01The $50 BOM inside the $10M rack

A single leak-detection cable run inside an AI training rack costs less than dinner for two in central London. The rack it protects can be worth several million dollars, drawing 60 to 142 kW, cooled by facility water pumped through quick-disconnect fittings. If a coolant drop reaches a live busbar before the sensor catches it, the outage that follows can wipe out days of training throughput and put a multimillion-dollar tray at risk.

Rack designs are increasingly shipping with dedicated leak detection integrated at build. NVIDIA's public GB300 NVL72 reference architecture documents three levels of leak detection: node and tray sensors on the cold plate and inner manifold, rack-level sensing ropes and spot sensors, and datacentre-level ropes and spot sensors around the CDU. Rack power is documented at up to 142 kW nominal. The response chain, including the NICo controller and the three severity classes, is documented in Nvidia's Mission Control admin guide and Infra Controller documentation (see §04c).

Figure L1. The economic asymmetry is the whole essay in one chart. A leak-detection kit for an NVL72-class rack lands somewhere between $540 and $1,290 all-in, a scenario range derived from the companion calculator. The consequence of that kit failing to catch a leak before it reaches live silicon runs into the millions. Every liquid-cooling deployment writes a small cheque for very asymmetric insurance.

The open question is who captures the economics as the category scales.

The Open Compute Project's active track on rope leak sensor specifications is pushing the physical sensing interface toward greater standardisation. Qualification costs, false-alarm suppression logic, and the hardwired control path from sensor to CDU isolation remain sticky. Both trends are compatible with a single outcome: the sensor commoditises while the qualification concentrates on the control layer.

02Where the liquid actually escapes

Before you can argue about detection, you have to know where the fluid can escape from. Direct-to-chip cooling has moved most of the leak surface area onto the rack itself. Facility-loop failures still happen and still trigger alarms; the failure modes with the tightest coupling to live silicon, though, are compute-side.

An NVL72-class rack has coolant pass through several mechanical interfaces before it reaches the die. Each one can fail in a different way: quick-disconnect coupling wear, manifold pinholes, hose or fitting failures, gasket or O-ring fatigue, cold-plate braze or weld defects, CDU heat-exchanger or pump-seal failures. The detection strategy has to respect the physical geometry of the leak.

Figure L2. Not every leak carries the same consequence. Facility-side failures involve larger volumes and more instrumentation, so they are often easier to detect. Compute-side failures between the rack manifold and the cold plate are harder to localise and couple more tightly to live silicon. The closer a containment failure gets to the protected asset, the more valuable localised detection and rapid intervention become.

Figure L2b. The full topology in one map. Left-to-right, the four panels are physical hierarchy from hall down to server tray, with sensor placement, sensor count and local controller called out at each layer. Below the panels, protocol labels show the data path from each controller into the software stack: Redfish carries BMC events from the tray into NVIDIA Mission Control, and Modbus / BACnet / SNMP carry facility, pod and rack events into the BMS. Both data planes converge at the DCIM, which classifies severity and hands the event to a policy engine. The policy engine fires two actions in parallel: electrical isolation via PDU or rack breaker, and coolant isolation via CDU valve (or, in Chilldyne's architecture, vacuum inversion). Observation closes the loop through the event log and the restart procedure. Every layer in this map is a vendor competition zone; the qualified path from sensor to isolation determines which vendors keep their own economics and which get absorbed into cooling and control.

The geometry forces two design constraints. The sensing has to be distributed: a single sensor at the bottom of the rack cannot tell you that a QD has started to weep three shelves up. The sensing also has to be selective: if every stray drop of condensate triggers a rack-level shutdown, the false-alarm rate destroys the value proposition faster than actual leaks do. Vendors earn their qualification premium on false-alarm suppression.

The other subtlety, easy to miss: detection is not the same thing as localisation. A resistive sensing cable can tell you fluid is present somewhere along its 12-metre run. An addressable cable can narrow that to a metre or two. A well-instrumented CDU can tell you the loop has lost 20 mL over an hour without knowing where. Whether the sensing architecture localises the leak to a specific rack unit or just tells you something is wet has a direct effect on the isolation strategy the BMS can execute.

03How you actually detect it

A leak sensor is only useful in the geometry and control system around it.

