800 VDC and liquid cooling are becoming one architecture
The systems remain physically separate. The design decision no longer is.
Rack density is becoming the common boundary condition for both power and thermal architecture. As AI rack power rises through 2024-2028, higher current pushes the power side toward 800 VDC distribution (fewer conversion stages, thinner busways, smaller cable), and higher heat flux pushes the thermal side toward direct-to-chip liquid cooling (air becomes uneconomic). The two systems remain physically separate. The design decision no longer is.
NVIDIA GB200 NVL72 is roughly 120 kW per rack. GB300 NVL72 is specified up to 142 kW with fully liquid-cooled tray/rack architecture, tray- and rack-level leak sensors, and Infra Controller / BMS integration that can trigger both electrical isolation and coolant isolation on severe leak events. Vertiv and NVIDIA are jointly developing 800 VDC platform architecture aimed at 1 MW-plus IT racks. That trajectory is what makes the coupling worth naming.
The DTC architecture, cooling-chain engineering and leak-detection specifics are covered in the dedicated cooling deep dives: The direct-to-chip cooling stack and the leak-detection deep dive. This essay focuses on the coupling with 800 VDC; the cooling architecture itself sits in a separate deep-dive.
Part 8/13 in the DC-DC Transition series. Independent analysis; no advisory conflict on any named party.
The DC-DC Transition · thirteen essays for data centre power architects
- 1/13. The real reason data centres are going DC
- 2/13. Two architectures wearing the same name
- 3/13. The architecture map
- 4/13. Arc behaviour + insurance
- 5/13. Grounding + ground-fault protection
- 6/13. Battery integration at 800 VDC
- 7/13. Retrofit vs greenfield
- 8/13. 800 VDC and liquid cooling coupling (you are here)
- 9/13. Power quality + grid interaction
- 10/13. Standards: OCP, IEC, NEC, IEEE
- 11/13. Vendor economics: who wins the transition
- 12/13. Commissioning, skills, operational readiness
- 13/13. The ten-year view
01 · Rack density is the common boundary condition
Rack density has been rising through 2022-2028 at rates prior data-centre planning assumptions did not accommodate. Enterprise racks in 2022 drew 5-15 kW. Modern hyperscaler AI training racks in 2025-2026 draw 60-130 kW, with GB200 NVL72 at roughly 120 kW/rack and GB300 NVL72 at up to 142 kW per NVIDIA reference designs. Purpose-built AI training racks announced for 2027-2028 target 400 kW to 1 MW per rack. That trajectory forces both the power delivery decision (see Part 2/13 on the two 800 VDC architectures) and the cooling architecture decision.
02 · Three ways in which 800 VDC and liquid cooling couple
"Coupled" is doing work here. There are three distinct couplings, and it is useful to keep them separate.
- Density coupling. Higher rack power creates both higher electrical current (which motivates 800 VDC to reduce conductor cost and conversion loss) and higher heat flux (which pushes past the air-cooling economic ceiling). The same trajectory drives both decisions.
- Interface coupling. The two systems remain physically separate but share the rack envelope, CDU location, busway and piping routing, service clearances and isolation zones. Boundary-condition disagreements between the power and cooling contracts surface at commissioning.
- Failure coupling. A failure on either side can require a protection response on the other. Cooling into electrical: a coolant leak that reaches an energised conductor, a coolant-loss event that starves rack thermal capacity while the electrical load is still drawing full power, a rising loop conductivity reading arriving alongside a falling insulation-resistance reading. Electrical into cooling: a fast load transient that drops rack draw from full to idle or reverses it in a few milliseconds, faster than variable-speed pumps and CDU flow controls can respond. Each is a single event that needs a coordinated response across two nominally separate systems.
- Control coupling. The BMS, insulation monitoring device, arc-fault detection layer and CDU control logic increasingly need coordinated event handling with defined latency and authority. The vendor moat and the standards agenda concentrate here.
Coordinated isolation resolves a severe mismatch in response times between electrical protection (fast) and hydraulic actuation (slower):
- Sub-millisecond electrical protection (roughly <1 to 10 ms). Solid-state circuit breakers or fast pyrofuses at the tray or power-entry module trip on overcurrent or ground-fault differential thresholds (around 30 mA differential DC current is typical for DC-GFCI designs, though exact thresholds are set per the protection standard) before a liquid film can sustain a thermal-runaway arc.
