AI Thermal Stack Map (junction-to-ambient)

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.

L0 — In-die / embedded cooling (the frontier that bypasses the TIM problem)

• Role: remove heat inside the silicon — microchannels etched into the die or interposer backside, on-die hotspot management. If heat never has to cross a die-attach interface, layers L1–L3 stop binding.
• Binding question: does the foundry/OSAT absorb the in-silicon primitive, or does a specialist hold a manufacturable one?
• Approaches / who: Corintis (EPFL; in-silicon microfluidics + Glacierware design software; Microsoft co-dev, ~3× heat removal / 65% lower peak temp, vendor-demoed); TSMC IMC-Si (microchannels fusion-bonded into CoWoS, >2.6 kW TDP, ~2027 commercialisation — the competitive fact); Jetcool SmartSilicon (on-die); Maxwell Labs (laser cooling, pre-product); imec/IBM-Zurich legacy; Adeia IP. See In Chip Microfluidic Cooling.
• Capture → route: value biases to foundry/OSAT capture (TSMC absorbing it ~2027); the narrow venture wedge is design software + hyperscaler co-dev, not the etch/bond primitive. Cross-link: BSPDN (backside power delivery, Intel 18A / TSMC A16) forces backside cooling and is the why-now. Listed: TSM, FLEX (JetCool), ADEA.

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.

Coverage read: thickest = L3 (commodity TIM2 incumbents, ~16) and L1 (the frontier, ~12). Thinnest = L0 (Corintis ~the only true entry), L5 (one real fluid line + dielectric secondaries), L8 (2–3). The thin layers are where either a genuine wedge (L0) or a dislocation (L5 PFAS) sits — built out as In Chip Microfluidic Cooling and Dielectric Coolant Pfas Transition.

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
• 2026 06 22 Thermal Stack Layer Landscape — the per-layer ground-truth source note