Who accepts a cold-plate retrofit? OEM support and leak responsibility

Who has authority to accept the retrofit?

Name the owner representative and the document delegating acceptance authority before the pilot starts. A commissioning witness signs for the result witnessed; that signature does not silently approve a commercial substitution, a tenant shutdown or a transfer of operating custody.

Four decisions that need separate evidence

The owner accepts contract deliverables, the OEM or expressly contracted support provider confirms configuration coverage, the engineer checks the specified interface, and the operator accepts custody and procedures. In colocation work, identify tenant authority over servers and facility authority over building systems, then name the delegate releasing their combined interface.

Track configuration support, pilot authorization, pilot results, operating custody and fleet release as separate dated decisions. A vendor commissioning certificate covers its named package; attach its exclusions.

Use the [direct-to-chip cooling guide](/field-guides/datacenter/data-center-direct-to-chip-liquid-cooling/) for architecture and [retrofit qualification](/field-guides/datacenter/qualify-data-center-liquid-cooling-retrofit/) for feasibility. ASHRAE TC 9.9's 2021 white paper provides system context. [S10]

Is the exact server and cold-plate kit supported?

A liquid-cooled factory option does not authorize conversion of an installed air-cooled server. Lenovo's ThinkSystem SR635 V3 Product Guide LP1609, updated October 18, 2026, lists the C1PM Processor Neptune Core Module as CTO only and expressly excludes field upgrade. CTO means configure to order. [S1]

Table A separates factory supply from field conversion. A [cold plate](/field-guides/glossary/cold-plate/) part number alone cannot establish the support status of the assembled server.

Separate field conversion, service replacement, and factory supply

Service replacement restores an approved configuration; conversion changes it. CoolIT's OEM relationship does not authorize conversion of a customer's installed server. [S7]

Extend each job row with asset and serial numbers, chassis/board revision, CPU/GPU and power limits, mounting hardware and thermal-interface material, kit revision, hoses, both quick-disconnect halves, wetted materials, coolant, firmware and installer authorization. Attach the support case, exclusions, approval date and change history.

Hold a CTO-only field-conversion order unless a documented OEM exception or expressly agreed alternative support contract changes its basis. For an integrator-backed conversion, identify what the alternative provider covers and secure owner acceptance of that support basis. An installer signature cannot turn it into OEM approval.

What warranty evidence must be secured before conversion?

Secure written configuration support before modifying the server, and keep server, kit, coolant distribution unit (CDU), workmanship and paid-service terms separate. Lenovo Limited Warranty L505-0010-02, August 2011, where incorporated into the purchase terms, excludes specified damage resulting from modification, unauthorized service and third-party products; it does not justify saying that every modification cancels every warranty. [S6]

That form starts coverage on the original purchase date shown on the receipt or invoice unless Lenovo specifies otherwise, with country-specific modifications in Part 2. Record the actual start event and expiry for each obligation; pilot acceptance does not reset an existing equipment warranty. [S6]

What counts as written configuration support?

Use a dated support case, signed statement, order configuration or contract attachment identifying the serial-number range and permitted work. Capture issuing organization, case ID, signer authority, required installer qualifications, restoration conditions, spares route and any requirement to restore the original configuration. A quote saying OEM compatible leaves these questions unanswered.

The warranty evidence checklist in the handover section connects each document to a release decision. Record jurisdiction, incorporated agreement, support entitlement, parts/labor coverage, onsite response commitment, exclusions, notice deadlines and claims contact. Keep collateral hardware damage, lost data, lost revenue, cleanup and service credits distinct from defective-part replacement.

Support evidence continues after handover: Lenovo LP1609 ties lifetime water-quality maintenance to warranty and support on affected components. Assign the sampling records to an operating custodian who can retrieve them when a claim arises. [S1]

Retrofit acceptance responsibility matrix

Anvilfield recommends one acceptance authority per deliverable, with executing, confirming and witnessing parties named separately. Table B is a proposed contract structure to complete with legal entities, package references, contacts and delegates. It is not an OCP liability allocation.

