Theses›Datacentre & AI compute›Datacentre thermal and power

Datacentre thermal and power

Differentiated thermal materials and power semis collect the per-rack tax; integration commoditises

  • Closing
  • $453B by 2035

The call, a draftHeat and power now limit how much AI compute a datacentre holds, so cooling and power-delivery hardware grows fast, probably faster than forecasts. The window is late: the system layer has consolidated, and the open question is who holds the in-package parts.

D1D2D3D4R1R2R3R4C1C2C3foragainst D1D2D3D4R1R2R3R4C1C2C3foragainst
Fig. 1The balance. Drivers and catalysts against restraints, each block sized by its weight; catalysts hang dashed because they have not happened yet, sized by their odds. It tips toward the drivers.
  1. D1TrendForce has liquid cooling in AI datacentres climbing from 14% to 33% as rack densities go from 40kW to 140kW, and GB200 NVL72 ships with no air-cooled version.
  2. D2Power SiC devices reach $11B by 2031 at 20% a year on Yole's numbers, with AI datacentres and 800VDC named as drivers.
  3. D3Microsoft and its partners have shown in-chip microfluidic cooling with up to 3x the heat removal of cold plates and 65% lower peak temperature.
  4. D4Water scarcity is starting to shape cooling choices, with operators planning desalination plants just to supply datacentre cooling water.
  5. R1Most of about $15.4B in completed acquisitions (Eaton and Boyd, Ecolab and CoolIT, Schneider and Motivair) went to the cold-plate and coolant-distribution layer said to be commoditising.
  6. R2Foundries absorbed backside power delivery themselves (Intel 18A shipping January 2026, TSMC A16 in production late 2026), and in-package cooling could follow.
  7. R3Datacentre interface materials are only $1.43B in 2024, against $14.8B for liquid-cooling integration by 2030, so the defensible part is the smaller one by revenue.
  8. R4Premium die-level thermal sells only into new builds, since removing a heat spreader voids the warranty on live servers.
  9. C1Liquid cooling becomes the majority architecture for new AI datacentre capacity by deployed rack power.31 Dec 2027
  10. C2TSMC, Samsung, Intel, GF or a leading packaging house offers in-package cooling or power delivery as a qualified package option.31 Dec 2027
  11. C3Wide-bandgap power content per AI rack at least doubles from 2025.31 Dec 2028
Rests onthe beliefs underneath
Supported by · confidence
What would prove itdated tests, open and resolved
202720282029today1234567predictionswatch 202720282029today1234567predictionswatch
Fig. 2Each ring is a dated test: dashed while open, struck through when overdue, ticked or crossed once resolved.
  1. 1Whether a second supplier qualifies for hyperscaler vertical power delivery next to Vicor.Jun 2027 · open
  2. 2Liquid cooling becomes the majority cooling architecture for new AI datacentre capacity (by deployed rack-power) by 2027, as air cooling exhausts above ~100kW/rack.Dec 2027 · open
  3. 3By end-2027 at least one top-tier logic foundry / OSAT offers an in-package thermal OR in-package power-delivery primitive as a qualified package option, signalling the margin for cooling/power is migrating INTO the package layer rather than staying with board/facility vendors.Dec 2027 · open
  4. 4Dell'Oro and TrendForce cooling mix by deployed AI rack power.Dec 2027 · open
  5. 5A foundry or packaging house offering in-package cooling as a qualified option.Dec 2027 · open
  6. 6GaN and/or SiC displace silicon in the datacentre power-delivery stage fast enough that wide-bandgap content per AI rack at least doubles 2025-2028.Dec 2028 · open
  7. 7An advanced thermal-material or microfluidic cooling maker winning a hyperscaler or accelerator design, beyond a pilot.Dec 2028 · open
Expressionslisted, private, technologies
Private · 5 private companies, 2 of them in the US.

Asetek · Jetcool Technologies · Zutacore · Submer · Chilldyne

6 more private companies tracked, not shown

Technologies
What changedthe record, newest first
  • 13 Aug 2026Cooling startups are increasingly positioning on capturing value at the chip and server level rather than at the facility level, which is where the incumbent thermal vendors have historically sat.
  • 3 Jul 2026A thermal vendor put real numbers on the per-rack tax: two to three thousand dollars of die-level thermal spend justified per hundred-thousand-dollar AI server. The catch is that premium thermal only sells into new builds, and the vendors actually monetising are selling whole servers, not materials.
  • 21 Jun 2026As chips get cooling and power delivery built into the package rather than bolted on at the board or facility, the profit for moving heat and delivering power shifts toward the chip foundries and packaging houses (and a thin layer of materials specialists), away from the data-centre cooling and power vendors. Whether a startup holds a piece of that or the foundry swallows it is the open question.
Reading
Show the reading