Chip Stocks Crashed 20% on Kimi K3 — Jensen Huang Says Buy the Panic

Here's a number that should stop you mid-scroll: $3.3 trillion. That's roughly how much global semiconductor market value got wiped out in a matter of days after a Chinese startup nobody outside the AI world had heard of dropped a model called Kimi K3. The Philadelphia Semiconductor Index (SOX) fell 20.2% from its June 22 record high, officially crossing into bear market territory on July 17 — its worst stretch since April 2025. And in the middle of that carnage, Nvidia's own CEO went on Axios and said the model everyone was panicking about is actually "excellent" and American companies should "absolutely" be using it.

Photo by Leeloo The First on Pexels

That's not a throwaway quote. It's Jensen Huang picking a fight with his own government while his stock gets hammered by the very thing he's defending. Worth digging into why.

What Kimi K3 actually did to the market

Moonshot AI's Kimi K3 is a 2.8-trillion-parameter open-weight model, released July 16 at the World Artificial Intelligence Conference in Shanghai, that performs competitively with top-tier Western models at a fraction of the per-token cost. That combination — near-frontier capability, open weights, and cheap inference — revived the same "DeepSeek moment" fear from early 2025: if AI intelligence is getting this cheap this fast, why do we need hundreds of billions of dollars in Nvidia GPUs?

Nvidia fell 2.2% on July 17 alone; Applied Materials dropped 5.6%. Coming off a 105% run in the SOX between March and late June, the sector was priced for perfection, so it didn't take much to trigger a sharp unwind. Treasury Secretary Scott Bessent added fuel by telling Fox Business the administration is investigating whether Chinese AI models were built on stolen US intellectual property — and separately, the Trump administration has accused Moonshot of illicitly obtaining Nvidia Blackwell chips to train the model in the first place.

Huang's bet: cheaper AI doesn't kill chip demand, it multiplies it

Huang's public defense of Kimi K3 puts him directly at odds with Bessent and the AI labs lobbying to ban Chinese models outright. But his logic tracks with something I laid out in SK Hynix's 'Efficiency Paradox': AI Savings Fuel More Memory Demand: when the cost of a resource drops, usage doesn't shrink — it explodes, because applications that weren't economical before suddenly are. Economists call it the Jevons paradox, and there's real data behind it here. Equivalent GPT-4-class inference cost roughly $0.14 per million tokens by August 2025, a 280-fold drop in two years thanks to DeepSeek, Qwen, and now Kimi. Over the same window, enterprise AI cloud infrastructure spending didn't collapse — it went from $11.5 billion in 2024 to an estimated $37.5 billion in 2026.

Cheaper tokens mean developers add deeper reasoning loops, bigger context windows, and multi-agent workflows that consume far more compute per task than the simpler queries they replace. If that pattern holds, a cheap, capable Chinese model isn't a threat to Nvidia's total addressable market — it's an accelerant, because it drags more use cases across the threshold of "worth automating."

Photo by valaymtw on Pixabay

Why the market isn't buying it yet

The problem is timing. Jevons paradox plays out over quarters and years; a 20% index drawdown plays out in days, and momentum funds don't wait around for the long-run thesis to prove itself. There's also a real, near-term risk Huang is glossing over: politics. The Trump administration's current export framework already caps Nvidia's China shipments at 50% of its US sales volume and routes every H200 chip through third-party verification before delivery — the so-called "50% Rule." If Bessent's IP-theft investigation escalates, or if Moonshot's alleged use of smuggled Blackwell chips gets confirmed, Washington has every incentive to tighten that framework further, not loosen it. Huang is arguing for openness at the exact moment his own government is leaning toward restriction, and Nvidia's China revenue sits right in the crossfire.

So is the selloff an overreaction or a real repricing?

I think it's a bit of both, and that's the honest answer. The SOX was up 105% in three months — some giveback was overdue regardless of what triggered it, the same dynamic behind the sharp reversal I covered in SOXL Fell 23% in a Day While Chips Dropped Just 8%. But the underlying fear — that cheap Chinese models permanently shrink the case for expensive US silicon — doesn't hold up against the token-consumption data so far. Enterprise infrastructure spend has kept climbing straight through every prior "cheap model" scare, not despite falling prices but partly because of them.

What would change my mind: if Kimi K3 usage stays confined to cost-conscious developers substituting away from GPU-heavy inference rather than expanding total task volume, or if Washington actually tightens China export rules in a way that dents Nvidia's realistic revenue base (not just the headline threat of it). Both are plausible. Neither has happened yet.

• Bull case: Jevons paradox holds, cheaper inference expands the market, Nvidia's non-China demand absorbs the slack.
• Bear case: China export rules tighten further, and cheap open-weight models genuinely cannibalize premium inference spend faster than new use cases appear.
• Watch: Q3 earnings from Nvidia, AMD, and the hyperscalers for actual capex commentary — not stock reactions to headlines.

This is not financial advice — always do your own research before making investment decisions.

If you're trying to figure out where you actually sit in this selloff, the simplest move today is to pull up your own chip exposure and separate what you own by "China revenue at risk" versus "domestic hyperscaler demand" — because this news hits those two buckets in opposite directions, and most portfolios aren't built to tell them apart.