The Free AI That Crashed Chip Stocks Just Sold Out

Key Takeaways

• The PHLX Semiconductor Index fell 20 percent from its June 22 close through July 17, but Nvidia lost less than 3 percent over the same stretch. The wipeout landed on memory and momentum names, with Micron down 30 percent from its June peak.
• The “free model” that triggered it is not free. Kimi K3’s API prices output at $15 per million tokens, its weights are not public until a promised July 27 release, and its paid consumer tier stopped accepting new subscribers three days after launch.
• Three weeks before the panic, Bloomberg reported that AI revenue had crossed the industry’s depreciation cost for the second consecutive quarter, an estimated $25 billion against $21 billion in the first quarter of 2026.
• The resource that ran short during the crash was computing capacity, not customers. Whether that makes the panic wrong is the open question of the next two quarters.

A Bear Market and a Waiting List in the Same Week

On Friday, July 17, the PHLX Semiconductor Index (SOX), the 30-stock benchmark for the American chip trade, closed at 11,673.89. That was a 10 percent loss for the week and left the index 20 percent below its June 22 close, the classic threshold of a bear market. The Wall Street Journal reported that the surprise release of a new Artificial Intelligence (AI) model from China’s Moonshot AI “intensified a selloff in chip stocks on Friday, fueling concerns about competition in AI and massive corporate spending that underpins its build-out.”

That Sunday, July 19, Moonshot announced it would stop taking new paid subscribers for the very model at the center of the panic. Since Kimi K3’s release, the company said, user requests had “sharply exceeded forecasts” and were approaching the limits of its existing computing clusters, a situation it called “unprecedented compute challenges.”

Hold those two facts next to each other. The story traders told was that a Chinese lab had made frontier intelligence too cheap to justify paying for the machines. The Chinese lab in question simultaneously ran out of machines. If you own an index fund, a retirement account, or any of the AI trade at all, the gap between those two sentences is where your money currently lives.

What Is Kimi K3?

Kimi K3 is the new flagship model from Moonshot AI, a Beijing startup whose backers include Meituan, China Mobile and the investment firm CPE. The company describes it as “a 2.8T-parameter model built on our Kimi Delta Attention and Attention Residuals, with native vision capabilities and a 1-million-token context window,” and calls it “the first open model to reach 2.8 trillion parameters.” The announcement landed on July 16, and by Friday’s close the reaction was global.

The performance claims are the reason markets moved. On Moonshot’s own evaluation suite, Kimi K3 “performed competitively with Fable 5” while it “substantially outperformed Opus 4.8, GPT 5.6 Sol, and GPT 5.5,” which would put a Chinese model shoulder to shoulder with the best American systems. Independent scoring is more measured but still remarkable: Artificial Analysis, the benchmark aggregator, gave Kimi K3 a 57 on its Intelligence Index at launch, a debut at third place overall, “comparable to Opus 4.8 and GPT-5.5” but “behind Fable 5 and GPT-5.6 Sol.”

Now the part the selloff narrative skipped. The model that supposedly made intelligence free is, as of this writing, neither free nor open. The weights, the downloadable files that would let anyone run the model on their own hardware, are promised “by July 27, 2026” and have not shipped. The API costs $3.00 per million input tokens and $15.00 per million output tokens. That is roughly four and a half times the $0.66 and $3.41 that OpenRouter lists for Moonshot’s own previous-generation Kimi K2.6, as of July 21. The consumer app’s paid tier is the thing that just closed to new customers. As of July 21, Kimi K3 is a proprietary, paid, capacity-constrained product with a press release about openness attached, part of a deliberate national strategy this site mapped in China’s Next Export Ban Is Its Own AI. The strategy became explicit the same week the market broke. On July 16, Beijing launched the World AI Cooperation Organisation (WAICO), signing up 29 member countries, and the next day Xi Jinping told the World Artificial Intelligence Conference in Shanghai that countries should seize the “historic opportunity” of open-source AI, a pitch Reuters described as “framing its open-source models as a global public good.”

Why Did Chip Stocks Crash in July 2026?

The one-word answer traders reached for was DeepSeek. In January 2025, a cheap Chinese model wiped out a record amount of Nvidia’s market value in a single day, and the Journal reported that Kimi K3’s release “raised concerns about a potential repeat of a market downturn similar to the DeepSeek incident.” The logic runs like this: American AI valuations rest on the assumption that frontier intelligence stays scarce and expensive. If a lab in Beijing gives away comparable intelligence, the pricing power evaporates, and with it the justification for hundreds of billions of dollars of data-center spending. The fear even hit Moonshot’s own compatriots, with Hong Kong-listed Chinese AI developers Z.ai and Minimax plunging as much as 28 percent on the announcement.

