Question: Following the money up Dell's supply chain — where did the money already go, is there alpha left (names that haven't run), and which layers are durable franchises vs cyclical/bubble? Verdict: The names Dell pointed at are all extended and consensus. MU / STX / WDC / AMD / SNDK are RSI 74–81, +100–184% 3m, all within ~3% of 52-week highs — the obvious trade already tripled. The alpha that's left is one layer upstream: the equipment / test / metrology that builds the "locked-through-2027" capacity — FORM, CAMT, ONTO, TER, ENTG sit RSI 39–53 and 12–28% off their highs while the chips they serve are at all-time highs. The durable end-state is the WFE + process-control oligopoly (ASML, AMAT, LRCX, KLAC, TER) plus consumable-materials (ENTG) — they win regardless of which memory/HDD/CPU brand wins, and they've run the least. The bubble risk concentrates in commodity memory at peak-cycle (MU/SNDK — the same locked-2027 supply that's bullish now is what historically breaks the cycle) and thin-margin assemblers getting AI multiple re-rates (DELL RSI 91 blowoff, SMCI).

Data discipline: every number below is from the scan summaries as-of 2026-06-01 (2 trading days stale at write time; structural read, not an entry trigger). Code computed the RSI/returns, not prose. Valuation/moat claims are qualitative + lean on the memory-supercycle perspective's contract-price data; a hard multiple pull is the named next step to fully harden the durable-vs-bubble call.

What we're asking

demian_ai's Dell read (inputs/2026-05-29-demian_ai-dell-upstream-suppliers-bottleneck-map.md) said Dell is "physically capped by components, supply-constrained not demand," and mapped the margin pools upstream: Memory → Microprocessors → Hard drives → cooling/power spillover. Three questions:

1. Follow the money — trace each Dell component to the actual chokepoint suppliers, and one layer past them (the picks-and-shovels behind the picks-and-shovels).
2. Is there alpha left — which names haven't run yet?
3. Durable vs bubble — where does this all end up; what's a franchise vs a cycle-top.

The chain, mapped (with named players per layer)

Dell sits at the box layer (assembler, thin margin, getting an AI multiple re-rate). Following the money up:

The alpha screen — where the money already went vs where it hasn't

Sorted by extension. The Dell-named bottlenecks are uniformly blown off; the equipment/test/metrology that serves them lags badly.

Already run — extended, high RSI, at/near 52w highs (the crowded trade):

Hasn't run / in pullback — the upstream laggards (where alpha may be left):

The pattern is unambiguous: +100–204% 3m at the demand-end vs +0–47% 3m (several negative-from-high) one layer up. If Dell + the memory makers are genuinely "capacity-locked through 2027," the spend to build that locked capacity flows to WFE and the qualification spend flows to test/metrology — and that layer hasn't been bid the way the chips have.

Durable vs bubble — the 2×2

Cross durability (oligopoly moat, recurring/consumable revenue, wins-regardless-of-brand) against extension (how much is already priced):

① Durable franchise + NOT yet extended → the best durability-adjusted alpha

• ASML (EUV monopoly), AMAT / LRCX / KLAC (deposition/etch/process-control oligopoly), TER (ATE duopoly), ENTG (consumable materials), and the inspection/metrology/probe cluster CAMT / ONTO / FORM. These have the deepest moats, recurring service/consumable revenue that smooths the cycle, win regardless of which memory/HDD/CPU brand wins the demand layer — and they've run the least (+0–38% 3m, many 12–28% off highs). This is the "picks-and-shovels behind the picks-and-shovels."

② Durable franchise + already extended → great business, rich/late stock

• STX / WDC (HDD is arguably a cleaner oligopoly than memory: 3 players, no new entrants, no greenfield capacity, HAMR transition extends the runway) — but the stock already ran +102–143% 3m. AMD (real share-gain story, +157% 3m). The "great business ≠ great stock from the peak" zone — you're paying up for a confirmed winner.

③ Cyclical/commodity + extended → highest bubble risk

• MU / SNDK — memory is the classic boom-bust. Our memory-supercycle perspective already flags it "late-cycle, consensus, extended" (contract prices +58–75% QoQ = mid-up-cycle). The bull case ("HBM consolidation + disciplined DRAM oligopoly makes this structurally tighter") is real, but the same "locked 2027 capacity" that's bullish today is exactly what floods supply and breaks the memory cycle in 2027–28. Memory carries the seed of its own top.
• DELL (RSI 91, +204% — blowoff) and SMCI — thin-margin assemblers getting AI multiple re-rates. Most bubble-shaped on the tape; their margins don't structurally expand the way a chokepoint's do.

