Weekly Deep Dive into Deep Tech #9
Cooling the Future: Heat Is the Last Barrier to Scaling Intelligence
GPUs today can draw more than a thousand watts each. Memory stacks run hotter than ever. Interconnect systems packed with SerDes and voltage regulators waste energy as heat. And as density increases, racks are hitting thermal ceilings long before they hit compute ceilings.
Cooling used to be a background detail in system design. In the age of AI clusters, it has become a primary architectural limit. The performance of a model now depends as much on thermal engineering as it does on silicon design.
Jetcool
Microjet Cooling at the Surface of the Die
Jetcool works at the opposite end of the thermal spectrum. Their microfluidic system cools the silicon die directly, targeting hotspots that traditional methods cannot reach. This is critical for GPUs and accelerators where heat is concentrated in very small areas.
Recent progress:
1. Direct-to-chip microjet cold plates
Capable of removing heat at extremely high flux, ideal for power-dense processors.
2. Efficient cooling for next-generation accelerators
Jetcool’s designs support hardware that exceeds one kilowatt per device.
3. Modular liquid cooling for data centers
The company is developing solutions that retrofit into existing racks without major infrastructure changes.
