Deep Decode: The AI Infrastructure Watchlist — One Stack, Two Markets
Thirty-two AI-infrastructure names, ten bottleneck layers. The top of the stack — chips, memory, servers — ran 30% to 100% in a month and is now overbought. The physical bottom — power, cooling, optics, grid — never moved. Where the Street still keeps targets, Nvidia has the most room left and AMD trades 44% above its. The watchlist that looks like one trade is actually two.
AI watchlist split: top of the stack vs the physical bottom (20-day return)
+70% / −9%
▲ +78.8 Servers ran; power, cooling, optics & grid never moved — and 9 of 32 names are overbought, all up top
It feels like one unstoppable AI trade where everything goes up together. The watchlist says otherwise. Sort thirty-two AI-infrastructure names into ten bottleneck layers and the board splits in two. The top — servers (Dell +100% in 20 days), memory (Micron +79%), custom silicon (Astera +69%), HPE +51%, AMD +43% — ran hard and is now stretched: nine of thirty-two names closed above RSI 70, and every one of them is top-of-stack. The bottom — power, cooling, optics, grid — sat the move out entirely (layer medians −6% to −9% on the month). Then the tells: where Wall Street still publishes targets, the momentum darlings ran PAST them (AMD trades 44% above its mean target, Cisco 24% above) while Nvidia — the name everyone calls expensive — has 30% upside left and an RSI under 50. And the only insiders buying in the entire watchlist are in the cold layers — Carrier, Vistra, Constellation. We walk the full watchlist layer by layer, then mark the one catalyst that matters before summer: Broadcom on June 3.
The contradiction
Build the AI-infrastructure watchlist the way the plan asks — every name sorted by the bottleneck layer it serves, from GPUs at the top to the power grid at the bottom — and one number ends the “it's all one trade” story. Nine of the thirty-two covered names closed the week overbought, above RSI 70. Every single one of them sits in the top of the stack: Dell, Super Micro and HPE in servers; Micron, Western Digital and Seagate in memory; Astera in custom silicon; AMD; Cisco. Zero are in power. Zero in cooling. Zero in optics. Zero in grid. The top of the stack ran 30% to 100% in twenty trading days and is now stretched; the physical bottom — the megawatts, the heat removal, the light, the generation — has a layer-median twenty-day return between minus six and minus nine percent and an RSI in the low 40s. Then the two tells that turn a screener into a thesis. First, where Wall Street still publishes a target, the hottest names ran past it: AMD trades forty-four percent above its mean price target, Cisco twenty-four percent above — while Nvidia, the name everyone calls expensive, has thirty percent upside left and an RSI under 50. Second, the only insiders BUYING anywhere in the watchlist are in the cold layers — Carrier, Vistra, Constellation. The watchlist that looks like one melt-up is two markets that stopped talking to each other. This Deep Decode maps both, and marks where they might reconnect.
What the headline says
What the AI trade looks like
One unstoppable melt-up — chips, memory, servers, power, the whole stack rising together
What the data says
What the watchlist actually shows
Two disconnected markets — an overbought top that ran 30–100%, and a cold bottom (power, cooling, optics, grid) that never moved
Chapter 01
One Stack, Two Markets
Sort the watchlist into ten bottleneck layers and rank them by twenty-day median return, and it splits in two. The top six layers are green and big: servers +70.1%, custom silicon +26.1%, GPUs +24.8%, memory +23.1%, foundry/equipment +13.0%, networking +11.7%. The bottom four are red and flat: grid −1.7%, power −5.9%, cooling −6.4%, optics −8.7%. The AI trade everyone describes as one melt-up is, on the watchlist, two markets — and the line between them is whether the layer is something you manufacture or something you have to physically build into the world.
Start with the frame the whole piece hangs on. Take the AI-infrastructure watchlist — thirty-two covered names — and sort it into ten bottleneck layers, top of the stack to bottom: GPUs, custom silicon, networking, optics, memory, servers, power, cooling, grid, and the foundry/equipment that makes the chips. Rank the layers by their median twenty-day return and a clean gradient appears. The top is green and steep: servers lead at a +70.1% median, dragged up by Dell's astonishing +100.3% month; custom silicon is next at +26.1%, GPUs +24.8%, memory +23.1%, foundry and wafer-fab equipment +13.0%, networking +11.7%. Then the gradient crosses zero and the bottom four layers are all red — grid −1.7%, power −5.9%, cooling −6.4%, optics −8.7%. This is the same split Saturday's Weekend Decode caught in a single week, now hardened over a full month and across a wider list. And the dividing line is physical: the green layers are the parts of AI you can fabricate and ship — chips, memory, assembled servers — while the red layers are the parts you have to build into the physical world: substations, chillers, fiber, generation. The market spent the month paying up for what ships and ignoring what has to be built. So the actionable read on the watchlist isn't “AI is hot.” It's that the watchlist is two lists with opposite setups — and you manage them differently.