The signature block that closes a decision

Each signature block identifies scope, assets, boundaries, governing revision, result, evidence links, exceptions, restrictions, signer capacity, delegated authority and timestamp. Record receiving, installation, testing, defect closure, custody and fleet-release dates.

Offer accept, reject, or accept with listed noncritical exceptions under the project contract. Under Anvilfield's proposed release rules, unresolved support, an active leak, unproved protection or an unapproved operating envelope holds production release. The [commissioning](/field-guides/glossary/commissioning-cx/) witness records technical evidence; the authorized delegate makes the acceptance decision.

Where is the facility-to-server demarcation?

Draw the physical boundary and assign each interface. Wiwynn's integration summary, updated April 27, 2026, describes the technology cooling system (TCS); ASHRAE's 2021 white paper explains the wider system context. FWS means facility water system. A heat exchanger separates fluids, not contract responsibilities. [S9, S10]

Separate fluid circuits: FWS supply → CDU primary side → FWS return. TCS supply → rack manifold → mating quick disconnects (QDs) → server hoses and cold plates → TCS return → CDU secondary side. Heat crosses the heat exchanger; the fluids remain separate.

Signal path: leak/temperature/flow sensors → server or CDU controller → building management system (BMS) / data center infrastructure management (DCIM) → operator and approved protective action.

The mating connection needs an owner

Record who supplies each QD half, verifies the exact pairing, makes the joint, tests it, isolates it and replaces each half. Include seal/material, part number, orientation, access, restraint, caps, cleanliness and authorized service method. Follow the [QD selection guide](/field-guides/datacenter/liquid-cooling-quick-disconnect-selection/); the marketing word dripless is not an installed-system leakage limit.

Assign fluid custody, fill equipment, temporary hoses, flushing, sampling, drains, valve settings and contaminated-fluid disposal. Preserve separate circuit limits and tests at the [CDU interface](/field-guides/datacenter/cdu-coolant-distribution-unit-commissioning/). A liquid-to-air CDU exchanges the FWS connection for a room heat-rejection interface; include the [residual air load](/field-guides/datacenter/liquid-cooled-data-hall-residual-air-load/) in acceptance.

Which limits belong in the acceptance record?

Keep model, module, circuit, measurement point, quantity, units, conditions and source revision together. Operating inlet pressure, pressure drop and proof-test pressure are different quantities. Record gauge or absolute reference and instrument location; obtain supplier resolution of an unmarked pressure basis before applying a limit.

Lenovo's SR635 V3 guide LP1609 specifies a maximum water pressure of 3 bar and per-server flow points of 1 L/min at a 45°C inlet and 1.5 L/min at a 50°C inlet; these are operating conditions, not a proof-test pressure. Do not interpolate unlisted conditions without supplier confirmation. [S1]

Resolve disagreements by matching scope

PG25 denotes 25% propylene glycol; DI means deionized water. Lenovo LP2018 requires biological and corrosion control as well as filtration; DI fill water alone is not a complete treatment program. [S4]

Choose the intersection of fluid approvals for the installed components, or revise the approved design. Service visits do not replace the specified sampling program. Preserve sample location, fluid batch, laboratory method and treatment history through the [water-treatment record](/field-guides/datacenter/liquid-cooling-loop-water-treatment-chemistry/).

Resolve source errors before testing. Lenovo LP1498, May 26, 2026, pairs 689.66 kPa and 6.9 bar with 20 psi under rear-door heat-exchanger water requirements, an inconsistent conversion. Obtain a corrected supplier limit rather than authorizing a test from that row. [S5]

What does the cold-plate pilot need to prove?

Test every distinct server/kit/firmware/coolant combination and hydraulically different position selected for release under the signed sampling plan. Table E is Anvilfield's proposed plan for 24 nodes in two racks. Document selection and exclusions so passing one configuration cannot silently release another.

The [ICT-Commercial](https://ict-commercial.com/liquid-cooling-retrofit-gpu-servers-2026/) and [BENQUAN](https://benquan.hk/article-liquid-cooling-retrofit-gpu-servers-2026.html) reseller articles propose 2 to 4 racks and 30 days. Those overlapping proposals are supplier pilot scopes, not an OEM, ASHRAE or OCP mandatory interval.

