Kimi K3 Triggers a Second DeepSeek Shock as Chip Stocks Slide Worldwide

Eighteen months ago, a Chinese lab called DeepSeek erased roughly $590 billion of Nvidia's market value in a single session and taught global markets a phrase they have not been able to forget. On Friday, a Beijing startup named after a Pink Floyd album made them say it again.

Moonshot AI's release of Kimi K3 — a 2.8-trillion-parameter open-weight model that its makers claim performs competitively with Anthropic's Fable 5 at roughly a third of the price — detonated across Asian trading floors before most of Wall Street had woken up. Taiwan's TAIEX plunged 2,953.71 points, or 6.47%, closing at 42,671.27 in one of the steepest single-day falls in the index's history. Japan's Nikkei 225 shed 4.03%. The MSCI Asia Pacific Index logged its worst day in three weeks.

By the time US markets opened, the damage had crossed the Pacific. The Nasdaq slid about 1.5% in its worst session of the week, per Decrypt, closing near the lows. Nvidia fell 1.2%, briefly forfeiting its crown as the world's most valuable company to Apple. Taiwan Semiconductor dropped 7% — on a day it reported a 77% jump in quarterly operating profit. SoftBank, widely traded as an OpenAI proxy, fell 9%.

The trade that unwound

The mechanics were less about Kimi than about positioning. The Philadelphia Semiconductor Index has now fallen more than 20% from its late-June record high — a technical bear market — even though it remains up roughly 68% year to date, a reminder of how much air had been pumped into the trade before Friday.

Rich Privorotsky, a Goldman Sachs partner and the bank's head of equities, characterized the decline as a "deleveraging event," and issued the line that ricocheted around trading desks all day: the "era of compute expansion" may be over.

Privorotsky's deeper point was not that scaling laws had broken. It was that they had stopped being the only game. How, he asked, could a Chinese lab unable to match the pre-training scale of Western rivals close the gap this fast through architectural innovation, synthetic data, reinforcement learning and post-training technique? His answer: this does not prove scaling failed, but it shows "scaling is no longer the only winning path."

That is a more corrosive claim than it sounds. The entire capex thesis — the roughly $700 billion hyperscalers are on track to spend on AI infrastructure this year — rests on the premise that spending more buys a durable lead. If a $31.5 billion startup can arrive within touching distance of the frontier and then give the weights away, the premise wobbles.

Andrew Tyler, head of global market intelligence at JPMorgan, put it plainly in a client note: K3's release "undoubtedly added fuel to the fire," and "concerns over a 'DeepSeek 2.0 moment' are weighing on both Asian and U.S. tech stocks."

A sharper shock in Asia, a milder one for Nvidia

The DeepSeek comparison is instructive precisely where it breaks down. In January 2025, the R1 shock was concentrated and violent at the epicenter: Nvidia alone lost about $590 billion in market cap in one session. On Friday, Nvidia's 1.2% decline was almost genteel by comparison — a rounding error against R1 day — while the pain in Asia was far sharper, with Taiwan and Japan absorbing losses that the US indices never approached.

That asymmetry is the story. The market has stopped treating a Chinese frontier release as an existential threat to one American chipmaker and started treating it as a repricing of the entire supply chain that feeds the buildout — foundries, memory, equipment, the Japanese and Taiwanese names whose order books assume compute demand compounds forever.

Friday's rout also had help. Netflix fell around 9% on earnings, TSMC's results came with a $64 billion 2026 capex plan that exceeded expectations, Alphabet's Gemini 3.5 Pro has slipped by months, and US strikes on Iran entered a sixth night, pushing Brent above $85. Kimi was a trigger, not the whole gun.

Analysis: the assumption that keeps eroding

Not everyone read Friday as a regime change. Matt Britzman, senior equity analyst at Hargreaves Lansdown, framed it as an "unwinding of crowded trades" rather than a collapse of the AI thesis, calling it a "healthy correction in the hottest corner of the AI market." His condition for a bounce is specific: "If large-scale AI spenders can confirm their investment plans remain unchanged, sideline capital may quickly come back in."

Sell-side analysts were even more blasé. Morgan Stanley's Gary Yu wrote that "K3 has received positive feedback globally, signaling an all-round catch-up of Chinese LLMs with U.S. leaders in model size, performance, and pricing" — progress, not rupture. Bernstein's Robin Zhu called the launch "confirmatory."

They may be right about the model and still be missing what the tape said. Moonshot prices K3 at $15 per million output tokens against Fable 5's $50. DeepSeek's V4 sells output at $0.87 per million. All five of the most-used models on OpenRouter this week are Chinese. Each launch chips at the same load-bearing assumption — that frontier AI must be expensive and American — and markets have learned to sell first and audit the benchmarks later.

The DeepSeek precedent cuts both ways, though. That selloff was also billed as the end of the US AI trade, and the capex kept flowing. A model existing is not the same as enterprises adopting it.

What to watch

July 27. That is when Moonshot releases K3's full weights under a Modified MIT license, and the benchmark claims stop being a press release and become a testable artifact anyone can download and self-host. If K3 holds up under independent scrutiny, the pressure on US labs to justify their cost structures — and on their suppliers to justify their multiples — intensifies. If it doesn't, Friday goes down as an overreaction with excellent timing.

Watch the hyperscaler earnings calls after that. Britzman's tell — whether the big spenders reaffirm their capex plans — is the cleanest read available on whether Privorotsky's "era of compute expansion" epitaph was prophecy or just a very bad Friday.