④ Cyclical + not extended → value/turnaround (different thesis)

• INTC (−17.6% from high, RSI 53) — a turnaround/foundry-policy bet, not a buildout-beneficiary bet. Different risk model; don't conflate with the chokepoint trade.

Verdict + reasoning

Is there alpha left? Yes, but not where demian/Dell pointed. The demand-end chokepoints (memory, HDD, server CPU) are consensus and extended — buying MU/STX/WDC/AMD here is paying all-time-high prices for a thesis the tape has fully discovered. The un-bid layer is the equipment/test/metrology that builds and qualifies the locked capacity (FORM, CAMT, ONTO, TER, ENTG, and the WFE oligopoly AMAT/LRCX/KLAC/ASML at moderate extension). That layer is a derivative of the same "capacity-locked-through-2027" fact, and it hasn't moved.

Where does it all go / durable vs bubble:

• Most durable end-state = the equipment + process-control oligopolies (ASML/AMAT/LRCX/KLAC/TER) + consumable materials (ENTG). Deepest moats, cycle-smoothing recurring revenue, brand-agnostic. They survive the cycle the memory names won't.
• Real but cyclical = memory (MU/SNDK). Durable up-cycle, but it is a cycle; the 2027–28 supply is the dated bear catalyst. HDD (STX/WDC) is a cleaner oligopoly but the stock already ran.
• Bubble-shaped = the assemblers (DELL/SMCI) getting AI re-rates without structural margin expansion, and commodity memory priced at peak-cycle multiples on peak-cycle earnings.

The honest caveat: "least extended" ≠ "cheap." This screen is RSI/return-based (what's run); it is not a valuation screen. ASML/AMAT can be un-bid on momentum yet still richly valued on earnings. The durable-vs-bubble call is qualitative + leans on the perspective's contract-price data; the next step to harden it is a multiple pull (EV/sales, fwd P/E). Resolved — see "Valuation overlay" below.

Valuation overlay — finishing the durable-vs-bubble call (2026-06-03)

The follow-up multiple pull, now run (forward P/E, trailing P/E, price/sales, EV/revenue, gross + net margin, code-pulled). Read the structure, not the absolute levels: absolute multiples run hot in this dataset (same price-inflation caveat as the memory-supercycle work). The two trustworthy real-world-shaped signals are (a) the absolute forward P/E — the multiple the market will actually pay for next year's earnings (a single-digit forward multiple on a stock sitting at all-time highs = the market refusing to capitalize peak-cycle earnings) — and (b) net-margin position — where in the value chain the structural profit sits.

The upstream basket — wfe-test-metrology watchlist (built this session):

The demand-end chokepoints (the consensus trade) for contrast:

What the overlay changes:

1. The memory "trap" is now quantified, not asserted. MU/SNDK carry a forward P/E of just ~10 — at all-time highs, on 34–41% net margins. A single-digit forward multiple on a stock at its high is the market refusing to capitalize peak-cycle earnings: it expects a big EPS print next year but will only pay ~10× for it because those margins mean-revert. Everything else in both tables — HDD and the WFE/test complex — sits at forward P/E roughly 25–46 (most clustered 30–43; KLIC 25 and ONTO 30 at the low end, FORM 46 at the high, BESIY has no positive forward number), i.e. the market does capitalize forward earnings at a durable-business multiple — an order of magnitude above the ~10 it will pay for memory. HDD specifically (STX fwd 35, WDC fwd 33) is not priced for an earnings collapse, corroborating the separate nearline-storage finding that HDD is density-led, not cycle-led. The bubble risk is specifically DRAM/NAND, not "storage" broadly. (Note the fwd-vs-trailing gap alone doesn't separate them — STX compresses 89→35 just like MU 50→10; what differs is the absolute forward level the market will pay.)
2. The WFE majors are the durable-value, not the small-cap metrology names. AMAT (fwd 30.9, 29% net, $397B liquid) and ASML (fwd 36, 30% net, monopoly) carry moderate forward multiples on structural margins. ENTG is the consumable-materials value (P/S 6.6, the cheapest quality name).
3. Refinement to bucket ① above: FORM/CAMT/ONTO lagged on price (low RSI) but the overlay shows they are low-margin (8–10% net) and not cheap (P/S 11–17, fwd 30–46) — momentum laggards, not quality-value. Don't conflate "hasn't run" with "durable bargain." The genuine durability-adjusted value is AMAT / ASML / ENTG, with KLAC/LRCX the rich-but-best-margin core.
4. BESIY (P/S 42) is the most extended-on-sales name in the entire complex — fully pricing the HBM4 hybrid-bonding inflection. High-conviction story, zero valuation cushion.

