The AI Infrastructure Scorecard: What the Last Two Weeks Proved
Six calls graded. Four confirmed, one emphatically wrong, one still pending. Dell broke the sell-the-beat rule with a +33% next-day move and a +100% twenty-day run. The cooling and power layers stayed cold. Broadcom prints Wednesday.
Dell’s 20-day return through Friday’s close
+100%
▲ +100.3 RSI 89.9, +43% in the 5 sessions around the print. The sell-the-beat call invalidated.
Two weeks ago we called sell-the-beat on Marvell and Dell, valuation risk on optics, structural mispricing in cooling, scarcity in memory, and a slow reprice in power. Today we grade. Marvell confirmed (−4.59% on the print). Dell did not — the stock added +33% the day after earnings and is now +100% over 20 days at RSI 89.9, the cleanest invalidation we have ever recorded. Optics confirmed (LITE −10%, COHR −4% on 20d). Cooling stayed cold (VRT −4%, ETN −6%, TT −7%, CARR −6%). Memory confirmed and overshot (MU +79%, mean PT $615 vs stock $971). Power confirmed-cold (NEE −10%, CEG −7%, NRG, DUK, SO all negative). The lesson Dell taught: when the guide-raise is larger than the run-up, sell-the-beat breaks. Marvell raised FY28 by $1.5B. Dell raised FY27 by $27B — from $140B to $167B. Broadcom reports Wednesday after the close.
The contradiction
Five of six chip-and-server names ran 24–100% in the last twenty trading days. Dell alone added $116 billion in market value on a single guide raise (FY27 from $140B to $167B). Marvell, Astera Labs, Credo, Micron and Super Micro all confirmed the demand thesis. But the layer that physically enables that demand — cooling, electrical infrastructure, power generation, transmission, grid equipment — stayed flat or got cheaper. Vertiv, Eaton, Trane, Carrier are negative on 20 days. NextEra is down 10%. Constellation Energy is down 7%. The market is paying for the silicon and paying for the box; it is not yet paying for the kilowatts and the heat removal. That is either a lag (which will close when Broadcom prints Wednesday and the next capex cycle is confirmed) or a structural verdict (the physical layer is a regulated, hyperscaler-bargained, low-multiple business). Broadcom’s Wednesday print is the next test.
What the headline says
The AI infrastructure trade is broadening — every chip-layer call confirmed
DELL +100% 20d, MU +79% 20d, ALAB +69% 20d, SMCI +70% 20d, MRVL +24% 20d
What the data says
But the physical infrastructure layer is still flat or down
VRT −4%, ETN −6%, TT −7%, CARR −6%, NEE −10%, CEG −7% on 20d
Chapter 01
The Scorecard: Four Calls Confirmed, One Invalidated, One Pending
Six MarketDecode calls from the May 17–30 arc. Marvell sell-the-beat — CONFIRMED (−4.59% on the print). Dell sell-the-beat — INVALIDATED (+33% next day, +100% 20d). Optics valuation risk — CONFIRMED (LITE −10% 20d). Cooling underpriced — PENDING (VRT, ETN, TT, CARR still negative). Memory scarce — CONFIRMED and overshot (MU +79%, stock above every PT). Power reprice — PENDING (NEE −10%, CEG −7%, only VST positive).
On May 27 we previewed Marvell as a sell-the-beat candidate: stock at $208, mean PT $174, +35% in 20 days. Marvell delivered a blowout print and fell 4.59% the next day. Call confirmed. On May 28 we extended the call to Dell: +48% in 20 days, RSI 80, stock $305 vs mean PT $240. Dell delivered the largest guide raise in AI server history — FY27 from $140 billion to $167 billion — and added +33% the next session. By today the 20-day return is +100% and RSI is 89.9. Call invalidated. On May 17–19 we flagged optics valuation risk after the COHR vs LITE post: COHR is down 4% on 5 days, LITE is down 10% on 20 days. Call confirmed. On May 24–26 we said the cooling layer was structurally underpriced; the four names are still negative on 20 days. Call pending; the thesis required a chip-layer print to act as the catalyst, and Broadcom on Wednesday is the next test. On May 29 we wrote that memory was the AI bottleneck no one was pricing; Micron is +79% in 20 days, RSI 78.9, stock $971 vs mean PT $615. Call confirmed; arguably overshot. On May 25 the Deep Decode argued the power layer would reprice on a multi-quarter lag; NEE is down 10%, CEG down 7%, DUK and SO and AEP all negative. VST is the only name positive (+3%). Call pending.