Two operational details rarely surface in vendor decks and matter to hyperscalers. First, facility PG25/water mixes typically carry biocides and corrosion inhibitors (tolyltriazole, glycol blends) that change surface tension and residual drying times on a sensing cable. The OCP rope specification does not settle how the base sensor interacts with these additives across suppliers. Second, cable "reset" after an event, whether a cleared segment can be dried, revalidated and returned to service versus replaced spot-by-spot, is a live opex debate inside hyperscaler operations teams. Both determine whether the sensing layer behaves like a consumable or an infrastructure line.

Figure L3. Detection combines three choices: sensing physics, geometry and control. TraceTek is resistive physics × continuous cable × dedicated controller. TTK dielectric is capacitive physics × continuous cable × dedicated controller. CoolIT-embedded leak telemetry is pressure/flow physics × CDU-loop geometry × CDU-embedded control. NVIDIA GB300 tray sensing is OEM-integrated cold-plate and inner-manifold sensing × compute-tray geometry × native BMC control. A single vendor typically owns one or two axes; hyperscaler AVLs qualify by combination.

03bWhat each vendor's technology stack actually is

The §03 map places the sensing approaches on axes; L3 and L4e place vendors on the qualified path. Neither drills into the proprietary technology inside each vendor's product. Two axes matter most for how vendors compete in the AI liquid-cooling era. Localisation (identifying where within a cable run the leak has occurred). Coolant chemistry (whether the system can sense the non-conductive dielectric fluids used in direct-to-chip and immersion architectures).

Figure L3b1. Detection architecture versus signal actionability. The specialist rope vendors cluster around the middle of the x-axis with locate-and-classify capability. The CDU vendors plot on the right (operational telemetry) and move upward as their analytics mature. Chilldyne uses plumbing architecture to reach the isolate-and-control level by reducing leak consequence rather than by sensor coverage. NVIDIA's Mission Control is intentionally not plotted: it is the compute-management and response layer rather than a physical sensor vendor. It provides centralised leak-detection aggregation, power control and event response, integrating with the BMC through Redfish and with the customer BMS. Author positioning based on public product documentation; exact coordinates are indicative rather than measured.

Pure-play specialists

TTK Leak Detection (France). A leak-detection specialist that has explicitly built product for the liquid-cooling transition. The FG-NET controller aggregates addressable digital sense cables that identify leak position along a circuit. TTK's public product literature covers the FG-DLC line, a point sensor purpose-built for direct-to-chip conductive coolant, alongside the FG-OD and FG-ODP range for hydrocarbons and non-conductive solvents. Qualification against the specific dielectric coolants used in AI immersion and two-phase DTC deployments remains a per-deployment diligence question rather than an assumed capability.

Parameter Technologies (RLE + Cellwatch). Parameter is the platform created by the merger of RLE Technologies and NDSL, the parent of Cellwatch. RLE's leak-detection and environmental-monitoring products (including the SeaHawk brand) continue within the portfolio. Continuous-distance leak localisation via the LD5200 controller family, with Modbus, BACnet and SNMP for BMS and DCIM integration. Parameter is explicitly positioning its technology for high-density and liquid-cooled data centres, actively repositioning sensing toward the CDU, manifold and connection-point layer.

Chemelex (TraceTek). A long-standing incumbent cable brand. Contrary to a common assumption, current Chemelex products do localise leaks: the TTDM-128 alarm panel identifies both circuit and location, and the TTSIM-2 sensor interface module carries a built-in three-digit leak-location display. TraceTek cable supports continuous location along circuits up to about 1,500 m. Chemelex became an independent PE-backed company when Brookfield completed its $1.7bn acquisition of the business from nVent in January 2025. The acquired business was primarily electric heat tracing; TraceTek is the relevant leak-detection portfolio. Whether that corporate independence translates into sharper AI-cooling investment is an open analytical question rather than an established fact. The demonstrated position is the mature detection-plus-localisation stack.

PermAlert (PPIH). The PAL-AT AT30/40 series uses pulsed cable-radar (time-domain reflectometry) to identify leak position along a sensing cable. The PHLR probe extends into containment applications. PermAlert's public documentation confirms the system detects both conductive and non-conductive liquids with continuous localisation, which is unusual in the specialist category and directly relevant to dielectric coolant sensing in AI DC applications.

