Modern AI racks reject an enormous amount of low-grade heat — typically 40–50°C water leaving a liquid-cooling loop. It's tempting to ask: can that heat be turned back into electricity with a small turbine, instead of just being dumped to a cooling tower? This article works through that question quantitatively, using real fluid thermodynamics (CoolProp/HEOS) to compare three working fluids and three expander technologies on a like-for-like basis.

The scenario — and why it's a hard problem

The test case: a heat source at 50°C (an AI rack pod's coolant return loop) delivering 100 kW of thermal power, rejecting heat to a 25°C sink (facility cooling water). That's a realistic scale — modern liquid-cooled AI racks commonly reject 80–130+ kW each, so a small pod of a few racks easily reaches 100 kW.

The catch is the temperature lift: only 25°C separates source from sink. After allowing a realistic 5°C pinch at the heat exchanger, the working fluid only sees roughly a 20°C swing between evaporation and condensation. The Carnot efficiency — the theoretical ceiling for any heat engine operating between these temperatures — works out to just 6.3%. Every number in this article sits below that ceiling, most of them well below it. This is fundamentally a low-grade-heat problem, and no amount of clever engineering changes the thermodynamic floor.

Comparing the working fluids

Three fluids were evaluated: R134a and R245fa (both familiar HFC refrigerants used in small ORC systems, including the radial-flow expander test rig this analysis is partly based on) and CO₂ (a natural refrigerant with a much lower critical temperature). All three were modeled through the same cycle: pump → evaporator/gas cooler → expander → condenser, with a 5°C pinch off the 50°C source, 3°C superheat, 2°C subcooling, and condensing at 25°C.

R245fa and R134a behave similarly — both are subcritical at these conditions (their critical points, 154°C and 101°C respectively, are far above the 50°C source), so they evaporate and condense at constant temperature in the conventional Rankine sense. R245fa edges out R134a slightly: its pressure ratio (1.98 vs 1.74) is a bit higher, giving marginally more specific work per kilogram of fluid expanded.

Why CO₂ is a special case

CO₂'s critical point is 31.0°C / 7.38 MPa — below the 50°C source temperature. That means CO₂ can't boil at constant temperature the way R134a or R245fa can; it has to run as a transcritical cycle, heated as a supercritical fluid in a gas cooler and expanded down from above the critical pressure. That's a well-understood, commercially normal way to run CO₂ heat pumps and refrigeration systems.

The problem in this specific scenario is the condensing side. Condensing at 25°C is only 6°C below CO₂'s critical temperature, which pins the low-side pressure at 6.43 MPa — already 87% of the critical pressure. There simply isn't much room left to raise the high-side pressure before the extra pump work needed to get there eats up whatever the turbine gains.

The practical conclusion: CO₂ is the thermodynamically wrong fluid for a 25°C heat sink. It becomes attractive again once the sink temperature drops well below ~20°C (ideally under 15°C), which is exactly the regime CO₂ heat pumps and the reference turbine design examined later in this article (condensing at 8.6°C) were built for.

Comparing the expander technologies

Three expander technologies were compared, each characterized by an isentropic efficiency applied to the same fluid cycles above:

• Radial-flow turboexpander (65% isentropic) — a velocity-type turbine built from a modified automotive turbocharger, per the reference paper's in-house design, validated at pressure ratios of 3.4–4.3.
• Gear-type (gerotor) PD expander (50% isentropic) — a positive-displacement machine, essentially a gear or gerotor pump run in reverse; a literature-typical placeholder efficiency for this simpler, lower-cost approach.
• Infinity cluster mesh turbine + compressor pump (80.78% expander / 74.57% pump) — efficiencies taken directly from a vendor EES cycle model for a purpose-built, 52,000 RPM, 2.5-inch CO₂ turbine/pump package, validated at its own (higher-lift) design point.

The Infinity package's higher expander efficiency (80.78% vs. 65% for the radial machine and 50% for the gear-type unit) makes it the clear winner for R134a and R245fa — roughly 25–30% more net power than the radial turboexpander, and 65–70% more than the gear-type expander. For CO₂, the picture is different: because Table 3 uses the Infinity package's own (isentropic-efficiency-based) pump model rather than the flat, generously "efficient" 0.5 kWe pump assumed for the other two CO₂ cases, its pump draws considerably more power near CO₂'s critical point — enough to erase almost all of the expander-side gain.