Mechanical, thermal, controls, and service evidence

OCP Liquid Cooling Integration and Logistics Revision 1.0, section 6.3, calls for recording FWS/TCS conditions, taking a secondary-fluid sample and proving prescribed alarm faults before handover. Use the signed supplier/project procedure for test medium, pressure, duration, component isolation and temperature correction. Isolate components outside that procedure before testing building piping; a server operating-pressure limit is not authorization for a pressure test. The [loop commissioning guide](/field-guides/datacenter/liquid-cooling-loop-commissioning/) covers the underlying flush/fill sequence. [S14]

Define full load by production workload and configured power limits. Record CPU/GPU watts, useful throughput, fan settings, air inlet, per-server coolant inlet/flow, chip temperatures and throttling. Fix settling criteria, sampling cadence, maximum data gaps, measurement uncertainty, thermal limits and protective-action times before testing.

Compare throughput and error rates with the recorded pre-conversion baseline at the same workload and power settings. A cool chip that throttles or produces less useful work has not met a performance acceptance criterion. Include memory, storage, voltage regulators and room air conditions in the agreed thermal checks; cold plates do not remove every component's dependence on airflow.

How many days and how much measured coverage?

Anvilfield estimates 5.5 loaded test hours in this plan: three 30-minute plateaus followed by four hours at final load, then 30 calendar days of operating observation. Setup, transitions, alarm/service tests, outage windows, corrections and retests add time. Anvilfield estimates 17,280 planned node-hours by multiplying 24 nodes by 30 days by 24 hours; replace that plan with actual logged exposure when nodes are offline or records are missing.

Retain raw trends, alarm logs and operator actions alongside service cycles and fluid additions. Agree which defects reset an observation window, which unchanged configurations retain evidence, and who authorizes restart. A leak-free pilot supports release within its recorded configuration, duty and period.

How do you prove leak response without creating a leak?

Use the manufacturer's authorized simulator, test feature or controlled procedure, then follow the event through detection, controller, operator receipt and approved action. Lenovo LP1609 describes detection of more than 0.5 mL along the module tube and an XClarity Controller (XCC) event; that sensor description is not an allowable coolant-loss limit or permission to wet electronics. [S1]

Dell's undated 15G DLC design note describes small-leak warning, large-leak critical and sensor-error warning events in iDRAC, with configurable alert or power-off actions. HPE DL325 Gen12 QuickSpecs V13 specifies shutdown and power-on inhibition for its identified closed-loop module until the leak event is cleared, with REST API recovery. HPE recommends replacing that module every five years and prohibits internal coolant replacement. Test the installed module, firmware and settings. [S12, S13]

Record whether the test exercised the physical sensor or injected a controller event. A software alarm alone proves neither sensor placement nor collection coverage; retain the separate supplier-approved detection check and installation inspection. OCP's Leak Detection and Intervention Revision 1.0 discusses sensor placement and separate intervention choices. [S17]

Fail-safe action includes the other servers

Timestamp detection, classification, receipt, valve/pump/server response, affected zone, unaffected-load behavior and reset. A shared pump trip can remove cooling from healthy servers; the approved cause-and-effect sequence must address those loads. Include sensor fault, communications loss, controller restart, power restoration and restart permission where the design includes them.

Set required action times in the signed procedure and compare actual results against them. Preserve source logs and as-left settings with the [leak-detection evidence](/field-guides/datacenter/data-center-liquid-cooling-leak-detection/), so alarm receipt cannot be mistaken for proof of isolation or shutdown.

Who responds, investigates, and pays after a leak?

Anvilfield's retrofit responsibility matrix assigns the emergency response lead separately from the parties investigating cause and paying component, collateral-damage, cleanup, and downtime claims. Operations coordinates the approved emergency operating procedure (EOP); installers and suppliers investigate their named boundaries. Table F records questions for the actual contract and cause finding.

