LEAK-PROOF AI COLD PLATES

The small machine that keeps a giant machine alive

A founder’s fieldbook on direct-to-chip cooling, qualification, and the hardware wedge before the factory

Research edition · 12 September 2026
A first-person opportunity narrative, checked against public engineering, company, standards, and filing evidence.

Working thesis I would not build a generic cold-plate factory. I would build the qualification layer for one or two real accelerator platforms, have the customer fund the work, and let a contract manufacturer make the metal while I own the evidence that makes it safe to install.

Evidence boundary AI cooling demand is real. The precise market size, 2026 adoption rate, supplier backlog, and procurement rules are not public enough to treat as facts. This is a plan for finding out.

01 · The heat problem we can feel

THE RACK IS A ROOM-SIZED HEATER

An AI server is not just a computer in a box. It is a dense electrical machine whose useful work arrives as heat. A rack that is easy to imagine as a row of servers can instead behave like a serious industrial load, with heat arriving at the same place where processors need to stay stable.

ASHRAE’s AI data-centre framework says air-cooling limits once thought to be around 25–35 kW per rack may extend to roughly 40 kW with better design. It also describes direct-to-chip liquid cooling as the mature path for high-density deployments. Those are design guides, not a universal physical ceiling. ASHRAE integrated design principles

The human-scale version is simple: a cooling failure can turn a valuable rack into an expensive service incident. The founder’s product is therefore not “a clever block of copper.” It is confidence that a particular platform can run, be serviced, and keep running.

02 · The water path

A COLD PLATE IS A CONTROLLED THERMAL LOOP

The cold plate sits on the heat-generating package, separated by a defined thermal interface material and a controlled mounting force. Inside it, channels or fins expose surface area to coolant. Supply and return ports connect to hoses, quick disconnects, a rack manifold, and a coolant distribution unit.

The coolant does not usually go straight into the building’s water pipes. The technology cooling system circulates through the rack; a facility water system and CDU move heat onward. The Open Compute Project cold-plate requirements separates those boundaries and names the variables a buyer must measure: heat transfer, flow, pressure, pressure drop, temperatures, active area, filtration, leakage, and material compatibility.

“Micro-fluidic” is not a magic word. Smaller channels can increase area, but they also increase filtering, contamination, pressure-drop, and manufacturing risk. I want the smallest geometry that meets the system requirement—not the smallest geometry that looks impressive in a rendering.

03 · Why now

THE COMPUTE PLATFORM IS MAKING THE QUESTION URGENT

NVIDIA’s public Vera Rubin NVL72 page describes a rack-scale system with 72 Rubin GPUs, HBM4, and a 45°C inlet specification for its stated 100 MW factory configuration. Its public page does not state a 2,300 W GPU TDP. NVIDIA Vera Rubin NVL72

That is enough to establish a direction: liquid cooling is being designed into the newest dense systems. It is not enough to turn every platform into one sizing number. The mating envelope, heat load, transient behaviour, coolant, allowable pressure drop, and service rules still belong to the actual customer design.

TrendForce’s public release forecasts liquid-cooling penetration in AI data centres at 14% in 2024 and 33% in 2025. It does not establish the supplied 53% in 2026. A BYD Electronic filing repeats the 33% 2025 figure and contains no 2026 penetration forecast. TrendForce · BYD Electronic HKEX filing

The timing is real; the precision is not.

04 · The corrected market reality

BIG DEMAND DOES NOT MEAN A BIG, OPEN MARKET

The supplied narrative makes six claims that I would not put in an investment memo as facts:

• 2,300 W for a Vera Rubin GPU is unverified in the public NVIDIA material reviewed.
• 225 kW for a Vera rack is not supported by the reviewed primary material; the platform definition and load term matter.
• 53% liquid-cooling penetration in 2026 is not publicly established. The inspected public evidence stops at the 33% 2025 forecast.
• Severe Taiwanese supplier backlog is plausible but not demonstrated by the public evidence reviewed.
• A blanket Western or hyperscaler ban on Chinese cooling hardware is overbroad. Contract-specific government restrictions do not prove a universal commercial policy.
• $30k–$80k is not a complete qualified-business budget. It might fund an outsourced prototype; it does not buy the full test, traceability, clean assembly, working capital, and liability envelope.

