Accelsius Austin Lab Expansion Puts Two-Phase Cooling Under Megawatt Pressure

Accelsius upgraded its Austin testing facility to multi-megawatt power, bringing the Accelsius Austin lab expansion into line with emerging megawatt-class AI rack designs. The company says the site can now test one rack above one megawatt or several racks at current high-density loads.

That is more consequential than a routine laboratory upgrade. Cooling vendors have spent years proving that two-phase systems can remove heat from individual processors. Accelsius now wants to validate the entire path, from cold plate to coolant distribution unit and facility loop, under full-rack loads.

The move also intensifies a contest between two-phase and established single-phase direct-to-chip cooling. Single-phase systems have a larger operating base and familiar service practices. Two-phase suppliers argue that boiling a dielectric refrigerant at the chip offers more thermal headroom as rack power rises.

Accelsius has not disclosed the precise electrical ceiling, equipment list, customer identities, or test schedule for the upgraded facility. Those gaps matter. A multi-megawatt test environment creates the conditions for credible validation, but it does not by itself establish commercial reliability at that scale.

The Accelsius Austin Lab Expansion Moves Testing Beyond Components

The central change is not a new cooling product. It is Accelsius gaining the electrical capacity to test complete systems at loads associated with future AI infrastructure.

The company announced the Springdale facility upgrade on October 12, 2026, during the Open Compute Project Global Summit. Its lab expansion adds multi-megawatt power for full-scale, warm-water testing of the NeuCool platform.

Accelsius also plans to add 18,000 square feet to the leased site in January 2027. The new space should support more equipment and customer qualification work, although the company has not detailed its intended allocation.

The distinction between component testing and full-system testing is important. A cold plate can perform well on a laboratory bench while the complete installation encounters pump, control, flow-distribution, or heat-rejection constraints.

Accelsius says Springdale can characterize the whole thermal chain. That includes the cold plates attached to processors, the internal refrigerant circuit, the coolant distribution unit, and the water loop carrying heat away from the rack.

A coolant distribution unit, or CDU, transfers heat between the equipment-side cooling loop and the building-side water system. At higher rack densities, CDU controls and heat exchangers become critical parts of the performance envelope.

The company says it can power a single rack above one megawatt or operate multiple racks at today’s highest densities. It did not say that a production NeuCool system is already commercially rated for a one-megawatt rack.

That difference deserves emphasis. The lab creates room to test emerging designs before those designs become standard products. It does not turn a development configuration into a field-qualified system overnight.

Accelsius links the investment to public roadmaps pointing toward 600-kilowatt racks in 2027 and later megawatt-class rack or pod designs. Those figures represent anticipated infrastructure requirements, not today’s typical production data center.

The facility also supports testing with warm facility water. This means the building sends relatively warm water to the cooling equipment, reducing reliance on energy-intensive mechanical chilling when outdoor conditions permit.

Central Texas gives Accelsius a useful setting for those trials because summer temperatures routinely exceed 100°F. Engineers can examine cooling behavior in a hot climate instead of extrapolating only from controlled, favorable ambient conditions.

The laboratory therefore addresses two variables at once: rising rack power and rising facility-water temperature. A cooling system that handles only one of them may still impose high costs or siting restrictions on operators.

The Accelsius Austin lab expansion is best understood as validation infrastructure for a future product envelope. Its value will depend on the test results, customer participation, and eventual translation into repeatable commercial deployments.

AI Rack Roadmaps Are Outrunning Today’s Cooling Test Beds

AI infrastructure planners increasingly need evidence at rack and row scale, because processor-level thermal results cannot expose every integration failure.

Modern accelerator clusters concentrate far more electrical power into each rack than conventional enterprise computing. Nearly all that electricity becomes heat, which must leave the processors without forcing them to throttle performance.

Air cooling becomes harder as power density rises because moving enough air demands larger fans, wider passages, and substantial facility space. Direct-to-chip liquid cooling moves heat through cold plates mounted on high-power components.

Single-phase direct-to-chip systems circulate water or a water-based mixture that remains liquid. Two-phase systems use a refrigerant that boils inside the cold plate, carrying heat through its phase change before condensing elsewhere in a closed loop.

The latter approach can move large heat loads with lower refrigerant flow, but it introduces its own engineering demands. Designers must control boiling behavior, pressure, vapor distribution, condensation, and refrigerant inventory across changing workloads.

Those interactions become more complicated when dozens of accelerators change power states together. AI training jobs can create sharp load changes, and a complete rack must respond without unstable flow or unacceptable temperature excursions.

