Kimi K3. Two point eight trillion parameters

I had the Bloomberg terminal open to six windows and none of them were saying anything I wanted to see. SOX down another 5.7%. Micron red. Marvell red. AMD bleeding out for the third straight session. And somewhere in Shanghai, a company I'd never heard of eighteen months ago just told the entire semiconductor complex that the emperor's data center might be missing some clothes.

Kimi K3. Two point eight trillion parameters, open-weight, and according to Moonshot's own benchmarks, beating GPT-5.6 on front-end coding while trailing only Fable 5 and GPT-5.6 overall. I don't care whether the number is exactly right. I care that the market believed it hard enough to erase roughly $3.3 trillion in global chip valuation since June 22, and that the full weights don't even drop until July 27. We are pricing in a model nobody outside Moonshot has actually run yet. That's not analysis, that's vibes with a Bloomberg subscription.

Here's the part that should keep every hyperscaler CFO up tonight, not just because of Friday's tape but because of what Friday's tape implies about the next eighteen months. The entire justification for roughly $700 billion in AI infrastructure spending — the number Apollo's Torsten Slok has been waving around like a warning flare — rests on a premise nobody actually stress-tested: that frontier intelligence stays expensive to build and therefore expensive to rent. Compute scarcity was the moat. Compute scarcity was the whole thesis behind every 40x-revenue multiple Nvidia and its downstream beneficiaries have carried since 2023.

Now a lab with a reported $31.5 billion valuation ships something that, if it holds up even at 70% of its claimed performance, is free to download, free to fine-tune, and free to self-host on whatever GPUs you already own. You don't need OpenAI's API. You don't need Anthropic's rate limits. You need electricity and a rack.

I keep hearing the DeepSeek comparison, and I keep wanting to throw something at whoever brings it up next, not because it's wrong but because it's incomplete. DeepSeek in January 2025 was a one-day gut punch — Nvidia lost $600 billion of market cap and then, weeks later, hyperscalers reported quarterly capex guidance that came in higher, not lower, and the whole scare got treated as a buying opportunity. That's the "DeepSeek precedent cuts both ways" argument, and I understand why bulls are clinging to it. TSMC just guided to higher capex this week, which on its face argues the same thing is about to happen again.

But sit with the asymmetry for a second. In January 2025, the market had to imagine what cheaper training might mean. In July 2026, we're eighteen months into an actual capex cycle where the hyperscalers have already spent the money, already signed the power purchase agreements, already committed to the depreciation schedules, and now a credible open-weight competitor is telling enterprise buyers they might not need to pay Anthropic or OpenAI rack rates to get most of the way there. This isn't a hypothetical discount to future training costs. It's a live threat to the monetization side of the ledger, which is the side that was always shakier than the spending side.

Andreas Steno Larsen says the selloff is overblown, that Kimi K3's pricing lands close to GPT-5.6 rather than dramatically undercutting it, that it leads on coding but is merely average elsewhere. Maybe he's right. Seven trading days from now the full weights land and either the benchmark claims survive contact with independent testers or they don't, and I genuinely don't know which way that breaks. But notice what already happened regardless of the outcome: SoftBank, the market's cleanest OpenAI proxy, dropped 9% in Tokyo on the announcement alone. Z.AI fell 28%. The KOSPI dropped over 6%, the Nikkei over 4%. That's not a rational Bayesian update on one company's benchmark table. That's a market that has been waiting for a reason to ask whether the AI capex story was ever fully priced for the possibility that intelligence gets commoditized faster than compute gets monetized, and Moonshot just handed it one.

And because nothing in this market travels alone anymore, Bitcoin dropped below $64,000 on the same news cycle, dragging Ethereum toward $1,800 and Solana toward $75. Whatever store-of-value mythology crypto still clings to, it moved like a Nasdaq beta trade the instant chip sentiment cracked. If your hedge only hedges when nothing's actually going wrong, it isn't a hedge, it's a correlation you haven't stress-tested yet.

Was the rout multi-causal? Sure — Netflix guided down nearly 9%, the Iran strikes hit their seventh consecutive night, industrial production disappointed. Kimi K3 was the trigger, not the whole story. But triggers matter precisely because they reveal what was already loaded in the chamber. What was loaded here was a market that had let "more compute always wins" go unquestioned for two straight years of parabolic capex guidance, and a Chinese lab just forced the question into daylight three weeks before the Fed even meets.

The full weights land July 27. Mark it. Whatever this market decides intelligence is actually worth, it's about to find out in public.