Every rack of GPUs or CPUs is, thermodynamically, a resistive heater. A 100 kW compute stack rejects 100 kW of low-grade heat, typically into a liquid loop running somewhere between 35°C and 65°C. That heat has to go somewhere — and increasingly, operators are asking whether it can do something useful on the way out, rather than just being dumped to a cooling tower or dry cooler. CO₂ (R744) ejectors keep coming up in that conversation because they're mechanically simple, use a cheap, non-flammable, zero-GWP refrigerant, and are already commercially proven in a different industry: supermarket refrigeration. This article reviews what the engineering literature actually supports, lays out a realistic strategy for a 100 kWth data center application, and runs the numbers against conventional cooling and against the alternative of just generating electricity from the same waste heat.

1. The problem: 100 kWth of waste heat, and what to do with it

A 100 kW rack today is typically cooled by direct-to-chip cold plates or rear-door heat exchangers feeding a facility water loop. Supply temperatures commonly run 30–45°C, with returns of 40–65°C depending on how aggressively the site runs "warm water cooling" to enable free cooling elsewhere in the plant. That return stream is the resource in question: 100 kWth of heat at a temperature that's too low to be very useful for power generation (see Section 6) but potentially useful as the driving energy for a heat-activated cooling cycle.

Two structurally different questions get conflated under the banner "CO₂ ejector cooling," and keeping them separate matters for everything that follows:

• Question A — Can an ejector make the electric cooling plant more efficient? Yes, and this is mature technology.
• Question B — Can the waste heat itself, with no compressor, drive an ejector cycle that produces cooling? Yes in principle, but at very low COP, and the published literature on doing this with CO₂ specifically at data-center temperatures (40–70°C) is thin.

2. How CO₂ ejectors actually work — two very different machines

2.1 Compressor-driven (transcritical) ejectors — the proven technology

The CO₂ ejector systems already running in thousands of supermarkets (Danfoss, Carel, Hillphoenix and others build these) are not heat-driven. In a transcritical CO₂ booster system, the ejector's "motive" flow is high-pressure CO₂ leaving the gas cooler (8–12.5 MPa, 35–55°C) — it uses the compressor's own discharge energy to recover the throttling loss that would otherwise be thrown away across a conventional expansion valve. That's the single largest exergy loss in a transcritical CO₂ cycle (commonly cited at >40% of total irreversibility), and CO₂ is unusually prone to it because it throttles across the critical point.

A comprehensive 2016 review of ejector refrigeration (Besagni, Mereu & Inzoli, Renewable and Sustainable Energy Reviews) tabulates COP improvements of roughly 7–18% from a single ejector, growing to 10–25%, and up to 42% in hybrid multi-ejector/parallel-compression configurations under high-ambient conditions, versus the same cycle with a plain expansion valve. This is a genuine, commercially deployed efficiency gain on an electrically-driven refrigeration cycle — it does not consume the waste heat, it consumes less electricity to produce the same cooling.

2.2 Heat-driven ejectors — the technology that actually "eats" waste heat

The architecture that matches the question "use 100 kWth of waste heat to make cooling" is the classic heat-driven ejector refrigeration cycle (sometimes called SERS/SoERS in the literature): a generator boils refrigerant using the waste-heat stream, the resulting vapor drives an ejector that entrains and compresses vapor from an evaporator, and a condenser rejects the combined flow — no compressor required, just a small liquid pump.

The Besagni review compiles generator temperatures across roughly 40 studies of this cycle type, spanning 60–140°C, with COP typically 0.2–0.5. The lowest-temperature demonstrations found in that review (solar-driven units using R600a, Śmierciew et al.) ran generator temperatures of 50–64°C — squarely in the data-center waste-heat band — but COP fell to 0.15–0.20 at those conditions.

2.3 A related option: absorption-ejector hybrids

A 2026 study (Hacıpaşaoğlu, J. Braz. Soc. Mech. Sci. Eng.) modeled an absorption-ejector-booster cycle (ammonia-water absorption cascaded with a booster compressor and ejector) against a conventional absorption + vapor-compression baseline. At a generator temperature of 87–105°C it delivered COP 0.22 (vs. baseline 0.21), a 17–20% reduction in electric compressor work, and roughly 15–18% lower CO₂ emissions at a 100 kW cooling scale. The generator temperature required (87–105°C) sits above typical data-center liquid-cooling returns, so this route only becomes attractive for sites running unusually hot coolant loops (some high-density GPU/immersion designs are pushing return temperatures this high) or where a heat pump is used to lift 60°C waste heat up to generator temperature — at a cost in electricity that has to be weighed against the cooling gained.