The Two-Week Scorecard: 20-Day Returns by Call
Source: MarketDecode scanner, 20-day returns as of 2026-05-31 close
Confirmed Calls
4 of 6MRVL, optics, MU, basket-wide rally
Invalidated
DELL+100% 20d; guide raise overwhelmed run-up
Still Pending
Cooling + PowerVRT −4%, NEE −10%, CEG −7% on 20d
Chapter 02
Why Dell Broke the Sell-the-Beat Rule: The Guide-Raise Was Bigger Than the Run-Up
Sell-the-beat requires the incremental surprise to be smaller than what the stock already priced. Marvell ran +27% pre-print and delivered a 3.85% guide beat — surprise smaller than run-up, sold off. Dell ran +48% pre-print and delivered a 19% guide raise (FY27 from $140B to $167B). The surprise was nearly twice as large as the run-up. The pattern survived; the condition got sharpened.
The sell-the-beat thesis says: when a stock has rallied sharply into a print, the price already reflects the beat. If the beat magnitude is smaller than the run-up, momentum unwinds and the stock falls. Marvell was the canonical case — it ran +27% in 20 days, delivered a 3.85% guide beat, and fell 4.59%. Dell was supposed to be the same setup: +48% run-up, mean PT $240, RSI 80. The pattern would have said: even a beat sells. Instead Dell delivered a guide raise that no model priced in. FY27 revenue guide went from $140 billion (prior outlook) to $167 billion — a $27 billion raise, or +19% above the consensus of $142.1 billion. That is roughly five times the magnitude of Marvell’s guide raise in percentage terms ($16.5B vs $15B prior = +1.5B = +10% for MRVL FY28; vs +19% for DELL FY27). The market’s pricing math was: at +48% pre-print, the stock had priced an 8–10% beat. Dell delivered nearly 20%. The incremental surprise was larger than the run-up. Sell-the-beat physics survived — the rule simply says the surprise has to be smaller than the run-up, and Dell’s surprise was not. The rule is now: pattern fires when expected; expect failure when the company is sitting on a multi-quarter visibility shift it hasn’t guided to yet.
Run-Up vs Guide-Raise Magnitude: Marvell vs Dell
Source: MarketDecode scanner + earnings releases, 2026-05-27 + 2026-05-28
Dell FY27 Raise
$140B → $167B+$27B, +19% above consensus $142B
Marvell FY28 Raise
$15B → $16.5B+$1.5B, +3.85% guide beat Q2
The Lesson
Surprise > Run-UpWhen true, the rally extends
Chapter 03
The Two-Tier AI Infrastructure Market: Silicon Wins, Steel Waits
The chip-and-server layer (DELL, MU, ALAB, SMCI, MRVL, CRDO) returned +24% to +100% in 20 days, every name above RSI 65. The physical infrastructure layer (VRT, ETN, TT, CARR, NEE, DUK, SO, AEP, CEG) returned −4% to −10%, every name below RSI 55. Same period, same AI thesis, opposite tape. The market is paying for compute. It is not yet paying for the kilowatts that make compute possible.
The decomposition is stark. Look at the chip and server layer over the last 20 days: Dell +100%, Micron +79%, SMCI +70%, Astera Labs +69%, Credo +28%, Marvell +24%, Broadcom +6%. Average return: roughly +54%. RSI distribution: ALAB 80, DELL 90, MU 79, SMCI 79, CRDO 69, MRVL 69, AVGO 64. Every single name is at or above the bullish-momentum threshold. Now the physical infrastructure layer over the same 20 days: VRT −4%, CARR −6%, ETN −6%, TT −7%, NEE −10%, AEP −7%, DUK −5%, SO −5%, CEG −7%, GEV −9%. Average return: −6%. RSI distribution clusters between 40 and 55 — the opposite of overbought. Vertiv’s composite score is 42, the lowest in the basket. NextEra is at 39. This is the cleanest two-tier market in the AI cycle. The chip layer is pricing a multi-year capex acceleration. The physical layer is pricing nothing. One of two things must be true: either the physical layer is structurally a lower-multiple business (regulated, capital-intensive, hyperscaler-bargained) and the gap is permanent, or the lag has not finished and the next chip-layer print (Broadcom Wednesday) is the catalyst that starts the physical reprice. Marvell + Dell already happened. Neither dragged the physical layer with them. The third print is the test.
