Cooling 100 kW and 150 kW GPU racks requires a coordinated liquid and air heat-removal strategy. The rack power number alone does not determine flow; the design must use the liquid heat fraction, coolant temperature rise, fluid properties and vendor limits.
The engineering basis begins with the actual GPU platform and thermal envelope. Rack density, liquid heat fraction and temperature limits should be confirmed before equipment is procured.
For W Land’s planned West Texas AI energy campus, this topic should be resolved through a documented basis of design, a commercial responsibility matrix and an evidence-based diligence package. Any public capacity, schedule, cost or performance statement should remain qualified until the relevant site, equipment, permit and tenant decisions are complete.
Key takeaways
- Obtain rack thermal design power and heat split.
- Calculate flow from heat and allowable delta-T.
- Size CDUs and manifolds for diversity and redundancy.
- Evaluate reliability, schedule, total installed cost and lifecycle operations—not a single headline metric.
- Keep the solution compatible with phased 25–50 MW deployment and a 100 MW Phase 1 campus.
What the decision really involves
The first step is to define the operating outcome. For an AI data center, the requirement is not simply to install equipment with sufficient nameplate capacity. The complete system must maintain acceptable voltage, frequency, thermal conditions and maintainability through credible faults, maintenance events and expansion work.
The project team should answer the following questions before design freeze:
- Obtain rack thermal design power and heat split.
- Calculate flow from heat and allowable delta-T.
- Size CDUs and manifolds for diversity and redundancy.
- Design residual air cooling.
- Confirm facility heat rejection on the design day.
The answers should be translated into single-line diagrams, thermal and hydraulic schematics, equipment data sheets, control narratives, operating modes and acceptance tests. That record is what allows a tenant, lender, insurer, owner’s engineer and permitting authority to evaluate the project consistently.
Decision matrix
| Decision factor | Configuration or reference | Alternative or practical implication |
|---|---|---|
| Rack power | 100–150 kW+ | Electrical input |
| Liquid heat fraction | Vendor-specific | Direct liquid load |
| Temperature rise | Design variable | Controls required flow |
| Residual air | Non-liquid components | Room cooling load |
| Redundancy | CDU/pump/loop topology | Availability |
The matrix is a screening tool, not a substitute for engineering. Site conditions, tenant specifications, equipment availability and the adopted regulatory framework may change the result. The preferred solution should be supported by net site performance, lifecycle cost and failure-mode analysis.
Practical planning example
If 120 kW of rack power has 90% liquid heat capture, about 108 kW enters the liquid loop and 12 kW remains for air removal. Multiplying by hundreds of racks creates both a major water-side load and a nontrivial residual air system.
A planning example should always state its assumptions. Electrical MW, thermal MW, MWh duration, gas heating-value basis, PUE, ambient condition, redundancy and end-of-life capacity are different metrics. Mixing them can make a concept appear more reliable or less expensive than it is.
For a phased campus, the example should also be tested at the first block, full Phase 1 and ultimate master-plan conditions. A solution that works for one 25 MW block may produce excessive fault current, pipe length, cable count, control complexity or maintenance exposure at 500 MW.
Engineering, schedule and commercial implications
Reliability and operations
The technology cooling system and facility cooling system must be separated by clear performance boundaries. Temperatures, flows, pressure, chemistry, heat-exchanger approach and allowable transients should be contractual.
The operator should be involved before the design is issued for construction. Maintenance access, isolation boundaries, alarm priorities, spare parts, staffing and recovery procedures influence the architecture. A design that is efficient at full output but difficult to maintain can reduce actual availability.
Procurement and delivery
High-density halls still reject residual heat to air. The design should quantify the liquid heat fraction by platform and preserve room conditions for networking, power supplies, storage and service personnel.
Long-lead procurement should use approved data sheets, witnessed factory tests, serial-number traceability and a controlled deviation process. The owner should receive editable drawings, calculations, configuration files, test data and operating manuals—not only scanned certificates.
Compliance and bankability
Cooling performance should be tested at the design envelope, including high ambient, degraded equipment and failure modes. Catalogue ratings at favorable temperatures are not sufficient.
W Land and CITC can integrate the powered shell, facility water system, CDUs, distribution piping and heat rejection around the tenant’s actual GPU platform.
The project should retain vendor neutrality unless a tenant or lender approves a proprietary standard. Equipment sourced through AiWB or CITC must satisfy the same U.S. technical, safety, cybersecurity, warranty and service requirements as domestic or European alternatives. The comparison should use landed, installed and risk-adjusted cost.
Common failure modes
- Assuming 100% liquid capture.
- Using average rack load for pipe sizing without diversity rules.
- No hydraulic balance across rows.
- Ignoring pump and CDU heat.
- Designing heat rejection only at annual-average ambient.
These failures tend to appear at interfaces: vendor versus EPC, factory versus site, electrical versus mechanical, power plant versus data center, and commercial promise versus permit condition. W Land should maintain one interface register and one integrated schedule across all parties.
W Land implementation approach
W Land should address cooling 100 kW 150 kW GPU racks through a gated process:
- Requirement definition. Confirm the tenant load, rack platform, reliability target, operating modes and expansion plan.
- Concept screening. Compare technically viable alternatives using the same site, ambient and commercial assumptions.
- U.S. engineering review. Assign licensed engineers and specialist consultants to validate code, protection, permitting, fire and cybersecurity requirements.
- Vendor qualification. Require complete performance data, deviations, factory capability, service support and contractual guarantees.
- Factory and site validation. Use FAT, SAT and integrated systems testing tied to objective acceptance criteria.
- Operational handover. Deliver training, spares, controlled configurations, maintenance plans and tested emergency procedures.
Final temperatures, flow, pressure, water chemistry, redundancy and controls must be approved by the GPU vendor, tenant, cooling OEM and licensed U.S. mechanical engineer.
Implementation checklist
- Rack OEM data obtained
- Liquid/air split modeled
- Flow and pressure budgets issued
- CDU redundancy defined
- Residual air capacity included
- Design-day heat rejection checked
- Thermal test plan prepared
Related W Land pages and articles
- Liquid Cooling & Thermal Management
- Cooling Equipment
- AI-Ready Powered Shell
- Request an NDA Briefing
- Direct-to-Chip vs. Immersion Cooling for AI Data Centers
- CDU Sizing for High-Density AI Data Halls
- Dry Cooling in the West Texas Climate
Frequently asked questions
How much flow does a 100 kW rack need?
It depends on liquid heat fraction, coolant properties and allowed temperature rise; the OEM should provide the operating envelope.
Can all racks share one CDU?
Large shared CDUs can be efficient but create larger fault domains; modularity and redundancy should guide the design.
What is residual heat?
Heat from components not connected to the liquid loop, plus system and room gains.
Why does supply temperature matter?
Higher temperatures can improve heat-rejection efficiency, but must remain within server and coolant requirements.
Next step
W Land is engaging with AI operators, hyperscale developers, energy partners, equipment suppliers and infrastructure investors regarding a planned West Texas private-power AI data center campus.
Request a 30-minute NDA briefing to review the 100 MW Phase 1 development concept, 500 MW+ expansion strategy, equipment architecture and U.S. qualification process.
Editorial qualification
This draft is educational and commercial content, not legal, engineering, permitting, fire-code or investment advice. Final public claims should be reviewed by W Land’s licensed U.S. engineers, permitting counsel, equipment vendors, tenant representatives and brand/legal teams. Standards, regulations, products and market conditions should be rechecked immediately before publication.