I therefore size the opportunity from named platforms, signed design inputs, current lead times, pilot quantities, and prices—not from a single market-report number. Until those arrive, the market is a hypothesis with a healthy physical signal and an unproven capture plan.

05 · The manufacturing bottleneck

THE LEAK IS HIDING INSIDE THE PART

The drawing is easy. The repeatable part is hard.

One route machines channels and joins a cover. Another uses controlled-atmosphere brazing. Friction-stir welding can join aluminium without melting it. Skived fins can raise surface area, but they do not solve the sealed liquid path. These are process options, not proof that a job shop can make a reliable product.

The minimum evidence stack is less glamorous than the CAD: alloy and temper traceability; mounting-face flatness and port dimensions; locked joining parameters; destructive cross-sections; 100% pressure/leak testing; 100% flow and pressure-drop testing; calibrated thermal testing; cleanliness and particle checks; coolant compatibility; pressure and thermal cycling; FMEA; gauge repeatability; calibration; and controlled engineering changes.

Boyd’s public technical material describes CAB brazing, friction-stir welding, and other cold-plate manufacturing methods. Its release claims five million delivered plates and 100% inline thermal/flow and leak testing. That is a company claim, but it shows the bar I would be competing against. Boyd metal fabrication · Boyd five-millionth cold plate

I would never promise “zero leak.” I would specify the leak rate, pressure, duration, detection limit, sample plan, and warranty boundary.

06 · Qualification is the product

THE PLATE HAS TO SURVIVE THE WHOLE RACK

Direct-to-chip cooling does not remove every watt from a rack. Vertiv’s guide says DTC systems can remove about 70–75% of equipment heat, leaving 25–30% for air cooling. Power supplies, memory, storage, networking, hoses, quick disconnects, manifolds, CDUs, sensors, and service access still belong to the system. Vertiv liquid-cooling deployment guide

That changes the sale. A plate with excellent thermal resistance can still fail because it has the wrong mounting load, a dirty channel, incompatible seals, a bad quick disconnect, too much pressure drop, or no safe response when a sensor detects a leak. OCP’s requirements make leakage intervention and component-level measurements explicit. OCP ACS cold-plate requirements

The buyer should provide the actual accelerator package, heat envelope, TIM, flow, pressure, coolant, life requirement, rack interface, service procedure, and acceptance test. Without those, “we need liquid cooling” is interest—not demand.

07 · The wedge

SELL THE EVIDENCE BEFORE THE COPPER

My first customer would be a Tier-2 server OEM, rack integrator, colocation operator, or AI infrastructure builder with a real platform slot and no mature cold-plate development team.

I would offer one narrow package: platform-specific design-for-manufacture, a controlled bill of materials, qualified contract manufacturing, serialised leak/flow/thermal records, material and coolant compatibility, service instructions, and a pilot lot. The customer pays $75k–$250k of NRE per platform across design, qualification, and pilot gates. These are operating assumptions to replace with quotes—not market prices.

The geography can be Southeast Asian manufacturing, but country is not assurance. I need audited process capability, a country-of-origin map, export-control review, and a second-source plan. The provenance message is “buyer-requested, auditable non-PRC manufacturing and component provenance” where a buyer actually requires it.

U.S. defense acquisition rules can restrict prohibited sources in a contract. They do not establish a blanket commercial hyperscaler ban on passive cooling parts. DFARS prohibited sources

08 · The first business is qualification

ECONOMICS BEFORE THE FACTORY

These are illustrative decision scenarios for one platform-specific plate, outsourced machining/joining, 100% testing, freight, and a warranty reserve. They are not a market-price forecast and exclude founder salary and corporate tax.

The equation is units × (ASP − variable cost) − fixed validation/operations. At low volume, I am running a qualification and evidence business. At the base case, a self-funded $150k programme takes roughly 2.1 years to recover before financing and founder compensation; paid NRE changes the risk materially.

The lean validation-first route is more plausibly $135k–$390k once test equipment, supplier audits, clean assembly access, scrap, working capital, and warranty reserve are counted. A $500k automation cell might be possible after repeat orders; it is not a safe all-in number for validated AI production.

