AI is changing the compute-side cooling loop much faster than the facility-side. This is where most of the developments, tech evolution and value creation are happening.
AI training workloads pushed rack thermal density through an order of magnitude in a few years (Fig C1). Water carries vastly more heat per unit volume than air, which is why air-side flow becomes impractical as rack density rises. Air-cooled racks cap around 30 kW in practical hyperscale designs; you cannot pump enough air through the rack to keep up with a GB200 tray. Liquid direct-to-chip can operate at package heat fluxes well beyond the range practical air cooling can support. For high-density hyperscale AI, air cooling ceased to be the economically dominant architecture as accelerator and rack densities moved from H100 toward Blackwell, and hyperscaler rack roadmaps from Blackwell forward have consequently converged on liquid DTC, with Rubin running single-phase 45°C warm-water DLC (Nvidia, "Vera Rubin driving performance per watt"). The DTC stack itself splits into a compute-side loop (cold plate, TIM, manifold, QDs) and a facility-side loop (chillers, dry coolers, towers), meeting at the CDU.
Deep Dive · independent analysis · Sep 16 2026
Three lines to hold onto
The question
As AI drives thermal density higher, does economic control migrate toward the compute-side, or does it stay with the largest dollar pool?
The flip
Dollar value ≠ economic value. The largest revenue pool is not necessarily the layer with the greatest economic control.
The moat
Qualification. The layer that controls thermal performance without becoming interchangeable holds economic control, even when the dollars are small.
01Why AI made heat a system problem
Rack thermal density rose faster in 2023-2027 than in the entire previous decade. For high-density hyperscale AI, air cooling ceased to be the economically dominant architecture as accelerator and rack densities moved from H100 toward Blackwell. Nvidia's rack roadmap moves from Blackwell through Rubin (NVL72) into Rubin Ultra (NVL576), with the Kyber rack architecture (NVL144) sitting inside the Rubin Ultra generation and now scheduled to ship in 2028 per SemiAnalysis reporting. The trajectory alone is why cooling stopped being a facility utility.
Figure C1
Rack thermal density and the cooling-technology transition, 2015-2030
Rack density trajectory (author's synthesis from Nvidia disclosures + OCP references + hyperscaler capex commentary). Cooling technology bands are approximate commercial ranges rather than hard physical limits; RDHx and enhanced air designs push above their nominal ceilings in specific SKUs. Rubin (Vera Rubin NVL72) is Nvidia's current-generation rack-scale platform, with production shipments starting in 2026 (Nvidia press release); Rubin Ultra (NVL576) is the higher-density variant. Kyber is the rack architecture designed to house Rubin Ultra GPUs at NVL144 density; SemiAnalysis reporting (July 2026) notes the Kyber NVL144 rack has slipped ~12 months to 2028 due to orthogonal-backplane PCB manufacturing challenges, though the generational association remains Rubin Ultra rather than Feynman (Nvidia Developer, Vera Rubin POD).
02The architecture taxonomy · air, liquid, hybrid
Cooling architecture splits at the top level between air-based and liquid-based. Within liquid, direct-to-chip and immersion are the two production paths, each with a single-phase and a two-phase variant. Rear-door heat exchangers are the hybrid: air on the compute side, water on the facility side. The rest of the essay focuses on direct-to-chip because that is where the AI-DC roadmap lives.
Figure C2 · Flagship
The cooling architecture taxonomy · air, liquid, hybrid
Air-based caps at ~30 kW / rack (precision air) or ~80 kW with RDHx retrofits. AI rack roadmaps from Blackwell forward assume liquid direct-to-chip. Single-phase DTC is the deployed path today, including Nvidia's Rubin 45°C warm-water platform; two-phase DTC remains a future economic threshold rather than an imminent architectural transition. Immersion coexists for edge, HPC, and niche use cases. The vendor scorecards, capex math, and value-migration analysis in the rest of this essay focus on liquid DTC first; RDHx and immersion return in §07.
Terminology note. This essay uses "compute-side" (the loop touching the cold plate) and "facility-side" (the loop rejecting heat to atmosphere). HVAC and data-centre engineering sometimes call these "primary" and "secondary" but the convention flips depending on system boundary; the functional distinction is the one that matters.