Restore service while the cause is being established

Preserve synchronized logs, as-found photographs, fluid samples with chain of custody, serials, configuration, last work and authorized recovery actions. Retain the suspected failed assembly until the claims procedure permits disposal. Leak location identifies where fluid appeared; manufacture, workmanship, pairing, contamination and upstream excursions remain separate causal questions.

Compare [dew point](/field-guides/glossary/dew-point/) with surface temperature when moisture appears without a failed pressure boundary. Preserve insulation and setpoint history alongside pressure trends. Condensation and a leaking coupling demand different investigations.

Name the first-call route, authorized isolations, spares custodian, restoration delegate and interim funding route while payment is disputed. Follow the [MOP/SOP/EOP procedures](/field-guides/datacenter/data-center-mop-sop-eop-procedures/) and retain notice recipients, claim deadlines and the reconciliation clause. The first responder does not inherit the repair bill by restoring service.

Obtain written cause and coverage decisions addressing replacement, collateral IT damage, cleanup, labor, shipping, retest, temporary cooling, downtime, service credits and deductibles. A denied component-warranty claim does not itself decide installer liability or insurance coverage.

Worked example: a 24-server support and flow decision

Start with the ordering decision: hold a proposal to convert 24 existing air-cooled SR635 V3 nodes using C1PM under Lenovo's published CTO-only route. A factory-configured alternative requires its own serial-linked support records and integration pilot. Cooling arithmetic cannot override the published supply path. [S1]

Anvilfield estimates that 24 SR635 V3 servers require an aggregate 24 L/min at Lenovo's published 45°C point or 36 L/min at its 50°C point, a 50% increase that does not change the module's CTO-only ordering restriction. Multiply per-node flow by server count, then divide the increase by the original total.

Method: Q_total = N × q_node; increase = (Q_B − Q_A) / Q_A × 100%.

How should measurement uncertainty affect a pass?

Anvilfield recommends agreeing a decision rule before testing. For a minimum-flow criterion, one conservative rule subtracts the stated measurement uncertainty from the reading before comparison. A branch reading 1.02 L/min with ±0.04 L/min uncertainty has a lower bound of 0.98 L/min, below the selected 1 L/min criterion: hold it for adjustment or a better measurement. Record the instrument range, calibration, fluid correction and uncertainty basis with the result.

Release the right decision

Hold the deficient branch, resolve restriction or balance, and remeasure every branch affected by the adjustment. Record the flow measurement method and location; a CDU header reading cannot establish individual server flow.

Method note: Anvilfield estimates use LP1609, checked November 5, 2026, and the selected 24-node count without diversity or hydraulic allowance. Sum branch inputs for aggregate flow; divide the deficient branch's 0.46 L/min shortfall by its 1 L/min requirement for the percentage. Manifold loss, pressure regulation, CDU capacity and workload verification remain separate design inputs.

How much time and acceptance work should the bid include?

Anvilfield estimates $4,700 for the stated acceptance attendance and records package, plus $1,500 for one defined retest, totaling $6,200 at Table H's selected US-dollar rates. Price conversion, equipment, design, fluid supply, temporary cooling, outage coverage and mobilization separately.

CoolIT's undated Global Professional Services FAQ, accessed November 5, 2026, gives 1 to 3 days for installation and commissioning of a single in-rack CDU deployment, and weeks to months for larger deployments with a secondary fluid network. Put configuration approval, operating observation and defect closure on separate schedule lines. [S8]

Price the failure path as well as the first test

Keep correction labor separate from retest attendance and post-correction proof. The [commissioning subcontract scope guide](/field-guides/datacenter/data-center-commissioning-subcontract-scope-retests/) helps assign return visits.

Method note: Anvilfield estimates sum labor products, samples and consumables using scope and pricing inputs selected November 5, 2026. Rates include burden, overhead and profit; tax, travel, equipment purchase, conversion and continuous staffing are excluded. The observation window uses existing operational monitoring under its agreed service scope. Obtain job quotes for additional work.