Tape-freshness correction: the RSI screen above (lines ~96–114) drew on the scattered source watchlists (monster-discoveries / optical-supply-chain / supply-chain-traces / ai-power), whose OHLC was stale. The focused wfe-test-metrology fetch (fresh 2026-06-03 intraday) shows the laggards already firming — FORM 39→48, ONTO 44→56, CAMT 44→55, TER 53→62 RSI. The "upstream lags the end-product" gap the investigation flagged is starting to close; the rotation is no longer purely prospective. (This is itself the argument for the focused watchlist: it surfaces the move while the scattered caches lag.)

Ranked read — where the durability-adjusted risk/reward sits now (2026-06-03 intraday)

Putting the two halves together — moat tier (from the chain map) × valuation (overlay above) × the fresh code-computed momentum below. Read against the focused fresh fetch, most of the cohort no longer screens as a laggard — the mid-caps and majors sit RSI 59–74 near highs — so the question shifts from "is there a laggard" to "which name is durable and reasonably valued and still un-extended." (How much of that shift is real two-session price movement vs. the earlier screen having run on stale scattered-cache OHLC is unpicked in the meta-lesson below; treat the standing of each name as the read, not the speed of the move.)

Fresh momentum — focused wfe-test-metrology fetch, scan summaries as-of 2026-06-03 intraday (code-computed RSI/returns, structural read not an entry trigger):

The ranked call (durability × valuation × momentum-room):

1. ENTG — the most interesting durability-adjusted name, but the value leg is unproven. It is the one name that screens as a durable franchise (consumable materials, recurring razor-blade revenue, brand-agnostic) and least-extended of the quality cohort (RSI 52, +8.9% 3m — the smallest 3m run in the basket, −11.8% off its high). The "cheap" claim, though, rests almost entirely on P/S 6.6 — and at an 8% net margin with trailing P/E 81, that is not self-evidently cheap. The working thesis is that the 8% net is depressed by CMC-acquisition-debt amortization and normalizes higher, but that requires a normalized / ex-amortization margin pull to confirm — until then ENTG is a candidate, not a conclusion (filed as a follow-up below).
2. AMAT / ASML — the durable end-state, now priced as such. Best quality-adjusted mega-caps (AMAT fwd 31 / 29% net; ASML the EUV monopoly, fwd 36 / 30% net) — but both sit RSI 74 / 68 at all-time highs, so the multiple already reflects the certainty. The franchise quality is the most defensible call in the complex; the momentum is the least favorable. The read flags that tension, not a timing instruction.
3. FORM — the lone un-rotated name, but the weakest on quality. It's the only genuine price laggard left (RSI 48, −20.8% off high) — but it's weak-down and hasn't turned (negative 7d and 30d), and it's the worst-quality / most-expensive name on the board (fwd 46, 8% net, small-cap probe cards). It's the cleanest illustration of "hasn't run ≠ bargain," not a durable-franchise candidate.
4. The through-rotated cohort has already re-rated — TER / ACLS / KLIC / AMKR / ONTO ran 34–84% in 3m and sit RSI 59–66 near highs. The un-run-laggard framing no longer applies to them; whatever upstream-rotation gap existed has substantially compressed here.
5. The contrast remains the trap — MU/SNDK at fwd P/E ~10 on peak margin is the cycle-top tell, not the bargain (overlay above). The whole point of going upstream was to read around owning the commodity layer at the top.