2.4 Vortex tubes — a simpler but weaker cousin

Vortex (Ranque-Hilsch) tubes turn up in the same literature as a passive, no-moving-parts alternative to an ejector for recovering CO₂ expansion work. The evidence is mixed: Nellis & Klein (Purdue, 2002) showed a vortex tube adds essentially nothing to a standard vapor-compression cycle (no temperature separation occurs beneath the vapor dome), but more recent transcritical CO₂ heat pump studies (Aghagoli et al. 2019; Mansour et al. 2022) report 8–31% COP improvement depending on discharge pressure and configuration — a wide range that is very sensitive to operating point. Like the transcritical ejector, this is a compressor-work-recovery device, not a waste-heat-driven cooling device.

3. A realistic cooling strategy for a 100 kWth compute stack

Given the technology honestly available, the sensible design is a hybrid, not a single "CO₂ ejector chiller." A strategy that reflects the literature above:

1. Primary plant — CO₂ transcritical booster with multi-ejector parallel compression, sized to the full 100 kWth load. This is the commercially mature piece. Expect COP roughly 10–25% better than the same architecture without an ejector, a non-flammable A1/zero-ODP/GWP-1 refrigerant (a real hedge against F-Gas and AIM Act phase-downs of HFCs), and mechanical simplicity relative to multi-stage HFC plants.
2. Supplemental heat-driven ejector loop, sized to whatever COP the actual return-water temperature supports. Bleed a fraction of the 45–65°C return stream through a small heat-driven ejector chiller (R600a, R1233zd, or water as the working fluid, not CO₂) to generate 15–30 kWth of supplemental cooling "for free," directly offsetting compressor electricity on the primary plant. Section 6 quantifies this.
3. If the loop runs hot (>85°C, e.g., some high-density GPU or immersion designs), add or substitute a single-effect or absorption-ejector hybrid instead of the low-COP heat-driven ejector — COP roughly triples (0.6–0.8 vs. 0.15–0.3) once generator temperature clears ~90°C.
4. Route any heat that can't economically be converted to cooling to a secondary use (domestic hot water preheat, district heating, absorption pre-heat) before final rejection to dry coolers, rather than treating it as pure waste.

4. CO₂ ejectors vs. chilled water vs. conventional refrigeration

The practical read: CO₂ ejector technology is a genuine, quantifiable upgrade to an electrically-driven cooling plant — comparable or better COP than a good water-cooled chiller, with a refrigerant that isn't facing regulatory extinction. But swapping CO₂ in as the "waste-heat-driven" component of the strategy is where the technology gets stretched past what's been demonstrated; that role is better filled by a heat-driven ejector or absorption cycle using a fluid suited to low-grade heat.

5. Cost comparison

Public cost data for CO₂ ejector systems specifically is sparse (the Besagni review explicitly flags this as a gap in the literature), so the figures below are assembled from vendor/industry and DOE-adjacent sources and should be treated as order-of-magnitude, not a quote.

6. Cooling vs. power: which is the better use of the waste heat?

This is the question worth the most rigor, because the honest answer is "it's close, and depends on your driving temperature" — not a landslide in either direction on pure thermodynamics. Here's the analysis, run against a 100 kWth waste stream at 60°C, an ambient/sink temperature of 25°C, electricity valued at $0.20/kWh, and a baseline water-cooled chiller at COP 5 handling whatever heat isn't converted.

6.1 The physics: why low-grade heat is a poor fit for making power

The maximum possible efficiency of any heat engine between 60°C and 25°C is the Carnot limit: 1 − (298/333) = 10.5%. Real organic Rankine cycle (ORC) units typically achieve 40–60% of Carnot, and published low-temperature ORC efficiencies bear this out: roughly 5–12% for 75–150°C sources, trending toward the low single digits as source temperature drops toward 60°C. A 100 kWth stream at 60°C is genuinely at the bottom of ORC's practical operating window — and at 100 kWth input, the gross electrical output (roughly 3–8 kWe before parasitics) is also below the scale where small ORC turbomachinery is cost-effective; published small-ORC capex runs $2,000–2,900/kWe even at 10–50 kWe, higher below that.

