What Supersonic Engines Can Teach Us About Cooling AI Racks

If you build AI infrastructure, sell it, or finance it, cooling is no longer a facilities detail. It is now part of the product stack. Rack density is rising, cooling is eating into system economics, and the wrong thermal architecture can turn deployment speed, uptime, and capex into a single problem.

That is why supersonic propulsion is worth looking at now. Not because data centers are becoming aircraft, but because both domains force engineers to deal with the same class of constraint: heat is arriving faster than simple airflow can remove it. Once that happens, thermal design stops being a background issue and becomes a decision about geometry, fluids, power, maintenance, and cost.

The useful question is narrower than the headline suggests: what thermal ideas transfer from supersonic engines to AI racks, and what does not?

Why the aerospace analogy matters

Supersonic propulsion forces engineers to confront a limit ordinary systems can ignore. Purdue Engineering notes that in subsonic engines, compressor bleed air can be cool enough to use directly for downstream cooling, but in supersonic engines that bleed air is too hot and has to be precooléd through a heat exchanger first. Purdue also shows how heat-sink choice matters: in some engine cycles, fuel can serve as a more compact heat sink than air. ¹

"In subsonic engines, compressor bleed air is cool enough to be used directly for downstream cooling purposes. However, in supersonic engines, compressor bleed air temperature is quite high and therefore a heat exchanger is needed to precool the air before it can effectively cool the downstream engine components."

— Purdue Engineering ¹

That is relevant to AI infrastructure because the modern rack has crossed into a regime where air alone is often not enough. Recent data-center sources in this set describe rack power densities in the 100–200 kW range and frame the industry as facing a thermal crisis rather than a routine facilities upgrade. ²

The overlap is not literal hardware transfer. It is the shared engineering reality that, once heat density gets high enough, the path heat takes through the system matters as much as the raw cooling capacity.

What transfers, and what does not

The transferable part is the thermal logic:

• move heat through a controlled path rather than just hoping ambient air will absorb it;
• reduce thermal resistance between the hot surface and the sink;
• use fluids and phase change when airflow becomes too weak or too expensive;
• treat the cooling loop as part of the overall power and reliability budget. ^{1, 3, 4}

The non-transferable part is the operating environment. A supersonic engine is optimized for weight, flight dynamics, and engine-cycle performance. A data center is optimized for uptime, serviceability, energy cost, and facility integration. Those are not small differences. A design that makes sense in a propulsion lab can become a bad data-center product if it requires exotic maintenance, rare parts, or a custom supply chain.

That distinction is the reason the aerospace analogy is useful but dangerous. It can clarify the physics while obscuring the business case.

The common thread: heat sinks, exchangers, and loop design

Aerospace thermal systems often begin with the question, “What is the heat sink?” Purdue’s supersonic-engine material describes modular air-fuel heat exchangers built from parallel mini-channels, and it notes that fuel can sometimes be a more compact and lightweight sink than air. ¹

Modern data-center cooling is asking a similar question, just in a different form factor. One 2026 study of pumped two-phase refrigerant-to-liquid cooling reports more than 160 kW of dissipation per rack, with flow around 0.5 LPM/kW and pump energy reduced by an order of magnitude versus single-phase liquid cooling. ³

"Pumped two-phase D2C leverages latent heat to cut required mass flow to ~0.5 LPM/kW and minimize temperature rise, reducing pump energy by an order of magnitude and enabling very low partial PUE values when paired with efficient heat rejection."

— Ut Dagor ³

That does not mean data centers should copy aircraft architecture. It means the best thermal systems in both fields tend to win in the same way: absorb, move, reject. They shorten the path between the hottest point and the heat sink, then spend less energy doing it.

The hard part is that data-center systems must do this under far less forgiving conditions. They need to work across mixed workloads, survive partial failures, fit into existing facilities, and stay maintainable by operators who do not have aerospace tolerances or aerospace budgets.

Why rack-scale cooling has become a strategic constraint

The infrastructure context is what makes this interesting for builders and investors. The a16z coverage in the source set says rack power is moving from 5–10 kW toward 150 kW, that new data-center facilities are projected to need 44 GW of power by 2028 versus 25 GW of expected grid expansion, and that only 2% of U.S. electrical contractors are certified for the DC power systems required for high-density AI clusters. ⁵

That makes thermal architecture part of deployment strategy, not just operations. If the cooling system cannot be installed, serviced, powered, or integrated with the rest of the site, the compute never becomes usable.

