Apheros and 3D Architech Target the Cooling Problem Inside AI Data Centers

Apheros and 3D Architech are pursuing two different answers to a conflict that now shapes AI infrastructure. Faster chips generate more heat, while existing cooling systems consume electricity, water, and valuable space. The startups want to redesign the hardware that moves heat away from processors.

Their timing is not accidental. AI servers place more computing capacity inside each rack than traditional cloud systems. That density turns cooling from a supporting utility into a limit on how much equipment operators can deploy.

The two companies approach the limit from opposite directions. Apheros develops porous metal structures for liquid cooling. 3D Architech prints intricate copper components intended to transfer heat more efficiently. Both are betting that better materials can delay much larger changes to data center buildings and power systems.

That promise still needs careful scrutiny. Laboratory performance does not guarantee affordable mass production, reliable operation, or straightforward installation. Data center operators also avoid technologies that introduce uncertain maintenance requirements.

The real contest is therefore not one startup against another. It is advanced thermal hardware against the cost, familiarity, and installed base of conventional cooling.

Two Startups Are Redesigning the Path Out of a Chip

Apheros and 3D Architech are changing the physical structures that carry heat away, rather than asking operators to cool entire rooms more aggressively.

The distinction matters because heat starts at the chip. Electrical resistance converts part of a processor’s energy into thermal energy. That heat must cross several interfaces before a cooling system can release it outside the facility.

Each interface creates resistance. A better fan cannot fully compensate for a poor connection between the chip, heat spreader, cold plate, coolant, and external heat exchanger.

Apheros focuses on porous metal foam. Its material contains an open network of small passages and a large internal surface area. Coolant can move through those passages while contacting far more metal than it would inside a plain channel.

The company says its manufacturing process produces open-pore structures suited to liquid cooling. Its current positioning centers on metal-foam cold plates, which sit close to heat-producing components.

A cold plate is a metal component containing coolant channels. It absorbs heat from a processor and transfers that energy into circulating liquid.

Apheros emerged from ETH Zurich and was founded in August 2023. Co-founder and CEO Julia Carpenter developed the underlying manufacturing approach during her academic work. Co-founder and CTO Gaëlle Andreatta brought experience in research, technology transfer, and startup development.

In August 2024, Apheros announced a CHF 1.6 million pre-seed round, described at the time as approximately $1.85 million. Founderful led the financing.

The company said the capital would support production expansion, research, and market deployment. Those plans remain company objectives, rather than independent evidence of commercial performance.

Its metal foam platform targets an important weakness in conventional cold plates. Smooth channels limit the surface available for heat exchange. Porous structures can create much more contact between coolant and metal.

That advantage comes with tradeoffs. Complex passages can increase pressure requirements, trap contaminants, or complicate manufacturing. Operators must evaluate those factors alongside thermal performance.

3D Architech takes a different route. The MIT spinout uses a lithography-based additive manufacturing process to create metal components with very small features.

Lithography uses patterned light or related processing methods to define detailed structures. Additive manufacturing then builds the designed object instead of cutting it from a larger block.

The company works with copper and copper alloys because copper conducts heat well. Its microscale copper printing can create porous architectures that would be difficult to produce through machining.

For cooling, the resulting structures can act as compact heat sinks or fluid-handling components. A heat sink increases the area available to transfer heat away from electronics.

The startup’s public materials also point toward energy conversion and flow-control applications. However, data center cooling presents the most immediate commercial story because AI hardware has made heat density a board-level concern.

Inc. highlighted both companies in its January 2025 examination of the cooling problem. That coverage framed smaller hardware changes as a practical bridge while more ambitious data center concepts face long development cycles.

The startups are not selling equivalent products. Apheros emphasizes metal foam integrated into liquid-cooling systems. 3D Architech emphasizes the manufacturing process and geometries that make compact metal structures possible.

Their common thesis is more important than their technical differences. Heat should be captured efficiently near its source, before operators spend more energy moving air around a building.

AI Infrastructure Has Turned Cooling Into a Capacity Constraint

The pressure comes from rising rack density, because operators cannot deploy more computing equipment if they cannot remove its heat reliably.

Traditional data centers often rely on air cooling. Fans move air through server enclosures, while larger systems transport warm air toward cooling equipment.

This approach is familiar and serviceable. Technicians can replace many components without handling facility-wide liquid loops near expensive electronics.