09 · The incumbents have the hard years

THIS IS NOT A COMMODITY COPPER MARKET

CoolIT offers cold plates, loops, manifolds, and CDUs, and publicly advertises a validated 15 kW cold plate. Boyd claims five million delivered plates and inline testing. Delta markets liquid-to-air and liquid-to-liquid CDUs and cold-plate loops. Asetek’s D2C ingredient cooler integrates a pump and cold plate, reducing connections. Vertiv sells system-level cooling, monitoring, and deployment services.

These are company claims and product positions, not a complete market-share census. They still tell me what not to do: do not begin by competing on volume, generic thermal performance, or a broad “we can machine copper” promise.

The opening is narrower. An integrator may need a second source for one platform, faster engineering iteration, an evidence pack it can hand to its platform customer, or a local service and provenance path. If none of those pains are real, the incumbent wins before my first purchase order. CoolIT cold plates · CoolIT server products · Delta data-centre cooling · Asetek D2C ingredient coolers

10 · The path of attack

FOUR GATES, TWO YEARS, MANY WAYS TO STOP

Months 0–3 · prove the buyer. Select one platform. Obtain a signed design-input sheet and paid feasibility NRE. Audit three manufacturers. Stop if there is only generic interest, no platform data, or no paid feasibility.

Months 3–6 · prove the process. Make 10–20 prototypes. Measure thermal resistance, flow, pressure drop, flatness, and repeatability. Lock the joining process and establish 100% leak and flow records. Stop if leaks remain unexplained or the partner cannot control the required process.

Months 6–12 · prove qualification. Run a 30–50-piece lot with destructive samples and environmental/pressure-cycle tests. Deliver the serialised data pack. Seek a paid 100–500-unit pilot with explicit warranty boundaries. Stop if the buyer wants system liability without NRE, test access, or a realistic price.

Months 12–24 · prove repeatability. Add a second source and only then consider a captive clean assembly/test cell. Expand to adjacent platforms only when the interface is genuinely reusable. Stop below 500 units with high engineering churn or incumbent price and lead-time parity.

The company earns the right to scale by meeting evidence gates, not by buying machines early.

11 · What the headline omits

THE PLATE CAN WORK AND THE SYSTEM CAN STILL FAIL

The true bottleneck may be the entire cooling loop: quick disconnects, manifolds, CDU redundancy, facility water, leak detection, commissioning, service labour, contamination, and field replacement. Direct-to-chip leaves residual air responsibility, so a plate startup does not own the whole rack’s thermal outcome unless its contract says so.

The bear case is serious:

• The platform owner bundles the plate and leaves no second-source slot.
• A two-phase, immersion, rear-door, or revised cold-plate architecture makes my tooling obsolete.
• CoolIT, Boyd, Delta, or a system integrator cuts price or lead time before I qualify.
• One leak creates customer downtime and an unbounded liability problem.
• Contract-manufacturer quality drifts across lots or after an engineering change.
• The non-PRC provenance premise turns out not to matter to the buyer.
• AI capital spending slows before the pilot becomes a repeat order.

My mitigations are equally concrete: platform-specific contracts, reversible tooling, independent insurance counsel, serialised records, dual sourcing, paid NRE, and adjacent HPC or power-electronics applications only after the first platform is profitable.

12 · The founder decision

WOULD I BUILD IT?

Yes—but only as a qualification-led company, not as a speculative factory.

I would start when one named integrator supplies a real thermal test vehicle, signs the design inputs, funds feasibility NRE, and agrees to staged acceptance. I would outsource non-critical machining first while keeping configuration control, test fixtures, data, and customer evidence in-house.

I would walk away if customers ask for free prototypes; if no buyer reports a meaningful second-source or qualification-delay problem; if a 30–50-piece lot cannot achieve repeatable leak, flow, and thermal results; if contribution stays negative at 1,000 units; or if the customer will not accept clear system-liability boundaries.

The opportunity is not proven by a market-size headline. It is proven when a real rack runs safely, a buyer signs the release, and the next lot repeats the result.

13 · The first 90 days

A SMALL TEAM CAN MOVE FAST—IF THE CUSTOMER OPENS THE INTERFACE

The asset-light version is not “buy no equipment.” It is “buy no irreversible equipment before a buyer supplies an interface.” I would outsource machining and joining, rent metrology, and keep the fixtures, test method, configuration control, serialised data and customer relationship.