- Rack-level firmware interlock (roughly 10 to 100 ms). Optical point sensors and resistive rope sensors trigger an emergency power-off signal via the tray Baseboard Management Controller or an OCP ORV3 Rack Management Controller, shedding the 800 VDC bus feed before fluid accumulates across creepage distances.
- Hydraulic containment and depressurisation (roughly 500 ms to 3 s). The rack Infra Controller commands CDU solenoid shut-off valves. Higher-end CDUs initiate a brief reverse-suction depressurisation on the secondary loop to prevent coolant head pressure from continuing to pump fluid into the energised volume.
If the electrical trip lags the fluid ingress by more than roughly 100 to 200 ms, the event can escalate from a cooling maintenance issue to a sustained DC arc across the energised volume.
Blackwell / Rubin AI workloads produce rack-level load transients of 20 to 100 percent of rated draw within milliseconds (speculative decoding, dynamic inference, synchronised training steps). Local rack capacitance and the fast voltage regulators inside each tray handle the electrical side.
The facility water loop cannot match those timescales. CDU pump control loops, thermostatic mixing valves and facility chilled-water setpoints operate on timescales of seconds, with millisecond response unavailable. On a fast drop from full draw to idle, the cold plate briefly over-cools (wasteful, not damaging). On a fast rise, the cold-plate microchannels can undergo localised cavitation, dry-out or coolant flash if pump ramp lags heat flux.
Modern CDUs handle this with feedforward from the rack power reading and reserve capacity in the secondary loop. CDUs sized on steady-state assumptions cannot. The design gate is the same as for the leak case: the electrical control layer has to hand a load-forecast signal to the cooling control layer, and both loops have to be provisioned for the transient profile.
The third coupling is the interesting one. The third and fourth couplings are the interesting ones. The vendor moat, the operational risk and the standards agenda concentrate around them.
03 · What does not converge
Before going further, it is worth naming what the coupling does not mean. The 800 VDC power circuit and the coolant circuit remain physically independent. The CDU does not become part of the DC power path. 800 VDC does not require dielectric coolant. Cold plates handle coolant, not current. The coupling occurs in the design boundary conditions, the protection logic and the operational response, not in the energy paths.
Where the two architectures meet is the failure response. A thermal failure in the cooling loop can require an immediate protective action in the electrical loop, and increasingly vice versa. NVIDIA's GB200/GB300 NVL72 platform already implements this pattern (see the case study below): severe leakage triggers coordinated electrical and coolant isolation through the BMS. The two systems remain physically separate; the failure response becomes a single coordinated action.
04 · The air-cooling ceiling
Air cooling, where computer room air handlers move chilled air through the raised floor and across the front of racks, has been the standard data centre cooling architecture for four decades. It can remain economically viable into the tens-of-kW/rack range under favourable containment and climate assumptions. Its facility and airflow requirements become increasingly punitive as density rises. The precise crossover with liquid cooling is site- and architecture-dependent, and depends on climate, chilled-water supply temperature, economisation regime, containment, fan configuration, redundancy tier and how PUE is measured.
The physics behind the ceiling is straightforward. Air has low specific heat compared to water or dielectric coolant. Removing a large heat load with air requires large volumetric flow rates, which require substantial fan power, which itself adds to the heat load.
Above roughly 60 kW per rack the air-handling infrastructure fights to move enough mass flow of air to remove the heat, and PUE degrades. Above roughly 100 kW per rack, air cooling becomes impractical at any reasonable PUE for current-generation AI hardware. That is the density at which DTC becomes the practical reference architecture for NVIDIA-class AI racks (see Figure 8.2).
05 · Direct-to-chip liquid cooling: the practical reference architecture above ~100 kW/rack
Direct-to-chip liquid cooling (also called cold-plate liquid cooling) routes coolant through cold plates mounted directly on the GPU, HBM and CPU packages. The coolant takes on heat at the source and moves it to a rack-level or row-level heat exchanger, which transfers the heat to a facility water loop that ultimately dumps it to ambient via cooling towers, chillers or heat-reuse systems.
The economics work above roughly 100 kW per rack because the coolant moves orders-of-magnitude more heat per unit volume than air. The engineering ask is substantial: coolant chemistry, leak detection, cold-plate integration with the silicon package, and pump-loop redundancy. Each is a known engineering task. What has changed in 2024-2026 is the vendor set that ships integrated solutions.