This is why Accelsius CEO John Hewitt described the industry as moving from proving physics to scaling products. The claim captures the company’s intended transition, although commercial evidence remains less extensive than laboratory evidence.

Accelsius already sells systems below the megawatt level. Its IR150 combines a two-phase CDU, 42U of usable IT space, and integrated liquid and vapor manifolds in an 800-millimeter-wide enclosure.

The IR150 design is rated for up to 150 kilowatts. Accelsius says it supports NVIDIA Blackwell hardware and was designed with thermal headroom for later platforms.

The row-based MR250 supports more than 250 kilowatts per rack in supported configurations. Accelsius has also announced a 375-kilowatt-capable unit, with more information promised in the coming months.

A jump from those ratings to megawatt-class validation is substantial. Pumps, manifolds, controls, heat exchangers, power delivery, and service procedures must all scale without introducing unacceptable failure modes.

Operators also need to understand how failures propagate. A rack-level CDU can confine a cooling problem to one rack, while a larger row-level design can share capacity more efficiently across several racks.

Neither architecture wins automatically. Operators must balance utilization, redundancy, maintenance access, floor space, capital equipment, and the consequences of taking any cooling unit offline.

The new lab gives Accelsius room to test these tradeoffs with physical equipment. It can also help customers evaluate reference designs before committing comparable hardware to a production site.

This matters because data center construction timelines are long. Cooling architecture, piping, water temperatures, and equipment clearances can become fixed well before every server configuration reaches final production.

A customer planning a 2027 or 2028 facility cannot wait until rack hardware arrives to investigate thermal compatibility. It needs credible models and representative testing while the facility remains adjustable.

The pressure therefore falls on cooling suppliers, server manufacturers, chip vendors, and data center operators together. Their designs must converge early enough to avoid stranded electrical or mechanical capacity.

Two-Phase Cooling Must Beat a Familiar Single-Phase Route

Accelsius is not competing against air cooling alone. Its harder opponent is single-phase direct-to-chip cooling, which already fits established water systems and operating practices.

Single-phase cooling has a straightforward proposition. Water or a treated mixture absorbs heat as its temperature increases, then transfers that heat to the facility loop through a CDU.

Data center teams understand pumps, filters, water chemistry, and leak management. Server manufacturers have also introduced liquid-cooled platforms around this model, giving buyers a broader integration path.

Two-phase direct-to-chip cooling replaces the server-side water mixture with a dielectric refrigerant. Heat from a processor boils the liquid inside a cold plate, and the resulting vapor moves toward a condenser.

The phase change absorbs substantial thermal energy without requiring the fluid to undergo a large temperature increase. The condensed refrigerant then returns to the processor loop.

Because the fluid does not conduct electricity, suppliers argue that leaks present less danger to electronics than water. However, operators still need reliable seals, refrigerant management, sensors, controls, and maintenance procedures.

Accelsius says its architecture keeps water outside the IT rack and supports facility water temperatures up to 45°C in the IR150. Higher water temperatures can extend the periods when a site rejects heat without compressor-based chilling.

The company has published results from a third-party systems integrator that tested an eight-GPU NVIDIA B200 server across roughly 40,000 operating points. The warm-water benchmarks compared the server’s original single-phase cooling with NeuCool.

According to Accelsius, its two-phase system produced NVIDIA B200 junction temperatures 9°C lower under one warm-water comparison. It reportedly matched the single-phase system’s performance at a 45°C inlet while using 54°C inlet water.

Increasing facility-water flow from 1.5 to 3.0 liters per minute reportedly widened the temperature advantage to 14°C. Accelsius says that configuration supported facility-water supply temperatures as high as 59°C.

These are meaningful results, but readers should keep the evidence boundaries clear. Accelsius published the findings, and the public summary does not identify the systems integrator that conducted the work.

The company also says each 1°C increase in facility-water temperature corresponds to about 4 percent in annual cooling-energy savings. Actual savings will depend on climate, facility design, workload, operating schedule, and heat-rejection equipment.

Independent analysis cited by Accelsius and its parent estimates lower operating expenses and five-year ownership costs than single-phase alternatives. Those estimates depend on project assumptions rather than a universal economic result.

The competitive field also extends beyond Accelsius. ZutaCore has developed two-phase direct liquid cooling and worked with SoftBank and Foxconn on a rack using NVIDIA H200 GPUs.

A 2025 rack demonstration recorded a partial power usage effectiveness of 1.03, according to the participating companies. The system also passed NVIDIA’s NVQual temperature testing.