AI Infrastructure: Chip Layer vs Physical Layer (20D Returns)
Source: MarketDecode scanner, 20-day returns as of 2026-05-31
Chip Layer Avg 20D
54%DELL, MU, SMCI, ALAB, CRDO, MRVL
Physical Layer Avg 20D
−6%VRT, ETN, TT, CARR, NEE, DUK, SO, CEG, GEV
The Gap
60 ptsCleanest two-tier split of the AI cycle
Chapter 04
The Broadcom Setup: Stock $446, Mean PT $508, Print Wednesday AMC
Broadcom reports Q2 FY26 after the close on Wednesday June 3. Unlike Dell and Marvell, AVGO sits BELOW its analyst tape: stock $446.77, mean PT $508.75. That is a 14% discount to consensus. The 20-day return is only +6%, RSI 64. Insider selling is −$79M (modest by AI standards), options P/C 0.43, premium $9.5M. This is the inverse of the Dell setup: low expectations, modest run-up, room to surprise upward.
After Marvell and Dell, Broadcom is the third leg of the AI custom-silicon trifecta. The setup is different. Where Dell sat $65 above its analyst mean PT going into the print, Broadcom sits $62 BELOW. The stock has only run +6% in 20 days; RSI is 64, well below the overbought threshold. Insider selling at −$79M over 90 days is modest — Dell’s was −$957M and Marvell’s was −$107M. Options put-call ratio of 0.43 is moderately bullish; total premium of $9.5 million is typical for a Tier 1 catalyst. The AVGO sell-the-beat exposure is therefore inverted: there is no run-up to unwind, but there is a $62 analyst gap that wants to close. A beat-and-raise of Marvell-or-Dell magnitude (5–20% guide raise) could trigger an analyst PT cascade that closes the gap upward, not downward. The risk: AVGO’s AI semiconductor business is dominated by two hyperscaler relationships (Google TPU and Meta MTIA); any color suggesting either is slowing the order pace could invalidate the entire custom-silicon thesis. The grade date is Thursday June 4 at the open. Triggers: (1) AI semi revenue >$5.5B for the quarter; (2) FY26 AI semi guide raised; (3) VMware/infrastructure software op margins >70%; (4) AVGO closes above $470 on Thursday. 3 of 4 = the AI custom-silicon thesis is reaffirmed and the basket extends. 0–1 of 4 = the trade is concentrating in NVIDIA and the custom-silicon premium fades.
Broadcom vs Analyst Tape Into Wednesday’s Print
Source: MarketDecode scanner (analyst.json), 2026-06-01
AVGO 20D Return
6.05%RSI 63.6 — not overbought
AVGO vs Mean PT
−$62Stock $446 vs mean $508 (14% discount)
Insider Net 90D
−$79MModest vs DELL −$957M, MRVL −$107M
Chapter 05
Bull vs Bear: Six Categories Across the Whole AI Basket
Bull: 5 of 6 chip-layer calls confirmed, Dell delivered the largest AI server guide raise in history, AVGO trades below mean PT into Wednesday print, NVDA composite 65 still leading, options remain call-skewed. Bear: physical infrastructure layer flat-to-down 20 days, RSI overextension across the chip layer (DELL 90, ALAB 80, MU 79, SMCI 79), insider selling cumulative −$1.2B across DELL+MRVL+MU+AVGO+VRT, no power-layer catalyst on the calendar before July.