Tatsuta (Japan, TSE:5809). Established Japanese cable maker. Leak-detection products sit inside a broader cable and sensor portfolio, with particular relevance in Japanese semiconductor and industrial applications and increasing exposure to water-cooled server and data-centre applications. Chemical-resistant sensor variants extend into the chemical-storage and chemical-tank segments. The Japan and APAC positioning is a distinct commercial vantage on the AI-DC build-out.

Andel, Aqualeak, Icon Process Controls, Dorlen (Water Alert). Regional specialists that compete largely through installed base, local distribution and application coverage. UK and EMEA for Andel and Aqualeak; North America for Dorlen and Icon.

Figure L3b2. Attribute matrix for the five named specialists. Author thesis based on the public product evidence reviewed here: TTK and PermAlert are the two clearest specialist positions on the coolant-chemistry axis in the public product evidence reviewed here. Chemelex and Parameter carry continuous localisation for conventional water-glycol architectures. Tatsuta is conventional with a Japanese and APAC-weighted footprint. Actual qualification for specific AI DC dielectric coolants remains a diligence question at deal level rather than an established capability at product level.

The specialists are one of several categories. Adjacent categories bundle leak detection into environmental / DCIM platforms, into CDU controllers, or into distributed-fibre interrogators. Each has a different competitive logic.

Figure L3b4. Three adjacent categories that bundle or absorb leak detection differently from the pure-play specialists. Environmental / DCIM platforms compete on the system, wrapping cheap sensors in dashboards. CDU vendors expose operational telemetry through mainstream industrial protocols. Distributed optical vendors carry oil-and-gas heritage and only earn their slot at facility scale.

Figure L3b5. Four emerging developments that shift the aftermarket-specialist model. NVIDIA is vertically integrating the response layer: physical sensing remains multi-vendor and enters the compute platform through OEM BOM design-in, while Mission Control integrates BMC and customer BMS data and coordinates the response layer (chassis shutdown capability, node DOWN status, 10-minute timeout to restart). OCP has published the R1.0.0 Rope Leak Sensor Base Specification, moving the physical rope interface toward an open specification. Chilldyne uses plumbing architecture (vacuum) to reduce leak consequence upstream. Predictive analytics on CDU telemetry aims to catch leaks before wet-detection triggers.

Figure L3b6. The response chain from sensor fire through transport, decision and action. Detection is stage 1 only. Transport (stage 2) uses BMC Redfish for device-level sensors and BMS protocols for controller-aggregated sensors. Decision (stage 3) is where severity is classified and policy applied. Action (stage 4) is a twin track: electrical isolation via PDU or rack breaker, coolant isolation via CDU valve or interlock. Architectures such as Chilldyne attack the consequence earlier by maintaining the loop under negative pressure so a breach tends to pull air inward rather than push coolant outward. That is a different architectural intervention from a conventional valve action rather than an equivalent one. Observation closes the loop with event logs, operator notification and the restart procedure. The commercial insight is on the bottom rule: sensor value migrates rightward toward the policy engine and the twin actuators.

Figure L3b3. Four movements are visible in the technology field at once. Rope hardware is standardising under the published OCP R1.0.0 base specification. Differentiation is moving upward from the cable to the controller, BMS/DCIM integration, CDU telemetry, policy engine and automated response. NVIDIA is vertically integrating the response layer (Mission Control integrates BMC and customer BMS data, classifies severity and coordinates the response up to and including chassis shutdown) while physical sensors move into the compute platform through OEM BOM design-in supply. Some architectures reduce the leak-risk problem upstream via plumbing design rather than downstream sensing coverage. The four movements converge on the same commercial insight.

What this maps to commercially

The published OCP R1.0.0 base specification is direct evidence that the physical rope interface is moving toward an open standard. That reduces the proprietary value of the sensor layer without automatically making sensing cable, controller algorithms, false-alarm performance, connectorisation, qualification history or installation methods interchangeable. TTK's product literature covers addressable localisation, direct-to-chip conductive coolant sensing (FG-DLC), and non-conductive hydrocarbon and solvent sensing (FG-OD, FG-ODP) on a common FG-NET controller. PermAlert's PAL-AT documentation covers TDR-based continuous localisation with both conductive and non-conductive liquid detection. Parameter Technologies (RLE + Cellwatch) covers continuous-distance localisation with full BMS and DCIM protocol coverage (LD5200, Modbus, BACnet, SNMP), positioned explicitly for high-density and liquid-cooled data centres. Chemelex's demonstrated position is a mature detection-and-localisation stack (TTDM-128, TTSIM-2, continuous location on circuits up to about 1,500 m); whether the post-Brookfield corporate independence translates into sharper AI-cooling investment is an open analytical question.