Figure C3 · Master stack
The DTC stack, heat-flow oriented · compute-side → CDU → facility-side
The two loops meet at the CDU. The compute-side loop is where AI creates new engineering. The facility-side loop changes less radically: it gets larger, warmer-water designs dominate, and heat-rejection choice shifts by climate, but the fundamental architecture is unchanged. Monitoring and controls run across both. Residual air heat (author estimate: roughly 15-30% of rack power still dissipated to air in a liquid-cooled rack, varying by rack design) is why air-cooling vendors keep selling into liquid-cooled halls. At the author's 15-30% assumption, a 140 kW rack leaves roughly 20-40 kW of air-side heat - equivalent to an entire legacy air-cooled rack's total power load; liquid cooling raises rather than eliminates demand for fan-wall and in-row precision cooling density.
Figure C3b · Compute-side physical schematic
Inside the rack · CDU to GPU, with sensors, manifolds, QDs, and drip trays
The compute-side of a Blackwell-generation liquid rack. Coolant leaves the CDU (cool, 35-45°C, ASHRAE warm-water class), enters the vertical supply manifold, flows through a quick disconnect to each cold plate on each GPU, absorbs heat, exits via the return manifold (hot, 50-60°C), and returns to the CDU. Liquid junctions are typically paired with drip trays and resistance-based leak-sensing cable underneath. Isolation valves at top and bottom of each manifold let the operator take one rack offline without shutting the whole row. Ownership: cold plates and QDs are compute-side components specified by the server OEM and Nvidia's reference design; manifolds and drip trays are rack-level, integrated by the rack builder; CDUs are their own category with their own vendor race (Fig C5). Leak-sense cabling is the low-price, high-attach-rate consumable that lands under every liquid rack.
Figure C4 · Flagship analytical visual
Who owns each layer · specification, manufacture, installation, service
Ownership fragmentation is the key insight. No single vendor controls the stack; specification, manufacture, installation and service split across different actors at each layer. This is what makes the "does the CDU become the control point?" question in §04 load-bearing. The MTDC row is where the SLA boundary between tenant and landlord creates the operational disputes that will define wholesale AI-DC contracts through 2027.
04The CDU · interface or control point?
The CDU straddles the boundary of two loops. It also potentially straddles the boundary of two industries. The testable proposition is whether that boundary position makes it the architectural control point of liquid cooling, or the largest standardised interface in the loop. That is a genuinely open question. Rittal now markets liquid-to-liquid CDUs from ~150 kW rack-mount to 1 MW+ in-row; Delta has a portfolio into the multi-MW range; Schneider/Motivair, Vertiv and CoolIT span similar envelopes. Capacity itself is becoming a specification rather than a differentiator. If that continues, CDU differentiation migrates toward controls, redundancy, serviceability, integration and qualification. That is closer to a category race between systems integrators than a pure hardware race.
Firmware is where Answer 1 stays alive. Hyperscalers (Meta and Google via OCP's Open Rack v3 and cooling-environments workstream on CDU standardisation) are aggressively pushing the CDU toward a swappable pumping and heat-exchange cart with standardised form factor, manifold pitch, and telemetry (Modbus / BACnet / Redfish). If that standardisation lands, the CDU follows the server power supply into commoditisation. The defensible moat comes from proprietary predictive pump and leak firmware integrated into the server orchestrator (BMC / Redfish), or from tight packaging integration with the rack power bus. The strategic upside is highest if CDU controls become integrated into the rack-management and orchestration layer rather than remaining a standalone appliance. A spec-compliant box competes on price; an integrated controls stack competes on switching cost.
Figure C5
CDU · what it is, the three proposition answers, and the vendor scorecard
The CDU proposition remains open. Figure C11 (value migration map) tests it. Figure C18 (bull-vs-bear grid) revisits it. Right-hand column lists the eight vendors most likely to define the CDU category over 2026-2028. Nvidia's reference CDU architecture is the most consequential entry: presence in the supplier bench signals control-point intent even if not backed by direct manufacturing.
The compute-side loop is where AI creates new engineering, and its four layers behave very differently for an investor. Cold plate carries both the growth and the standardisation pressure; its rent depends on the gap between what the reference design specifies and what a specific vendor can differentiate on. QDs are trivial on a dollar chart yet catastrophic when they fail, which is why hyperscaler qualification can extend over multiple design cycles and switching is rare. TIM never appears on a market map but sets the ceiling on how much heat the cold plate can pull, giving small chemistry vendors outsized influence at the reference bench. Manifolds face the strongest commoditisation pressure. Fluid chemistry, when handled properly, becomes a lifecycle-services annuity most investors currently model at zero.