Air also has physical limits. It carries less heat than liquid, and large airflows require fans, ducts, floor space, and careful containment. Those requirements grow as more high-power accelerators occupy each rack.

Accelerated servers contain specialized processors, including graphics processing units, that handle highly parallel AI workloads. They account for a growing portion of new data center demand.

The International Energy Agency estimates that data centers consumed about 415 terawatt-hours of electricity worldwide in 2024. That represented roughly 1.5 percent of global electricity consumption.

Its global electricity outlook projects about 945 terawatt-hours in 2030 under the base case. Accelerated servers account for almost half of the projected increase.

Cooling is not responsible for all that demand. Servers remain the largest load, and the cooling share varies across facility types, climates, and operating conditions.

Still, every unit of server electricity eventually becomes heat. The facility must remove that heat continuously, including during periods of high outside temperatures or grid stress.

Power usage effectiveness, or PUE, compares total facility electricity with the electricity consumed by computing equipment. A lower PUE indicates less overhead from cooling, power conversion, lighting, and other infrastructure.

PUE is useful, but it does not settle the cooling question. A facility can report an efficient ratio while still consuming enormous amounts of electricity. The metric also says little about water use or component-level temperatures.

The United States illustrates the scale of the constraint. Berkeley Lab estimated that data centers consumed 176 terawatt-hours in 2023, or about 4.4 percent of national electricity.

Its U.S. energy forecast projects consumption between 325 and 580 terawatt-hours in 2028. That would equal between 6.7 and 12 percent of U.S. electricity.

The range is wide because chip shipments, server utilization, efficiency, and AI adoption remain uncertain. Nevertheless, even its low end implies substantial growth.

This is where cooling startups gain leverage. A better thermal component cannot create electricity, accelerate transmission construction, or secure a grid connection. It can help operators fit more computing capacity inside the power and thermal envelope they already have.

That opportunity pressures established equipment suppliers as well as data center operators. Existing cooling vendors must adapt products designed around lower-density racks. Operators must decide whether to retrofit buildings or reserve liquid-ready capacity in new construction.

Chipmakers face related pressure. Processor performance loses commercial value when customers must reduce clock speeds or leave rack positions empty because of thermal limits.

Cloud providers must balance speed against standardization. A custom cooling architecture can improve density, but it can also restrict hardware choices and complicate maintenance across multiple facilities.

Enterprise buyers encounter the consequences through availability and operating costs. They rarely choose a cold plate directly, but constrained infrastructure can affect where AI services run and how quickly capacity becomes available.

Apheros and 3D Architech therefore occupy a narrow but consequential layer. They do not need to build an entire data center. They need their materials to solve a painful constraint without creating a larger operational problem.

Better Geometry Is Competing With Familiar Cooling Systems

The startups’ main challenge is proving that thermal gains outweigh manufacturing, pumping, integration, and maintenance costs over years of operation.

Advanced geometry is the mechanism behind both approaches. More surface area creates more opportunities for heat to cross from hot metal into air or liquid.

Yet surface area alone does not determine system performance. Engineers must also consider flow resistance, temperature uniformity, corrosion, pressure, fouling, and the energy used by pumps.

Apheros says its foam contains open pores and a very large internal surface. That structure can promote contact between coolant and metal throughout a compact volume.

The desired result is a cold plate that transfers more heat without becoming dramatically larger. That could support dense AI accelerators while preserving valuable space around the server board.

However, narrow and irregular flow paths can raise pressure drop. Pressure drop is the loss of fluid pressure as coolant moves through a component.

A larger pressure drop can require stronger pumps. Those pumps consume electricity and can reduce some of the system-level energy benefit promised by better heat transfer.

Designers can adjust pore size, thickness, flow rate, and coolant chemistry. The final advantage depends on the complete loop, not a material sample tested under ideal conditions.

Manufacturing consistency also matters. Two cold plates intended for identical servers should behave predictably. Small differences in internal geometry can alter flow distribution and temperatures.

3D Architech’s process offers another form of design freedom. Engineers can create heat sinks with fine copper features and complex channels that conventional machining cannot reproduce economically.

Copper presents its own manufacturing difficulty. It reflects common laser wavelengths and conducts heat away from the active printing area. Both characteristics can complicate metal additive manufacturing.

A lithography-based approach can potentially avoid some limits associated with direct laser processing. It also allows many small structures to be patterned together.

The commercial question is throughput. A process that creates impressive samples may still be too slow, expensive, or variable for data center volumes.