The interesting investment question shifts from "who makes the best leak cable" to who controls the transition from physical event to physical intervention. Value stays with the controller intelligence, DCIM integration, CDU telemetry, policy engine and automated response layers. Two different forces squeeze the aftermarket specialist model from opposite directions. OEM integration is absorbing sensing upward into the compute platform: physical sensors procured through design-in supply into the OEM BOM, response layer owned by NVIDIA's Mission Control policy engine. Architectures such as Chilldyne reduce the amount of leakage risk that downstream sensing has to manage in the first place.

A last point on category economics. Detection capex is a rounding error in an AI campus budget: on the reference topology (§04b) the whole installed instrumentation set is roughly $1.26M for an 85 MW IT campus, or ~$14.8/kW IT, against total campus capex in the low billions. The category is strategically critical and commercially tiny at the same time. A missed leak from a $50 cable takes down a $10M rack. The moat is qualification history, installed topology, control integration and switching friction. The physical sensor may commoditise. Qualification, installed base and integration protect the control layer.

04The leak detection deployment stack

Sensor → Controller → Interlock → BMS

Sensing hardware (rope cable, spot puck, dielectric probe or optical fibre) feeds a controller that applies false-alarm logic and localises the event. The controller signals a rack CPLD or CDU interlock that executes the physical isolation. The BMS records the event and drives the operator response.

Figure L4d. CLASSIFY is where the sensor and its controller judge the severity of the event. AUTHORISE is the policy decision to trigger an interlock, shutdown or valve closure. The physical rope interface is moving toward an open specification under OCP R1.0.0, which reduces proprietary differentiation at the sensor layer without making the entire sensing stack interchangeable (controller algorithms, false-alarm performance, connectorisation, qualification history and installation method continue to differentiate). The classification logic, the authorisation path, and the BMS integration remain proprietary, requalification-expensive, and installed-base-locked. Whichever column a vendor actually controls decides the outcome: staying standalone, becoming a bundler's ingredient, or exiting the category.

Figure L4e. Named entities across the qualified path. The specialists (rows 1-3) control the left side of the chain (DETECT / LOCALISE / CLASSIFY) and stop at the AUTHORISE boundary. They contribute a dry-contact signal into the authorisation input; they do not control policy authorisation or physical isolation. The cooling-bundled vendors (rows 6-7) control AUTHORISE, ISOLATE and OBSERVE via the CDU/BMS platform and are absorbing DETECT from the specialists. NVIDIA GB300 (row 8) integrates four of the six stages within the compute and rack control architecture (Mission Control integrates BMC events, classifies severity and triggers power-shelf shutdown; the customer BMS shuts rack breakers and liquid valves). Physical coolant isolation remains downstream in the CDU and BMS domain. Sensor supply remains multi-vendor via OEM BOM design-in. The bundling model is already shipping.

04bA sample topology: how many sensors for a 100 MW AI campus

The abstract discussion of layers is useful. The concrete arithmetic is more useful. Take a 100 MW gross campus, 85 MW IT after PUE and non-IT allocations, 90 kW average per rack, twelve racks per pod, row-based CDUs at N+1 redundancy. The reference topology in this configuration looks like this.

Figure L4b. A reference 100 MW AI campus in the reference topology, with row-based CDUs at 1.25 per pod. Total detection instrumentation and localised controllers, inclusive of a 35% installation uplift, come to approximately $1.26M or $14.8/kW IT at the reference topology, with a labour variance band of roughly $14.2-16.4/kW IT depending on whether the site is greenfield or a retrofit. The four primary topology inputs (gross MW, IT load factor, kW/rack, racks/pod) drive the reference geometry; installation uplift, redundancy assumptions and coverage tier convert that geometry into installed cost. Native tray sensing is modelled as OEM baseboard telemetry. Aftermarket leak-detection SKUs do not appear at this layer. Switching the calculator to in-rack (1:1) CDU mode roughly doubles $/kW IT because the CDU count scales to the rack count.