"The market spent the month paying up for what ships — chips, memory, assembled servers — and ignoring what has to be built into the world: substations, chillers, fiber, generation."
AI Infrastructure Watchlist — Median 20-Day Return by Bottleneck Layer
Source: MarketDecode scanner — median(price.ret20dPct) by layer, 2026-05-31
Hottest layer
Servers 70.1%Dell +100.3% · Super Micro +70.1% · HPE +50.6%
Coldest layer
Optics −8.7%Power −5.9% · Cooling −6.4% · Grid −1.7%
Top-vs-bottom spread
78.8 ptsServers median minus optics median
Chapter 02
Why Now: Two Clouds, Not One
Plot all the watchlist names by how far they ran in twenty days against how overbought they are now, and they fall into two separate clouds. The top-right is the hot cohort — Dell, Micron, Super Micro, Astera, HPE, AMD, Cisco — big runs and RSI 73 to 90. The bottom-left is the cold cohort — Vertiv, Eaton, GE Vernova, Lumentum, Applied Optoelectronics, plus Nvidia — flat-to-negative months and RSI in the 40s. The two clouds barely touch. Nine of thirty-two names are overbought, and every one is in the top-right.
A two-factor view turns the layer gradient into something you can act on. Put twenty-day return on the horizontal axis and RSI-14 on the vertical, draw the 70 line for overbought and the zero line for a flat month, and the watchlist resolves into two distinct clouds with almost nothing in between. Top-right — ran hard and now stretched — is the entire hot cohort: Dell at +100% and an almost unheard-of RSI 89.9, Micron +79% at RSI 79, Super Micro +70% at 79, Astera +69% at 80, HPE +51% at 83, AMD +43% at 76, Cisco +31% at 76. Bottom-left — didn't run and isn't hot — is the cold cohort: Vertiv, Eaton, GE Vernova, Carrier in the power-and-cooling block, Lumentum and Applied Optoelectronics in optics, all sitting around a flat-to-negative month with RSI in the low-to-mid 40s. Here's the teach-one-thing, and it's the reusable part: when a name is BOTH up a lot and overbought, the move is being driven by positioning and momentum, not by something fundamental that just changed — there's no new earnings number under most of these twenty-day runs. That combination tends to mean-revert. The single most interesting dot is Nvidia, and it's not where you'd expect: it sits in the calm bottom-left at RSI 47.9, a +6% month — the one mega-cap that did NOT join the melt-up. Hold that thought; section five is about exactly what that calm is worth.
"When a name is BOTH up a lot and overbought, the move is positioning and momentum, not something fundamental that just changed — and that combination tends to mean-revert."
20-Day Return vs RSI-14 — The AI Infrastructure Watchlist
Source: MarketDecode scanner — ret20dPct & rsi14, 2026-05-31
Most stretched
DELL 100%, RSI 89.9A twenty-day double, deep into overbought
Overbought count
9 of 32All top-of-stack; none in power/cooling/optics/grid
The calm mega-cap
NVDA RSI 47.9+6% month — sat out the melt-up
Chapter 03
The Names That Ran: Inside the Hot Cohort
The top of the stack didn't just drift up — it went vertical. Dell is up 100.3% in twenty trading days, Micron +79.1%, Super Micro +70.1%, Astera +69.2%, HPE +50.6%, AMD +43.1%, Cisco +31.1%, Credo +28.0%, Marvell +24.3%, Western Digital +23.1%. These are the watchlist's compute, memory, and box layers — the parts that turn capex into shippable product — and they priced a quarter's worth of good news in a month.
Look at the magnitudes, because they're the why-now. Dell rose one hundred percent in twenty trading days — a mega-cap doubling on the back of a blowout AI-server quarter that Yahoo Finance summarized as $60 billion in AI-server revenue and a wave of enterprise demand. Micron is up seventy-nine percent, having crossed a trillion dollars in market cap as the Financial Times and others marked memory chipmakers joining the trillion-dollar club. Super Micro +70%, Astera +69%, HPE +51%, AMD +43%, Cisco +31% — the WSJ's framing of a “$5.7 trillion and counting” chip rally captures the mood. Every one of these is a top-of-stack name: the GPUs and custom silicon that compute, the memory that feeds them, the servers that assemble them. They share a property the cold layers don't — they convert AI capex into a product you can build in a fab and ship this quarter, so a single strong order cycle reprices them fast. That's the bull case and the warning in the same sentence. The good news is real and recent; it's also now IN the price, which is what an RSI in the high 70s and 80s across the whole cohort is telling you. The actionable point: these are not names to chase here — they're names to have a re-entry level on. The move already happened; the question is what holds it.