Neither startup can win through peak thermal performance alone. Data center components operate continuously, and failures can interrupt costly computing workloads.

Operators will ask how materials behave after repeated heating and cooling cycles. They will examine corrosion, vibration, clogging, coolant compatibility, and leakage risk.

They will also ask whether technicians can inspect and replace the component using established procedures. A cooling system that requires specialized service may remain confined to a small group of high-value installations.

Conventional heat sinks and cold plates begin this contest with major advantages. Suppliers understand their manufacturing tolerances, expected lifetimes, and failure modes. Operators already have procurement and maintenance processes built around them.

Air cooling also continues to improve. Better airflow management, rear-door heat exchangers, and facility controls can extend its usefulness in some environments.

Direct-to-chip liquid cooling does not necessarily eliminate air. Memory, storage, power supplies, and networking components can still require airflow. Hybrid systems may therefore persist across many facilities.

Immersion cooling presents another competing route. It places electronic equipment in a dielectric liquid, which does not conduct electricity under normal operating conditions.

Immersion can absorb heat across a large portion of the server. However, it requires compatible hardware, specialized tanks, fluid management, and altered maintenance practices.

Apheros and 3D Architech fit most naturally into direct, component-level cooling. Their opportunity rests on offering meaningful improvement without demanding a wholesale redesign of the server room.

That positioning is commercially sensible. Operators generally prefer changes that can enter existing supply chains and validated reference designs.

It also narrows the measurable value. A metal foam or printed heat sink must show how much it improves temperatures, flow, pumping energy, or computing density against a standard component.

Company claims about superior cooling should remain claims until customers publish comparable results. Independent tests need matched workloads, coolant temperatures, pump power, and environmental conditions.

Without those details, percentage improvements can mislead. A component can perform well in isolation while delivering a smaller benefit at rack or facility level.

The Data Center Cooling Problem Extends Beyond Heat Sinks

Component innovation can reduce thermal resistance, but it cannot resolve grid delays, water disputes, or the emissions tied to rapid computing growth.

The focus on cooling hardware risks making the infrastructure problem appear smaller than it is. Data centers need electricity generation, transmission, substations, backup systems, land, water, and construction capacity.

A more efficient cold plate does not guarantee that a proposed facility receives a grid connection. It also does not determine whether local residents accept the project.

The environmental balance depends on the cooling design. Some facilities use evaporative systems that consume water to reject heat. Others use air-cooled chillers with different electricity and performance tradeoffs.

Liquid cooling inside a server does not automatically remove water use. The internal loop still needs an external system that transfers heat into the surrounding environment.

Warm-water systems can improve that equation. If coolant leaves the server at a higher temperature, operators may use outside air more often and run mechanical chillers less frequently.

The result depends on climate and design. A strategy that works well in a cool region can perform differently during a hot, humid summer.

Heat reuse offers another possibility. Data centers can transfer waste heat to nearby buildings, industrial processes, or district-heating networks.

That option requires customers close enough to use the heat. It also needs temperatures, contracts, and infrastructure that align with local demand.

This broader context explains why ambitious cooling experiments attract attention. Microsoft placed a sealed data center containing 855 servers underwater near Scotland’s Orkney Islands in 2018.

The company retrieved the unit in 2020 and reported a lower server failure rate than its comparable land-based equipment. It attributed part of the difference to stable temperatures and a dry nitrogen atmosphere.

The underwater cooling trial showed that unusual architectures can work technically. It did not establish a widely adopted commercial model.

Underwater deployment introduces difficult maintenance, cabling, environmental, and permitting questions. The project also demonstrates why operators value less dramatic improvements that fit conventional buildings.

That is the reversal behind the current startup interest. Data centers appear digital, but their hardest limits now involve materials, fluids, construction, and power systems.

Software can improve workload scheduling and cooling controls. It cannot repeal the thermal properties of air, water, aluminum, or copper.

The startups also face a timing problem. AI infrastructure spending creates urgency, but data center qualification cycles reward patience.

Operators cannot deploy an untested thermal component merely because new accelerators are arriving. They need reliability evidence, supplier capacity, and clear responsibility when something fails.

Large incumbents can bundle cooling equipment with warranties and service coverage. A small materials company may depend on manufacturing partners or established equipment vendors to reach the same buyers.

That dependence can shape who captures the value. A startup might supply a critical structure while a larger company owns the customer relationship and complete system.

