l0grisk intelligence · english

// analysis

The chip relapse: a bear market on record profits

The semiconductor index has lost more than 20% since its late-June record and entered a bear market, while TSMC posts profit up 77% and analysts raise their Nvidia estimates. The correction is not hitting profits, it is hitting the multiple. Anatomy of a derating: Meta selling its surplus compute, a delayed Gemini, a hawkish Fed again, and a rotation into memory that reshuffles the AI bet instead of closing it.

12 min read#ai#markets#semiconductors#macro#valuations
dated revision: July 17, 2026French originalprimary sourcesno tracker

In Friday trading, the Philadelphia semiconductor index shed as much as 5.7%, taking its fall past 20% from its late-June record, the technical threshold of a bear market, after a 105% surge between its March low and its peak. The day before, TSMC, the AI world’s factory, had posted record quarterly revenue of $40.2 billion, up 33.7% year on year, and earnings per share up 77.4%, its eighth consecutive quarter above expectations. The stock fell anyway. That is this correction’s singularity, and the reason to dwell on it: profits are not faltering, estimates are rising, and the market is nonetheless paying less and less for each dollar of earnings. What is deflating is not the E in the P/E, it is the P: a derating, in the strict sense.

Records, sold

The Nvidia case sums up the moment. The stock has lost about $1 trillion of market value in under two months, down 16% from its 14 May record, reached days after becoming the first company in history to cross $5.5 trillion. Over the same stretch, according to data compiled by Bloomberg, analysts raised their profit estimates by 13% in three months, and the consensus projection for fiscal 2027 stands at $228 billion of profit on $393 billion of sales. The mechanical result: Nvidia now trades at 18 times expected earnings, its lowest multiple since early 2019, that is, since before the AI boom. More striking still: that is less than the S&P 500, above 20 times, and the Nasdaq 100, at nearly 23 times. The emblematic stock of the AI revolution trades at a discount to the index it carried.

At TSMC, the same grammar. The 16 July results beat the top of the company’s own guidance, and the sanction came anyway: the stock fell, and the selling spread to Asia. Strategist Andrew Jackson of Ortus Advisors summed up the market’s verdict in the morning press: results judged not strong enough to justify another leg higher, and growing concerns over the sector’s excessive spending. When beating expectations no longer lifts a stock, the problem is not in the accounts: it is in the price.

July's paradox: profits rise, the price falls Left, the operating reality; right, the market's sanction. Data as of the dates cited. The accounts TSMC, Q2 revenue: +33.7% year on year TSMC, Q2 EPS: +77.4% 8th consecutive quarter above expectations Nvidia: profit estimates +13% in 3 months FY2027 consensus: $228bn of profit The market SOX: more than -20% since late June (bear market, after +105% March-June) Nvidia: -16%, ~$1trn erased Nvidia multiple: 18x, lowest since 2019 Marvell, ARM, Intel: -30% and more The correction bears on the valuation multiple, not on published earnings or estimates. Sources: TSMC (6-K of 16 July 2026), Bloomberg (8 and 17 July 2026).
A bear market triggered without an earnings warning: the accounts beat expectations while the sector loses a fifth of its value. Sources: TSMC, Bloomberg.

The catalysts of the derating

The correction has not one trigger but four, spread over three weeks, and their nature reveals this market’s particular fragility.

The first carries the most meaning. On 1 July, Reuters and then Bloomberg revealed that Meta is preparing “Meta Compute”, a cloud business designed to sell its surplus AI computing capacity to outside customers, raw or packaged with its Llama models. The decisive word is surplus: Meta is spending $125 to $145 billion on AI infrastructure in 2026, against about $72 billion last year, and has just expanded its CoreWeave contract by roughly $21 billion, while building enough to compete with that same CoreWeave and the big clouds. A hyperscaler looking to resell its surplus validates, from the inside, the hypothesis the sector feared most: overcapacity.

The second catalyst touches demand. On 16 July, Alphabet lost more than 4% after Bloomberg reported the delay of Gemini 3.5 Pro, its most advanced model. If model roadmaps slip, the chip orders that serve them slip too. The third is positioning: the rally’s most speculative names, Marvell, ARM and Intel, have lost more than 30% from the peak, the classic mechanics of a saturated trade unwinding.

The fourth is macro-financial, and we dissected it as it happened: Kevin Warsh’s first FOMC moved the median rate projection from 3.4% to 3.8% for end-2026, with nine members now projecting at least one hike and the 2026 PCE inflation forecast raised from 2.7% to 3.6% under the oil shock. Dallas Fed president Lorie Logan argued on 16 July for further hikes. Tech valuations are long-duration assets: when the discount rate climbs, the most stretched multiples compress first. The link between the Hormuz blockade, imported inflation and Nvidia’s P/E is indirect, but it is real.