But here is an inconvenient detail for the tidy version of that story. On the morning of the global rout, under the headline “‘Bloodbath’: Analysts react to Asian shares sinking on tech selloff,” Reuters collected reactions from portfolio managers across Asia and Europe as benchmarks in Japan and Taiwan fell as much as 6 percent, with the Nikkei confirming a correction, down more than 10 percent from its June 25 all-time high. The striking thing is what the analysts blamed: hawkish signals from the Federal Reserve, doubts about whether hyperscalers “can make returns that justify their massive investments,” a poorly received SpaceX debut, and profit-taking. Pinpoint Asset Management’s Zhiwei Zhang called the move “largely technical rather than fundamental,” noting “there doesn’t seem to be any major change in the tech capex expectations. It is more of an adjustment of crowded positions that led to a certain state of stampede.”

The crowding was extreme by any historical measure. Chip stocks had staged what a Reuters column describing the Bank for International Settlements’ warnings called “a record 75% rally in the second quarter of 2026,” even as the rest of big tech deflated, a divergence covered in The Mag 7 Lost $2.3 Trillion. Chip Stocks Doubled. By July 20, UBS’s prime brokerage desk reported that hedge funds had unwound long positions in momentum and semiconductor stocks by about 5 percent of gross market value, “one of the largest reductions on record,” leaving net positioning back at April levels. A crowded theater does not need a big fire to produce a stampede. Kimi K3 was the shout, not the flames.

The Tape the Panic Ignored

If the market genuinely believed a free Chinese model destroys the economics of AI compute, the selling should have centered on the company that sells the compute. It did not. Here is the arithmetic of the drawdown, from each stock’s late-June peak close to the July 17 close.

Read that table bottom to top. Nvidia, the purest proxy for “the world needs more AI compute,” barely participated in its own bear market. The violence concentrated in memory, the hottest corner of the crowded momentum trade UBS described, and in the parallel crash in Seoul, covered in Two Chips Are Half of Korea’s Market. Then It Crashed. The selloff sorted stocks by how crowded they were, not by how exposed they are to free Chinese intelligence. That is the signature of a positioning unwind wearing a narrative as a costume.

By Tuesday, July 21, the SOX had bounced more than 5 percent intraday off Monday’s close, and UBS was publicly advising clients that the unwind was nearing its end. None of that guarantees the bottom is in. It does tell you the market has not actually voted that AI demand is dead.

Is AI Actually Profitable Yet?

Here is the number the crash talked over. In late June, Bloomberg reported analysis from the research firm Exponential View finding that global AI sales for hyperscalers and neoclouds, excluding China, reached $25 billion in the first quarter of 2026, exceeding the estimated $21 billion in depreciation costs tied to their data centers and chips, and doing so “for the second consecutive quarter.” Depreciation is the accounting cost of hardware wearing out and aging into obsolescence, and it is the number that decides whether the buildout ever pays for itself. Analyst Azeem Azhar’s own summary was appropriately modest: “It just about clears the depreciation hurdle, and roughly speaking, it’s improving over time.”

That crossing matters because it is precisely the test this site set in January, when The $600B Gamble argued that 2026 would be AI’s valley of death, the year exploding depreciation charges collided with revenue that had not yet arrived. Half of that prediction is aging well: depreciation is exploding on schedule. The other half just failed in the bulls’ favor, because revenue showed up faster than the melting ice cube could melt. Generative AI revenue excluding China reached $110 billion over the trailing twelve months and is scaling roughly three times faster than any previous technology wave Bloomberg’s sources measured, including the internet and the cloud.

Before anyone declares victory, the sober rejoinder comes from the Bank for International Settlements (BIS), the central bank of central banks. In a late-June report it warned that competitive pressure is driving capital spending so high that “the sector’s overall payoff after investment costs could shrink or even turn negative in adverse scenarios,” and that disappointment could “turn the capex boom into a protracted investment bust.” Both statements are true at once. The hurdle has been cleared, and the hurdle is about to get much taller.

The Hurdle Is About to Double

The bears’ best argument is not that the current math fails. It is that the current math is measuring yesterday’s spending. Depreciation charges lag capital expenditure by design: a server bought in 2024 spreads its cost over five or six years of accounting, so the $21 billion quarterly hurdle the industry just cleared mostly reflects hardware purchased before the current spending wave peaked.

That wave is enormous and still building. On April 29, all four hyperscalers reported the same quarter, and the numbers were uniform in direction. Microsoft spent $31.9 billion on capital expenditures and finance leases in three months and guided to $190 billion for 2026, up 61 percent from 2025. Alphabet spent $35.7 billion in the quarter and raised its 2026 guidance to $180 billion to $190 billion. Amazon spent $44.2 billion, an increase that squeezed its trailing free cash flow down to $1.2 billion, and had earlier projected roughly $200 billion for the year. Meta spent $19.8 billion and raised its full-year range to $125 billion to $145 billion. Add the midpoints and the big four alone plan roughly $700 billion of 2026 capital spending, a sum CNBC noted “could approach $700 billion in 2026.” The BIS’s annual report frames the same buildout on a two-year scale: the five largest hyperscalers are “set to spend over a trillion US dollars on AI-related capital expenditure from 2025 through 2026.”