6.2 The physics: why the same heat is a comparatively better fit for making cooling

A heat-driven ejector isn't trying to extract high-quality work — it's moving heat from an evaporator to ambient, assisted by the driving heat. That's a fundamentally less demanding conversion at small temperature lifts, which is why heat-driven refrigeration COPs of 0.2–0.8 are achievable at temperatures where a heat engine can barely clear 5–10% efficiency. The catch is that a unit of cooling is worth less in dollar terms than a unit of electricity: cooling only offsets the electricity a chiller would otherwise have spent to produce the same effect (COP 5 chiller → a kWh of cooling is worth about $0.04, not $0.20).

6.3 Running the numbers

Those numbers aren't a typo: at matched efficiency assumptions the two routes land close together, because the formula reduces to a simple breakeven condition — the ejector wins on pure dollar terms whenever its COP exceeds (ORC electrical efficiency × chiller COP). At a COP 5 baseline chiller, that's a breakeven ejector COP of 0.15 against a 3% ORC, 0.25 against 5% ORC, or 0.40 against 8% ORC. Given that realistic 60°C-class ejector COPs (0.15–0.30) and realistic 60°C-class ORC efficiencies (3–8%) overlap almost exactly around that breakeven line, the raw thermodynamic accounting alone does not decisively favor either route at typical data center waste-heat temperatures.

6.4 What tips the balance toward cooling

Four practical factors, none of them captured in the $/kWh table above, favor the cooling route once you move past the spreadsheet:

• Guaranteed local demand. A data center needs cooling 8,760 hours a year, at essentially the full capacity the ejector can deliver. Generated electricity, by contrast, needs somewhere to go — behind-the-meter offset is easy, but exporting a 3–6 kWe trickle to the grid usually isn't worth the interconnection paperwork, and most sites will simply self-consume it as a fractional demand offset.
• Equipment simplicity. A heat-driven ejector is a fixed or adjustable nozzle with no moving parts beyond a small circulation pump. An ORC unit needs a working-fluid pump, evaporator, condenser, and a turboexpander or scroll/screw expander with a generator — more capital, more maintenance, and a minimum viable scale that a 100 kWth/3–8 kWe case sits right at the bottom edge of.
• Scale mismatch. Commercial small-ORC pricing ($2,000–2,900/kWe) is quoted for 10–50 kWe units; a 100 kWth data center stack produces only 3–8 kWe of ORC output, meaning the unit is undersized relative to where that pricing applies and is likely to see materially worse $/kWe economics in practice.
• System-level thermodynamics. Converting part of the heat to electricity via ORC doesn't reduce the amount of heat that still needs full mechanical cooling — the ORC's own condenser rejects nearly all of the original 100 kWth (minus the small electrical fraction extracted) at close to the same temperature it came in at, so the site's cooling plant sees almost no relief. A heat-driven ejector, by contrast, directly reduces the compressor-electric cooling duty by the exact amount of cooling it delivers — it's substituting for the load the site actually has to serve, not adding a parallel, only loosely coupled revenue stream.

7. Bottom line

CO₂ ejector technology has a well-proven, commercially deployed role in this picture: bolted onto an electrically-driven transcritical CO₂ refrigeration plant, it delivers a real 10–25% COP improvement with a refrigerant that isn't facing an HFC phase-down clock. Using the ejector concept to run directly off the waste heat itself — no compressor, CO₂ as the working fluid — is thermodynamically sound in the abstract but isn't well demonstrated in the published literature at 40–70°C; the honest strategy substitutes a low-temperature working fluid (R600a, R245fa, ammonia-water, or water) for the heat-driven portion of the cycle, sized for the COP the actual loop temperature supports (0.15–0.3 typical), while CO₂ does the job it's actually proven for on the compressor-driven side. Compared to generating electricity from the same 100 kWth at $0.20/kWh, the cooling route comes out ahead in practice — not because the thermodynamics are lopsided, but because a data center's cooling load is a guaranteed, full-capacity, year-round sink for exactly the kind of low-grade, small-scale heat a 100 kWth compute stack produces.