Days 0–30: buyer and interface. Interview thermal/mechanical, rack/CDU integration, procurement/supplier-quality, field-service and CM-quality owners. Ask for a platform revision, heat envelope, fluid, flow/pressure limits, QD/manifold, service procedure, reliability test and country-of-origin policy. A redacted drawing or test vehicle is evidence; “liquid cooling is coming” is not. The goal is one named design owner and one written paid-feasibility proposal.

Days 31–60: process and test. Audit three contract manufacturers for alloy traceability, joining records, cleanliness, CMM access, pressure/leak, flow, thermal testing, calibration and engineering-change control. Send a coupon and sacrificial plate through each candidate. Freeze the first acceptance matrix with the buyer.

Days 61–90: paid feasibility. Build 3–5 geometries only after the interface is frozen. Measure thermal resistance versus flow, pressure drop, flatness, leak rate, cleanliness and connector fit. Deliver a serialised report and ask the buyer to fund the next NRE gate. Stop if the result is technically interesting but nobody will pay for the next lot.

These are proposed operating gates, not market benchmarks. The first 90-day output is not a factory; it is a buyer-owned requirement sheet, a controlled test method and a paid reason to continue.

14 · Who actually buys

INTERVIEW THE PEOPLE WHO CAN SAY “YES,” “NO,” AND “THIS WILL LEAK”

I would not invent an exact person or pretend a famous robotics company is accessible. I would find the current role through a warm introduction, a target integrator’s public team, or the buyer’s own contact.

The evidence bar is a signed design-input sheet, buyer-supplied drawing or test vehicle, written acceptance method, identified supplier gap and paid next step. A LinkedIn reply, public RFP search, or enthusiastic call is not customer evidence.

15 · NRE buys learning; the test stack buys trust

THE SOW MUST SAY WHAT “PASSED” MEANS

A paid NRE statement of work should name the platform revision and define the inputs: package/TIM, heat-load points, coolant chemistry, supply/return temperatures, flow/pressure, QD/manifold, mechanical loads, environmental profile, applicable standards and exclusions. It should deliver controlled CAD/BOM, DFM review, fixture drawings, material certificates, process flow/control plan, FMEA, test procedure, raw data, serialised reports, nonconformance log, service instructions and change history.

Acceptance must be numeric: thermal resistance and uniformity; pressure-drop/flow curve; leak detection limit, pressure and duration; hydrostatic and cycle conditions; cleanliness/particle limit; corrosion and compatibility; connector fit; sample sizes and re-test rules. The OCP qualification guide calls for a representative processor stack and TIM, stable bubble-free flow, recording case temperature, inlet-liquid temperature, applied power, flow and pressure drop, plus X-ray or equivalent inspection, hydrostatic testing and fluid-compatibility work.

The minimum bench stack is:

• CMM/flatness and connector-fit inspection, with X-ray or destructive coupons for joining defects.
• Buyer-selected pressure-decay, bubble or helium leak method; report the detection limit rather than “zero leak.”
• A calibrated heater or representative board that sweeps flow and computes R=(Tc−TL)/Q, alongside pressure-drop/flow curves.
• Named-fluid corrosion, elastomer compatibility, leach/particle and post-test inspection.
• Pressure/thermal cycling, vibration/handling, QD cycles, shipping/flush/dry and a controlled leak-response drill.

The Gemini suggestion of a roughly $1,500 accelerated-life rig is a useful learning hypothesis, not a production-life certificate. Fatigue and system reliability need a buyer-defined load spectrum, destructive analysis and—where the contract requires it—an independent lab. The NRE quote is customer-specific; a universal $25,000 prototype or easy margin is not evidence.

16 · The plate is a component inside somebody else’s safety system

EARN THE RIGHT TO SELL THE LOOP

NVIDIA’s public GB300 NVL72 reference describes a liquid-cooled MGX rack, integrated tray/rack leak detection, eight 33 kW power shelves and a full rack up to 142 kW. That is a useful reminder that cooling, power, management and service are one integration problem. NVIDIA NVL72 AI Factory components

NVIDIA documents three leak-detection layers for GB200/GB300 NVL72 systems: cold-plate and inner-manifold sensors at tray level, sensing rope/spot sensors at rack level, and CDU/datacentre sensors. Rack-level handling belongs to the customer’s BMS, while the compute system exposes signals through its management path. NVIDIA Mission Control BMS integration NVIDIA leak detection operations