Qualification is the moat. What defends a cold plate, CDU, QD or manifold vendor over time is the qualification record around the hardware, more than the hardware itself. Getting qualified for production scale by Nvidia, the ODM, and the hyperscaler is the actual barrier. Meeting the spec on thermal performance, pressure drop, materials compatibility, leak rate, blind-mate cycles, contamination tolerance, vibration, manufacturing consistency and field reliability is table stakes; passing the requalification process is what keeps a supplier on the reference bench. This is the mechanism behind the economic-control-versus-TAM distinction: layers where qualification is deep hold economic control even when the dollars are small; layers where qualification is thin lose economic control even when the dollars are large.
Figure C6
Compute-side vendor scorecard · four layers, per-layer character
Four sub-systems, four different investment characters. TIM is a small-dollar market with disproportionate technical importance. Cold plate is contested between engineering differentiation and reference-design standardisation. My base case is that manifold commoditises first under reference-design pressure. Less semiconductor-adjacent IP, rack-integration lock-in, and low switching cost per unit. QD carries the failure consequence and the serviceability moat. Figure C7 sizes the disproportion between QD dollar value and QD failure consequence.
Figure C7
QD paradox · dollar value versus failure consequence
Quick disconnects, TIM, and leak sensors occupy the "paradox" quadrant: small dollar value per rack, catastrophic consequence on failure. Assessing them by TAM × market share understates their strategic importance. Chillers and CDUs sit in the high-dollar quadrants, but redundancy architecture (2N or N+1) reduces per-unit consequence relative to what pure dollar value would suggest. The disproportion in the top-left is where the essay's investment thesis on serviceability + qualification barriers lives.
Coolant chemistry, metallurgy, and the O&M annuity
Compute-side hardware gets treated as capex. The recurring revenue in the compute loop is lifecycle services: commissioning a new hall to the required loop-cleanliness spec, monitoring pH, conductivity, particulate and corrosion inhibitors, changing filters, responding to contamination events, managing biocide chemistry, and, when the loop degrades or the site changes hands, flushing and refilling. Nvidia's Rubin design describes a closed loop where the coolant can last years in appropriate conditions, so the aggressive 18-36 month flush cadence some vendors historically assumed does not hold universally. What holds is that the lifecycle-services layer becomes strategically important. Vertiv's acquisition of PurgeRite (mechanical flushing, purging, filtration services) is a clear recent signal that strategics view this as a durable revenue category.
Figure C7b
Fluid chemistry and lifecycle services · the compute-loop annuity
Fluid management is the compute-loop's lifecycle-services layer. The revenue is not principally periodic coolant replacement. Nvidia's Rubin closed-loop design targets a decade-scale service life under appropriate conditions. It is commissioning, filtration, chemistry management, contamination control, monitoring and lifecycle service where operators cannot tolerate degraded thermal performance or contaminated loops. Vertiv's acquisition of PurgeRite and Ecolab's acquisition of CoolIT are read correctly only if this services layer is read as durable. If OCP or Nvidia specs standardise loop metallurgy toward sealed, long-life designs, the flush revenue compresses but the monitoring and commissioning revenue expands.
Single-phase vs two-phase DTC · an economic threshold, not a physics deadline
Commercial DTC deployments today are overwhelmingly single-phase. Nvidia's Rubin platform stays on single-phase warm-water DLC at 45°C inlet, and its published rack designs describe a closed-loop coolant intended to last years without a full flush. Single-phase can keep scaling further than the two-phase proponents imply. Two-phase has a higher heat-transfer ceiling; the system trade-offs are the harder question. What matters now is when the required flow rate, pressure drop, pumping power and cold-plate footprint make single-phase uneconomic. For this analysis, ~100 W/cm² package heat-flux and ~1,500-2,000 W package power are the threshold at which single-phase economics become materially harder. Read it as an analytical marker for this essay; the industry has no single fixed limit. The actual number moves with cold-plate geometry, coolant chemistry, allowable ΔT, and package design. Two-phase becomes an investment story when that economic threshold is crossed at scale. Rubin does not force the fork. A later generation might.