DTC is currently the more mature reference architecture for mainstream NVIDIA-class rack-scale deployments (GB200, GB300 and the emerging Rubin platforms are DTC in the reference designs). Immersion cooling remains a credible alternative where extreme density, heat-rejection constraints, service model and fluid economics justify the different operating model. Immersion offers higher heat-transfer efficiency and lower facility PUE at very high density, and simpler in-rack electrical interfaces because there is no need for cold-plate integration with each package. Immersion constraints: coolant chemistry qualification with silicon vendors is still catching up, servicing requires draining the tank, and the trained-technician base is smaller than for DTC. Which architecture wins the 400 kW-plus band stays unsettled.
06 · Coolant chemistry and DC electrical isolation
Coolant chemistry choice at 800 VDC interacts with the electrical protection design. The primary coolants in use in 2026 are water-based and water-glycol mixtures (common in direct-to-chip), engineered dielectric fluids (the legacy 3M Novec category is now under supply-chain transition after 3M's exit from PFAS manufacturing at end-2025; successor chemistries from Solvay, Chemours and others are qualifying with silicon and cooling vendors), and specialty low-conductivity fluids designed for compatibility with high-voltage DC operations.
Water-based and glycol-based coolants carry different conductivity profiles from an electrical-conductivity perspective. Conductivity depends on formulation, glycol concentration, additives, contamination and fluid condition. OCP's current cooling-fluid work explicitly distinguishes water/additive, glycol-based and dielectric approaches as three coexisting families, and notes that glycol formulations have different conductivity characteristics depending on glycol content. Commercially available PG25-class DLC fluids are marketed as low-conductivity by design. That makes conductivity monitoring and fluid-quality management part of the electrical-system risk architecture; a simple "conductive versus dielectric" binary choice stays weaker.
The 800 VDC dielectric reality: why 54 V assumptions fail
At standard 54 VDC rack distribution, a minor propylene-glycol-water weep produces milliamp-level leakage, minor galvanic corrosion or a localised logic-rail shutoff. These are routine facility-maintenance events. At 800 VDC nominal (approximately ±400 VDC to ground in midpoint-grounded designs), the same fluid contact triggers a different failure sequence:
The escalation depends on both fluid chemistry and busbar geometry. Coolant conductivity in PG25 / water loops typically starts under 5 μS/cm on fresh fill and climbs past 50 to 100 μS/cm over service life, driven by metal-ion leaching from wetted copper, aluminium and brass surfaces, and by inhibitor decomposition. At elevated conductivity, a thin-film meniscus of leaked coolant can bridge standard 800 VDC busbar creepage distances (roughly 6.3 to 10 mm per IEC 60664-1 for common pollution degrees and material groups). Surface tracking initiates from that conductive film. From there, the three-step escalation runs:
- Electrolytic flash-boiling. A fluid bridge across adjacent busway conductors or connector pins, at coolant conductivity in the tens of μS/cm range, conducts substantial resistive current. Ohmic heating vaporises the coolant within milliseconds.
- Steam ionisation and dielectric collapse. The expanding vapour bubble expels ionised droplets, which lower the air-gap breakdown voltage below the Paschen-curve minimum for the geometry. For the mathematical treatment of how liquid mist reduces breakdown voltage and changes arcing energy (I²t), see Part 4/13 · Arc behaviour and insurance.
- Unquenched DC arcing. AC systems get an assist from natural voltage zero-crossings 100 or 120 times per second, which give an arc a natural extinction point. An 800 VDC arc self-sustains through its own plasma column until upstream solid-state or hybrid switchgear forces clearance.
Coolant-loop deionisation becomes an electrical safety specification, not only a fluid-cleanliness housekeeping task. CDUs serving 800 VDC environments run continuous resin-bed ion-exchange bypass cartridges and real-time in-line conductivity monitoring. Trip thresholds are set jointly with the facility's insulation monitoring device: low-single-digit μS/cm for an alert and mid-teens μS/cm for a ground-integrity warning would be a typical author-scenario range; specific numbers depend on the IMD design, coolant chemistry and midpoint-grounding architecture. A rising conductivity trend on the coolant loop and a falling insulation-resistance reading on the DC bus are one event and belong on the same control platform. The interaction between falling coolant resistance and midpoint-grounded architectures (HRG vs IT systems) is covered in Part 5/13 · Grounding and DC ground-fault protection.