Meta Compute, the detail that changes the thesis

Among these catalysts, Meta’s surplus sale deserves a freeze-frame, because it threatens three floors of the AI financial scaffolding we have been documenting for months. The first floor is the scarcity narrative: the sector’s entire valuation premium rests on the idea that compute is scarce and will remain so. A surplus put up for sale by one of the world’s largest buyers says otherwise, at least at the margin, and markets are set at the margin. The second floor is the price of compute: if Meta discounts its excess capacity, neocloud pricing and the revenue assumptions backing their debt compress with it. The third floor is the most fragile: the residual value of GPUs, on which a growing share of AI infrastructure credit rests. A secondary market for compute fed by hyperscaler surpluses is exactly the scenario that tests those guarantees.

Add a structural irony that readers of our work on AI’s circular financing will recognise: Meta expands its CoreWeave contract by $21 billion while preparing to become its direct competitor. The same capex dollar feeds the supplier’s order book and the future supply that will weigh on its prices. This is not fraud, it is a loop; but loops amplify in both directions.

Rotation is not exit

Reading this correction as the end of the AI bet would nonetheless be a misreading, and this is where the picture gets interesting. At the very moment logic chips deflate, memory is on fire. Micron has gained 229% in 2026 after 239% in 2025; SK Hynix, up about 248% this year, has joined Micron in the $1 trillion market cap club, with Samsung gaining some 165%. The engine is physical: high-bandwidth memory (HBM), the bottleneck of inference, is sold in advance, with Micron’s capacity booked through 2027, and Bank of America projects the HBM market growing from $34.6 billion to $54.6 billion in 2026. Capital is not leaving AI: it is migrating from compute to memory, from Nvidia to Micron and SK Hynix, from the segment that got expensive to the segment still rationed.

That migration has its own fragility, and it is historical: memory is the most cyclical corner of the entire industry, the land of repeated booms and busts, as fund managers were warning back in May. Investment plans announced at the top of the cycle, hundreds of billions of dollars of new fabs on the Korean and American side, replay the pattern that has always turned memory shortage into memory glut. The crowd is not leaving the theatre: it is moving to the part of the room where the floor has given way most often.

The AI bet's internal rotation 2026 performance (Bloomberg data as of 8 July, before the correction accelerated). Nvidia (compute) +5.6% SOX index +74% Samsung (memory) ~+165% Micron (memory) +229% (after +239% in 2025) SK Hynix: ~+248% in 2026, joining Micron in the $1trn market cap club. Sources: Bloomberg (via Yahoo Finance, 8 July 2026), Yahoo Finance, IG.
Capital is not exiting the AI theme, it is migrating from compute to memory, historically the industry's most cyclical segment. Sources: Bloomberg, Yahoo Finance, IG.

The benign reading

There is a benign interpretation of everything above, and it is defensible. A derating that happens while earnings rise is the least painful way to deflate a valuation excess: time and growth do the work a crash would otherwise do. Nvidia at 18 times earnings projected to grow 90%, cheaper than the S&P 500, is a statistical anomaly; of 82 analysts, 78 remain at buy with an average target 50% above the price. If the projected profits materialise, this correction will enter the textbooks as an entry point.

The counter-argument, which we developed in the bubble within the bubble, fits in one question: what is the E worth on which this reasonable multiple is computed? Part of the sector’s revenue is fed by circular flows between players financing one another, and by hyperscaler capex that Meta has just shown exceeds its own needs. A P/E of 18 on peak-cycle, partly self-fed earnings is not necessarily cheap; it is in fact the classic configuration of valuation traps in cyclical industries. The market is not choosing between these two readings, it is pricing both: hence a sector worth a fifth less than in June on record accounts. The gap between the two readings will close on numbers, not on narratives.

Markers for what comes next

Five appointments will separate a healthy correction from a turn. Hyperscaler capex in the late-July earnings round: a downward revision, even cosmetic, would change the episode’s nature, turning Meta Compute from an isolated signal into the start of a series. Nvidia’s results in late August, the first test of the $228 billion FY2027 profit consensus. TSMC’s monthly revenue, the fastest thermometer of real demand. Memory prices and 2027 HBM contracts, which will say whether the rotation rests on durable scarcity or on the umpteenth top of a cycle. And the spreads on the debt financing data centres, because that is where contagion would show, more than in equities: stocks absorb multiple compression, credit only absorbs defaults. As long as the latter do not follow the former, the chip relapse remains a story about price. The day the E joins the P, the story changes genre.

Sources

This article is journalistic analysis and does not constitute investment advice. The 17 July price moves are intraday data, subject to change by the close; the cited earnings projections are consensus estimates, not facts. Data as of the dates of the cited sources.

This analysis is not investment advice.

// cite this analysis

l0g, “The chip relapse: a bear market on record profits”, l0g.fr, published July 17, 2026, updated July 17, 2026, https://l0g.fr/en/analysis/chip-relapse-bear-market-record-profits/

$ cd ../analysis