The rise of artificial intelligence accelerators is pushing data centers to unprecedented thermal limits. Training and inference workloads generate heat densities that traditional cooling systems can barely handle. In this context, startup Accelsius has introduced a two-phase cooling solution that could redefine the operational efficiency of bit barns. But beyond the technology, the key lies in how companies can integrate these innovations with custom software to optimize their infrastructures.
Accelerators like the Nvidia B200 or upcoming 72-GPU racks reach thermal powers of up to 250 kW per rack, expected to hit 600 kW by 2027. Air cooling is no longer enough; single-phase liquid systems fall short. Accelsius proposes using two-phase fluids (similar to air conditioning refrigerants) that evaporate in cold plates, extracting more heat per surface area. According to their tests, they can operate with facility water at 54 °C (and even 59 °C by doubling flow rate), compared to the typical 45 °C. Each degree increase in facility water temperature yields 4% annual energy savings.
This approach not only reduces electricity consumption but also allows waste heat to be used for applications like building heating or domestic hot water, since above 50-60 °C the heat has practical value. Accelsius’ technology, called NeuCool, uses skived fin cold plates with small channels where the fluid boils in flow, without requiring large immersion baths. This makes it compatible with standard racks, though it requires specific coolant distribution units (CDUs).
However, mass adoption faces challenges. Current cold plates are customized for each accelerator; Accelsius is working on universal designs so OEMs like Dell or Supermicro can offer servers compatible with both coolant types. Additionally, CDUs must scale: the company already has a 350 kW unit and is developing a 2.4 MW one. On the other hand, two-phase fluids historically contained PFAS (perfluoroalkyl substances) linked to health issues; Accelsius uses an ASHRAE A1 refrigerant blend, non-toxic and non-flammable.
For companies managing data centers, the decision to migrate to two-phase cooling must be accompanied by a comprehensive monitoring and control strategy. This is where services like AI agents come in, optimizing cooling demand in real time based on workload. It is also essential to have cybersecurity solutions to protect thermal management systems, and BI platforms based on Power BI to analyze energy efficiency KPIs. Q2BSTUDIO, as a software development company, offers cloud services on AWS and Azure that enable scalable deployment of these architectures, integrating sensors, APIs, and custom dashboards.
In short, two-phase cooling is not just a technical improvement: it is an enabler for data centers to keep pace with AI demand. But to get the most out of it, organizations need an ecosystem of custom applications, process automation, and advanced analytics. The combination of innovative hardware and intelligent software will make the difference between a data center that survives and one that thrives.
