TMTB Asia: Overnight Wrap

Meta gatecrashes the cloud, MoMo has its worst day since ‘vax-day’, and Asia takes the memory complex to the woodshed. Fade or follow?

Sample issue — a format prototype, not a live publication. Index levels are illustrative; single-stock moves and datapoints are drawn from real 2 Jul desk flow.

Good evening from Asia. If you were long anything with a memory chip in it, we just finished the job the US started overnight. Meta saying it’ll rent out AI compute detonated the “overbuild” trade in the US session, and Asia sold the follow-through hard — Hynix −14.6%, Samsung −9%, Kioxia −13.5%. US futures are trying to steady into NFP, but SOX is still heavy. Let’s get to it.

Our call: this is a positioning purge, not the end of the memory cycle. The tape is pricing a glut; the contract-price curve (below) is pricing a shortage — and it isn’t the curve that’s wrong. We’re fading the Hynix panic, not chasing it.

|1| Markets — Oh Gosh

Full-blown momentum unwind. Goldman’s High-Beta Momentum basket — these days mostly chip and memory names — melted −9%; L/S High-Beta Momentum −10%, its worst day since ‘vax-day’ 2020. Momentum is now ~20% off the highs (historically where it’s bottomed) and ~11% on the day alone. The Bloomberg Mag7 index vs the SOX just printed its best day on record back to 2015 — a ~6-sigma +8%. Rotation into cyclicals and value; breakevens lower. And it’s July: the momentum factor has averaged −4-5% every July for five years — after a +33% three-month run into this, that seasonal air-pocket is very real.

Asia AI — today’s sector scorecard. The Meta headline questioned whether compute is really that tight, and Asia sold the answer by theme. Korea memory got nuked; Japan semicap and Kioxia followed; Taiwan was a split tape — foundry, substrate and ASIC red, but packaging, probe/test and power caught their own bids. China went green across the board on the anti-memory rotation and the domestic-substitution trade.

Sub-sector baskets map to MS thematic baskets on Bloomberg. Mock levels for the prototype — single-stock spot moves (Hynix, Kioxia, Sumco, ASE, MPI) are real 2 Jul prints; aggregates to be wired to live quotes.

Hardware pockets bucked the tape on their own news: ASE reportedly pushing a ~20% price rise; probe-card play MPI ripped on May sales +58% yoy (NT$1.9bn), Winway alongside it.

|2| Key Data — the datapoint with memory (US read-through)

The print: TrendForce sets 3Q26 DRAM contract prices +8% QoQ. NAND guided +5–10%.

The streak: fourth straight quarterly hike, cumulative +38% off the 3Q25 trough — the longest unbroken up-run since the 2017-18 super-cycle.

The read: MU / SK Hynix / Samsung are all geared to this, yet sell-side still models only +5% 3Q ASP — numbers are too low if the contract print sticks. The counter-tell: GigaDevice’s 30 Jun risk filing warned prices are “at historically high levels” and could “fall sharply.” So the debate is set — the tape is short a glut vs the curve is long a shortage. We side with the curve into the prints.

|3| Sell-Side Desk (US read-through)

KGI · GlobalWafers — re-init Outperform, NT$1,500. Calls a 12" silicon-wafer super-cycle from 2H26 (memory ramp + TSMC advanced-node starts + CoWoS interposer demand outrunning top-5 supplier capacity); GWC 12" ASP +10%/+10% QoQ into 3Q/4Q, then +39% in ’27. Read: wafers are the cleaner, less-crowded second-derivative of the memory trade — Sumco +17% today says the tape already smells it. On a MoMo-purge day, this is where you hide.

Jefferies (Thill) · META — BUY, PT $825. Pushes back hard on today’s panic: overbuild is backward — Meta is ROIC-maxing the early-AWS playbook, renting ~35% idle capacity “at a premium to what we bought it at,” with the industry sold out through ’27. Read: if that framing is right, today’s Asia memory bloodbath is a gift, not a top.

|4| A Few Quotes

What the buy-side group chats were actually chewing on today — the bear case that got louder the second Meta blinked:

“The market is conflating an AI capex bubble with the underlying industry. The two labs that matter are private and deeply unprofitable — ~$200bn raised in six months, and they don’t look like past cycle winners.”

“A $200 sub reportedly costs one lab ~$8k and the other ~$14k a year to serve. The unit economics are hidden, not solved.”

“The AWS/Uber comps aren’t fair — Amazon’s inflation-adjusted cumulative capex was ~$29bn. AI dwarfs that, and unlike cloud, model costs go up, not down.”

“Open source is the threat neither lab has a real answer to. And the hyperscalers carry ~$748bn of remaining obligations — watch for the first real capex pullback.”

Quotes are paraphrased and anonymized from buy-side chatter and reflect individual personal views only. Not investment advice, not a recommendation, not attributable to any named person or firm.

|5| Single-Stock Spotlight — SK Hynix

Down 14.6% — its worst session of the cycle — on a narrative, with not a single fundamental datapoint attached. HBM remains sold out through ’26, HBM4 is sampling, and the 3Q DRAM contract print just went up. The honest read: this was a crowded-positioning unwind wearing a fundamental-reversal mask — Hynix was the fattest long, in the fattest basket, on the fattest MoMo day in five years.

Lean: buyers of the panic into the July print, sized for a second gap lower if NFP runs hot and the factor keeps bleeding. The thesis breaks only if the contract curve rolls — and it hasn’t.

|6| Supply-Chain Watch — Power & Thermal

Today’s cut of the living map — the AI power/cooling names the English-language desk never sees:

Kaori (fuel cell) — order visibility extended from 6 months to 1 year; Zhongli capacity to triple by 2027, 6-7× by 2028. Fuel-cell + liquid-cooling now ~70% of revenue, heading to 80%.

Vinatech (supercap) — disclosed a supercapacitor contract with Bloom Energy worth ~50% of FY25 revenue.

GlobalWafers / Sumco (wafers) — Sumco +17% today; the KGI super-cycle note is the driver.

MPI / Winway (probe cards) — MPI May sales +58% yoy; probe-card demand tracking AI test intensity.

|7| Policy Watch (US read-through)

New York’s Responsible Data Center Development Act would place a one-year pause on permits for new large data centers pending a community-impact review — now awaiting the governor’s call. Read-through: a marginal negative for DC-buildout and power names if it becomes a template other states copy — but it also tightens the very capacity today’s “overbuild” sellers claim is loosening. Another brick in the shortage wall, not the glut.

|8| On the Radar

• Tonight — US June NFP; the whole tape is leaning on it.
• ~Jul 10 — Taiwan June monthly sales: TSMC, Foxconn, Quanta, PCB/ABF substrate.
• Jul 17 — TSMC Q2 earnings; the Asia tentpole.
• This week — Kimi K3 rumored; watch China open-source model cadence.

|9| The Deeper Cut — the ABF-substrate bottleneck

Nobody’s pricing the substrate constraint into the memory panic. ABF layers per GB-class package are rising ~3× generation-over-generation, and top-5 supplier capacity is effectively locked through 2027 — as HBM supply gets solved, the bottleneck migrates from the die to the package itself.

The trade: Kinsus and Nan Ya PCB screen cheapest on 2027 substrate content, with Unimicron the quality name at a premium. On a day the tape is dumping everything memory-adjacent, the substrate names are the wrong things to sell — they’re geared to unit complexity, not DRAM ASP. The pair we like: long ABF substrate vs short the crowded Hynix long — same theme, cleaner positioning.