Buying the July semis crash — memory, optics, power

Distilled from @ParadisLabs' August watchlist — all of them his own holdings, re-rated by the July crash. Six of his nine names are listed here. I checked his numbers against primary sources and corrected three of them in place.

Why now

July 2026 was the SOX's worst month since October 2008: -21% on the month, -28.6% peak-to-trough (14,634.72 on 6/22 to 10,447.49 on 7/29), ~$2.2T of market cap gone, every index member red, more than half down 25%+. The proximate causes were narrative, not order books — Meta reselling surplus compute, the first real wave of Amazon/OpenAI custom silicon, TSMC guiding 2026 capex to $60-64B, SK Hynix deferring HBM4 capacity. The demand disclosures since have not deteriorated. This basket buys the disclosures, not the tape.

The book

NVDA — the anchor, and the one place I depart from the source. This is not a drawdown buy: NVDA rose through July and sits ~5% off its 52-week high ($225.16 vs $236.54, 8/14 close). The case is purely relative. Forward P/E ~23-25x against AMD at ~46-77x depending on vendor and EPS base — vendors disagree wildly, but the 2-3x gap holds under every methodology, on $253B TTM revenue growing 71% and TTM EPS +111%. Earnings 8/26.

SKHX + KIOXIA — memory priced for a cycle the contracts are slowly removing. SK Hynix Q2'26 (7/29): revenue ₩79.3T +257% YoY, operating profit ₩60.5T at a 76% margin, ₩69.4T net cash, and LTAs now finalised with ~10 customers, typically ~5 years, explicitly to damp price swings. It still missed — consensus wanted ₩64T OP — and the stock fell 9.6% on HBM4 revenue recognition slipping right. That miss is the entry. Kioxia (7/31): revenue ¥1.77T, non-GAAP operating income ¥1.33T (75% margin), net cash after repaying ¥407.5B, guiding ¥2.39T / ¥1.90T for Jul-Sep, plus a ¥800B buyback running 8/3 to 10/30, a 3-for-1 split effective 10/1, and a 50% total-shareholder-return commitment. Correction to the source: Kioxia's LTAs are a stated target to cover ~50% of CY2028 shipment volumes, terms undisclosed — not present-tense earnings cover.

CIEN — the inter-DC bandwidth constraint, with a backlog you can date. Q2 FY26 (6/4): revenue $1.57B +40% YoY, adjusted EPS $1.64 (~4x YoY), backlog $7.7B, up $600M sequentially. Correction to the source: the ~80% one-year conversion applies to the $6.4B hardware slice only — about $5.1B inside 12 months, not 80% of the $7.7B. Still chunky against a $6.3B FY26 revenue guide. Direct cloud customer revenue +70% YoY, routing & switching +88% on the DCOM ramp, and the first multi-rail order from a hyperscaler. Q3 prints 9/3 — the single most dateable catalyst in this book.

BE — the rare AI-adjacent name management says is not capacity or order constrained. Q2'26 (7/28): first $1B+ quarter at $1,065.4M, +165% YoY; product revenue +215%; GAAP operating income $182.2M against a small loss a year ago; FY26 guide raised to $3.9-4.2B. Bloom adds US capacity in "copy exact" increments ahead of committed orders (Fremont 1 GW to 2 GW by end-2026). Two honest caveats: the not-constrained claim is management framing, and Hunterbrook (7/8) disputes the 5 GW path on scandium-oxide single-sourcing. I am deliberately not quoting the marketed $20B "backlog" — the marketed number and the audited GAAP remaining performance obligations are not the same object, and that gap is live short-seller territory.

NOK — a legacy telco multiple on a business quietly repositioning into DC optics. Two corrections. The drawdown is deeper than the source's ~36%: -38.3% from the $17.45 52-week high at 8/14's $10.76. And the +105% growth is Nokia's AI & Cloud customer cohort, not a reporting segment — Network Infrastructure itself grew 12% cc, with Optical +20% and IP +16%. What looks underpriced is the optics build-out: €2.8B of AI & Cloud order intake in Q2, roughly half converting within 12 months; the San Jose fab ramping Q4'26; Pennsylvania test and packaging capacity going 10x from Q3'26; the NXP Chandler campus acquired for indium phosphide; and NVIDIA on the register with a $1B stake struck at $6.01. Smallest weight because the operating leverage is real but slow, and reported operating margin is still negative on restructuring.

Not listed on the board: ALAB, BESI, GFS — the other three names in the source post.

What confirms it

NVDA's 8/26 print holding data-center growth without a guide-down. CIEN's 9/3 backlog converting ahead of the $6.3B FY guide. Kioxia actually pushing the ¥800B into the market through 10/30. SK Hynix's HBM4 recognition catching up in 2H26. Nokia's Q3 AI & Cloud order intake holding near Q2's €2.8B.

What kills it

Any hyperscaler trimming 2027 capex. Custom silicon taking share faster than the Amazon/OpenAI ramps imply. Memory ASPs rolling over before LTAs cover volume — Kioxia's cover is a 2028 target, not today's book. Bloom's capacity claim failing on single-source inputs. And the ugliest version for the anchor leg: the NVDA/AMD valuation gap closing by NVDA de-rating rather than AMD.

Mechanics

Unlevered on every leg, deliberately. The SKHX contract flash-crashed ~20% to ~$900 in under two minutes on 7/27-28 when one thin ₩1,272,000 print on Korea's NextTrade venue reached the oracle, liquidating ~$60M across 960 accounts. Leverage on a 24/7 perp tracking a Korean equity is a materially different risk from owning the equity.