Why this map exists — the comparison logic
Heat leaves an AI accelerator along a path, junction → ambient, through a stack of layers. Each layer is a distinct engineering problem with its own physics, its own competing approaches, and its own value-capture answer. The single most useful discipline:
The meta-thesis over the whole map is Thermal Stack Arms Race: the stack stays segmented — no single material or layer wins all of it. This page is the substrate that thesis reasons over; the demand-side spine is Ai Power Thermal Binding and the buy-side market is Datacentre Thermal Power Market.
The stack, layer by layer
Heat flux at the hotspot now exceeds ~1 kW/cm²; package TDP is passing 1 kW and heading to 1.8–2.3 kW (Nvidia Rubin, CES Jan 2026). Each layer below is where some of that heat is conducted, acquired, transported, or rejected.
Off-ladder — the power-converter plane (added 2 Sep 2026)
L0-L8 is one path: the accelerator’s own junction to ambient. An AI assembly has a second heat source that does not sit on it. The 48V-to-sub-1V conversion plane is mounted beneath the package, on the far side of the die from the cold plate, dissipating 3-4kW in one plane on a ~23kW part with 2mm or less of height for switches, magnetics and any thermal solution combined. It has its own junction-to-ambient path and that path terminates nowhere.
Consequence for anyone reading this map as a coverage check: a layer can be fully covered here and the assembly still be thermally bound, which is what a Cerebras architect reported in Aug 2026 (cooling the wafer is easy, cooling the converters is very difficult). Bet page: Converter Plane Thermal. Evidence: ref.
L3 — TIM2 (lid/spreader → cold plate)
- Role: second interface, lid to cold plate. Lower flux than TIM1, larger area.
- Approaches / who (incumbent-dense — the thickest, most commoditised material layer): grease, gap pads, phase-change pads — Henkel, Dow, Shin-Etsu, Momentive, Wacker, Fujipoly, Laird, Parker/Chomerics, Honeywell/Solstice PTM, T-Global. See Phase-Change Materials (Thermal).
- Capture → route: commodity incumbent material houses. Public/incumbent.
L4 — Heat acquisition (cold plate / direct-to-chip)
- Role: carry heat into the coolant — direct-to-chip cold plate, microconvective, heat pipe, or (legacy) air heatsink.
- Approaches / who: single-phase DTC (Jetcool→Flex, CoolIT→Ecolab, Asetek-exited) — the AI reference: NVIDIA GB200/GB300 spec single-phase DTC; two-phase DTC (Zutacore $100M-C Jun 2026, accelsius $65M-B) — the >2 kW future bet.
- Capture → route: bare cold plate commoditising to assembly margins; value pools one layer up (integrated CDU + chemistry + services) and was rolled up in a 2025–26 M&A sweep (Boyd→Eaton $9.5B, CoolIT→Ecolab $4.75B). Public read. Listed: ETN, SU, VRT, NVT, ECL, FLEX.
L5 — Coolant / dielectric working fluid
- Role: the fluid that carries heat away — single-phase water-glycol, two-phase dielectric, immersion fluid.
- Approaches / who: the 3M PFAS exit (manufacturing ceased end-2025, ~80% of PFAS coolant supply) created a two-phase supply vacuum; fillers are chemical majors (Chemours Opteon, Solvay/Syensqo Galden, Daikin) and oil majors for single-phase PFAS-free (Shell, bp/Castrol). See Dielectric Coolant Pfas Transition.
- Capture → route: no ownable molecule; moat is plant + qualification + regulatory capex. Public/incumbent-chemical read. Listed: CC, SYENS, 6367, SHEL, BP.
L6 — Distribution (CDU, manifolds, plumbing)
- Role: move coolant between cold plate and heat rejection — CDU, manifolds, pumps, quick-disconnects, rear-door HX.
- Approaches / who: Vertiv, alfa-laval, motivair→Schneider, chilldyne→Daikin, rittal, nvent, foxconn.
- Capture → route: plumbing — competes to assembly margins; incumbents roll it up. Pass / public.
L7 — Heat rejection (system → ambient) / immersion
- Role: reject heat to the environment — immersion tanks, dry coolers, chillers, evaporative.
- Approaches / who: single-phase immersion (Iceotope, Submer, GRC), two-phase immersion (LiquidStack→Trane Mar 2026 — pivoted off two-phase after 3M exit).
- Capture → route: HVAC-major margins — public read. NVIDIA DTC reference keeps immersion a <15% specialty (the Thermal Stack Arms Race prediction). Listed: TT, VRT, NVT, SU.
Adjacent thermal domains (not the AI datacentre stack — cross-link anchors)
- Power-electronics modules (sintered-Ag die-attach, double-sided cooling) → Gan Power Datacentre. Photonics / CPO (laser-junction thermal, CTE-matched layers) → Photonic Packaging Cpo. EV / SiC traction → Sic Ev Traction Dominance. Cryogenic / quantum (dilution fridges; genuinely active VC — Maybell Quantum ~$73M Mar 2026) → Cryo Computing. Space / satellite (radiative-only; orbital DCs — Starcloud put an H100 in orbit Nov 2025).
Company placement table
The cohort, placed. L# = primary layer (secondary in parens). Status from the company page decision:. Garbled/duplicate auto-map slugs and demand-side (server OEMs, power devices) dropped — see the source note for the full raw classification.
Where the genuine early-stage venture wedge is (vehicle-agnostic, ranked by where value is held)
Frontier
See frontmatter frontier: block.
Reading list
- TIM Approach Map (every thermal-interface-material class, in-market + R&D) — the orthogonal axis: every TIM material class (in-market + R&D), the comparison catalog
- Thermal Stack Arms Race — the thesis this map serves (no single substrate wins)
- Ai Power Thermal Binding / Datacentre Thermal Power Market — demand spine + buy-side market
- Cnt Tim Billion Revenue / Liquid Metal Tim / Sintered Silver Die Attach — the L1 material slices
- In Chip Microfluidic Cooling (L0) / Dielectric Coolant Pfas Transition (L5) — the thin-layer build-outs
- — the per-layer ground-truth source note