Recent reports from late 2024 through 2025 have revealed rising tensions between Nvidia and Microsoft over the deployment of Blackwell GPUs, the most power-hungry and advanced AI chips Nvidia has ever produced. The friction began after an internal Nvidia memo surfaced in which an engineer criticized Microsoft’s cooling method as “wasteful.” This has sparked widespread discussion about the future of AI data center design, efficiency, and scalability.
The Source of the Criticism: “Wasteful” Cooling Strategies
The controversy centers around Microsoft’s installation of GB200 NVL72 racks, which pack 72 Blackwell GPUs into a single liquid-cooled unit generating up to 120kW of heat. According to the leaked email, Microsoft’s data center intended to support OpenAI used a hybrid cooling setup, relying on both liquid cooling at the rack level and air-cooling at the building level.
Industry experts suggest the “wasteful” label refers to the significant electricity required to push massive volumes of air through the facility. While hybrid cooling reduces water usage, it increases power demand and reduces overall thermodynamic efficiency compared to full liquid-to-liquid systems.
Microsoft’s Response: Speed Over Perfection
Microsoft defended its infrastructure strategy, stating it uses a closed-loop liquid cooling heat exchanger, enabling high-density AI racks to operate inside existing air-cooled buildings. The company argues this approach helps them:
- Scale AI capacity quickly
- Avoid multi-year construction delays
- Maximize their global data center footprint
- Maintain sustainability goals without excessive water usage
According to Microsoft, this is the fastest and most practical way to support exploding AI workloads from partners such as OpenAI.
Installation Challenges and Operational Friction
The Nvidia employee’s memo also highlighted installation hurdles, noting that Microsoft required extensive on-site guidance to validate and verify the new systems. Documentation and handover protocols needed more “solidification,” pointing to growing pains in deploying a generation of hardware that pushes the limits of power and heat density.
This followed earlier reports from 2024 showing Nvidia had already redesigned Blackwell racks multiple times due to overheating and mechanical stress during early tests.
Nvidia’s Public Position: Everything Is Normal
Nvidia has downplayed the dispute publicly, framing the challenges as normal engineering iterations. They emphasize that Blackwell systems are already deployed at scale and deliver exceptional performance and efficiency once installed.
The situation underscores a deeper truth in modern computing: AI performance is now limited more by cooling and power than by chip design. As the industry shifts from air-cooled servers to fully liquid-cooled infrastructure, hyperscalers like Microsoft face the tough balancing act of retrofitting older data centers while racing to meet unprecedented AI demand.
