Deep Decode: The AI Infrastructure Stack After Nvidia
The AI trade is no longer one stock. It is a stack — compute, networking, optics, memory, power, cooling, servers, and the grid. Five of those layers got paid this week. Three are still waiting.
AI infrastructure layers already repriced
5 of 8
▲ +62 after Nvidia’s $91B Q2 guide
Nvidia confirmed the largest capex cycle since the internet buildout: $81.6B in Q1 revenue, $91B guided for Q2, Data Center networking up 199%. Past five days of MarketDecode data show capital has already repriced four of the eight infrastructure layers — connectivity (ALAB +28%), compute (AMD +6%), memory (MU +5%), networking (MRVL +8%). Optics reversed Friday (LITE +11% in one session). The cold layers — Vertiv −13%, Eaton −4%, Trane −4%, Carrier −4% — face the longest revenue conversion lag AND the only structural margin compression risk in the stack. This is the map. And by mid-June we’ll know which cold layers are mispriced and which are correctly cheap.
The contradiction
Nvidia just proved the largest capex cycle since the internet buildout is happening, not optional. Capital has repriced five of the eight AI infrastructure layers in the five trading days since. But the three layers furthest from the GPU — power, cooling, and grid contractors — still trade as if the buildout might pause. Either the lag resolves in months and they reprice too, or the buildout is real but the physical layer is where the margins get compressed by external constraints (permits, interconnection, construction timelines) that no amount of AI demand can shorten.
What the headline says
Nvidia confirms the buildout
$81.6B Q1, $91B Q2 guide, Data Center networking +199%
What the data says
But the physical-layer trade is still selling
VRT −13%, ETN −4%, TT −4%, CARR −4% on the week
Chapter 01
The Eight-Layer Map
After this week, the AI trade is no longer a stock. It is a stack. Eight distinct layers, each with its own revenue timing, margin profile, and capital cycle.
Until last weekend, the consensus AI trade was: own Nvidia, then maybe own one or two derivatives. After Nvidia’s Q1 FY27 print this week — $81.6B in revenue, $75.2B from Data Center, Data Center networking revenue up 199% year over year, and a $91B Q2 guide — that frame is no longer enough. Nvidia is now the trigger, not the trade. The trade is the stack it pulls behind it. Eight layers: (1) GPUs themselves, (2) the connectivity silicon that lets data move between the GPU and everything else on the same board, (3) the high-bandwidth memory that sits next to the chip, (4) the optical links that move data between server racks, (5) the servers that host the GPUs, (6) the electrical infrastructure that delivers power to those servers, (7) the cooling that removes the heat those servers produce, and (8) the grid contractors and generators that supply the underlying electricity. Each layer has a different revenue lag from the GPU shipment date. Each has a different margin profile. Each has a different capital cycle. Past five days of MarketDecode data give us the first clean snapshot of how the market is now pricing those layers separately, in the light of confirmed demand. The picture is not subtle: layers closest to the GPU rallied hard. Layers furthest from it kept selling.
Eight-Layer AI Infrastructure Map: Past 5 Days of Returns
Source: MarketDecode scanner aggregation across layer leaders, 5-day returns ending 2026-05-22
Hottest layer
ConnectivityALAB +28% — on-board with the GPU itself
Already paid
5 of 8Connectivity, networking, generation, compute, memory
Still waiting
3 of 8Optics (reversing), power, cooling
Chapter 02
What Nvidia Actually Confirmed
Nvidia’s Q2 guide is the demand confirmation that lets you separate the AI trade into eight layers in the first place. Without it, the stack collapses back into one stock.
Three numbers in Nvidia’s Q1 FY27 print did the work. First: $81.6B in revenue — a single-quarter record, on its own already removing any “maybe AI demand is slowing” thesis. Second: $75.2B of that came from the Data Center segment. The split tells you Nvidia is not a chip company with a side AI business; it is now an AI infrastructure company with a smaller chip business attached. Third — the one that did the most stack work: Data Center networking revenue grew 199% year over year. That number is not about GPUs. It is about Spectrum-X switches, NVLink interconnects, and the software stack that ties multi-rack clusters together. A 199% number tells you the buildout has moved past “buy more GPUs” and into “build more cluster fabric.” That is the moment the entire downstream stack becomes investable as separate layers, because each layer now has its own confirmed demand stream, not just a derivative bet on Nvidia’s sell-through. The $91B Q2 guide ratifies the cycle for at least one more quarter. That is enough runway for capital to do what it did this week: pick the layers, rank the layers, and reprice them.