Figure L4c. Physical topology in plan view. Dashed line around the hall perimeter is the facility-scale addressable cable, catching CHW riser and CRAH-row leaks before they migrate. Each pod carries its own overhead cable; the small dark circle shows the primary row-based CDU per pod. The reference case allocates ~1.25 CDUs per pod (99 CDUs across 79 pods) to model an N+1 redundancy pool, so about one pod in four carries a redundant unit not drawn here. Every pod uplinks Modbus/BACnet to master locator panels and the BMS/DCIM stack in the plant room. The full 100 MW campus is roughly seven of these tiles. Rack-level loops sit inside each pod rectangle; too small to draw at this scale.

Detection capex is a rounding error in an AI campus budget. At about $14.8/kW IT in the reference topology, the detection stack for an 85 MW IT campus lands around $1.26M installed. That is against a total campus capex measured in the low billions. The category is strategically critical and commercially tiny at the same time. The absolute spend is small enough that hyperscalers can push aggressively on price, unless qualification or single-source exposure gives the specialist leverage.

Pod-level cable dominates the physical BOM. The row/pod layer accounts for two-thirds of total cable metres. That layer is also the OCP standardisation target. If the pod cable becomes interchangeable under an open spec, two-thirds of the cable BOM stops being a proprietary lock.

Tray-level sensing is native to the compute platform. GB300 NVL72's reference architecture places sensors on the cold plate and inner manifold, with the signal terminating at the BMC, discovered by NICo, and forwarded to Mission Control (§04c). Nvidia has stated Mission Control on the Vera Rubin platform will provide rapid leak detection; the tray-level sensor topology has not been published for Rubin SKUs as of publication. The sensor itself remains multi-vendor through OEM BOM design-in, while the controller and response logic run inside Nvidia's platform. Specialists supply the sensor. The response chain around it stays on Nvidia's platform.

04cNvidia's documented response reference

Nvidia documents NVL72 leak detection across two products. Mission Control's 2.2 admin guide names the sensors and the Redfish paths. The Nvidia Infra Controller (NICo) documentation names the automated handling by severity.

Figure L4nv1. Cold-plate and inner-manifold sensors on every NVL72 compute tray, exposed via Mission Control 2.2 admin guide Redfish paths. NICo discovers BMC endpoints and queries the sensors. Rack rope, drip pan and CDU sensors belong to the BMS.

NICo owns tray-level sensor discovery via Redfish and severity classification. BMS owns the actuators: rack AC breakers and coolant valves. General severity means NICo shuts down leaking trays and blocks new allocations. Severe hands off to BMS for tray-level electric and liquid isolation. Critical is a direct BMS action on rack AC and coolant.

Figure L4nv2. Nvidia's three severity classes and their automated actions per the Infra Controller documentation. Mission Control adds a coarse-grained leak:large node annotation for the scheduler with a 10-minute timeout before restart-required.

05Who controls the qualified BOM?

Before the vendors matter, the specification chain matters. Five parties touch the leak-detection line on the BOM. Only one of them writes the cheque, and it is almost never the same one that writes the spec.

Figure L5. Twenty-four vendors compressed into four archetypes evaluated on Adi's essay four moat pillars: qualification history, installed topology, control integration, switching friction. Standalone specialists hold the classic moat; OEM captive emerges strongest on control integration through Nvidia Mission Control.

Figure L5. The leak-detection specialist rarely controls the specification, the vendor selection, or the approved-vendor list. It usually acts as the terminal supplier in a five-party chain, with the decisions that matter made above it. Specialists can and do influence reference designs and get named directly at qualification time; the base case, though, is that the small absolute spend keeps procurement pressure high while qualification cycles run long.

06The vendor map

The named players group by strategic position. The winner takes the largest slice of the hyperscaler qualification stage.

Figure L6. Six specialists positioned on legacy-water to AI-coolant-native and sensor-SKU to full-system axes. Only vendors with a branded standalone leak-detection SKU appear as circles. Cooling-system OEMs enter as external threat vectors, not competing products.

07The second-source problem

A hyperscaler qualifies a single vendor because switching costs kill the alternative. Then a big-enough outage or supply squeeze arrives, and the same hyperscaler decides that being single-sourced is more expensive than qualifying a second. The break-even between qualification cost and the expected cost of remaining single-sourced sets when a second qualifies.