Now run that number through the depreciation machine. AI servers are depreciated over five to six years at these companies; Amazon shortened a subset of its fleet to five years in 2025, citing “the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.” Straight-line the 2026 cohort over five and a half years:

$\frac{\$700\text{B (2026 capex)}}{5.5\text{ years}} \approx \$127\text{B per year} \approx \$32\text{B per quarter}$

That is the eventual quarterly depreciation from the 2026 purchases alone, before counting the 2024 and 2025 cohorts still working through the books. Against a hurdle headed from $21 billion toward something in the $30s and beyond, the $25 billion in quarterly AI revenue that just cleared the bar would need to roughly double within about two years merely to stay above water. The strain is already visible in the accounts: Microsoft’s quarterly depreciation expense jumped from $5.8 billion to $9.0 billion in a year, and its gross margin narrowed to 67.6 percent, the thinnest since 2022, “as depreciation costs mounted in connection with the company’s data center infrastructure build-out.” The five biggest data-center spenders have added roughly $350 billion of debt over five years to fund all this, doubling their load.

One more detail from the filings deserves daylight, because it shows how soft this hurdle really is. Meta moved in the opposite direction from Amazon, extending the useful life of most servers to 5.5 years in January 2025. That single accounting change cut its 2025 depreciation expense by $2.92 billion and added $2.59 billion to net income. The depreciation hurdle is not a law of physics. It is an estimate, and the companies being measured hold the pencil.

The Sellout Is the Tell

Now return to the event that closes the loop. On July 19, Moonshot froze new paid signups “immediately” and said it would “allocate available computing power to current paid users.” A healthy skepticism is warranted here, because scarcity is excellent marketing and Moonshot is preparing to sell shares. The company is unwinding its offshore structure ahead of a Hong Kong Initial Public Offering (IPO), has engaged Goldman Sachs and China International Capital Corp as advisers, and reached a $30 billion valuation in June after raising more than $5.5 billion in its lifetime. A viral capacity crunch the week before an IPO process is convenient.

But notice what the skeptical reading and the credulous reading have in common: in both, the binding constraint is compute. Either demand for Kimi K3 genuinely outran Moonshot’s clusters, which means a supposedly deflationary model is generating inflationary demand for hardware, or Moonshot is manufacturing the appearance of a compute shortage because it knows that is what investors treat as proof of value. Nothing about a model being open or cheap reduces the number of graphics processing units (GPUs) required to answer the world’s prompts. If intelligence gets cheaper and demand explodes, the value migrates to whoever owns power, land, and depreciated hardware, the dynamic already visible in Bitcoin Miners Are Quietly Becoming AI’s Landlords.

The people with the strongest incentive to reject that conclusion are the model labs themselves. Per Crunchbase, OpenAI and Anthropic alone absorbed $217 billion in the first half of 2026, some 43 percent of all startup funding on the planet, and the exit door is already in use after SpaceX’s IPO took the most valuable private company off the board. A funding pipeline that concentrated needs investors to believe model intelligence is the scarce, defensible asset. A Chinese lab handing out frontier weights, even a week late, is a direct attack on that story, which is exactly why Beijing subsidizes the giveaway. The margin does not disappear; it moves down the stack, away from model IP and toward the physical layer.

The DeepSeek Rerun, Eighteen Months Later

The market has run this exact experiment before, and recently enough that the results are still legible. On January 27, 2025, DeepSeek’s cheap R1 model convinced traders that Chinese efficiency had broken the AI capex story. Nvidia “lost close to $600 billion in market cap on Monday, the biggest drop for any company on a single day in U.S. history,” with the stock down 17 percent to $118.58, its worst session since the Covid crash of March 2020. The thesis was identical to July 2026’s: free Chinese intelligence means nobody needs expensive American compute.

What followed was the opposite. Cheap models multiplied usage, usage multiplied compute demand, and the hyperscalers raised their capex guidance instead of cutting it. Nvidia closed at a fresh record of $154.31 on June 25, 2025, fully reclaiming the DeepSeek crater in roughly five months. That history does not guarantee a rerun; the depreciation hurdle is taller now, and the BIS’s warning that “the demand bottleneck becomes the binding constraint” is the scenario in which this time genuinely differs. But a market that watched cheaper intelligence produce more hardware demand eighteen months ago just sold the hardware complex on the theory that it works the other way.

The evidence for which way it works is sitting in plain sight, and it is not subtle. The company whose model started the panic spent the weekend rationing access because its clusters cannot handle the demand, charges premium API prices while its weights remain a July 27 promise under a license nobody has seen, and has bankers preparing an IPO at a $30 billion valuation built entirely on the proposition that this business is valuable. As of Tuesday morning the chip index had clawed back 5 percent, and somewhere in Beijing a subscription page was still telling new customers, politely, that the revolution is sold out.