AI’s Physical Limits: Why Power, Heat, and Distance Matter
The AI trade is no longer just about chips. It is now bounded by three physical limits — power, heat, and distance — and the eight layers of companies that solve them are the next dollar of AI capex.
Heat dissipation per GB200 NVL72 rack
132 kW
▲ +16 16× a traditional server rack — ~100 average U.S. households of heat in one rack
Nvidia just confirmed the largest AI capex cycle on record: $81.6 billion in a single quarter, with Data Center networking up 199% year over year. The same week, Google and Blackstone committed $5 billion to bring 500 megawatts of new AI capacity online by 2027. Reuters Breakingviews projects Big Tech will spend roughly $630 billion on data centers and AI chips in 2026. The constraint is no longer demand. It is physical: every watt of AI compute becomes heat, every cluster becomes a power load, and every meter of distance between chips becomes a networking cost. This article maps the eight layers of the AI factory, the physical limits at each, and the public tickers that own each tier — with the first scorecard of which bottleneck binds next.
The contradiction
Demand is no longer the constraint. Nvidia confirmed it. Google and Blackstone signed a $5 billion check for 500 megawatts of new capacity. Reuters projects Big Tech will spend roughly $630 billion on AI in 2026. The constraint is physical delivery. Between an ordered GPU and a running AI training run sit cooling vendors, electrical contractors, optical-transceiver suppliers, grid generators, water permits, and physical floor space. Each of those constraints has its own quarterly clock, its own gross margin profile, and its own public tickers. Five of the eight layers of the AI stack already repriced this month — memory ran +44%, servers +37%, optics +24%. Cooling and grid generation sat flat or down. That gap is the trade: the layers that physically have to scale next are the ones the equity market has not yet bid up.
What the headline says
AI capex is unlimited
$81.6B NVDA Q1 print, $91B Q2 guide, ~$630B projected Big Tech 2026 AI spend (Reuters)
What the data says
AI capex is physically rate-limited at every layer
Memory and servers ran 30–45% in 20 days; cooling and grid sat flat or down — carrying the most binding physical constraints
Chapter 01
What Changed This Week: The News Pulse
Five dated, public-source events from the past two weeks confirmed that AI capex is real, large, and accelerating — and that the physical-infrastructure layer is now where the binding constraints live.
Five events between mid-March and last Friday redrew the AI infrastructure map. First, on May 21, Nvidia reported Q1 FY27 revenue of $81.6 billion, with Data Center revenue of $75.2 billion (+92% year over year), and Data Center networking revenue of $14.8 billion alone (+199% year over year). That last number is the cleanest signal that the AI buildout is no longer just GPUs — it is now visibly a networked-cluster business at scale. Second, on May 19, Reuters reported that Google and Blackstone are forming an AI cloud venture with an initial $5 billion equity commitment, to bring approximately 500 megawatts of new data-center capacity online by 2027. Five hundred megawatts is the energy footprint of a mid-sized city; bringing that online in 18 months is a physical-execution test, not a chip-procurement one. Third, Reuters Breakingviews on March 26 projected the largest tech firms will spend roughly $630 billion on data centers and AI chips in 2026 — but framed the bigger issue as the physical infrastructure: grid connections, permitting, transformers, cooling, and construction may limit how quickly that capital becomes live compute. Fourth, on April 6, Reuters reported investor pressure on Amazon, Microsoft, and Google over data-center water and power use, citing a market research estimate that North American data centers used nearly one trillion liters of water in 2025. That trillion-liter figure is what turns water from an ESG footnote into a permitting risk for any new AI build in the Western U.S. Fifth, Reuters Breakingviews noted that Blackwell-class and Rubin-class chips generate enough thermal density to force data centers from air cooling toward more complex direct-to-chip liquid systems — the engineering shift that creates the cooling-vendor revenue cycle this article maps in Section 3. These five events, in sequence, are the news pulse. They are why the article you are reading exists today rather than three months ago.
Recent AI Capex Commitments: This Cycle’s Dollar Scale (USD billions)
Source: NVIDIA Q1 FY27 press release 2026-05-21 (revenue, DC, networking, Q2 guide); Reuters 2026-05-19 (Google + Blackstone); Reuters Breakingviews 2026-03-26 (2026 Big Tech capex projection)
Single-quarter print
$81.6BNVDA Q1 FY27 — record; +92% YoY on the Data Center segment
Pure networking
$14.8B+199% YoY — the cleanest signal AI is now a cluster business
2026 horizon
~$630BBig Tech AI + data-center spend projection (Reuters Breakingviews)
Chapter 02
The Mechanism: A Physics-to-Profit Chain in Six Links
The dollar travels through six physical links between an ordered Nvidia chip and a paying AI customer. Each link binds at a different severity right now. The severity ranking IS the investability ranking.