07 · Spatial segregation: the rack envelope contract
Co-locating 800 VDC busbars and high-flow liquid manifolds inside a 600 mm or 800 mm rack frame imposes physical layout constraints that operate at the same design authority as the electrical protection architecture:
- Opposition placement. Vertical 800 VDC power backplanes and busways route on the rear centre or opposite corner from the coolant supply and return manifolds. Liquid does not route directly over uninsulated high-voltage copper.
- Drip-plane isolation. Every blind-mate quick-disconnect (QD) plane (per OCP Open Rack V3 specifications) has continuous drip trays underneath, with integrated fluid-collection channels and resistive leak-sense cables.
- No fluid lines over horizontal conversion stages. Power supply units and 800 V-to-48 V intermediate bus converters carry drip-shielded top enclosures at roughly IP54 equivalence. Condensation, micro-weep and disconnect splatter are physically diverted away from PCB creepage paths.
- Dry-break quick-disconnects. Quick disconnects meet non-spill dry-break vendor specifications (approximately 0.05 ml or less spillage per cycle from mainstream OEM cold-plate QDs). Standard industrial compression fittings that require bleed-down or catch-buckets during service are unacceptable inside an energised 800 VDC frame.
These are physical-design constraints, and they should be enforced at the same review gate as arc-fault detection, insulation monitoring and ground-fault protection. A layout failure to segregate propagates into an electrical protection event.
08 · The leak event
A coolant leak becomes a potentially electrical event when leaked fluid reaches energised conductors. At 800 VDC, the higher system voltage and available fault energy make the speed and authority of detection and isolation part of the protection architecture. The severity of any given leak depends on where the coolant lands, coolant chemistry, insulation condition, protection response, isolation architecture, proximity to energised conductors and leakage volume. It is a variable-consequence event whose worst-case handling should be assumed at design time.
09 · Case study: the GB300 NVL72 architecture
The concrete case for the coupling thesis is already deployed in the NVIDIA reference platform. GB300 NVL72 is the current 142 kW/rack, fully liquid-cooled AI training platform, and its documentation shows exactly the architecture this essay is arguing for.
The reference platform shipping in 2026 makes the abstract "coupling" thesis concrete. Per NVIDIA reference architecture documentation:
- The rack is specified up to 142 kW across compute + switch trays.
- It is fully liquid-cooled with tray-level (cold plate, inner manifold) and rack-level leak sensing.
- Facility-level leak detection integrates with the tray and rack sensors.
- The NVIDIA Infra Controller integrates with the facility BMS.
- On severe leak events, the controller can request the BMS to authorise both electrical isolation of the rack and coolant isolation as one coordinated response.
The rack keeps a power system and a cooling system as two physically separate loops, with a single control layer holding coordinated authority over both. That is the coupling.
10 · The vendor value chain
AI density is expanding the addressable value chain around power, thermal management and fluid services. Eaton's own announcement of the Boyd Thermal completion describes the combination as an "integrated grid-to-chip solution" for data centres, which is a useful primary-source frame for the broader direction. Power equipment, cooling hardware, fluid management and commissioning services now belong to a single project scope from the operator's point of view, even when the components come from separate vendors. That creates incentives for the major infrastructure vendors to assemble those capabilities under one platform, and the M&A activity through 2024-2026 reflects that.
Three named transactions illustrate the pattern.
Vertiv acquired PurgeRite. announced 3 November 2025, completed 4 December 2025, approximately $1.0B at closing plus up to $250M contingent (per Vertiv press release). PurgeRite is principally a fluid-system flushing, purging, filtration and commissioning services company. PurgeRite is adjacent to the cooling-hardware value chain; it sits outside the hardware itself. The acquisition adds a commissioning and fluid-management capability alongside Vertiv's existing cooling and power portfolio.
Eaton signed to acquire Boyd Thermal. announced November 2025, completed 12 March 2026, approximately $9.5B (per Eaton company press). Boyd is a cooling-hardware and thermal-manufacturing scale asset that complements Eaton's power side. This is a hardware-plus-hardware combination.
Schneider Electric acquired Motivair. completed early 2025 for approximately $850M for a 75% controlling interest, with the remaining 25% expected in 2028 (per Schneider Electric disclosures). Motivair's liquid cooling equipment complements Schneider's power portfolio. This is also a hardware-plus-hardware combination.
