.