China's Kimi K3 Triggers a Second 'DeepSeek Moment' — and a Global Chip Selloff

A Chinese AI startup valued at roughly $20 billion released a model overnight, and by the end of the week the global semiconductor trade was in retreat. Moonshot AI's Kimi K3 — a 2.8-trillion-parameter open-weight model — landed with enough force that investors began openly relitigating whether America's AI capital spending boom can be justified. Bloomberg's China correspondent Minman Lo, reporting from the World AI Conference in Shanghai, called it the largest model China has produced to date; a Greylock general partner interviewed on Bloomberg Tech went further, describing it on air as the largest open-weights model ever released.

The numbers that moved the market

The pricing figures driving the reaction — roughly $3 per million input tokens and $15 per million output tokens — were cited on Bloomberg Tech by Greylock general partner Saam Motamedi, framing the argument others were making rather than as Bloomberg's own reporting. Moonshot says the model outperforms all rivals except Anthropic's Fable 5 and OpenAI's GPT-5.6.

The market reaction was immediate and international. Bloomberg Tech put the Nasdaq 100 down 1.7 percent intraday and the Philadelphia semiconductor index — the SOX — down almost 3 percent, on track for its biggest weekly decline since April 2025 and trading at its lowest level in months. The anchor noted on air that Bloomberg's team had written the SOX was set to fall into a bear market, though the broadcast did not confirm it reached one. Nvidia was the biggest points drag and Intel fell significantly. Apple at one point briefly overtook Nvidia as the world's most valuable company — a move Bloomberg attributed to a broader rotation out of the big AI spenders.

In Asia the damage was concentrated in memory. On Bloomberg's Washington broadcast, the Kospi was described as down more than 6 percent as investors reassessed whether the more than $600 billion Samsung and SK Hynix are committing to new fabs can be justified "given that China is doing it better and cheaper." A separate Bloomberg Asia broadcast covering the following Monday session showed a messier picture: SK Hynix off 2.1 percent, Samsung weaker by almost 3 percent, and the Kospi on what the anchor called "another wild ride" after a wild week, while the Hang Seng Tech index rose. These are distinct sessions, not a single move.

Not one model — two

Kimi K3 did not arrive alone. Alibaba unveiled a preview of an upgraded Qwen model at 2.4 trillion parameters, which Lo said ranks second only to Anthropic's Fable 5. Two frontier-adjacent Chinese models inside a weekend is what produced the "second DeepSeek moment" framing repeated across every Bloomberg broadcast reviewed here. The structural point Bloomberg senior editor Mike Shepard emphasized is not benchmark scores but distribution: these are open-weight releases you can download, modify and tune, which he said is much harder to do with the frontier models US labs are developing.

The counter-argument the selloff skipped

The most useful dissent came from Motamedi, who called the release "amazing" and praised Moonshot's efficiency techniques while arguing the 24-hour reaction was "perhaps a little bit premature." His three objections are worth isolating. Benchmarks are imperfect, and he cautioned against jumping to the conclusion that Kimi is now competitive with the best. Kimi K3 is, he said, more expensive on a per-token basis than Kimi K2 — directly contrary to the cheapening narrative driving the selloff. And K3 is "not particularly token efficient," meaning that for a given task it uses many more tokens than an OpenAI or Anthropic model, so token price and task cost diverge.

The Bloomberg anchor added a caveat that deserves more weight than it received: nobody outside Moonshot knows what K3 cost to train. There is an idea of the figure for K2. For K3 there is not. Every claim that China achieved frontier performance for a fraction of American spending currently rests on an unpublished number.

A compute ceiling, and an accusation

There is also evidence China's own constraint is real. Lo reported that Moonshot has paused new subscriptions after a surge in demand over 48 hours, prioritizing existing customers — a decision she tied to a domestic GPU shortage stemming from US export controls, and one that could curtail the deployment rate of these very large models. That framing was contested on-air by Kai-Fu Lee, who told Bloomberg export control "was an imaginative but unsuccessful effort to contain China. It has failed," attributing any shortage instead to insufficient supply or to businesses being frugal or choosing not to buy.

Running underneath all of it is the distillation fight. Shepard reported that US companies have pressed the case that their model outputs are being harvested to build a rival generation of chatbots in China on the cheap. No one has pointed the finger specifically at Kimi K3; Bloomberg reported weeks earlier that Anthropic accused Alibaba of a large-scale distillation campaign. Shepard noted distillation is itself a commonly accepted technique in AI. China's deputy foreign minister publicly rejected the accusations — what Shepard called the first and most prominent public statement from China's government on the issue — and the dispute is expected to surface if a Trump-Xi meeting proceeds.

What it signals

The selloff is best read as a repricing of moat, not of demand. TSMC said this week it sees demand for AI chips and hardware running through 2030 that will be hard to meet, and Bloomberg reported SK Hynix's CEO echoed that message. What changed this week is investor confidence that American labs hold a durable lead — not the evidence that the chips will be bought.