That project demonstrates that Accelsius does not own the two-phase category. Multiple suppliers are pursuing similar physics, and customers can compare architectures, fluids, service models, manufacturing capacity, and server support.

Meanwhile, single-phase suppliers continue improving cold plates and CDUs. Their technology does not remain static while two-phase vendors work toward higher densities.

The relevant contest is therefore not advanced cooling versus an obsolete alternative. It is a competition between different liquid-cooling architectures, each improving against the same processor roadmaps.

Springdale can help Accelsius make its case, especially if customers test complete racks under their intended conditions. The decisive evidence will come from repeatable operation, maintainability, and economics outside a vendor-controlled lab.

Multi-Megawatt Capacity Does Not Equal Field-Proven Reliability

The biggest uncertainty is whether high-power laboratory performance will translate into dependable operation across years, sites, and changing server generations.

A laboratory can reproduce power loads and water temperatures, but it cannot compress every field condition into a short qualification program. Long-term seal behavior, contamination, component aging, and maintenance quality emerge over time.

Two-phase systems also depend on stable refrigerant behavior throughout a complex loop. Differences in elevation, manifold geometry, workload distribution, and control response can affect real deployments.

The industry must consider refrigerant availability and environmental characteristics as well. A fluid can be nonconductive and thermally effective while still creating sourcing, handling, regulatory, or end-of-life questions.

Accelsius describes its refrigerant as having an A1 safety classification and low global-warming potential. Buyers will still need detailed fluid documentation, supply assurances, service procedures, and jurisdiction-specific compliance reviews.

Vendor scale presents another issue. Hyperscale operators assess whether suppliers can manufacture, install, and support equipment across multiple regions without creating a fragile dependency.

Accelsius raised substantial financing in 2026 and has expanded its product line, but it remains smaller than many established data center infrastructure providers. Its manufacturing and service footprint must grow alongside customer demand.

The company says several hyperscalers are participating in its HyperStart qualification program. It has not named those operators or disclosed the number, scale, duration, or results of their evaluations.

That confidentiality is understandable during product development. It also means outsiders cannot yet distinguish exploratory testing from a committed production pipeline.

Commercial deployments offer another validation layer. Accelsius announced an IR150 installation at the EdgeConneX Global Training Center in Houston, where technicians can practice deployment and servicing.

The training deployment is useful because unfamiliar maintenance requirements can block otherwise capable technology. However, a training center does not reproduce the uptime demands of a large production AI cluster.

Accelsius has also discussed deployments linked to larger planned data center projects. Those announcements indicate customer interest, but planned capacity should not be treated as installed and operational capacity.

The Springdale facility can narrow several evidence gaps. It can exercise full racks, simulate changing workloads, compare water temperatures, and test control responses before customers deploy equipment.

It can also expose undesirable system interactions earlier. A failed test inside a development lab costs far less than a thermal problem inside an accelerator cluster running valuable training workloads.

Yet the company must publish enough methodology for customers to evaluate the findings. Test duration, server configuration, measurement uncertainty, failure simulations, and comparison baselines all affect credibility.

The strongest results would involve named independent laboratories, server manufacturers, or data center operators. Repeatable testing across more than one facility would provide greater confidence than a single-site demonstration.

Customer reference designs would add another signal. If hyperscalers formally include NeuCool in approved architectures, the technology has passed a deeper operational and procurement process.

Service data will matter just as much as thermal data. Operators need replacement times, alarm behavior, spare-parts requirements, technician training, and documented recovery procedures.

A cooling system can maintain excellent chip temperatures and still disappoint if maintenance frequently interrupts compute. Conversely, familiar operations can compensate for a modest laboratory disadvantage.

Accelsius should therefore be judged on a larger reliability equation. Thermal headroom is one variable among availability, serviceability, supply resilience, environmental compliance, and total facility cost.

The expanded lab is a credible response to that challenge. It creates a place where more of those variables can be tested together, but it does not settle them.

The Austin Facility Becomes a Customer Qualification Tool

The commercial value of Springdale lies in shortening the path from promising thermal data to an approved data center design.

Cooling choices affect building systems, server procurement, rack layouts, and operating procedures. A buyer cannot treat them as interchangeable accessories late in a project.

Accelsius wants HyperStart customers to bring their expected conditions into the lab. Those conditions can include rack power, facility-water temperatures, server layouts, redundancy requirements, and anticipated load changes.

This approach can help both sides. Customers receive performance data before construction decisions become irreversible, while Accelsius learns which configurations create integration or service problems.