The Hot Cohort — 20-Day Return (Top of the Stack)
Source: MarketDecode scanner — price.ret20dPct, 2026-05-31
The move
DELL 100% / 20d$60B AI-server quarter (Yahoo Finance, May 29)
Memory milestone
MU 79% / 20dCrossed $1T cap — FT, May 27
The cohort
Compute · memory · boxesWhat turns capex into shippable product
Chapter 04
The Full Map — and the Layers Nobody Bought
Score the hot cohort against the cold cohort on the factors the scanner tracks and the fingerprints are opposites in three places and identical in one. The hot names lead big on momentum (70 vs 34), RSI heat (76 vs 46), and twenty-day return. They tie on composite quality (52 vs 45). And they LOSE on the one factor that's supposed to be smart money: insider buying tilt, where the cold cohort scores more than double (35 vs 14). The layers nobody bought — power, cooling, optics, grid — are the only ones insiders are quietly accumulating.
Lay the two cohorts on the same radar and the shape tells the story a return number hides. On momentum and RSI, the hot cohort's polygon balloons out — momentum 70 against the cold cohort's 34, RSI 76 against 46 — and on normalized twenty-day return it's not close. But slide to the composite-quality axis and the two nearly meet: 52 versus 45, because the scanner's underlying value and quality factors barely separate these names. So most of the hot cohort's advantage is the two momentum axes — exactly the factors that can unwind without a single fundamental changing. Now the axis that flips: insider tilt. Across the hot cohort, insiders are net sellers almost everywhere — Dell, Micron, Super Micro, AMD, Cisco, Marvell, Western Digital all carry bearish insider signals — for a cohort tilt of about 14 on our 0–100 scale. The cold cohort scores 35, more than double, because the only bullish insider signals anywhere in the watchlist live here: Carrier in cooling, Vistra and Constellation in grid. This is the part of the watchlist the screen scrolls past — the physical layers MarketWatch dubbed “the Nvidias of power” in its $700 billion energy-grab piece — and it's the only part where the people who run these companies are buying their own stock. The reusable read: when momentum and insider tilt point in opposite directions, trust the insiders for the next quarter and the momentum for the next week.
"When momentum and insider tilt point in opposite directions, trust the insiders for the next quarter and the momentum for the next week."
Hot Cohort vs Cold Cohort — Factor Fingerprint (0–100)
Source: MarketDecode scanner — cohort median scores, 2026-05-31 (*20-day return & insider tilt normalized to 0–100)
The momentum gap
Momentum 70 vs 34And RSI heat 76 vs 46 — the hot cohort’s real edge
Where insiders buy
Cold tilt 35 vs 14CARR, VST, CEG — the only bullish signals in the list
Quality is a tie
Composite 52 vs 45The split is momentum, not business quality
Chapter 05
Bull vs Bear: Where the Street Still Sees Room
Only four watchlist names carry a full, live set of Wall Street price targets in our data — Nvidia, Broadcom, Cisco, and AMD — and they invert the popular story completely. Nvidia, the “expensive” one, has +30.0% upside to a $274.53 mean target and a 48 RSI. Broadcom is fair at +1.0%. But the two momentum darlings have run PAST consensus: Cisco trades 23.7% above its $91.90 target, AMD a stunning 44.0% above its $289.04 target. The names that ran the most have the least support; the name that sat still has the most room.
This is the section the whole piece has been setting up, and it comes with a caveat worth stating plainly: of the thirty-two names, only four carry a complete, current Street price-target set in our data — Nvidia, Broadcom, Cisco, and AMD — so read this as the four with live targets, not the whole list. Even so, the inversion is striking. Nvidia, the stock retail investors are most likely to call overheated, has the most upside left: a $274.53 mean target against a $211 price, thirty percent of room, with thirty-seven of thirty-nine analysts at buy — and it's the name that DIDN'T run, RSI 47.9. Broadcom sits right at fair value, one percent of upside, the calmest read on the board going into its print. Then the two that ran: Cisco, up thirty-one percent on the month, now trades twenty-four percent ABOVE its $91.90 mean target; AMD, up forty-three percent, trades forty-four percent above its $289 target. Strong-buy ratings, yes — but the price has sprinted past where the analysts setting those ratings will defend it. Barron's caught the setup in mid-May: “The chip-stock rally has stalled. Nvidia will prove it still has room to run.” The watchlist agrees. The actionable read on these four: of the names with targets, Nvidia is the one where the chart and the Street point the same direction; AMD and Cisco are where you'd want a pullback before adding, because they're already trading on momentum the analysts haven't underwritten.
