As AI investment races toward an estimated **$1 trillion annually** for the foreseeable future, the industry faces a growing constraint in its ability to **cool** the massive infrastructure powering this expansion.
Mainstream chips are becoming much hotter after years of incremental increases in wattage. Suddenly, we're seeing a J-curve in terms of the heat these chips produce. With Rubin architecture coming as soon as **H2 2026**, power density will double compared to the previous generation.
Today's datacenters using the latest GPUs use single-phase direct-to-chip liquid cooling. However, as newer GPUs go past 200–300 kW per rack, **single-phase starts running into physical limitations**.
This means that the real challenge of the next decade remains unchanged: efficient heat dissipation at massive scale and doing it responsibly from an environmental standpoint. This is precisely where **dual-phase direct-to-chip cooling** comes in, because it can serve higher intensity GPUs, compared with single-phase.
**How do both systems work?**
* **Single-phase liquid cooling:** moves heat away by warming liquid;
* **Dual-phase liquid cooling:** removes heat by using the liquid-to-vapor transition as part of the cooling engine. Water can remain warmer than in single-phase.
**Better economics**
Cooling accounts for approximately **35%-40% of a data center's total energy consumption**, or **1.2% of total US energy consumption**. Reducing this spending is critical for every hyperscaler and neocloud. Dual-phase cooling reduces the CapEx required for chillers by reducing the number of chillers, significantly lowering annual OpEx, because all the racks in the data centers can operate with warmer water temperatures.
**Lower insurance costs**
Current data centers that use single-phase cooling face leak risk. However, with two-phase cooling, which uses specialized, non-conductive (dielectric) fluids, it significantly reduces the risk of equipment damage from leaks.
The question I leave is: Who is better positioned to capture the two-phase liquid cooling market and force hyperscalers to adopt it and include it in AMD and Nvidia GPUs reference architectures?