That creates system-supplier risk. The platform owner or integrator may bundle the plate, own the approved-vendor list, require one accountable warranty, or change the QD/fluid/manifold before the startup is qualified. NVIDIA’s release notes even document firmware-update edge cases for leak-detector settings and recommend standard Redfish leak-detection endpoints rather than raw sensor voltage. NVIDIA DGX GB300 release notes

My migration path would be explicit:

1. Platform plate: one package and fluid; outsourced machining/joining; paid NRE; serialised thermal, hydraulic, leak and cleanliness records.
2. Qualified plate family: reuse only proven mounting/TIM/flow interfaces; add a second CM and repeat lot.
3. TCS subassembly: integrate plate, hose/QD and manifold only when the integrator supplies the rack interface and acceptance test, and pays for the added pressure-drop, leak, service and field-return responsibility.
4. Thermal subsystem: add CDU controls, leak telemetry, commissioning and residual-air coordination only with a system partner under a new system-level SOW. This is a different liability and support business, not a natural next SKU.

Captive assembly/test is an option after two paid platform programmes, two repeat lots with stable distributions, measured test-cycle economics, a second-source plan and customer acceptance of the startup’s records. Those are internal gates, not public benchmarks. Until then, buy evidence and flexibility. Cost-down comes from measured yield, rework, cycle time, material use, tool life, test automation and committed volume—not from an unverified Chinese-price target, a $500k factory assumption, or a promised 80%→99% yield jump.

The Gemini thread’s useful leads were hands-on validation, buyer-persona interviews, paid platform NRE, and treating provenance as a requirement to verify. Its exact robotics specifications, “desperate buyers,” universal NRE price, cheap fatigue certification, labour-rate comparisons and named-person access were downgraded or rejected. The unknowns remain the point: actual SKU and thermal envelope, approved-vendor gap, paid pilot quantity, accepted tests, CM capability and yield, country-of-origin scope, field liability, and whether the buyer pays for a plate, a tested subassembly or a full subsystem.

Research receipts

• NVIDIA Vera Rubin NVL72 — 72 GPUs, HBM4, 45°C inlet for the stated 100 MW factory configuration; no public 2,300 W GPU TDP on the reviewed page.
• NVIDIA Vera Rubin platform announcement — rack-scale platform context and liquid-cooled infrastructure.
• ASHRAE integrated design principles — air-cooling transition range and DTC guidance.
• ASHRAE introduction and purpose — liquid-temperature class vocabulary.
• Schneider Electric reference design — a 40 kW air-cooled reference scenario, not an industry-wide maximum.
• OCP ACS liquid-cooling cold-plate requirements — metrics, pressure/leakage, filtration, compatibility, and hybrid-cooling boundaries.
• Vertiv liquid-cooling deployment guide — DTC heat-removal guidance and system integration.
• Boyd five-millionth cold plate release — company-reported scale and testing claims.
• Boyd metal fabrication — CAB, FSW, skived-fin, and fabrication options.
• CoolIT cold plates — incumbent product and validation benchmark.
• TrendForce public penetration release — public 14% 2024 / 33% 2025 estimate; no 53% 2026 figure.
• BYD Electronic HKEX filing — repeats the 33% 2025 figure and contains no 2026 penetration forecast.
• DFARS prohibited sources — narrow government/defense provenance context, not a blanket commercial ban.
• OCP cold-plate development and qualification — representative thermal, hydraulic, X-ray, hydrostatic, cleanliness and compatibility methods.
• OCP cold-plate workstreams — current fluid, quick-disconnect and interoperability workstreams; evidence that interfaces continue to evolve.
• NVIDIA NVL72 AI Factory components — GB300 rack architecture, power shelves and integrated leakage detection.
• NVIDIA Mission Control BMS integration — tray-, rack- and datacentre-level leak sensing and BMS boundary.
• NVIDIA Mission Control leak detection — cold-plate/manifold leak metrics and rack liquid-isolation signals.
• NVIDIA DGX GB300 release notes — Redfish leak-detection guidance and firmware-update caveat.
• ASHRAE Handbook Chapter 20 — hybrid DTC/air architecture and liquid-loop redundancy considerations.
• OCP advanced liquid cooling design — pressure-drop/thermal-resistance trade-off and flow-dependent cold-plate design.