Figure L7. A framework only. No filled-in worked example. Single-source economics beat second-source economics until the risk side outweighs the qualification side. Once a hyperscaler faces one supply squeeze, one critical-defect finding, or one price hike, the second-source cheque gets written. That is when a decade-long qualification moat starts to unwind. Whether any given hyperscaler is already in mid-flight second-source qualification is not public.

08OCP is standardising one sensor layer of single-phase DTC

The Open Compute Project has an active track on rope leak sensor standardisation, narrow in scope: a base specification for the rope sensor and its interfaces in single-phase DTC using water-based coolant. Named contributors include CoolIT, Thermal Control Technology, PermAlert, Envicool, Gredmann, and Tatsuta. The stated goal is to reduce design complexity, improve availability, and enable more vendors to participate. The sensor still differentiates. OCP's base specification pushes commoditisation into the physical interface, on purpose. Controller logic, actuation and BMS integration remain vendor terrain.

Figure L8. Unit attach rate progression and sensor topology migration. Sub-floor perimeter era to in-rack, above-rack and CDU-integrated coverage. Attach curve is authors' estimate anchored to Nvidia GB300 rack BOM inclusion and Supermicro / Dell / Vertiv public specs.

Figure L8. The specification standardises the two layers where the leak-detection specialist historically added the most differentiation. The controller retains vendor-specific IP. The BMS integration remains sticky. Isolation and intervention become the layer worth owning. The specialists know what they are signing up to when they join the OCP process, which is why so many of them are participating actively rather than resisting.

09What actually happens when the sensor trips

If OCP is right about the sensor, the response path decides who wins. That path runs from sensor tripping to coolant-valve closure. Every vendor in this category competes for it. NVIDIA's public GB300 architecture actually shows how the modern chain runs in practice: the tray-level sensor is read by the system BMC (NICo), which classifies severity and requests the BMS to perform both electrical and liquid isolation. The BMS receives the leak event and coordinates the configured electrical and liquid-isolation response, including rack power and coolant isolation where those controls are integrated. Detection happens in the compute platform. Isolation authority happens in the BMS. The qualification value is moving into the classification-to-authorisation-to-isolation path, and away from whoever sold the sensor.

Figure L9. For failure modes that can threaten live electrical hardware, mitigation has to be deterministic enough that detection, classification and isolation complete before the escape volume overwhelms drip-pan capacity. The hardwired loop cannot beat a droplet already in freefall onto a live busbar; what it can do is depressurise the loop and cap the leak volume before it becomes a flashover event. A slow weep is a different problem from an aerosol pinhole; the same architecture handles both by acting on the fastest independent path available. The hardwired mitigation path from sensor to CDU solenoid via the rack CPLD can act independently of the BMS alarm path (typically Modbus/BACnet polling loops that add seconds of latency). That independence makes the control architecture matter as much as sensing accuracy.

10Can cooling vendors absorb the category?

CDU and BMS majors already integrate spot sensing into their reference designs (Vertiv Liebert, Schneider Netbotz, CoolIT). Ecolab's July 2026 acquisition of CoolIT for ~$4.75bn signalled how far this consolidation goes: it combines CoolIT's CDUs and cold plates with Ecolab's water, chemistry and digital monitoring, creating a fluid-management platform that reaches from coolant chemistry through sensing to CDU actuation. NVIDIA's public GB300 NVL72 reference architecture goes further inside the rack, documenting leak detection at three levels (node/tray, rack, and datacentre-scale around the CDU) and BMS-controlled electrical and liquid isolation on severe events. If a compute-platform vendor is specifying sensing at every level of the physical stack and a strategic buyer like Ecolab is bundling detection into a fluid-management platform, the aftermarket detection SKU is being squeezed from two directions. Rack integrators (Foxconn, Quanta, Wiwynn) add a third pressure: they push back on non-AVL leak detection because a failure to isolate transfers warranty liability to the system integrator. Hyperscaler AVL mandates exist partly to shield the ODM from system-level liquid-damage claims, which is why the AVL is a durable gate on which vendors even get considered.