When the fork does happen, the risk shape shifts alongside the physics. A single-phase water-glycol leak is a short-circuit event on contact. A two-phase dielectric leak evaporates without wetting the electronics. Absolute risk does not fall for two-phase; the pressure-containment envelope introduces its own failure mode. What changes is the accident distribution facility owners underwrite. Insurance premiums and MTTR profiles rebuild around a different fault tree.
Figure C7c
Single-phase vs two-phase DTC · the economic threshold that decides the fork
The economic threshold, not a physics deadline. Single-phase DTC absorbs heat by warming the coolant (sensible heat, Q = m · cp · ΔT). For this analysis, the ~100 W/cm² marker is the point at which flow rate and pressure drop economics become materially harder, not a hard physical wall. The actual threshold moves with cold-plate geometry, coolant chemistry, allowable ΔT, and package design. Nvidia's Rubin architecture demonstrates single-phase can keep scaling at 45°C warm-water DLC. Two-phase DTC pushes the accessible ceiling higher via latent heat (Q = m · h_fg · x), which can carry substantially more heat per unit mass of working fluid, at the price of vapour-pressure containment, a different QD architecture, and potential PFAS / fluorinated-fluid regulatory exposure. ZutaCore is the clearest two-phase DTC pure-play; JetCool extends single-phase DTC via microjet impingement; Chilldyne differentiates through a negative-pressure architecture that is leak-safe by design (see Fig C7e). Motivair and CoolIT have two-phase development programmes. The investment question is when single-phase economics stop winning at package heat-flux levels a post-Rubin generation might demand, not whether two-phase eventually arrives.
Figure C7c-i
Single-phase vs two-phase DTC · side-by-side
Dimension
Single-phase DTC
Two-phase DTC
Thermodynamic path
Sensible heat. Fluid warms in liquid state. Q = m · cp · ΔT.
Latent heat. Fluid boils at the plate and condenses in the return loop. Q = m · hfg · x.
Working fluid
Water-glycol or treated water. Mature chemistry, cheap.
Engineered dielectric with a low boiling point. Higher unit cost, PFAS scrutiny on some formulations.
Mass flow
High. Pumping power scales with heat load. Cold-plate footprint and pressure drop become the binding constraint above the economic threshold.
Low. Vapour carries multiples of the sensible-heat energy density per unit mass. Pumping load drops; vapour-management load rises.
Leak-fault envelope
Short-circuit exposure on contact. Managed by dripless couplings, drip trays, and closed loops.
Non-conductive fluid. Leaks evaporate. The failure mode moves upstream to pressure containment and vapour return.
Deployment maturity
Standard plumbing. Multi-generation Nvidia reference designs. Cost-optimised vendor stack.
Precision containment and controls. Vapour-return loops require condensation management. Vendor stack thinner and more concentrated.
Side-by-side comparison. The physics differs; the deployment call is economic. Two-phase becomes an investment story when the flow-rate and pressure-drop economics of single-phase stop winning at higher heat flux.
When does two-phase actually arrive?
The essay's frame turns the two-phase question from a physics prediction into an economic timing question. The chart below plots peak package heat flux across GPU generations against three illustrative analytical zones: an illustrative single-phase zone (below ~100 W/cm²), an illustrative transition zone (~100-150 W/cm²), and an illustrative two-phase case (~150 W/cm² and above). The zones are analytical markers for this essay; they are not an industry classification. On the author's illustrative heat-flux model, Rubin falls within the illustrative transition band. Post-Rubin could cross into the illustrative two-phase case later this decade, depending on package heat flux and how far single-phase cold-plate design continues to advance.
Figure C7d
Two-phase timing · when a mainstream package crosses the ~150 W/cm² illustrative threshold
The two-phase question reframed as an economic timing question. Package heat-flux values are the author's synthesis from Nvidia disclosures, OCP references, and industry reporting. Read them as illustrative, not sourced. What is durable: Rubin stays single-phase at 45°C warm-water DLC by Nvidia's own design point, and post-Rubin package density could cross the ~150 W/cm² illustrative threshold later this decade, depending on how far single-phase cold-plate design continues to advance. That is when ZutaCore (two-phase DTC pure-play) and the two-phase development programmes at Motivair and CoolIT become an investment story rather than a technology story. JetCool (single-phase microjet) and Chilldyne (negative-pressure) are competing routes that can push the crossing later. Three blockers can delay the crossing. First, cold-plate geometry innovation extending single-phase further. Second, PFAS regulation eliminating leading dielectric fluids before the economic threshold is crossed. Third, and often under-modelled: vapour-line routing and condenser pressure drop inside a 1U / 2U tray. Bending high-volume vapour return lines within standard rack tray envelopes while keeping dry-out risk near zero during transient load spikes is a mechanical yield risk that can slow two-phase adoption independently of the physics or the fluid chemistry.