There is a tendency to talk about AI as a software story with a hardware footnote. The physical reality is opposite. The AI dollar travels through a six-link mechanical chain, and every link must operate physically before the next one earns revenue. Link one: a chip is ordered. Nvidia takes the order, TSMC manufactures the die, the connectivity silicon (Astera Labs, Marvell’s retimers) gets co-loaded on the same printed circuit board, and the memory (Micron HBM) gets stacked alongside. Link two: the chip is racked. Dell, Supermicro, or a hyperscaler’s ODM assembles the chip into a server, and then assembles servers into a 132-kilowatt rack — the Blackwell GB200 NVL72 unit. Link three: the rack is powered. Eaton sells the busway and switchgear, Vertiv sells the uninterruptible power supply, and the data center’s mains connection has to support roughly 132 kilowatts continuously per rack. Link four: the heat is removed. Vertiv, Trane, Carrier, and a few specialist names supply rear-door heat exchangers, direct-to-chip liquid loops, and the chillers that send the heat to the building exterior. At 132 kilowatts per rack, air cooling alone is no longer an option — this is where the physics binds. Link five: water and electrical pass the regulator. Permits, water-rights filings, environmental review, and the utility interconnection queue. This is the link with the longest tail — 18 to 36 months in restrictive regions. Link six: the grid generates and delivers. Constellation, Vistra, GE Vernova’s gas turbines and Quanta’s transmission contractors collectively determine whether enough firm megawatts arrive at the right voltage on schedule. The chart below scores each link’s severity right now — zero is "fully provisioned, not binding," 10 is "binding hard, project schedules slipping." The shape of the chart is the shape of the trade.
Physics-to-Profit Chain: Severity Score per Link (1 = slack, 10 = binding)
Source: MarketDecode editorial framework, 2026-05-25 — severity scoring per link aggregates 90-day news velocity, supplier lead-time disclosures, hyperscaler capex earnings commentary, and regulator filings; methodology in private/methodology/physics-chain-scoring.md
Hardest binding
Water / permitsScore 9/10 — 18–36 month interconnection queues in restrictive regions
Hard binding
Heat & gridScore 8/10 each — these are the links Section 3 maps to tickers
Slack today
Chips & racksScore 2–3/10 — fully provisioned, the dollar moves through fast
Chapter 03
The Investable Map: Eight Layers of the AI Factory, Tickers Attached
The eight-layer AI infrastructure stack repriced unevenly this month. Five layers ran 10–45% in 20 days; three did not. The flat layers are not behind — they are next. The chart pairs each layer with its strongest public ticker.
The mechanism in Section 2 maps directly to a market structure: eight layers, each with at least one publicly tradable name. The 20-day return per layer is the single best one-number proxy for "how much has the market already paid for this layer’s role in the AI buildout?" High return = layer already discovered. Flat or negative = layer not yet rewarded, but still on the physical critical path. Below, each layer is anchored by its highest-conviction or largest public name and ranked by 20-day return. Compute (NVDA) is essentially flat at -0.59% — the news is fully priced into the chip-maker, even though Q1 was a record. Memory (MU) ran +44.77% in 20 days — the HBM-supply cycle is being aggressively bid as the next obvious-after-Nvidia layer. Networking silicon (MRVL) ran +24.09% and Servers (DELL) ran +37.12%, both confirming the cluster narrative. Optics (COHR) ran +16.81%, with the broader optics cohort (LITE +10%, AAOI +24%, GLW +15%) confirming the silicon-photonics theme without crowding into a single name. Then the chart breaks. Cooling (VRT) sat at +1.29%, despite carrying the heat-removal constraint with the steepest physical scaling problem (132 kW per rack rising to 200 kW with Rubin). Grid generation (CEG) is down 6.80%; Vistra down 6.24%; GE Vernova down 7.26% — the entire generation cohort sold off while utilities-as-AI-trade became briefly fashionable a month ago and then unwound. The one positive name in the bottom three layers is Quanta Services (PWR) at +12.98%, which sits in transmission and grid-construction services rather than power generation. The pattern is straightforward: the layers closest to Nvidia’s revenue recognition repriced first; the layers furthest from it — cooling, grid, contractors — did not. The unevenness is the alpha. The mechanism in Section 2 is what closes it.