Other useful comparisons

Second-law efficiency: how close to the theoretical limit?

Raw thermal efficiency is misleading when comparing scenarios with different temperature lifts, because it doesn't say how much of the available (Carnot-limited) potential was actually captured. Second-law efficiency — thermal efficiency divided by the Carnot limit — is a fairer way to compare technologies against each other.

This view reinforces the same story: the Infinity turbine on R245fa captures nearly three-quarters of the theoretical maximum available at these conditions — a genuinely good result for compact hardware. CO₂ tops out under 40% of Carnot in every configuration tested.

How much does expander quality matter?

Since two of the three expander efficiencies used here are literature placeholders rather than site-specific measurements, it's worth seeing how sensitive the results are to that single number.

The relationship is close to linear for all three fluids, and CO₂ never catches up to the HFCs anywhere on this curve — confirming that its shortfall is a property of the fluid and the 25°C sink temperature, not something a better expander alone can fix.

Validating the numbers against real hardware

The Infinity turbine efficiencies weren't assumed — they came from a vendor EES cycle model for an actual CO₂ turbine/pump package, which reports a self-consistent design point: 104.4°C gas-cooler exit, 8.6°C condensing, 12.51/4.34 MPa, 52,000 RPM, and 13.99 kW net output at 6.5% thermal efficiency. That works out to a Carnot limit of 25.4% at their conditions and a second-law efficiency of 25.6% — a believable, respectable number for small high-speed hardware, and a useful sanity check that the 80.78%/74.57% isentropic efficiencies extracted from that model aren't unrealistic.

Full results table

Bold green = best result in its fluid group; bold red = lowest result overall. All figures at 100 kW thermal input, 50°C source, 25°C sink.

Takeaways

• Fluid choice matters less than expander choice, but isn't free. R245fa beats R134a by 3–7% across every expander technology; both comfortably beat CO₂ at these specific temperatures.
• CO₂ needs a colder sink to be worth it. Its low critical temperature (31°C) is fundamentally mismatched to a 25°C condenser; every CO₂ configuration tested here underperforms its HFC counterparts, regardless of expander quality.
• Expander quality swings results by 2×. Going from a 50%-efficient gear-type expander to an 80.78%-efficient purpose-built turbine roughly doubles net output for the HFCs — the single biggest lever examined in this study.
• Every number here is well under the Carnot ceiling. With only a 25°C source-to-sink lift, no combination of fluid and hardware escapes a fundamentally low absolute efficiency (under 5% in every case). This is a low-grade-heat problem first, an engineering-optimization problem second.

Methodology & assumptions

All cycle states were computed with CoolProp v8.0 (HEOS backend, NIST REFPROP-equivalent equations of state), not textbook correlations. Shared assumptions across all fluids and expanders: 5°C pinch off the 50°C source, 3°C superheat at the evaporator/gas-cooler exit, 2°C subcooling at the condenser exit, condensing at 25°C, and 100 kW thermal input (representative of a small AI rack pod on liquid cooling). CO₂'s high-side pressure was fixed at 8.0 MPa — the efficiency optimum under a standard isentropic pump model — and held constant across all three expander comparisons rather than re-optimized per case, since letting it float against a fixed pump-power assumption produces an unphysical, unbounded result.

Two different CO₂ pump treatments were used depending on the table: the radial and gear-type expander comparisons assumed a flat, user-specified 0.5 kWe "efficient pump" (a generous simplification), while the Infinity comparison used that package's own measured isentropic pump efficiency (74.57%) via the standard thermodynamic formula — which is why CO₂'s pump penalty looks larger in the Infinity case despite it being the "better" pump on paper.

Radial turboexpander and gear-type PD expander isentropic efficiencies are model assumptions, not site measurements — 65% is the mid-point of 60–74% reported experimentally in the underlying reference paper (at higher pressure ratios than this scenario sees), and 50% is a literature-typical placeholder for a reversed gear/gerotor pump. The Infinity figures (80.78% expander, 74.57% pump) come from a vendor EES cycle model for an actual CO₂ turbine/pump package and were verified by reproducing its reported net power output from first principles (12.33 kW pump work, 26.32 kW expander work, 13.99 kW net — matched to the watt after unit conversion).