Escape Velocity

Enabling technologies / enablers/heat-removal

High-flux heat removal

scopingTRL 4 (4 of 9)horizon 2030s

Cooling that removes the heat of dense chips and stacked dies at the device, at heat fluxes well above those of air and cold-plate cooling, in a form that fits in a package and can be made in volume.

Scope

In scope: embedded and near-junction cooling (microfluidic channels, manifolds, jets, two-phase flow, diamond spreaders), measured by heat flux removed at a stated temperature rise. Out of scope: the data-center plant and heat rejection; low-temperature refrigeration, covered by enablers/dilution-refrigeration.

Readiness
TRL 4 (4 of 9)
Embedded cooling is shown on thermal test vehicles and small GaN-on-diamond devices; a review warns that records are not transferable to package-compatible, large-area, deployed platforms.
Serves
Industry, innovation and infrastructure
Last reviewed
2026-10-04
Curators
none yet: volunteer

Metrics

Heat flux removed headline1.0 orders of magnitude to go

Heat a cooling solution removes per unit area of the device surface, at the stated temperature rise. Higher is better.
Heat flux removed: log scale, one tick per order of magnitude; better to the righttargetnow
Current (2026-01-01)10⁷ W m⁻²
Target10⁸ W m⁻²
Limit–
Conditions. Heat flux a cooler removes per unit area of the device surface. Heated area, coolant, temperature rise and pumping cost differ across sources and must be stated.
Why this target. Atlas-set reasoning: 10 kW/cm2 over large package-compatible areas. DARPA's near-junction program reached above 40 kW/cm2 (4e8 W/m2) on small GaN-on-diamond devices, so a quarter of that record on deployable, area-scaled hardware would give stacked logic a tenfold margin over today's microfluidic range.
Note. The upper end of the 10^2 to 10^3 W/cm2 range stated by the review, taken as 1e3 W/cm2. as_of is the year of the review; the abstract gives no date.

Gaps

No gap recorded yet.

Dependencies

Requires

Nothing recorded.

Required by

  • Energy-efficient AI inference Delivered power and cooling capacity bound how many accelerators fit in a rack and therefore the tokens produced per site.

Holds open

Arrows point from a technology to what it requires. Select a node to open it.

Evidence

In The Alan Machine

Source TOML · Page on GitHub · Suggest a correction