Nvidia Quarterly Revenue Trajectory ($B)
Source: Nvidia Q1 FY27 press release, 2026-05-21 (investor.nvidia.com)
Q1 FY27 revenue
$81.6BSingle-quarter record
Q2 FY27 guide
$91BImplies +11.5% sequential acceleration
Networking growth
199% YoYThe number that unlocks the downstream layers
Chapter 03
The Conversion Map: How Long Each Layer Waits for Revenue
Each layer of the stack has a different revenue-conversion lag from the GPU shipment. Capital sorted itself this week along that exact axis.
There is a clean physical logic to which layers repriced first and which are still waiting. It is the same logic that organizes the inventory cycle in any deep supply chain: how many fiscal quarters pass between the trigger (here, Nvidia shipping the GPU) and the layer’s revenue showing up? Layer 1 — connectivity silicon — sits on the SAME printed circuit board as the GPU. The retimers and CXL chips ship the same week as the GPU. Revenue lag: effectively zero quarters. Layer 2 — networking silicon and HBM — ships into the same server SKU. Lag: one quarter. Layer 3 — optical transceivers, servers, generation contracts — ship into the cluster within one to two quarters of GPU delivery. Layer 4 — power distribution, cooling units, electrical infrastructure — must be specified, ordered, manufactured, shipped, and installed in physical buildings. Lag: two to four quarters at best, often longer when grid permits and interconnection studies are involved. Layer 5 — the grid contractors and generators — face multi-year permitting cycles. Capital this week did not invent this ordering. It just recognized that with demand finally confirmed, lag is now the primary variable that decides which layer gets paid first.
Revenue Conversion Lag from GPU Shipment to Layer Revenue
Source: MarketDecode editorial framework based on supplier disclosures, 2026-05-22
Zero-lag layer
ConnectivityShips same board as GPU — ALAB +28% reflects it
Mid-lag layer
Optics~2 quarters — explains Friday’s reversal beginning
Long-lag layers
3+ quartersPower, cooling, grid — physical-build timelines
Chapter 04
Where Capital Already Won: The Five Hot Layers
Five of the eight layers repriced this week. The winners cluster at the proximity-to-GPU end of the stack. ALAB +28% is the cleanest signal of the entire week.
Astera Labs is the headline of the week. Up 28% in five trading days, more than double the next-best performer. ALAB makes PCIe retimers and CXL connectivity chips that sit on the same printed circuit board as the GPU. When Nvidia ships more Blackwell, Astera Labs ships more retimers. The revenue lag is measured in weeks. That is why the market repriced it harder than anything else in the stack — not because the bull thesis is stronger but because the lag is the shortest. Marvell at +7.8% is the same story one layer out: custom AI silicon and optical DSPs that ship into the same server SKUs. AMD at +6% is the compute layer, with HBM bundled in. Micron at +5.2% is the pure memory layer. Dell at +4.5% is the server layer, where the new Blackwell-class systems ship. Constellation Energy and Vistra at +7% each are a special case: a PJM capacity-procurement timeline acceleration this week shifted their contracted revenue forward by twelve months. These six tickers across five layers all share one structural feature: the revenue lag from Nvidia’s confirmed shipment to their own confirmed shipment is short enough that the next two quarterly prints will be visibly larger. That is what “already paid” means.
The Five Hot Layers: Tickers That Repriced This Week (5-Day %)
Source: MarketDecode scanner, 5-day returns ending 2026-05-22
Standalone winner
ALAB 28%RSI 78.8 — stretched but momentum intact
Layer cluster
4% to +8%Networking, compute, memory, servers, generation
PJM catalyst
CEG 7%, VST +7%Capacity procurement moved up 12 months
Chapter 05
Where Capital Still Waits: Power, Cooling, Contractors
Vertiv −13%, Eaton −4%, Trane −4%, Carrier −4%, Quanta −7%. The longest-lag layers are also the only ones with structural margin compression risk — capital is pricing both at once.