"The names that ran the most have the least support; the name that sat still has the most room. That is the whole watchlist in one chart."
Upside to Mean Price Target — The Four Names With Live Street Targets
Source: MarketDecode scanner — analyst.upsidePct (price vs mean PT), 2026-05-31
Most room left
NVDA 30.0%$274.53 target vs $211 · 37/39 analysts buy
Ran past consensus
AMD −44.0%$516 price vs $289 mean target
Fair value
AVGO 1.0%Calmest read into its June 3 print
Chapter 06
What to Watch: The Heat Dial and the One Catalyst Before Summer
The hot cohort enters June with a median RSI of about 76, deep in overbought — Dell at 89.9, Nvidia the outlier at 47.9. There is exactly one near-term catalyst that can re-rate the watchlist: Broadcom reports fiscal Q2 after the close on Wednesday, June 3, the quarter's key custom-silicon read-through for Marvell, Astera, and Credo. After that, the watchlist goes quiet into the late-summer earnings cluster. Until something cracks the cold layers' RSI back above 60 on rising prices, the two markets stay split.
Half the value of a watchlist is knowing what would change it, so end on the dial and the calendar. The heat dial first: the hot cohort's median RSI sits near 76, with Dell at an extreme 89.9 — the kind of reading that historically precedes a pause, not an acceleration. Nvidia is the lone hot-cohort-adjacent name nowhere near it, at 47.9, which is the entire reason section five flagged it. Now the calendar. The one catalyst that matters before summer is Broadcom: it reports fiscal Q2 after the close on Wednesday, June 3 — a public date from Broadcom's investor-relations page — and it's the cleanest custom-silicon read-through of the quarter, the natural sequel to Marvell's print and the first real test of whether the AI-networking-silicon layer (Marvell, Astera, Credo) keeps its momentum. After June 3, the watchlist's own earnings calendar thins out dramatically: the next dense cluster doesn't arrive until the late-summer window, which means for most of June and July there's no fundamental catalyst to either justify the overbought top or wake the cold bottom — the tape will be driven by positioning, not numbers. So here's the watch list, specific and datable. One: does the hot cohort cool — do Dell, Micron, and Super Micro pull back toward their moving averages, or does the bid hold? Two: do the cold layers finally catch a bid — does power, cooling, or grid push RSI back above 60 on rising twenty-day returns, with Carrier, Vistra, and Constellation (the insider-buy names) leading? Three: Broadcom on June 3 — a strong custom-silicon number pulls Marvell, Astera, and Credo with it; a soft one takes the air out of the most overbought corner of the market. Until one of those resolves, you manage the watchlist as what it is: two markets, one stretched and one ignored, waiting to find out which way they reconnect.
"For most of June and July there’s no fundamental catalyst to justify the overbought top or wake the cold bottom — the tape will be driven by positioning, not numbers."
The Watchlist Heat Dial — RSI-14 Into June
Source: MarketDecode scanner — rsi14, 2026-05-31
Heat reading
Hot cohort RSI ~76Dell 89.9 — overbought into a thin tape
The one catalyst
AVGO — Wed Jun 3 AMCCustom-silicon read-through (MRVL/ALAB/CRDO)
The asymmetry
CARR · VST · CEGCold layers, insider-bought — watch for RSI > 60
Resolution window — 1 month
What would confirm or invalidate this read
Confirmation
By end of June 2026: the two markets reconnect toward the bottom — the cold layers (power/cooling/grid: VRT, ETN, GEV, CARR, VST, CEG) push median RSI back above 60 on rising 20-day returns AND/OR Broadcom’s June 3 print pulls the whole stack (not just custom silicon) higher, while the overbought top (DELL, MU, SMCI, ALAB above RSI 70) cools. That is the watchlist re-broadening into one market.
Invalidation
By end of June 2026: the split persists or widens — the cold layers stay flat-to-negative with RSI in the 40s while the hot cohort holds its gains, confirming the move was a top-heavy, positioning-driven melt-up rather than a broad re-rating of the AI-infrastructure stack.