The recent cooling papers in the source set reinforce that point. A phase-change air-assisted liquid-cooling study reports stable operation under a 50 kW cabinet heat load, a 39.85°C average loop temperature, up to 30% energy savings in one configuration, and annual PUE as low as 1.16 in the modeled cases. But it is still a prototype-oriented result, not proof that every production deployment is solved. ⁶

"The proposed AALC systems provide a practical and scalable solution for thermal management in high-density data centers."

— Lu WANG, Xinyi WANG, Bo ZHANG, Xiaoxuan CHEN, Zhen LI ⁶

That claim is directionally useful, but “scalable” hides the real work: maintenance procedures, redundancy, facility-side compatibility, and supply-chain depth. In practice, the cooling system has to be judged as a whole product, not as a promising thermodynamic result.

Where the aerospace-to-data-center story is already real

The clearest concrete example in the source set is Karman Industries. Los Angeles Times coverage says the startup is using technology derived from SpaceX rocket engines to build a data-center cooling system that lowers electricity use and eliminates water. Its compressors are aerospace-inspired, but the target is ordinary AI chip cooling. ⁷

That is the right kind of crossover to watch. It is not rocket hardware being dropped into a server room. It is a transfer of compressor and heat-management ideas into a problem where efficiency and footprint matter enough to justify hardware risk.

The same report says high-end AI chips need to stay below 150°F for optimal operation and can slow down or shut off above 200°F. ⁷ That makes thermal design a direct constraint on usable compute, not an afterthought.

Karman has also raised $20 million and planned to begin manufacturing its first compressors in Long Beach in 2026. That is a signal that investors see a market for hardware-level cooling innovation, but it is not proof that aerospace-derived approaches will scale cleanly across the industry. ⁷

The part people overstate

This is where the conversation usually gets too big too fast. Supersonic engine lessons are real, but they do not automatically translate into a universal data-center answer.

In propulsion, engineers can accept more complexity because the system is built around a narrow duty cycle and an integrated vehicle. In a data center, the thermal stack has to survive mixed workloads, maintenance windows, site-specific plumbing, and operational teams that value uptime over elegance. A heat exchanger that looks excellent in a test rig can still fail commercially if it is hard to service or too expensive to deploy at scale.

That is why the strongest takeaway is not “aerospace is the future of data centers.” It is that aerospace pushes thermal systems into design territory data centers are only now entering: compact exchangers, phase-change loops, tighter control of thermal resistance, and more attention to the shape of the heat path itself.

The engineering frontier is moving toward shorter thermal paths

The recent data-center studies in the source set show a common direction of travel. They move away from broad air cooling and toward systems that reduce intermediate thermal resistance.

One direct-on-die two-phase jet-impingement study on an NVIDIA Tesla V100 reports a junction-to-coolant thermal resistance of 0.056°C/W, compared with 0.1709°C/W for conventional air-cooled cold plates. It also reports theoretical pumping power of 0.172 W versus 39.9 W for fan-based cooling, and stable operation over a 200-hour reliability test. ⁴

That is not aerospace hardware transplanted into a rack. It is a data-center design borrowing the same instinct aerospace engineers use when they want heat to leave the system faster: remove bottlenecks and shorten the thermal path.

A similar pattern shows up in another 2026 open-access study, which reports stable operation under 50 kW cabinet load and a two-phase loop temperature of 39.85°C. ³ The details differ, but the design instinct is the same.

What builders should take from this

For AI companies, the strategic implication is straightforward: thermal architecture is becoming part of product strategy. The source set on Lightning AI shows a company shifting toward self-owned infrastructure and specialized, smaller models to reduce dependence on third-party frontier providers and unpredictable pricing. ⁸

That is useful context only if you keep it in its lane. The relevant parallel is not the enterprise AI strategy itself. It is that control over the infrastructure layer is increasingly a competitive issue, and cooling is one of the layers that cannot be abstracted away.

For builders and investors, the practical questions are more grounded:

1. What is the actual heat path from chip to sink?
2. What happens at partial load, not just peak load?
3. What specialized labor, water, power, or refrigerant infrastructure does the system require?
4. Does the architecture improve usable compute, or only improve a facility metric?

Those questions matter more than whether the system sounds futuristic. A cooling design inspired by rocket engines is not automatically better than a conventional liquid loop. But if it lowers pumping energy, reduces footprint, or makes high-density racks operational where air cooling fails, then it is solving a real constraint.

"Future fighter aircraft are expected to manage heat in the megawatt range. This necessitates the implementation of an efficient thermal management system (TMS) capable of intelligent heat dissipation through the utilization of multiple heat sinks."

— German Aerospace Center (DLR) ⁹

The lesson for AI infrastructure is narrower than the metaphor suggests. Supersonic engines are not the model for data centers. They are a reminder that when heat density rises enough, the winning systems are the ones that treat cooling as architecture, not accessory.