The negative-pressure alternative
The two-phase-versus-single-phase choice is not the only fork. Chilldyne's approach runs the compute-side loop under vacuum rather than positive pressure. In a conventional positive-pressure loop, a fitting failure sprays coolant into the rack. In a negative-pressure loop, the same failure draws air in and cannot leak coolant out. That changes the leak-failure economics enough to matter to insurers and hyperscaler qualification teams.
Figure C7e
Negative-pressure architecture · what changes when the loop runs under vacuum
The negative-pressure architecture is the third distinct route to defending high-density DTC performance without moving to two-phase. Chilldyne's design keeps the compute-side loop below atmospheric pressure. A fitting failure draws air in rather than pushing coolant out, which changes the qualification calculus for hyperscalers and insurers. In the very-high-heat-flux regime it will still cede ground to two-phase; below that, it stands alongside JetCool (single-phase microjet impingement) as a competing way to push out the point at which two-phase becomes economically necessary.
Facility cooling is the biggest dollar pool in this stack, but AI changes its architecture less radically than it changes the compute-side. The base business is HVAC incumbents selling into datacentres they have served for two decades. What is changing is the supply-water temperature envelope, dry-cooling and heat-reuse designs, and the PUE-vs-WUE geographic constraint.
Figure C8
Facility-side vendors · facility revenue vs AI-specific participation
Facility cooling revenue and AI-DC-specific participation are two different axes. Trane, Johnson Controls, Carrier, Daikin are dominant in facility HVAC but their AI-DC-specific participation is much narrower than the raw revenue implies. The AI-liquid-native quadrant (top-left, top-right) is where the strategic buyers (Vertiv, Schneider, Ecolab via CoolIT, Eaton via Boyd) have positioned. PUE-vs-WUE geography now shapes vendor selection at the facility-water and heat-rejection layers as much as capex does.
Figure C8b · Facility-side loop schematic
How the coolant is cooled · CDU → facility water → heat rejection → atmosphere
The facility-side loop. Everything downstream of the CDU is what removes heat from the datacentre building and hands it to the atmosphere. Modern AI DC design uses warm supply water (32-45°C, ASHRAE warm-water class) so heat rejection can be dry-cooler or adiabatic in many climates without a mechanical chiller. Chillers become less necessary as supply-water temperatures rise, but remain in the architecture where climate, redundancy or peak-condition design requires them. Compute-side design sets the thermal duty; local water availability, permitting and climate largely determine how that duty is rejected. In this stack, the facility-side vendor base (Trane, JCI, Carrier, Daikin, Munters, Baltimore Aircoil, Modine, SPX, Evapco) remains the biggest dollar pool (Fig C11). Nvidia's 45°C Rubin envelope pushes this layer toward dry-cooling and heat-reuse designs, but the change is directional rather than architectural.
07The alternatives · RDHx and immersion
The DTC stack lands between two other liquid paths. RDHx keeps air on the compute-side and adds water only at the rack door: the retrofit and mixed-density path. Immersion replaces the compute-side loop entirely with a dielectric bath: the extreme-density path. Both coexist with DTC rather than competing head-on across the whole market, but each pulls a specific class of hyperscaler use case away from DTC. The figure below sizes what each gives up and what each buys.
Figure C9
Three architectures side-by-side · capex, operational, vendor stack
Three architectures with distinct economics and vendor stacks. DTC dominates the AI-DC roadmap. RDHx wins where retrofit or mixed-density halls make cold-plate integration impractical. Immersion has architectural elegance and extreme density headroom but faces retrofit friction, regulatory exposure on dielectric fluids, and slower reference-design adoption. All three keep the facility-side loop and heat-rejection infrastructure · the difference is at the compute-side and CDU layer.
08Monitoring, safety and controls · the sub-industry inside the stack
Whichever compute-side architecture wins (DTC, RDHx, immersion), liquid inside the IT environment creates a new failure surface. The unresolved question is whether that failure surface generates a new control layer inside the data centre, or whether it stays instrumentation.