Meta-lesson (filed): the central call — "the alpha left is one layer upstream" — looks right, but I can't cleanly claim "the rotation played through in two sessions." The 06-01 screen ran on stale scattered-cache OHLC (FORM 39, ONTO 44, CAMT 44); the 06-03 read is a fresh focused fetch (mid-caps/majors RSI 59–74). Stale-screen-vs-fresh-fetch can't separate real two-session price movement from the earlier cache simply being behind — a same-source 06-01-vs-06-03 comparison (filed below) is what's needed to make a timing claim. What is defensible: by the time we looked with fresh focused data, only FORM (and arguably ENTG/CAMT/ONTO) was still genuinely un-rotated, and the durable read (franchise quality) is the part that holds regardless of exactly where the momentum window sits — which is the real lesson over the momentum read (which name is temporarily un-run).

Coverage gaps + follow-ups (filed)

• HDD deep-upstream — RESOLVED by the follow-on investigation. The initial TDK/Resonac/Nidec guess was superseded: 2026-06-03-nearline-hdd-chokepoint-corroboration-supply-chain-durable-vs-bubble found that heads/media/HAMR-writer are captive in-house at STX/WDC, so TDK/Resonac/Nidec are too diluted. The real externally-investable upstream is Hoya (glass/HAMR, ~70%) + NHK Spring (suspensions duopoly) — now added to a nearline-storage watchlist (security-families in _securities.json). See that investigation for the corrected map and coverage decision.
• WFE + advanced-packaging-test watchlist — BUILT (2026-06-03). wfe-test-metrology (FORM, TER, CAMT, ONTO, ENTG, ACLS, KLAC, AMAT, LRCX, ASML, KLIC, BESIY, AMKR), OHLC and scan summaries live and verified clean. Building it immediately surfaced that the scattered source watchlists were carrying stale OHLC — see the tape-freshness correction in the Valuation overlay. Next: wire into a scan slot so the rotation stays scannable.
• Valuation pull — DONE (2026-06-03). EV/sales + fwd/trailing P/E + net margin pulled for both buckets — see Valuation overlay above. Net finding: memory (MU/SNDK fwd P/E ~10 at peak margin) is the quantified peak-cycle trap; the WFE majors (AMAT/ASML) + consumable-materials (ENTG) are the durability-adjusted value; the small-cap metrology laggards (FORM/CAMT/ONTO) are momentum-cheap but margin-thin, not bargains. 10y-multiple-percentile considered and dropped — .info gives spot multiples only, there's no clean historical multiple series to build the percentile from, and the feasible proxy (price/return percentile vs own history) is a momentum read, not valuation. The spot overlay already makes the durable-vs-bubble call; not worth a constructed-series build.
• Nearline-HDD-allocated-through-2027 claim — CLOSED / corroborated by the follow-on investigation (Seagate Q3 FY26: "almost fully allocated through calendar 2027"; WDC: lead times "about a year," "no plans to add unit capacity"; TrendForce >52-wk). No longer open.
• ENTG normalized-margin pull — OPEN. The ranked read above hinges ENTG's "value" leg on the claim that its 8% net margin is depressed by CMC-acquisition-debt amortization. Confirm with a normalized / ex-amortization (or segment-EBIT) margin pull before treating ENTG as durability-adjusted value rather than a P/S-cheap-but-thin name. Until done, ENTG is a candidate, not a conclusion.
• Same-source 06-01 vs 06-03 rotation comparison — OPEN. The "window closed fast" read compares a stale scattered-cache 06-01 screen to a fresh focused 06-03 fetch — which can't separate real price movement from cache lag. Re-pull the cohort from the same watchlist at both dates (or accept the timing claim is unproven). The standing read (which names are extended now) does not depend on this; only the speed claim does.
• Hoya/NHK quality ranking — POSITION-BASED, financials owed. The companion nearline-storage work ranks Hoya "highest-quality upstream" on supply-chain position (glass/HAMR ~70%), but both Hoya and NHK Spring are conglomerate/foreign wrappers; a segment-level read (what fraction of each company is actually HDD-exposed) is owed before "best risk/reward" is more than a layer-position inference. Tracked in that perspective.
• These feed the memory-supercycle perspective (storage-vs-memory split) and ai-power-bottleneck (the equipment layer). Producer-only; no watchlist mutation or perspective edit made here.

Sources

Price, RSI and trend figures read from the desk's pre-computed scan summaries (summaries). No number in this note was computed in prose.