How this relates to the stack map (two orthogonal axes)
AI Thermal Stack Map (junction-to-ambient) answers where in the heat path (L0 in-die → L8 heat reuse). This page answers which material fills the interface. They cross: a TIM class (say liquid metal) occupies a stack position (TIM1 die-attach). Use this page to be exhaustive about material options; use the stack map to place and compare them. The governing thesis over both is Thermal Stack Arms Race (the stack stays segmented).
Stack-position shorthand: TIM1 = die-attach (junction→lid); TIM1.5 = lid interface; TIM2 = spreader→cold plate. k in W/m·K (vendor-datasheet typicals, method-dependent).
A. Commercial / in-market classes (24)
B. R&D / emerging approaches (14) — maturity, players, wedge
| Approach | Maturity (TRL) | Key players (startup / lab) | Binding constraint | Wedge verdict |
|---|---|---|---|---|
| Next-gen VACNT + encapsulation | 8-9 prod / 3-5 frontier | Carbice, Nawah, Trimtabs; Cola (GA Tech) | off-chip growth + transfer, $/cm² | incumbent-captured (Carbice) |
| Vertically-aligned graphene (VG) | 7-8 | Sht Smart High Tech, Thermal Rex; Liu (Chalmers) | through-plane alignment + shorting | narrowing (SHT public + Henkel) |
| Cu/metal-nanowire arrays | 6-7 | Novolinc; Berkeley/UCSD/Stanford/CMU | dense uniform low-R arrays at scale | YES — process moat |
| Sintered nano-copper | 5-7 | Kuprion, Nano Join, Elephantech | pressureless void-free oxidation-free | incumbent-captured (Heraeus/MacDermid) |
| Self-healing / reworkable | 2-4 | none pure-play (Arieca gets it free) | k vs mobility trade-off | no standalone wedge |
| LM containment composite | 6-8 | Boston Materials, Arieca, Hymet; UT Austin (Yu), CMU | pump-out, Ga corrosion, reliability | YES — strongest wedge |
| 2D beyond graphene (hBN/MXene) | 3-5 / 2-4 | MXene Inc (Drexel), BNNT Materials; Murata | k + alignment vs supply scale | aligned-hBN = white space; MXene = supply play |
| Nanocomposite hybrid | 5-7 | Novolinc, Boston Materials; Shen (CMU) | continuous nano-architecture mfg | YES |
| Phononic / metamaterial | 1-3 | CMU, MIT (no TIM startup) | nanofab at scale; mostly insulating | no TIM wedge (science) |
| Direct-bond / TIM-less (Cu-Cu hybrid, TLP/SLID) | 8-9 / 5-7 | TSMC/Intel/Samsung, BESI, Adeia | sub-nm flatness, fab-level | foundry-absorbed |
| Printed / jettable TIM | 6-8 | Novolinc, Arieca, fabric8labs; Indium | HVM line-qual, BLT control | YES — selective (formulation + process) |
| Bio-inspired / structured | 2-4 | none (nanoGriptech adhesion pivot) | manufacture at yield; unproven vs LM | weak/none |
| Boron-arsenide / θ-TaN / BP | 2-4 | no startup; UCLA (Hu), Houston (Ren), UCSB | scalable defect-free crystal growth | open but unbuilt (sourcing target) |
| Diamond (GaN-on-diamond, nanocrystalline) | 6-8 | Akash Systems, Diamond Foundry, Diamfab; Element Six De Beers, Coherent | cost, CVD speed, CTE/TBR | yes-but-late/capital-heavy (EU seed: Diamfab) |
Interconnect-layer diamond spreader (adjacent frontier, NOT a TIM slot; ECTC 2026). Purdue / U. Aveiro / UCLA showed Cu/Sn microbumps embedded in nanocrystalline diamond, giving an effective in-plane thermal conductivity of 500-600 W/m·K (~20× conventional microbumps + underfill) — but it spreads heat laterally through the interconnect layer of a 3D stack, it is not a TIM1 replacement (single-sided test structure only, no assembled stack). It bounds the map from the packaging-interconnect side rather than the die-interface side: if it matures it substitutes for in-stack lateral spreading, not for the die-to-lid / die-to-coldplate TIM the classes above serve. Watch, don’t source as a TIM wedge. Source: 2026 07 02 Semianalysis Ectc2026 Advanced Packaging.
Metal-TIM note. Solder TIM (sTIM), indium foil, and PCMA are the genuine volume die-level path for >1kW AI accelerators today (Indium-led) — promoted 22 Jun 2026 to its own slice, Metal Solder Pcma Tim. They are the incumbent that is already winning the volume, distinct from the speculative carbon classes and from consumer-led Liquid Metal Tim.
C. Venture-stage startup roster (the live sourcing list)
D. State of the art for >1kW AI accelerators (vs commodity)
Deployed/designed-in on bare-die GPUs/ASICs where commodity grease/pads pump out: 1. solder TIM (sTIM) · 2. liquid metal + LM composites · 3. PCMA · 4. polymer PCM (PTM7950) · 5. VACNT pads (Carbice). Next ring: indium foil (immersion), sintered-Ag (power stages, not the GPU die), carbon-fibre (Dexerials). Commodity = silicone grease / gap pads / gels (bulk of TIM2 by volume, just not the thermal-limited interface).
E. Value capture & routing (vehicle-agnostic, ranked by where value is held)
F. KB coverage status (the 100% check)
Frontier
See frontmatter frontier: block.
Reading list
- AI Thermal Stack Map (junction-to-ambient) — the orthogonal axis (where in the path)
- Thermal Interface Materials (TIM) — the TIM physics/concept anchor
- Thermal Stack Arms Race — the governing thesis (segmented stack)
- 2026 06 22 Tim Approaches Full Sweep — the exhaustive source sweep (vendors, R&D, startups)
- material slices: Liquid Metal Tim · Sintered Silver Die Attach · Cnt Tim Billion Revenue · In Chip Microfluidic Cooling