Figure C10
Monitoring, safety and controls · the evolution ladder
Six rungs of the monitoring / safety / controls ladder. Rungs 1-2 are the sensor + zone-controller commodity business. Rungs 3-4 are where existing BMS incumbents (Schneider EcoStruxure, Siemens, Honeywell, Johnson Controls Metasys) currently sit. Rungs 5-6 are open · this is the sleeper sub-industry inside AI-DC cooling. The upside is a thermal-control platform rather than another sensor OEM.
09Who captures the value?
Figure C11
Value migration map · where each incremental AI-cooling dollar accrues
Author's central estimate of where each incremental AI-DC cooling dollar accrues in 2027-2030 hyperscale build. Facility HVAC remains the largest single slice by raw dollars. Compute-side new engineering (cold plate + TIM + CDU + controls + QD + fluid annuity) is where the fastest AI-specific growth compounds. Hyperscaler-retained share (Nvidia reference CDU + Meta/Microsoft in-house controls) is the fastest-moving line item and the one the essay's bear case (§14) most directly threatens.
10Economic control × growth · where value actually accrues
The value-migration map shows where the dollars go. This section shows why those dollars accrue unevenly. TAM × market share is a bad way to assess strategic value in the cooling stack. Economic control (qualification barrier + switching cost + failure consequence + installed base + integration dependency) is the better axis, and it correlates weakly with dollar TAM. The framework holds for the rest of the essay.
Figure C12
Economic control × growth · where each sub-system lands
Author's placement. The upper-right quadrant is where the framework identifies potential high-control × high-growth positions: QD and qualified cold-plate suppliers are the clearest current cases (concentrated supply, deep qualification, high failure consequence); CDU and monitoring move there only if the control-point hypotheses from §04 and §08 resolve in their favour; TIM has technical importance but weaker commercial control at scale. Chillers hold the biggest bubble by TAM but land in the mid-quadrant on control - they carry the volume story rather than the multiple-expansion story. Cooling towers face the greatest substitution pressure in water-stressed markets.
11The consolidation race
The 2025-2026 M&A wave puts eight strategics into the cooling stack directly. Six via full acquisition (Ecolab, Vertiv, Schneider, Eaton, Trane Technologies, Daikin Applied); one via minority investment (Johnson Controls); one via multi-generation product-line integration (Delta Electronics at rack level). The rationale is strategic repositioning rather than incremental vertical differentiation. Large HVAC and industrial-water incumbents have decided AI-specific liquid cooling is the future of the mission-critical cooling business, and the way in is to acquire the qualified hardware supplier before the design-win cycle closes.
Turnkey modular DC cooling. Sits inside Trane's Commercial HVAC Americas unit.
Ecolab
CoolIT Systems
Acquisition (~$4.75B, from KKR)
Mar 2026 signed · Jul 2 2026 closed
Cold plate + CDU + manifold
End-to-end DTC hardware layer that plugs into Ecolab's Nalco water-chemistry annuity.
Trane Technologies
LiquidStack
Acquisition (builds on 2023 minority stake)
Feb 2026 signed · Mar 2026 closed
DTC + immersion
Central-plant-to-chip capability. Positions Trane against Vertiv, Schneider and Ecolab as a full-stack thermal management strategic.
Verified against primary sources (buyer investor-relations releases, target company disclosures, independent trade coverage). Acquisition, minority investment, and commercial partnership are treated as distinct transaction types. Live list; expect additional consolidator moves through 2027.
12Capex spend by rack density
Cooling capex per rack shifts shape with density. At 40 kW, air remains competitive and DTC is overspec. At 100 kW, DTC and RDHx compete on retrofit friction. At 250 kW, DTC is the base-case economic choice and immersion enters the conversation.
Figure C15
Capex waterfall across three rack densities
Author's estimates for a fully commissioned rack including equipment, installation, and commissioning at the given power density. Real deployments vary materially with hyperscaler-specific reference designs, negotiated CDU pricing, and site geography. What the shape shows is defensible: the composition shifts toward compute-side sub-systems as density rises, but facility-side capex remains a large fixed pool because the physical heat still has to be rejected somewhere.
Interactive companion
Run your own DTC capex scenario
Change the rack density, rack count, redundancy, phase and coolant supply temperature and the CDU count, cooling capex per rack, total campus capex and compute-side vs facility-side split recompute in real time. Interpolates between the three anchor points from Figure C15 above.