The one-day attendance allowance requires completed preparation and a test sequence that fits around 5.5 loaded hours. Schedule setup, alarm/service tests and records explicitly; add attendance when they do not fit. The 30-day observation period uses separately assigned operations staff. State who supplies workload generation, instruments, lab analysis and reporting, and what triggers a return visit.

Public upgrade packages should separately identify conversion authorization, test attendance, observation monitoring and corrective retests, naming who supplies each record and funds each return visit. Estimators use PorchRocket Bid Control to find past public contract awards and award amounts when scoping the next cooling upgrade bid.

What must be in the handover before fleet rollout?

Release only the serials, configurations and operating envelope covered by signed evidence. Use Checklist W with status, evidence link, responsible organization and due date for every item. Missing support, a failed test, active leakage or unproved protection holds the affected release; noncritical exceptions need owner authorization, restrictions and a closure date.

Checklist W. Converted-server warranty and handover evidence

• Exact server asset IDs, serial numbers, machine types, CPU/GPU variants and power limits match the supported-kit register.
• Kit, mounting hardware, thermal-interface material (TIM), hose, QD, wetted-material and coolant revisions match the written approved configuration.
• Published field-conversion route or documented exception is attached; CTO-only configurations are not represented as field upgrade kits.
• Written support provider statement identifies included components, permitted work, installer qualification, exclusions and support case/contract ID.
• Server, coldplate, CDU and workmanship warranties have their own actual start events, dates, terms, contacts and claim instructions.
• Consequential damage, cleanup, temporary cooling, labor, shipping, retesting, service credits, interim recovery funding and notice deadlines have named contract references.
• FWS/TCS/server/control boundaries show port tags, valves, sample points and custody owners.
• Both QD halves and the mating-joint installation owner are identified.
• Tenant/server authority and facility authority each sign their scope, with one named delegate for integrated release.
• Mechanical integrity tests carry approved limits, procedure revision, isolated components, instruments, readings and witnesses.
• Fluid fill batch, baseline sample, approved analytes, filter record and treatment history are attached.
• Pilot matrix lists configurations, workloads, actual power/throughput, operating states, raw trend locations, calibration and uncertainty records, data-gap rules and pass criteria.
• Actual observation exposure, service cycles, fluid additions, defects and restart/retest scope are recorded; planned calendar days alone do not close the pilot.
• Leak sensor, source controller, BMS/DCIM route, required protective action, sensor fault and restart have signed results.
• Authorized service/reconnection demonstration and post-service checks are recorded.
• Blocking deficiencies are corrected and retested; accepted noncritical exceptions have operating restrictions, owners and dates.
• As-left firmware, thresholds, delays, setpoints, alarm recipients and versioned cause-and-effect are exported.
• Operators have the EOP/MOP, incident contacts, spare inventory, fluid handling instructions and training record.
• Sampling/maintenance calendar identifies actual module and service-provider intervals rather than a generic annual visit.
• Operational custody and fleet release are signed by the expressly authorized owner representatives, with serial scope and change triggers.

Name the next operator and the next claim owner

Transfer usable files and access to operations, indexed by asset and serial number. Preserve the [proof packet](/field-guides/datacenter/proof-packet-assembly-turnover/) and [closeout calendar](/field-guides/datacenter/punch-list-closeout-turnover/) with separately dated conversion authorization, pilot completion, custody, warranty start and fleet release.

CoolIT's undated Global Professional Services FAQ describes typical preventive visits every six months and SLAs of 2 to 5 years in one-year increments. These are service offerings; record the purchased term and duties separately from Lenovo's chemistry program and HPE's module replacement recommendation. Name sampling points, methods, limits, treatment, spares, contacts and due dates. [S8]

Reopen affected acceptance after changes to kit/server revision, CPU/GPU power, firmware or protective settings, fluid, wetted materials, QD pairing, CDU controls or topology. Name the rollback configuration, decision owner and required outage before deploying the change. OCP Integration and Logistics Rev 1.0, section 6.4, supports keeping lifecycle records; preserve the change, its support decision and affected retests together. [S14]

Reference records: supported kits, pilot matrix, demarcation, incident responsibility, warranty checklist.