Intellectual property offers some protection, especially when the manufacturing method produces a structure competitors cannot easily copy. Scaling knowledge can provide another barrier.

Those advantages still require commercial discipline. Founders must decide whether to sell materials, components, manufacturing licenses, or engineering partnerships.

Each model carries different capital needs. Direct manufacturing requires equipment, quality systems, and inventory. Licensing offers lower production exposure but can reduce control over implementation.

Market growth does not ensure that either company becomes a major supplier. Large demand often attracts established manufacturers with strong purchasing relationships and substantial engineering teams.

The cooling market may also fragment. Different processors, server designs, rack densities, and facility standards can require distinct thermal solutions.

Apheros could find its strongest position in a specific cold-plate design rather than across all data centers. 3D Architech could gain traction in specialized applications before reaching hyperscale volumes.

Such outcomes would not invalidate the underlying technology. They would show that commercialization follows integration boundaries, not the breadth of the original vision.

What Independent Validation Must Show

The next stage depends on repeatable customer evidence, not another demonstration of an unusually shaped metal sample.

The first signal to watch is qualification inside production server hardware. A credible test should use representative processors, power levels, coolants, and operating cycles.

It should also compare the new component against an established alternative. The results should include chip temperature, coolant flow, pressure drop, and pump electricity.

If Apheros secures a disclosed server or cooling-equipment partner, its commercial case becomes stronger. Such a relationship would suggest that the foam can meet integration and manufacturing requirements.

The absence of named partners would not prove failure. Data center supply agreements often remain confidential during development.

Still, public deployments matter because company laboratory results cannot reveal every operating condition. Production environments expose components to contamination, maintenance variation, and long duty cycles.

The second signal is manufacturing yield. Yield measures how many produced parts meet the required specifications without rework or disposal.

A high-performance process can struggle commercially if too many components contain defects. Fine internal features can be especially difficult to inspect with ordinary methods.

3D Architech must show that its process can repeat microscale copper structures across meaningful production volumes. Customers will want stable dimensions and thermal behavior from batch to batch.

Cost comparisons must cover more than the part itself. A more expensive heat sink can still make sense if it enables additional processors per rack or reduces cooling infrastructure.

Conversely, a low component cost cannot compensate for difficult installation or higher pumping requirements. Buyers will evaluate total ownership and operational risk.

The third signal is adoption by established system vendors. Cooling specialists, server manufacturers, and chip companies define many of the interfaces a startup must use.

A design win with such a vendor would strengthen the thesis that advanced materials can enter mainstream data centers without forcing a complete architectural shift.

It would also clarify where these startups sit within the supply chain. A component supplier operates differently from a company responsible for the entire cooling loop.

Standards will influence adoption. Common connectors, coolant specifications, leak-detection systems, and facility interfaces reduce the risk of committing to one vendor.

Operators should also demand environmental accounting at system level. Lower chip temperatures do not automatically mean lower electricity or water consumption.

A strong assessment should include pumps, fans, heat rejection, coolant production, component manufacturing, and expected service life. It should also examine whether higher density merely encourages more total computing.

That rebound effect matters. Efficiency can lower the resources required for each unit of computation while total consumption continues climbing.

The IEA expects data center electricity demand to more than double by 2030 under its base case. Component efficiency can moderate that trajectory, but it does not reverse the demand behind it.

For enterprise buyers, the practical lesson is to ask infrastructure providers specific questions. Buyers should examine where workloads run, how capacity constraints affect availability, and which environmental metrics are independently reported.

For developers, cooling may appear distant from application design. Yet model choice, inference frequency, and workload scheduling determine how much computing infrastructure an application ultimately requires.

Knowledge workers should care for a different reason. AI services often hide their physical supply chain behind a simple interface. Availability and cost still depend on chips, cooling equipment, grid connections, and construction schedules.

Apheros and 3D Architech make that physical layer visible. Their work suggests that some of AI’s most valuable advances will come from quieter changes in materials and manufacturing.

Their success is not guaranteed. Both must bridge the distance between a technically interesting structure and an ordinary, trusted infrastructure component.

Watch for production partners, comparable thermal tests, and evidence of repeatable manufacturing. Those signals will reveal whether advanced metal structures become standard cooling hardware or remain specialized engineering achievements.

The next data center winner may not build a model or design a processor. It may remove heat more efficiently from hardware that everyone else is racing to install.