All three transactions belong to the same broader pattern of value-chain assembly around AI infrastructure. It is worth being careful about pushing the interpretation further than that. PurgeRite is a services capability; Boyd and Motivair are cooling-hardware manufacturers. They serve different parts of the same project scope. What has changed is the boundary of the buyer's project, not that the three deals reduce to one pattern.
11 · Control authority, not just telemetry
The convergence this essay describes is operational, not physical. The 800 VDC power loop and cooling loop remain distinct systems, but a thermal failure can now require a coordinated electrical and coolant response. Merging two telemetry streams into one dashboard is a modest gain in that direction. What matters at high density is control authority: a single controller that can authorise both electrical and coolant isolation on a shared failure event, with defined latency and a defined policy. The NVIDIA GB300 Infra Controller integrated with the facility BMS is the current reference for that pattern.
Two different control domains today: electrical (voltage, current, insulation resistance, protection, BMS, breakers) and thermal (flow, temperature, pressure, conductivity, leak detection, valves). The emerging architecture is a coordinated event layer that owns both. A rising conductivity reading on the secondary coolant loop and a falling insulation-resistance reading on the DC bus arrive on the same platform within seconds and are treated as one event, with pre-authorised isolation actions on both sides.
The operator has to pre-decide who acts when the IMD reads a falling insulation-resistance trend from safe steady-state (typically several thousand Ω/V) through an alert threshold (around 500 Ω/V as commonly configured under IEC 61557-8) toward a trip threshold (around 100 Ω/V). Two design choices are available. The IMD can trip the upstream 800 VDC solid-state circuit breaker (SSCB) autonomously. Or it can hand off to the Rack Management Controller or BMC for a graceful GPU workload migration onto a redundant rack before shedding.
The answer belongs in the design specification and the runbook alike. A silent runbook here produces a shared-failure event where two systems try to act on the same signal with different latencies and cancel each other.
The corollary is that the interlock has to be deterministic. Standard building-management protocols (Modbus, BACnet) operate on polling cycles of seconds, which is orders of magnitude too slow for a fault-clearing interlock. Modern designs route the critical trip loop through hardwired dry-contact EPO wiring or a low-latency deterministic path (PCIe sideband or low-latency Redfish over a dedicated management fabric), with BMS / EMS receiving downstream event notification for logging and correlation.
12 · What the operator should do
The coupling implies a design sequence that puts rack-scale physics first and vendor procurement last:
- Compute architecture. Which platform generation (GB200, GB300, Rubin, custom silicon), which precision profile, which workload mix, which upgrade cadence.
- Target rack density. The output of decision 1 sets the density band, which determines whether 800 VDC is warranted and whether liquid cooling is required.
- Thermal envelope. Air, DTC or immersion; CDU location; facility water tie-in temperature; heat rejection strategy.
- Power architecture. 800 VDC topology (sidecar vs native, see Part 2/13); protection architecture (Essay 4); grounding (Essay 5); battery integration (Essay 6).
- Protection and isolation architecture. How electrical and coolant isolation coordinate on a shared failure event. The operational chain is detection latency, then localisation, then isolation authority. Those three variables in that order determine whether a leak becomes a maintenance ticket or a protection event on the 800 VDC bus. NVIDIA's severe-leak workflow (see §09) is the working reference for what "isolation authority" means: the Infra Controller can request the BMS to actuate both power and coolant isolation, not merely raise an alarm.
- Facility distribution. Perimeter, building, hall and row-level distribution laid out consistent with the power and cooling architectures.
- Vendor procurement. Now the operator can specify RFPs consistent with the above.
The rule that follows is the sentence a data-centre architect should remember: do not procure power and cooling independently against different rack-density assumptions.
The commissioning risk: Division 23 versus Division 26
The design sequence above only holds if the commercial contract structure supports it. Traditional CSI MasterFormat splits mechanical (Division 23) and electrical (Division 26) into separate contract packages with distinct scope, schedule, warranty and change-order authority. On a conventional data centre, that separation is workable because the boundary is a well-understood conduit-versus-piping wall.
On an 800 VDC + DTC rack with an integrated busway plus quick-disconnect coolant manifold plus shared drip tray, the boundary walks straight through the same physical hardware. When a QD flange weeps onto an 800 VDC busbar and trips an SSCB, whose warranty is compromised: the mechanical contractor who sourced the QD, the electrical contractor who sourced the busway, or the rack integrator who married them?