The cold side of the heat map is concentrated in two related layers: power and cooling (Vertiv, Eaton, Trane, Carrier) and grid contractors (Quanta Services, GE Vernova). Vertiv is down 12.8% in five days — the worst single-stock drawdown anywhere in the AI infrastructure complex this week. Eaton is down 4.5%, Trane Technologies down 4.1%, Carrier down 3.9%. On a 20-day basis the picture is even harder: Eaton −10%, Trane −8%. These companies build the power distribution units, uninterruptible power supplies, liquid cooling systems, and precision air handling that determine how many GPUs you can physically rack per square meter. Without them, the chips sit in warehouses. The question is why the market won’t bid them, given Nvidia just confirmed demand and given every layer above them has rallied. Two non-mutually-exclusive answers. First, revenue conversion lag: a Q2 spike in GPU orders does not translate into a power-distribution-unit order book until Q3, and not into installed revenue until Q4 at the earliest. Most of the underperformance is timing, not story. Second, structural margin compression: cooling and power are still partly commodity equipment categories where hyperscaler buying power compresses gross margins back toward 25–35%, well below Nvidia’s 70%+. Capital is asking whether the volume win is worth the margin compression. Carrier’s composite score of 64 with insider buying is the early anomaly — the one cooling name where internal buyers are stepping in ahead of the broader market, suggesting the answer for at least one of these four names is "yes, eventually."
The Three Cold Layers: 5-Day and 20-Day Drawdowns (%)
Source: MarketDecode scanner, 5-day and 20-day returns ending 2026-05-22
Worst in stack
VRT −12.8%Vertiv 5-day — deepest of any AI infra name
Month-long bleed
ETN −10% (20d)Eaton sold for a month while everything above it rallied
The anomaly
CARR composite 64Insider buying — first cold-layer name with internal conviction
Chapter 06
The Catalyst Chain: What Grades the Cold Layers
Three dated catalysts grade the cold-layer thesis in the next 14 days. Marvell May 27, Dell May 28, PCE May 28. By mid-June we know whether the gap closes or the cold layers are correctly cheap.
The stack map is a hypothesis. The next two weeks grade it. Marvell reports Q1 FY27 on Tuesday May 27 after the close. Marvell is the next layer downstream from Nvidia in networking silicon and custom AI ASICs. A confirmation of accelerating networking orders — anything close to the Nvidia networking +199% growth read-through — propagates immediately into the optical layer (LITE, COHR, AAOI), where Friday’s sharp single-session reversal already suggests the market is positioning for that confirmation. Dell reports Q1 FY27 on Wednesday May 28 after the close. Dell is the server layer. A clean print there confirms or denies the volume-versus-margin question — specifically, whether the AI server unit growth comes with maintained gross margin or with the compression that capital is currently pricing into power and cooling. On the same Wednesday morning, the Bureau of Economic Analysis releases April PCE inflation data; PCE matters for AI infrastructure because power, cooling, electrical, and grid contractors are long-duration physical-build capex, and the rate at which their forward revenue is discounted depends on the inflation print and the 10-year yield response. The macro layer and the micro layer collide on the same day. By Saturday June 14 — the next Deep Decode — the catalyst chain has resolved enough that the cold layers are either repricing or confirmed structurally cheap. The article you are reading is the map. The next two prints are the test.
Composite Scores Across the Stack (2026-05-22)
Source: MarketDecode composite scoring, 2026-05-22
May 27 catalyst
Marvell Q1Confirms networking demand — lifts optical layer
May 28 catalyst
Dell Q1 + PCEServer margin test + rate path for physical-layer discounting
Next grade
Sat June 14Deep Decode follow-up grades the cold-layer thesis
Resolution window — 1 month
What would confirm or invalidate this read
Confirmation
Marvell’s May 27 print shows networking and custom-silicon revenue growth above 80% year-over-year (read-through to optical and downstream layers); Dell’s May 28 print shows AI server gross margin within 100 bps of consensus (read-through to whether physical-layer margin compression is structural or cyclical); and the power/cooling layer (VRT, ETN, TT) recovers at least 30% of its 5-day drawdown within 10 trading days. If two of three confirm, the stack thesis is intact and the cold layers are mispriced.
Invalidation
Marvell misses on networking revenue OR Dell’s AI server gross margin compresses by more than 200 bps versus the prior quarter, AND the power/cooling layer extends its drawdown past −20% over the next 10 sessions. That combination would say the AI buildout is real but the physical-layer margin profile is structurally lower — the cold layers stay cheap and the volume goes to a different set of beneficiaries entirely.