Six evidence categories. Bull: (1) Call accuracy — 5 of 6 chip-layer calls confirmed in two weeks. (2) Dell’s $27B guide raise validates that AI server demand is accelerating, not plateauing. (3) Broadcom enters Wednesday below its analyst tape, the cleanest setup for an upside surprise we have seen. (4) NVIDIA composite score remains 65, anchoring the demand signal upstream. (5) Options tape across the chip-layer basket is call-skewed (P/C ratios cluster around 0.20–0.45). (6) Memory has confirmed it as a structural shortage, not a one-quarter print. Bear: (1) RSI exhaustion — Dell at 90, ALAB at 80, MU at 79, SMCI at 79 — these are levels historically associated with 5–15% pullbacks within 3 weeks regardless of fundamentals. (2) The physical layer has not moved. If it doesn’t reprice on Broadcom, the thesis that capex translates into power/cooling stock returns is in serious doubt. (3) Cumulative insider selling across the basket is −$1.2B over 90 days. (4) The macro calendar this week is heavy: Broadcom Wed AMC, Jobs Friday AM. (5) No power-layer earnings catalyst until late July. (6) The custom-silicon trade is structurally concentrated in 4 names; if any one underperforms, the others sell off in sympathy regardless of their own setup.
Bull vs Bear: AI Basket Going Into Broadcom
Source: MarketDecode analyst scoring framework, 2026-06-01
Bull Score
25/30Calls + demand + AVGO setup
Bear Score
−24/30RSI + physical lag + insiders
Net
1Most balanced setup of the cycle
Chapter 06
The Read: Three Triggers for the Week and the Sharpened Rule
Three triggers this week: (1) Broadcom Wednesday AMC — AI semi revenue >$5.5B + raised FY26 guide; (2) cooling/power names start to reprice on Thursday/Friday if Broadcom confirms; (3) May jobs report Friday — a hot print pressures the AI multiple via yields. The sharpened sell-the-beat rule: the pattern only fires when the beat magnitude is smaller than the pre-print run-up. Dell’s 19% guide raise vs 48% run-up failed the test; Marvell’s 3.85% guide raise vs 27% run-up succeeded.
The MarketDecode forward watchlist. Wednesday June 3 after the close: Broadcom Q2 FY26. The grade date is Thursday June 4 at the open. The four triggers (AI semi revenue, FY guide, software margins, stock holds $470) decide whether the custom-silicon trade extends or concentrates. If 3 of 4 fire, the chip-layer rally continues and the physical layer gets a second chance to reprice on a multi-week lag. If 0–1 of 4 fire, the trade is sitting on NVIDIA alone and the custom-silicon premium fades. Friday June 5 at 8:30 AM ET: BLS May employment report. A hot print (>200K jobs, wages above 4%) pressures the AI multiple by lifting yields; a cool print (<150K, wages soft) supports valuations and gives the cooling/power layer cover to participate. The sharpened sell-the-beat rule is now production: the pattern fires when the beat magnitude is SMALLER than the 20-day run-up. Both halves of the comparison have to be measured. Marvell satisfied: 27% run, 3.85% beat — sold. Dell did not satisfy: 48% run, 19% beat — rallied. Broadcom is the next test: 6% run, expected guide raise unknown. If Broadcom raises FY26 by more than 6%, the rally extends. Less than 6%, sell-the-beat for AVGO. The grading is binary and dated.
This Week’s Three Triggers
Source: MarketDecode grading framework, to be scored 2026-06-05 close
Next Grade Date
Thu Jun 4Broadcom pre-market reaction
Macro Risk
Jobs Fri 8:30Hot = yields up = AI multiple compresses
New Rule
Beat > Run-UpSell-the-beat only fires when False
Resolution window — 1 week
What would confirm or invalidate this read
Confirmation
AVGO Q2 FY26 print on Wednesday June 3 AMC delivers 3 of 4 triggers (AI semi revenue >$5.5B, FY26 AI semi guide raised, software op margin >70%, AVGO closes above $470 on Thursday June 4). If all three hit, the chip-layer thesis extends and the cooling/power layer gets a second chance to reprice. Additionally, BLS Friday jobs report comes in <200K with wages <4% — cool enough that yields do not pressure the AI multiple.
Invalidation
AVGO Q2 FY26 delivers 0–1 of 4 triggers AND closes below $440 on Thursday June 4 — indicating the custom-silicon trade is concentrating in NVIDIA alone and the broader basket sell-the-beat dynamic returns. Alternatively, jobs print >250K with wages >4.2% pushes 10Y yields above 4.6% and the AI multiple compresses across the basket regardless of AVGO results.