Question: The user's standing unease, raised twice (2026-07-05): SPY has run four years straight, hard shorts keep getting burned, and his own instinct is "we're burning a ton of DRAM doing what we're doing now." Is the AI melt-up real end-demand or a bubble top — turned honestly on OUR OWN book, not as perma-bull confirmation and not as a crash call.

Verdict: BOTH/AND — real, audited physical undersupply AND bubble-grade pricing coexist, and they are breaking separately. The physical-shortage evidence is the strongest-sourced thing on the desk; the multiple risk is real but concentrated, and the melt-up regime's own tailwind signals cracked in June — weeks before the July semi derate. The two failure modes have different tapes, different falsifiers, and different casualties. Posture: hold, with the multiple-risk carriers named per thesis below.

The demand case (what's audited)

Integrity correction, logged here deliberately: the task row that commissioned this read cites "MU rev +196%" as the validated anchor. It is NOT validated — the reconciliation ledger entry revenue-accel:MU:2026-05-28 carries agreement: disputed / QUARANTINE (EDGAR 48.9% YoY vs Massive 167.0% YoY, and the EDGAR asOf is stamped to FY2025 year-end, not FQ3). The demand case above rests on the EDGAR quarterly ramp, which reconciled; the +196% figure is unusable until re-reconciled. Source: research/api-data/fundamentals-wire.json.

The multiple case (what's priced)

• The derate is sectoral and deep, the market is not: SMH −15.7%/30d (RSI 41), MTUM −10.8%/30d, memory cohort −25% to −38%/30d and −31% to −42% off highs — against SPY −0.5%/30d and RSP (equal-weight) +2.0% with rising breadth. research/issues/2026-07-18-ai-semi-derate-turn-signature-armed.md, data/summaries/market-pulse.json.
• The melt-up's own regime signals broke first: BUFR cap-saturation (the low-vol melt-up tell) flipped breaking-down 2026-06-11; the growth-leading/value-lagging signal broke the same day; single-name dispersion has been elevated since 06-12. The multiple-expansion tailwind was already fraying BEFORE this leg — July is a sector re-rate on top of a fraying melt-up, not the melt-up's first crack. research/signals/2026-05-11-bufr-cap-saturation.md, 2026-05-11-value-vs-growth-inflection.md, 2026-06-12-dispersion-regime-elevated.md.
• A new failure channel opened this week — credit, not sentiment: Oracle's rating downgrade explicitly tied to AI-funding/leverage. A financing shock can cut capex even where end-demand is real; this is the channel that would transmit a multiple problem into a demand problem. research/perspectives/2026-06-26-ai-capex-digestion/log.md.
• Earnings-quality asterisks on the demand champion: MU's ramp is contract-pricing-driven with shipments flat/down in some BUs (price rolls unwind faster than volume rolls); QoQ growth decelerated at the margin (+73.75% vs +74.89%); insiders 1 buy / 54 sells; the memory cohort traded 120–152% above lifetime VWAP at the 06-26 froth scan. research/perspectives/2026-05-19-memory-supercycle/README.md, the 06-29 validation investigation.

Per-thesis map: demand-anchored vs multiple-anchored

The mechanism test applied throughout: if AI multiples compress 30% but physical demand holds, does the thesis still work? Memory passes (its falsifiers are wafer/TB-denominated: the capex-glut watch fires on contract-price deceleration + fab volume landing + cohort break + capex cut, ≥2 co-confirming — none is "NVDA's P/E is lower"). Electricity passes at the generation/equipment/fuel/regulated layers and fails at the speculative tier. Physical-AI passes only at the arms-dealer core.

Does the posture change?

No — hold, but know which legs carry which risk. The book is already demand-anchored by construction: every PRIMARY's investable expression deliberately excludes its own speculative tier. What this read changes is vocabulary and watch-order:

1. The honest answer to "bubble top?" is: the multiple leg is ALREADY compressing — that's what July is — and breadth says it is so far a rotation, not a reckoning. The resolution is calendar-dated, not open-ended: GOOGL capex language Jul 22, K3 open-weights Jul 27, storage prints Jul 28–29, SNDK Aug 5. The turn-signature note carries the armed tripwires.
2. Watch the credit channel, not just the tape: the ORCL downgrade is the first financing-side tell. A second AI-capex-linked downgrade or a funding-cost story at a hyperscaler outranks any RSI reading on this question.
3. Follow-up filed: re-reconcile MU revenue YoY (revenue-accel:MU:2026-05-28 quarantine) so the desk stops citing a disputed number.

Sources

Lanes A/B/C of this session read: the 06-29 memory validation investigation, memory-supercycle + ai-power-bottleneck + physical-ai-robotics + ai-capex-digestion perspective READMEs/logs, .brief/operating-book.md, the 07-17 EOW scan set (market-pulse, monster, ai-scan), the 07-18 turn-signature note, regime/BUFR/dispersion/capex-glut signal files, and data/summaries/*.json at the 2026-07-17 close. All numbers verbatim from those artifacts; none computed in prose.