AI Factory Stack — 20-Day Layer Returns and the Primary Ticker (2026-05-22)
Source: MarketDecode scanner, 20-day returns ending 2026-05-22 (scanner_universe.json); each bar uses the layer’s highest-conviction public ticker as the proxy — see methodology in private/methodology/stack-layer-tickers.md
Best repriced
MU 45%Memory — the obvious-after-Nvidia layer the market found first
Mid-stack confirmation
DELL 37%Servers — the cluster business is now visible at the assembler
Flat/down despite binding
CEG −7%Generation — the longest-tail physical constraint with the weakest tape
Chapter 04
The Bottleneck Scorecard: Severity, Timing, News Velocity, Investability
Each of the six physical bottlenecks is scored on a 0–100 composite of severity, timing, news velocity, and investability. The composite score IS the priority ranking for the next 90 days.
Section 2 showed the mechanism. Section 3 showed the market reaction. Section 4 is the synthesis — which of these bottlenecks is the highest-priority watch over the next 90 days? Six bottlenecks are scored on a 0–100 composite. Severity is the same severity from Section 2 — how hard does this bottleneck bind today? Timing is when the bottleneck resolves or expires — a 3-month bottleneck is more investable than a 24-month bottleneck because the equity repricing is faster. News velocity is the rate at which wire and trade press are covering the bottleneck — a rising news cycle is itself a signal of accelerating investor attention. Investability is whether there are clean public tickers that translate the bottleneck into a tradeable position. Below, the composite is plotted on a 0–100 range chart with three editorial bands. The orange band (50–70) is "watch this quarter." The red band (70–100) is "this is binding now and the equity has not yet caught up." Cooling and heat removal score 78 — binding right now (Section 3 confirmed the equity has not rewarded it). Grid generation scores 72 — binding but slower-to-resolve; the equity is correctly cautious. Water and permits score 88 — the most binding, but with the least investable proxy (no pure-play public name); read-through is via the contractors (PWR) and the utilities. Optical interconnect scores 65 — still binding (every cluster grows distance), but the equity has already repriced (COHR/LITE +10–17% on 20 days, Section 3). Compute scores 35 — not binding (the chip is there; the question is what runs through it). Memory scores 42 — binding earlier in the year, now the equity has caught up (MU +45%, Section 3). The ranking gives the next quarter’s priority order: cooling and water first, grid second, then optical to maintain. Compute and memory are now a hold.
Bottleneck Composite Scores — Severity + Timing + News Velocity + Investability
Source: MarketDecode editorial framework, 2026-05-25 — composite scoring combines mechanism severity (Section 2), the 20-day equity reaction (Section 3, inverted), 30-day wire-press article counts on the theme, and the existence of clean public proxies. Methodology in private/methodology/bottleneck-composite.md
Top priority
Water / permits 88Most binding, least investable — watch via PWR + utilities
Highest-conviction trade
Cooling / heat 78Binding + investable (VRT, ETN, TT, CARR) + equity not yet caught up
No longer the bottleneck
Compute 35GPUs are there — the question is what surrounds the GPU now
Chapter 05
Bull Case vs Bear Case: The Evidence Parallel
Both the bull and bear cases on the cooling/power/grid layer have specific, dated evidence. The chart shows the parallel — four bullish data points, four bearish data points, same time window, same standard of proof.
A research call without an explicit bear case is a one-side projection, and one-side projections do not survive a real catalyst window. The bull case for the under-priced bottom three layers (cooling, power, grid) rests on four data points. First, the demand is confirmed — NVDA Q1 print and Q2 guide are no longer forecasts (Section 1). Second, the physics is real — 132 kW per rack is the GB200 datasheet, not an estimate (Section 2 mechanism). Third, the capex commitment is dated and specific — $5 billion / 500 MW from Google and Blackstone signs the order book directly to grid generation, transmission, and cooling vendors (Section 1). Fourth, the 2026 horizon is $630 billion of projected Big Tech spend (Reuters), which mathematically cannot all land in the layers that already repriced — some of it has to flow to the layers that haven’t. The bear case has four equally specific data points. First, hyperscaler buying power compresses cooling and electrical gross margins to 25–35% (vs. NVDA’s 75%, Section 4 of Sunday’s heat article); even if volume scales, margin may not. Second, the order-book lag is real — cooling sits at quarter three of conversion (Sunday’s heat article, Section 3); VRT’s next print is October, five months away, so the equity has no near-term forcing function. Third, grid interconnect queues are 18–36 months in restrictive regions — some of the projected 2026 spend simply will not become live compute in 2026. Fourth, the equity sell-off in CEG/VST/GEV over the past 20 days is not a mispricing; it may be correctly pricing the slow conversion. Below, both cases are plotted on the same axis with the same evidence-quality standard.
Bull / Bear Evidence Parallel — Same Time Window, Same Standard of Proof
Source: MarketDecode evidence stack, 2026-05-25 — each bar tags the evidence type that supports it. Magnitude is the editorial confidence on that single piece of evidence (1–10). Total bullish weight: 33/40. Total bearish weight: 31/40. Spread: +2 — the call leans bullish but the bear case is real.