The cooling-choice equation · five inputs, one architecture pick
The cooling-choice equation as five separable inputs. Thermal density (Fig C15 waterfall) drives the compute-side pick. The other four inputs (facility constraints, energy cost, space value, retrofit cost) determine the facility-side pick and the retrofit vs greenfield path. The essay's investment framing then rests on the observation that DTC's total-cost advantage arrives at a lower rack density than the standalone cooling-equipment comparison suggests, because space value and avoided hall expansion swing the equation.
13Sizing and disagreement
Where the market boundary gets drawn matters more than the headline TAM. Three commonly-used definitions give three very different numbers, with incremental AI-cooling spend being the one practitioners actually underwrite against. On top of the sizing question, five specific claims about where value lands split reasonable analysts down the middle. The bull-vs-bear grid below traces each.
Figure C16
Three TAM definitions · Narrow / Broad / Incremental AI · 2024-2030
Author's synthesis of Yole + Dell'Oro + Grand View + vendor commentary, illustrative only. Broad TAM includes CRAH, precision cooling, and all facility HVAC serving DCs generally. Narrow TAM includes only liquid-cooling equipment (CDU, cold plate, manifold, QD, RDHx units, immersion tanks). Incremental AI TAM is the additional cooling spend that AI compute drives above the baseline DC growth trajectory. That third number is the one investors care about: it is what a strategic buyer is really acquiring exposure to when they buy a cooling asset.
Illustrative 2028 sub-system revenue mix inside the Narrow TAM (~$22B, author's triangulation, not a reported industry market size). Chillers + facility equipment remain the largest slice; CDU is among the fastest-growing concentrated slices; monitoring is the platform bet with least clear leadership; QD supply is concentrated among a small number of qualified specialists (Parker Hannifin, CPC, Stäubli lead); cold plate is fragmented across ODM channels. Concentration commentary uses Herfindahl-Hirschman intuition; specific HHI values are directional.
Figure C18
Five contested claims · bull view vs bear view
All five claims are unresolved as of September 2026. Each has observable triggers within 12-24 months (see §15). Together, the five claims determine how the value-migration thesis should be underwritten - which layers deserve a premium and which do not.
14What breaks the thesis
Six bear cases that would reset the value migration story. The last one is the most consequential because it is the one hyperscalers are actively pursuing.
Figure C19
Six bear cases · what could compress the compute-side value migration
Six bear cases. Bear case 6 is the load-bearing one: Nvidia standardises the thermal stack faster than independent suppliers can differentiate it. Nvidia's own reference CDU alongside a supplier bench, and the 80+ MGX partner ecosystem around rack-scale architecture, are the leading edge. If reference designs become prescriptive (as they did in switch fabric with NVLink), cooling suppliers keep growing enormously. The economic rent moves to Nvidia and hyperscaler retained share. The bull case assumes the M&A / build cost of full cooling vertical-integration is higher than the strategic benefit relative to compute; the bear case assumes it is not.
A note on institutional friction outside the bear-case frame. Two implementation risks slow adoption without compressing the compute-side value migration, which places them alongside the six cases above rather than inside them. First, the physical mass of a fluid-cooled rack is materially higher than the air-cooled predecessor it replaces. CDU cabinets and hall-level coolant reservoirs add live load. The combined stack moves outside the envelope legacy data centre floors were designed to underwrite. Brownfield conversions that plan around plumbing and power without a structural live-load survey and permit revision run into months of engineering delay. At some sites the answer is floor reinforcement or slab replacement, which are the most under-priced items in brownfield conversion budgets. Second, servicing a DTC installation pulls in the chip vendor, the server OEM, the CDU vendor, and the coolant-chemistry provider. A poorly isolated failure lets warranty responsibility get disputed across four contracts. Neither risk compresses value migration; both slow adoption and shape which brownfield conversions get underwritten at all.
15Five triggers to watch
Bear cases become live risks only when specific things happen. Below are the observable signals over the next 12-24 months that will resolve the essay's open questions. Trigger 1 is the highest-value signal because it directly reveals where economic control lands.
Figure C20
Five triggers · what to watch and what each resolves
Five triggers with different information value. Trigger 1 has the highest information value because it directly reveals where economic control lands. Trigger 3 is the M&A signal that would confirm the monitoring-platform thesis. Trigger 4 is the geopolitical fork on Chinese supply. Trigger 5 is architecture-defining. Trigger 2 is important but slower-moving.