The upgraded electrical supply also allows more realistic comparisons between architectures. Engineers can hold workload and facility conditions constant while changing cooling equipment or control strategies.

Such comparisons are difficult when test sites cannot power a complete high-density rack. Scaling results from one server can miss system-wide constraints, especially in manifolds and shared heat-rejection equipment.

Springdale may also support server and chip partners. A server maker could evaluate cold-plate interfaces and internal routing before finalizing a platform around an emerging accelerator.

Facility designers can test warmer water set points and estimate whether a site can reduce chiller operation. Those decisions influence electrical demand, equipment sizing, and local water requirements.

The lab’s Central Texas location strengthens this use case. Hot-weather trials can reveal whether warm-water operation remains practical when outside conditions become least favorable.

Still, a vendor site controls the equipment, instrumentation, and test procedures. Customers should request raw data, documented calibration, repeat runs, and clear separation between measured results and modeled savings.

They should also test abnormal conditions. Useful scenarios include pump failures, blocked flow paths, sensor faults, rapid workload changes, partial rack shutdowns, and maintenance while neighboring equipment remains active.

The Accelsius Austin lab expansion will be most persuasive when customers use it to try to break the system. Passing only nominal performance tests would leave the most expensive operational questions unresolved.

The company’s planned 375-kilowatt unit provides an early opportunity. Public specifications, availability, independent qualification, and customer testing will show whether the roadmap is keeping pace with the lab.

A later megawatt-capable commercial design would face an even higher standard. At that density, a short cooling interruption can threaten an enormous concentration of compute hardware.

Operators may respond by dividing workloads across smaller failure domains instead of placing a full megawatt behind one cooling unit. The lab should help compare those architectural choices rather than assume maximum density is always best.

This is where the event becomes strategically important. Accelsius is investing not only in thermal capacity, but also in a venue for influencing future reference designs.

If customers validate NeuCool at Springdale and incorporate it into planned facilities, the lab becomes part of the company’s sales and qualification infrastructure. If evaluations remain private experiments, its market impact will be smaller.

Three Signals Will Show Whether the Expansion Matters

The next test is evidence, not additional claims about the theoretical advantages of boiling refrigerant at the chip.

The first signal is detailed performance from the 375-kilowatt NeuCool system. Accelsius has announced the capacity but has not yet provided complete specifications, independent comparisons, or a commercial schedule.

Results should include inlet temperatures, processor loads, coolant flow, control behavior, and sustained test duration. Data from failure and recovery scenarios would strengthen the case more than peak-capacity demonstrations.

Successful independent qualification would support the argument that Accelsius can scale from its 150-kilowatt and 250-kilowatt products. Delays or limited disclosures would weaken that interpretation.

The second signal is a named customer moving from evaluation to a production reference design. Accelsius says several hyperscalers already use HyperStart, but anonymous engagement does not establish adoption.

A meaningful announcement would identify the operator, server platform, rack density, facility-water conditions, and deployment stage. It should distinguish a test agreement from an ordered production system.

Production operation across multiple racks would be stronger evidence than another showcase installation. Published availability and service data would be stronger still.

The third signal is credible megawatt-class testing at Springdale. Accelsius should show how it powers the load, rejects the heat, measures performance, and manages rapid changes across a complete system.

Independent observers should be able to determine whether the test represents a realistic AI rack or a collection of thermal load banks. Both have engineering value, but they answer different questions.

The results should also clarify which NeuCool components are production hardware and which remain developmental. That distinction will prevent readers from confusing laboratory capability with product availability.

These signals will arrive against active competition. Single-phase suppliers will continue raising capacity, while other two-phase vendors pursue similar customer qualifications and partnerships.

No single benchmark will determine the market. Data center operators usually favor architectures that combine adequate performance with predictable maintenance, diverse suppliers, and manageable integration risk.

That reality gives Accelsius both an opportunity and a burden. Megawatt-class roadmaps make its thermal argument more relevant, but they also raise the cost of any failure.

The Accelsius Austin lab expansion is therefore a serious infrastructure commitment, not proof that two-phase cooling has already won. It gives the company a better stage on which to prove its claims at realistic scale.

For buyers, the practical next step is to ask for test conditions that match their own planned facilities. Compare sustained performance, service procedures, failure domains, water temperatures, and supply-chain support before choosing an architecture. Watch whether named operators move beyond trials, whether the 375-kilowatt system reaches independent testing, and whether Springdale produces transparent megawatt-scale results. Those outcomes will determine whether Accelsius has built a bridge to production or simply a larger laboratory for an unsettled cooling contest.