The operator discipline is to specify the Division of Responsibility at the same design gate as the protection and isolation architecture (item 5 above), so that the question is resolved before commissioning begins. Three options: single-contract EPC delivery of the rack perimeter; a written interface matrix between Division 23 and Division 26 with a joint-hold clause on the shared hardware; or a rack integrator with a single-throat-to-choke warranty across both sides. The choice is a commercial one, but it has to be made before the RFP goes out.
For an existing facility, the cooling-readiness question flips the retrofit-versus-greenfield decision covered in Part 7. A facility that cannot be re-piped for liquid cooling falls outside the candidate set for the density band where 800 VDC becomes relevant.
Vendors that ship integrated power and cooling with a shared control layer (Vertiv, Eaton, Schneider under the current M&A footprint) can internalise more of this interface engineering as a project deliverable. Vendors that ship only one side (CoolIT, JetCool, Submer, LiquidStack on the cooling side; AC UPS-only players on the power side) can still be part of the answer, but the operator shoulders the interface engineering.
- ← PREVIOUS · Essay 7 · Retrofit vs greenfield economics
- → NEXT · Essay 9 · Power quality and grid interaction
- FRAMEWORK MAP · The AI Power Chain framework
Frequently asked
Does 800 VDC require liquid cooling?
No. 800 VDC and liquid cooling solve different constraints. 800 VDC reduces current and conversion / distribution burden at high electrical density; liquid cooling removes heat from increasingly dense compute. They become coupled because the same rack-density trajectory drives both decisions. A facility can deploy 800 VDC at densities where air cooling remains viable, and liquid cooling can be deployed without 800 VDC.
Does the liquid coolant need to be dielectric for 800 VDC?
No. Many direct-to-chip systems use water-based or glycol-based fluids. Their electrical characteristics depend on formulation, additives, contamination and fluid condition, not on any inherent property of water or glycol. The electrical design controls the consequences of a fluid failure through physical separation, insulation-monitoring architecture, leak detection and coordinated isolation. OCP's current cooling-fluid work explicitly treats water / additive, glycol-based and dielectric approaches as three coexisting families with different trade-offs.
Why isn't 800 VDC by itself enough to trigger liquid cooling?
It is not. What couples them is the target rack density: the density that motivates 800 VDC power distribution is also the density that pushes air cooling past its economic limit. The operator can specify 800 VDC without liquid cooling if the target density stays below the air-cooling ceiling. And liquid cooling can be deployed on lower-voltage racks. The two decisions become one design decision at the density where both individually become attractive.
Where does the operator's cooling decision belong in the architecture stack?
At perimeter and building-distribution architecture, alongside the power topology. Cooling as a rack-level procurement decision late in the design cycle produces inconsistent boundary conditions with the electrical protection and grounding architecture from Essays 4 and 5.
What breaks first at commissioning when power and cooling are procured separately?
Interface documentation. The rack-input specification, CDU location, facility-water tie-in, cable tray routing and coolant-conductivity thresholds form boundary conditions between the power and cooling contracts. When those boundaries were negotiated in separate procurements against different density assumptions, they can end up inconsistent with the arc-fault and grounding design. That inconsistency only becomes visible during commissioning testing.
What does control authority mean in this context?
A single controller that can authorise both electrical and coolant isolation on a shared failure event. Not a data plane that merges telemetry from two systems after the fact, but a control layer that owns the decision on which isolation actions to take, with what latency and against what policy. The NVIDIA GB300 Infra Controller + BMS integration is the current reference for this pattern.
When does immersion make sense over DTC?
DTC is currently the more mature reference architecture for mainstream NVIDIA-class rack-scale deployments. Immersion remains a credible alternative where extreme rack-density, heat-rejection constraints, service model and fluid economics justify the different operating model. Immersion advantages: higher heat-transfer efficiency, lower facility PUE at very high density, simpler in-rack electrical interfaces. Immersion constraints: coolant chemistry qualification with silicon vendors is still catching up, servicing requires draining, and the trained-technician base is smaller than for DTC.
Which vendor category has the most to lose from consolidation?
Cooling-only equipment suppliers (CoolIT, JetCool, Submer, LiquidStack) and power-only suppliers on the AC UPS side, unless they build a credible integration protocol with the emerging control layer. Fluid-management and commissioning specialists (like PurgeRite before Vertiv) carry the same risk from a different direction: their services are still needed, but the buyer prefers to procure them from an integrated platform, and a separate contract fragments the interface.