The Chip Industry Just Had Its Best Quarter Ever. Wall Street Sold It Anyway.

In the same week that TSMC, Samsung, and ASML each reported the strongest results in their history, the Philadelphia Semiconductor Index fell into a bear market. That's not a contradiction that resolves itself with a little more context — it's the story. The companies making AI's physical infrastructure have never been more profitable, and investors have rarely been more nervous about what that profitability is actually worth.

Start with the numbers, because they're genuinely startling. TSMC posted second-quarter revenue of $40.2 billion, up 36% year over year, with net income up 77.4% — a record for the fifth consecutive quarter, driven by full utilization of its advanced-node capacity and an AI-heavy product mix that now makes up 77% of revenue. Gross margin hit 67.7%, an all-time high. CEO C.C. Wei used the earnings call to announce another $100 billion for TSMC's Arizona buildout, bringing the company's total committed U.S. investment to $265 billion. Samsung's preliminary Q2 operating profit came in at 89.4 trillion won — roughly $59 billion, about 19 times what it earned a year earlier, beating analyst estimates by more than 6% on the back of simultaneous price increases across DRAM, NAND, and HBM. And ASML, the sole supplier of the extreme ultraviolet lithography tools that make advanced chips possible, raised its full-year sales guidance for the second time this year, to €43–45 billion, and said its order book is essentially full through 2027 with meaningful demand already visible for 2028.

There was a genuine engineering milestone tucked into that same week, too. On July 15, Intel became the first company to ship high-volume logic chips patterned with ASML's High-NA (0.55 numerical aperture) EUV scanners — years ahead of TSMC, which has ruled out the technology through 2029. The chips are select layers of Intel's Panther Lake laptop processors, built on the 18A node at Intel's Hillsboro, Oregon fab, and ASML says those layers are already yielding at parity with the older NXE EUV platform. It's a real technical win for Intel's foundry ambitions — worth noting precisely because it's narrow. It applies to specific layers of chips already shipping to consumers, not a verdict on 18A's overall profitability for the external customers Intel is trying to win back; separate reporting this month suggested those customers may not see profitable yields out of 18A until late 2026 or 2027. Genuine progress and unresolved risk are sitting right next to each other in the same node.

None of that record-breaking is what moved the stock market. On Friday, July 17, the SOX benchmark shed as much as 5.7% in a single session, extending a slide that has now erased more than 20% of the index's value since its late-June record — the technical threshold for a bear market — and roughly $3.3 trillion in global chip-sector market value since June 22 alone. It was the worst weekly rout for the sector since April 2025. Marvell, Arm, and Intel are each down more than 30% from that June peak. This is the same index that had rallied 105% from its March low just a few weeks earlier — so the reversal isn't a slow deflation, it's a fast one.

The trigger was a piece of software, not a piece of hardware: Kimi K3, an open-weight model released July 16 by Moonshot AI, one of China's "Six AI Tigers." Kimi K3 has 2.8 trillion parameters — the first open-weight model to reach that scale, nearly triple its predecessor — a million-token context window, and a mixture-of-experts design that routes each request through just 16 of 896 available expert subnetworks to keep inference costs down. It trails Claude Fable 5 and GPT-5.6 on most benchmarks, but beats Claude Opus 4.8 and GPT-5.5, at a price of $3 per million input tokens and $15 per million output tokens. Full weights land publicly on July 27.

Why would a chatbot rattle a chip index that just posted its best quarter in years? Because the entire trillion-dollar hardware buildout — TSMC's $265 billion Arizona bet included — is a wager that frontier AI will keep demanding exponentially more compute, indefinitely. A capable open-weight model built more efficiently, out of China, at a fraction of the assumed cost, is the specific scenario that wager is exposed to. If good-enough intelligence gets cheap enough, the argument goes, maybe the industry doesn't need quite so much silicon after all.

It's worth being skeptical of that logic, though, and not just because it triggered a $3.3 trillion round-trip in three weeks. Cheaper, more efficient inference has historically increased total compute consumption rather than shrinking it — more people running more queries more often, a semiconductor-flavored version of the Jevons paradox that showed up when GPT costs fell in 2024 and usage exploded rather than contracted. There's already a market signal pointing that direction: General Compute, an inference-focused cloud startup, just closed a $400 million loan from Upper90 using its inference chips as loan collateral — reportedly the first deal of its kind. That's a bet that inference capacity specifically, not training capacity, is the durable, financeable asset — a more precise read on where AI compute demand is actually heading than "more of everything, forever."

The chip industry's fundamentals didn't get worse this week. What got clearer is that the market pricing those fundamentals is still trying to figure out whether it's underwriting a training arms race, an inference boom, or both — and a single well-timed model release from Beijing was enough to make it flinch. That's the part worth watching over the next earnings cycle, not the headline number on any one company's income statement.

Sources

TSMC Q2 2026 earnings release: https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2026-07/a80d7933be643644081584087731f73b22ea5a2c/2Q26%20EarningsRelease.pdf

Yahoo Finance coverage of TSMC's Q2 results and Arizona investment: https://finance.yahoo.com/technology/articles/tsmc-adds-100bn-us-arizona-081156641.html

The Next Web coverage of Samsung's preliminary Q2 2026 profit: https://thenextweb.com/news/samsung-q2-2026-operating-profit-record-ai-memory

Yahoo Finance coverage of ASML's Q2 2026 guidance raise: https://finance.yahoo.com/technology/articles/asml-raises-2026-guidance-second-114432617.html

Tom's Hardware coverage of Intel and ASML's High-NA EUV milestone: https://www.tomshardware.com/tech-industry/semiconductors/intel-becomes-the-first-company-to-ship-high-volume-logic-chips-made-with-asmls-high-na-euv-select-panther-lake-layers-on-18a-are-now-dual-qualified-for-0-55-na-scanners

MarketWatch coverage of the SOX bear-market move: https://www.marketwatch.com/story/chip-stocks-enter-bear-market-territory-a-bofa-analyst-says-not-to-panic-4d34df17

Tom's Hardware coverage of Moonshot Kimi K3: https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3

TechCrunch coverage of General Compute's $400 million inference-chip-backed loan: https://techcrunch.com/2026/07/17/why-the-first-gpu-financiers-are-turning-to-inference-chips-in-a-400-million-deal/