16The geopolitical layer · sovereign manufacturing risk
The compute-side stack is not manufactured evenly across the world. Cold plates and manifolds are more geographically concentrated in East Asian manufacturing clusters than QDs and CDUs, which have a more diversified supplier base. That asymmetry makes cold plates and manifolds the more exposed layers in a cross-strait disruption scenario. It also matters for the qualification moat: a strategic that qualifies suppliers on multiple continents holds a durable risk-pricing advantage over one that concentrates in a single geography.
Figure C21
Compute-side supply-chain concentration · manufacturing share by sub-system and geography
Manufacturing share by sub-system and geography (author's synthesis, illustrative rather than sourced). Cold plates and manifolds are more concentrated in East Asian manufacturing clusters than QDs and CDUs. That asymmetry makes cold plates and manifolds the more exposed layers in a cross-strait disruption scenario, and diversified strategics hold a durable risk-pricing advantage over concentrated ones. This is the macro overlay the essay's compute-side investment thesis requires; it is not load-bearing on the core argument.
17Decision map
Where to look, given who you are. The map crosses investor profile with sub-system layer and points toward an action.
Investor decision map. Where the framework points by investor type: for PE strategic buyers, the framework points toward monitoring + CDU tuck-ins where holes remain (Delta); for public-equity growth investors, the framework points toward diversified strategics such as Vertiv, Schneider and Ecolab as leveraged exposures to compute-side value migration; for private growth / VC, the framework points toward monitoring and thermal-control platforms as the highest-uncertainty and highest-optionality layer; for infrastructure funds, it points to the facility HVAC and fluid-services annuity rather than the compute-side.
Glossary
DTC
Direct-to-chip. Liquid coolant flows through a cold plate attached to the accelerator die.
CDU
Coolant Distribution Unit. Heat exchanger + pumps + controls that isolate the compute-side loop from the facility-side loop.
RDHx
Rear-door heat exchanger. Air-to-water heat exchanger mounted on the back of the rack; captures hot air leaving the servers.
QD
Quick disconnect. Drip-free or blind-mate coupling between rack manifold and server cold-plate loop.
TIM
Thermal interface material. Thin layer between silicon package and cold plate that determines actual heat transfer rate.
Compute-side loop
Cooling loop that touches the cold plate. Cold plate → manifold → QD → CDU.
Facility-side loop
Cooling loop that rejects heat. CDU → facility water → chiller / dry cooler / cooling tower.
PUE
Power Usage Effectiveness. Total facility power divided by IT power. Lower is better.
WUE
Water Usage Effectiveness. Water consumed per unit of IT energy. Lower is better; matters in water-stressed sites.
OCP
Open Compute Project. Industry consortium that publishes reference specifications for DC hardware including cooling.
ORv3
Open Rack v3. Current OCP rack spec including blind-mate liquid-cooling manifold references.
PFAS
Per- and polyfluoroalkyl substances. Family of chemicals including many dielectric coolants; under regulatory restriction in EU and US.
Two-phase
Cooling architecture where coolant boils inside the cold plate or tank, absorbing latent heat. Higher heat-flux capacity than single-phase.
Single-phase
Cooling architecture where coolant stays liquid throughout the loop; heat transfer is sensible (temperature-rise) only.
Dielectric
Electrically insulating fluid, used in immersion cooling and two-phase DTC to allow direct contact with electronics.
Kyber
Nvidia's NVL144 rack platform positioned for the Feynman era beyond Rubin Ultra, per Nvidia's rack-roadmap disclosures.
Blind-mate
Rack coupling design that mates liquid connections automatically when a tray slides into a rack.
MTDC
Multi-tenant data centre. Wholesale colocation facility with multiple tenants sharing common infrastructure.
W-class water cooling
ASHRAE classifications for liquid-cooled IT based on supply-water temperature. Higher W-class means warmer supply water and enables chiller-less designs.
Method and sources. Author's synthesis from Yole Développement, Dell'Oro Group, Grand View Research, Omdia, DCD reporting, OCP specifications (Open Rack v3, Open CDU), vendor press releases and technical datasheets, and hyperscaler earnings-call commentary through Q2 2026. Cooling M&A facts verified against AI infrastructure M&A tracker; specific deal values not verified in primary sources are labelled directional. Regional splits, TAM breakdowns, and sub-system revenue mixes are the author's central estimates for illustration; specific numbers vary by source. Analytical framework, not investment guidance.
