Broadcom reported $22.19 billion in revenue and $2.44 EPS against $2.39 expected, and the stock fell 12% in a day.
Read that again. A beat, and a company hemorrhaged up to $80 billion in market value in a single session. That is not the market discovering that Broadcom's business is weak. That is the market discovering that its own positioning was.
Two weeks ago, the story was Huawei's parallel stack — a China-compute problem, a question about whether Beijing could route around export controls with a homegrown silicon supply chain. This week the same shock crossed a border it wasn't supposed to cross. Kimi K3 didn't stay a China story. It became a Western valuation event.
Here is the tell. TSMC printed record second-quarter revenue north of $40 billion and its shares fell roughly 4%, dragging the Nasdaq 100 down about 1.4%. Records, and the stock drops.
Markets do not pay for performance. They pay for future cash flows under credible assumptions. A bear market is the moment the price stops containing the story it was built on — and by the time TSMC reported, the price already contained every good thing TSMC could say.
The Philadelphia Semiconductor Index hit 14,655 on June 21, then fell more than 20%, erasing roughly $2.1 trillion in market value. The median constituent is down 21%. That is a full bear-market entry, and the surface narrative — "AI hype is over" — is too lazy to survive contact with the mechanism underneath.
Decompose the drop and you find layers, not a verdict. There was a trigger, a capital-flow mechanism, and a structural signal. Only one of those three is about earnings, and it is the smallest one.
Start with the mechanism, because it is where the real story lives. The sellers were not long-only pension funds rotating into value. They were hedge funds — four consecutive weeks of net hedge-fund selling in chip stocks through early July.
Now put that against the run they were unwinding. The Q2 2026 SOX gain of nearly 90% was its best quarter on record. From 2025 lows to this year's highs, semiconductors returned more than 300%.
A trade that runs 300% and posts a record quarter is not under-owned. It is the most crowded book on the Street. The hedge fund late to the trade at the end of Q1 had no choice but to chase; the fund full at the peak had no room to add. Everyone was already long, which means the only trade left was to sell.
That is a positioning flush, not a fundamental repricing. But positioning flushes are precisely what create bear markets, because when everyone is on the same side, there is no marginal buyer to catch the fall.
The trigger, then, was almost incidental. Broadcom's number told you the buildout is real — $22.19 billion in revenue, a genuine beat on EPS. What broke the chain was that Broadcom declined to raise its 2026 AI semiconductor forecast, leaving a figure the market had already capitalized "sitting above what management could commit to."
That is the entire event, mechanically. The price had run ahead of what management would underwrite, and the moment management refused to underwrite it, the price had to come back to meet the guidance.
The price had run ahead of what management would underwrite, and the moment management refused to underwrite it, the price had to come back to meet the guidance.
Nvidia shed roughly $740 billion on a 6% move — a smaller percentage than Broadcom's, but off a vastly larger base. The dollar figures are terrifying; the percentage moves are ordinary de-risking. That gap is the whole thesis: this looks like an earnings apocalypse and reads like a book being flattened.
Kimi K3 arrived into exactly this setup — a 2.8-trillion-parameter open-weight model, weights downloadable, claiming parity with the strongest US systems. It was the second Chinese-model shock in a year, and traders reached for the "DeepSeek moment" label because it was lying right there.
But a model existing is not a demand event. The selloff was overdetermined — weak Netflix, the Iran war, rate fears, risk-off — and Kimi was one author among several. It did not refute a single earnings number. It detonated a trade that Broadcom's flat guide had already cracked.
So where do the bears actually have a case? Not chip volume. Pricing power.
This is the one place the sell-off got something right, and it is worth being precise about what. The threat from open-weight Chinese models is not that the compute stops getting bought. It is that the software layer sitting on top of the compute loses its ability to charge for access.
Look at the numbers. DeepSeek-V4-Pro charges $0.87 per million output tokens against $50 for Anthropic's frontier Claude. When a near-frontier model can be self-hosted and priced 50-fold below the closed incumbent, the API margin at the top of the stack is not defensible — and that is what the market repriced.
Then look at what happens to demand at the same time. Coinbase runs GLM-5.2 and Kimi in production and cut its overall AI spend nearly in half — even as its actual token consumption spiked.
Hold those two facts together, because they are the entire argument. Price-per-token collapsed. Total tokens consumed exploded. Spend fell in the middle only because the price fell faster than usage rose — for now.
That is the Jevons dynamic, and agentic workloads are its engine. A single innocuous request can spend millions of tokens through amplification, because a model with no memory reprocesses the entire exchange at every step. One report scheduled every fifteen minutes can run $2,880 a month; a tricky multi-stage version, ten times that.
So the compute keeps getting bought. TSMC records, sold-out memory, sold-out GPUs — none of that is fiction, and cheaper models make agents cheaper to run, which makes people run vastly more of them. The hardware layer stays constrained even as the token price falls through the floor.
The sell-off correctly repriced who captures AI's economics. It wrongly indicted whether the compute gets bought. Those are different claims, and the market collapsed them into one because a positioning flush needs a narrative to wear on the way out.
Which is where the efficiency-versus-demand debate reveals itself as a post-hoc rationalization. The numbers — record foundry revenue, spiking token consumption, sold-out accelerators — flatly contradict a demand-destruction story. The debate persists because it dignifies a mechanical de-risking as an intellectual repricing.
The US ban push tells you which layer is actually threatened. The administration is reviving its move against Chinese models on cybersecurity grounds — and critics note the two labs that benefit most are OpenAI and Anthropic, the closed duopoly. David Sacks put it plainly: the leading closed labs "want the government to eliminate their open-source competition." You do not lobby to ban a product that threatens your chip supply. You lobby to ban one that threatens your pricing.
Now the consequence, because multiple compression is never neutral. When the sector's multiple resets, it does not shrink demand evenly — it selectively starves whoever cannot self-fund.
Watch where the pressure lands. TSMC, on record revenue, raised its 2026 capex to $60–64 billion and shrugged; the record-revenue incumbents absorb a re-rating and keep spending. Intel, still loss-making with accelerating losses, reliant entirely on higher-risk funding, having just diluted shareholders — Intel cannot shrug. A higher hurdle rate is fatal to the balance-sheet-weak and a rounding error to the cash-rich.
So the "reset" does not shrink the buildout. It rewrites the financing hierarchy of it — concentrating the ability to keep building inside the handful of players who never needed the market's permission in the first place.
That is the fiction the $2.1 trillion exposed. Not that the chips get bought. That the money buying them was ever as broadly distributed as the stock prices pretended.