How to cite this page: Anvilfield, Who accepts a cold-plate retrofit? OEM support and leak responsibility, 2026, https://anvilfield.com/field-guides/datacenter/cold-plate-retrofit-oem-acceptance-responsibility/

Related tools

FAQ

People also ask

• What is direct-to-chip liquid cooling?
• Can an existing air-cooled data center be retrofitted for liquid cooling?
• What is a punch list in construction?
• What is direct-to-chip liquid cooling?
• What is a CDU in a data center?

Sources checked

• S1. Lenovo ThinkSystem SR635 V3 Server Product Guide, LP1609, updated October 18, 2026 Processor module, ordering and water requirements. Sources checked and page reviewed November 5, 2026. Update log, November 5, 2026: clarified separate fluid circuits, treatment duties, pilot measurement decisions and change control; added the body citation line.
• S2. Lenovo ThinkSystem SR630 V3 Server Product Guide, LP1600, updated August 27, 2026 BXBC ordering restrictions. Accessed November 5, 2026.
• S3. Lenovo ThinkSystem SR650 V4 Server Product Guide, LP2127, updated October 20, 2026 Processor and Compute Complex module distinctions and water requirements. Accessed November 5, 2026.
• S4. Lenovo Neptune Direct Water-Cooling Standards, LP2018, September 13, 2024 Water quality, filtration, treatment and monitoring. Accessed November 5, 2026.
• S5. Lenovo 42U and 48U Heavy Duty Rack Cabinets Product Guide, LP1498, updated May 26, 2026 RDHX water requirements contain inconsistent pressure units; corrected supplier limit required. Accessed November 5, 2026.
• S6. Lenovo Limited Warranty, L505-0010-02, August 2011 Part 1 coverage, exclusions and start event; Part 2 country terms. Applies where incorporated. Accessed November 5, 2026.
• S7. CoolIT, Direct Liquid Cooled Servers Undated. OEM-installed supply route. Accessed November 5, 2026.
• S8. CoolIT, Global Professional Services Undated. Deployment timing, preventive visits, SLA duration and PG25. Accessed November 5, 2026.
• S9. Wiwynn, OCP Liquid Cooling Integration and Logistics White Paper landing page, updated April 27, 2026 Updated April 27, 2026. Contributor summary, not a new technical revision. Accessed November 5, 2026.
• S10. ASHRAE TC 9.9, Emergence and Expansion of Liquid Cooling in Mainstream Data Centers, 2021 white paper 2021. Architecture and service-interface context. Accessed November 5, 2026.
• S12. HPE ProLiant Compute DL325 Gen12 QuickSpecs, a50009232enw, Worldwide V13, October 5, 2026 Worldwide V13, October 5, 2026. Core Options and change history reopened through the web reader; curl timed out. Checked November 5, 2026.
• S13. Dell, Dell Technologies Direct Liquid Cooling Support for New PowerEdge Servers, undated 15G design note Undated 15G design note. Leak Sense classifications and configured actions. Accessed November 5, 2026.
• S14. OCP, Liquid Cooling Integration and Logistics White Paper, Revision 1.0 Revision 1.0, sections 6.3 and 6.4. Commissioning, handover and lifecycle evidence. Accessed November 5, 2026.
• S15. OCP, Cold Plate project current resource listing Live resource index; distinguish released revisions from drafts. Accessed November 5, 2026.
• S16. OCP, Cold Plate Cooling Loop Requirements, Rev 2 (undated) Rev 2; issue date not established. Single-phase CDU guidance, SR 5.1-1 and telemetry. Accessed November 5, 2026.
• S17. OCP, Leak Detection and Intervention White Paper, Revision 1.0 Sensor placement, collection coverage and separate intervention choices. Accessed with curl November 5, 2026.
• D1. ICT-Commercial retrofit article Supplier proposal for pilot scope; not technical or warranty authority. Accessed November 5, 2026.
• D2. BENQUAN retrofit article Overlaps ICT-Commercial; supplier proposal, not independent validation. Accessed November 5, 2026.