AMD's Helios rack-scale platform — the double-wide chassis built around 72 Instinct MI455X accelerators — lands at roughly 140 kW of rack-level power draw. In a liquid-cooled AI rack, essentially all of that electrical input leaves as heat, so the thermal design point is the same number: 140 kWth of waste heat that has to go somewhere. The three questions below walk through three different ways of getting rid of it, using published performance data for each technology class rather than nameplate marketing numbers.

1. Transcritical CO₂: electrical power to cool the stack

A transcritical CO₂ (R744) chiller compresses the refrigerant above its critical point (31.1°C / 73.8 bar) and rejects heat through a gas cooler rather than a condenser. For data center duty with direct-to-chip liquid cooling — where the loop runs warm (typically 25–45°C supply) — CO₂ systems perform well because the elevated evaporating temperature favors the cycle. Published COPs for transcritical CO₂ data center cooling range from about 3.5 in hot, above-critical ambient conditions up to 5–7+ in favorable climates or with gas-cooler enhancements like adiabatic/evaporative assist.

~31 kWe at the design point — call it 22% of the rack's own power draw as a mechanical-cooling "tax." That's a reasonable, literature-consistent estimate for a warm-water, direct-to-chip CO₂ loop; it gets meaningfully better with a well-designed gas cooler and worse on a hot day with no evaporative assist.

2. Ejector cooling from the stack's own waste heat

An ejector cooling cycle uses a nozzle and a heat-driven "generator" instead of a mechanical compressor: waste heat boils a working fluid, the resulting high-velocity vapor jet entrains and compresses low-pressure vapor from the evaporator, and cooling is produced without electrical compression work. The tradeoff is a much lower coefficient of performance than a compressor-driven cycle — typically 0.2–0.5 depending on the generator (driving) temperature, condensing conditions, and working fluid. Ejector cycles are known to run effectively off relatively low-grade heat, around 60–80°C, which lines up well with a rack's warm liquid-cooling return.

~49–56 kW of cooling from the rack's own reject heat at a design-point COP of 0.35–0.4. That's meaningful — on the order of a third to 40% of the 140 kW cooling demand — but on its own, the rack's waste heat isn't hot enough or plentiful enough to fully self-cool. It needs a second heat source. That's where the gas-fired mesh comes in.

3. Natural gas cluster mesh: sizing and the combined ejector cascade

Sizing the mesh

At 25 kWe per cell, powering a 140 kWe Helios stack requires 6 cells (140 ÷ 25 = 5.6, rounded up), for 150 kWe of installed capacity — about 7% headroom over the load, which is a sensible margin rather than a redundant spare.

Waste heat from the mesh

Small natural-gas gensets/microturbines in this size class typically run 25–30% electrical efficiency, with exhaust temperatures in the 500–600°C range and total (electrical + recovered thermal) CHP efficiencies commonly cited at 60–85%. Using a representative 25% electrical efficiency and ~50% recoverable thermal efficiency (consistent with published CHP fact sheets, leaving ~25% as unavoidable stack/radiation loss):

The gas mesh throws off roughly 300 kWth of recoverable exhaust/jacket heat — more than double the Helios stack's own 140 kWth, and at a notably higher temperature (raw exhaust in the hundreds of °C vs. a ~45–65°C liquid-cooling return), because it's coming straight off high-temperature exhaust rather than a warm chip-cooling loop.

Combining both heat streams in the ejector system

Because the gas-mesh heat is higher grade, it can drive an ejector stage at a better COP (~0.55) than the lower-grade rack waste heat (~0.35). Cascading both streams through a CO₂ ejector cooling system:

~214 kW of additional cooling capacity from the combined ejector cascade — a surplus of about 74 kW over the 140 kW the Helios stack actually needs to reject. In practice, that means the two-stream ejector system alone can plausibly cover the entire cooling load thermally, turning the transcritical CO₂ compressor loop from the primary cooling system into a trim/backup unit for humidity control and off-design conditions, rather than the ~31 kWe workhorse it would otherwise need to be.

Key assumptions and caveats

• Helios rack power/heat figures (140 kWe / 140 kWth) reflect early public estimates for the double-wide MI455X rack-scale platform; AMD has not published a final thermal design spec as of this writing, and production units (targeted 2H 2026) may vary.
• Transcritical CO2 COPs of 3.5–7+ are drawn from published data center CO2 chiller performance; the high end reflects optimized systems with internal heat exchangers or evaporative gas-cooler assist, not a generic install.
• Ejector thermal COPs of 0.3–0.5 are representative of waste-heat-driven ejector refrigeration cycles at 60–150°C generator temperatures; actual COP is sensitive to working fluid choice, condensing temperature, and ejector geometry.
• Gas mesh efficiency assumes 25% electrical efficiency and ~50% recoverable thermal efficiency, consistent with published microturbine/CHP fact sheets (60–85% overall CHP efficiency in that class).
• These are engineering planning-level estimates, not a substitute for a full thermodynamic model or vendor-specific performance data. Real deployments should validate COPs against actual ambient design conditions, loop temperatures, and equipment selections before sizing capital equipment.