Strongest bull
NVDA Q1 + datasheetBoth score 9/10 — confirmed demand + confirmed physics
Strongest bear
Grid 18–36m queuesScore 9/10 — the constraint that defeats any 2026-only thesis
Spread
2 / 40Bullish lean is real but narrow — the catalysts in Section 6 do the work
Chapter 06
The MarketDecode Read — and What Would Prove Us Wrong
The thesis: the bottom three layers (cooling, power, grid) are under-priced relative to the physical constraints they own. Four dated conditions confirm or invalidate this in 30–90 days. The chart is the watchlist.
The MarketDecode read: AI demand is no longer the constraint, physical delivery is, and the equity market has rewarded the layers closest to Nvidia’s revenue line while leaving the layers furthest from it flat. That gap is the central trade of the next 90 days. The bull case (Section 5) carries narrowly. The bear case is real. The thesis grades against four dated conditions, each on its own resolution window. First, Marvell’s May 27 Q1 print: a Data Center networking print growing 80%+ year over year confirms the cluster-buildout pace is accelerating, which directly extends the order book for the cooling layer two to three quarters downstream. A Marvell miss is the first bear-case confirmation. Second, Dell’s May 28 Q1 print: AI-server gross margin within 100 basis points of consensus denies the structural margin compression thesis for the cooling layer; a worse-than-100-bps compression confirms it. Third, by Q3 2026 (utility capex disclosures, August–October): Constellation, Vistra, GE Vernova, and the regional utilities confirm or deny that grid-interconnect investment is accelerating ahead of the $630 billion 2026 projection. Acceleration confirms the layer; flat confirms the 18–36 month queue is real and the layer stays priced for slow conversion. Fourth, by Q4 2026: a single under-priced cooling or grid name closes a meaningful portion of its 30–40 percentage-point gap to memory and servers (Section 3) within 60 trading days of any of the above confirmations. That is the visible repricing. If three of four conditions resolve in the bullish direction, the thesis is confirmed and the bottom three layers reprice on a 90-day horizon. If three of four resolve in the bearish direction, the equity market is correctly pricing the bottom layers as slow-conversion, the trade is to stay in the upper layers (memory, servers, networking, optics), and the bottom three remain a research thesis rather than an investable one for another quarter. The chart below is the watchlist — each condition with its resolution window and the direction of confirmation.
What Would Prove Us Wrong — Four Conditions, Four Windows
Source: MarketDecode editorial framework, 2026-05-25 — each condition with its resolution window in months from today. Confirmation criteria detailed in the `followUp` block of this story.
Earliest grade
May 27–28Marvell + Dell prints — 48 hours to read the first two conditions
Slowest grade
Q4 2026Bottom-3 equity repricing — the actual outcome to track
Confidence threshold
3 of 4Three confirmations = thesis confirmed; otherwise the call remains research
Resolution window — 3 months
What would confirm or invalidate this read
Confirmation
Three of four conditions resolve in the bullish direction within 90 days: (1) Marvell’s May 27 Q1 FY27 print shows Data Center networking revenue growing 80%+ year over year, confirming the cluster-buildout pace is accelerating and the cooling order book extends; (2) Dell’s May 28 Q1 FY27 print holds AI-server gross margin within 100 basis points of consensus, denying the structural margin compression thesis for the bottom three layers; (3) by Q3 2026 (August–October utility capex disclosures), Constellation, Vistra, GE Vernova, and the regional utilities confirm grid-interconnect investment is accelerating ahead of the projected $630 billion 2026 Big Tech AI spend; (4) at least one bottom-three layer name (Vertiv, Eaton, Trane, Carrier, Constellation, Vistra, GE Vernova, or Quanta) closes a meaningful portion (~50%) of its 30–40 percentage-point 20-day return gap to memory and servers within 60 trading days of any of the above confirmations.
Invalidation
Three of four conditions resolve in the bearish direction: Marvell misses on Data Center networking revenue; Dell’s AI-server gross margin compresses more than 200 basis points versus the prior quarter; Q3 utility capex disclosures show grid-interconnect investment is FLAT versus 2025 (despite the projected 2026 capex acceleration); and the bottom three layer names (VRT/ETN/TT/CARR/CEG/VST/GEV) extend their 20-day flat-to-negative pattern over the following 90 days without closing the gap to memory and servers. That combination says the bottom three layers are correctly priced for slow conversion, the equity beneficiaries of the AI buildout remain upstream (compute, memory, networking, optics, servers), and the physical-delivery layer stays a research thesis rather than an investable one for another quarter.