Figure L10. Era 1 belongs to standalone specialists selling into sub-floor perimeter zones. Era 2 is the transition, where detect-to-isolate collaborations (Parameter-Rotork) and OCP R1.0.0 compliance keep specialists in play. Era 3 is OEM captive dominance, when cooling OEMs (CoolIT, Motivair, Boyd) absorb sensor selection into their factory-line CDU and rack packaging. Category revenue climbs the whole way; specialist share compresses at each transition. Qualification depth slows that compression on the specialist side.

Figure L10. In Scenario A the specialist retains its slot. In B the CDU vendor absorbs the sensing. In C the compute platform itself owns the detection. All three are technically plausible; qualification decisions across the next few AI-rack generations will determine which captures the economics. The category outcome depends on architecture more than on cable volume.

11Sizing the market

The addressable spend, its growth, and its geographic split all matter for an investment decision. The numbers below come from a scenario model built on the essay's topology arithmetic, not from surveyed market research. Regional splits reflect published AI-DC capex distribution; vendor-group shares within each region are the author's structural allocations. Change any assumption and the numbers change; the shape carries more weight than the point estimate.

Figure L11c0. Every input in this chain is exposed in the companion calculator. The $260M annualised figure is not a market-research number. It falls out of four scenario assumptions: cumulative liquid-cooled AI DC capacity, IT load factor, reference-topology $/kW IT, and the deployment window. Change any of the four and the $260M moves proportionally. Read the mekko below in that context.

Figure L11c. This is a scenario-derived installed instrumentation cost. It is deliberately not a market-research size. The $260M mid-year figure is what the topology arithmetic produces when 85 GW of cumulative liquid-cooled AI DC capacity is spread across a 2026-30 deployment window at $14.8/kW IT (see L11c0). Regional splits follow published AI-DC capex distribution; vendor-group shares within each region reflect the author's structural allocation, informed by named vendor presence rather than surveyed market data. Shape carries more weight than the point estimates.

Figure L11a. The realistic addressable spend for dedicated leak-detection specialists across 2026-30 is roughly $400-500M cumulative. That is a fraction of a single hyperscaler's annual AI capex. The category is sub-scale for most sponsor-scale platforms and viable for strategic tuck-ins, or as a durable $40-90M annual-revenue specialist business. There is no unicorn hiding here.

Figure L11b. Facts of the January 2025 Brookfield acquisition of nVent Thermal Management (rebranded Chemelex). The 2.85x sales multiple is the headline arithmetic. Segment EBITDA was not broken out at close, so this datapoint alone cannot fix a multiple range for AI-DC leak detection. Treat it as one entry in a comparables set that will grow as more thermal-management platforms and leak-detection specialists change hands.

Figure L11d. Greenfield 2026 leak-detection TAM with two denominator views: JLL commissioning-year (12 GW) and Cushman under-construction pool (30 GW). Category MW split anchored to Meta Prometheus, Oracle Stargate, CoreWeave Q2'26 and Nebius Q1'26 disclosures. Liquid share by category is authors' estimate.

Figure L11h. Brownfield leak-detection TAM waterfall on 103 GW installed base. Four pools: sensor + panel replacement, air-to-liquid retrofit, DCIM + software upgrade, services + monitoring contracts. Pure leak-detection scope; broader cooling services excluded.

Figure L11f. 2024-2030 TAM evolution. Greenfield flow compounds at ~36% CAGR; brownfield stock at ~18%. Combined ~22%. GW growth (Panel A) drives $ growth (Panel B). Base case with bull and insurance-mandate scenarios shown for reference.

Figure L11e. Installed-system TAM to specialist vendor revenue opportunity. Two-stage waterfall separates the cost stack (installation, commissioning, integration) from market structure (OEM captive absorption). Specialist pool $88-134M is roughly 40% of the installed system TAM.

Figure L11g. Sensitivity matrix: liquid GW × standalone attach rate. Base at 6 GW × 90% attach = ~$73M greenfield hardware. FM Global underwriter mandate progression is the biggest lever on the attach-rate axis.

12Commodity sensor or qualified infrastructure-control market?

The falsifiers are straightforward. Watch them over the next four quarters.

Figure L12. Signal watchlist through 2027-28. Most rows currently point toward bundler and control-plane consolidation. The Chemelex transaction is the specialist-platform exception on the list, and even that transaction was primarily electric heat-tracing rather than leak detection. Hyperscaler second-source qualification and dielectric-coolant adoption at scale are the two rows that remain open, and they are the ones with the most information value if either moves.