Amazon builds custom liquid cooling for Nvidia’s (literally) hottest AI chips

Amazon Web Services has developed its own in-house liquid cooling solution called In-Row Heat Exchangers (IHRX) to address the cooling requirements of Nvidia’s next-generation AI chips, particularly the Blackwell series. The move allows Amazon to upgrade its existing data center infrastructure without the time and cost of building new facilities, positioning the company to deploy more advanced AI chips while maintaining its competitive edge in cloud computing.

What you should know: Nvidia’s latest AI chips generate significantly more heat than previous generations, requiring liquid cooling rather than traditional air cooling systems.

• The Blackwell chips and newer models run so hot during AI tasks that they cannot rely on conventional air cooling methods without risk of overheating.
• Amazon wanted its current infrastructure ready for the Blackwell rollout but faced constraints with existing third-party solutions that couldn’t scale effectively.
• Building entirely new data centers would have taken too much time to meet deployment timelines.

How it works: Amazon’s IHRX system can be installed directly into existing data centers to make them compatible with liquid-cooled AI chips.

• The company developed this hardware internally rather than relying on external vendors, following Amazon’s typical approach of creating proprietary solutions.
• This enables Amazon to deploy more Blackwell chips without being limited by air-cooled infrastructure constraints.
• The solution addresses both timing and scalability issues that plagued third-party alternatives.

Market implications: The development represents a positive outcome for both Amazon and Nvidia while having mixed effects on data center equipment suppliers.

• Companies like Eaton, which specializes in electrical equipment including transformers and battery systems, could benefit from increased demand for uninterruptible power systems needed for liquid cooling pumps.
• Dover, which manufactures thermal connectors that serve as “plumbing” in liquid cooling systems, may see increased demand.
• The in-house development shouldn’t materially impact data-center-focused industrial companies since it addresses cooling specifically rather than broader infrastructure needs.

The big picture: Amazon’s move reflects the broader infrastructure challenges facing cloud providers as AI workloads become more demanding and heat-intensive.

• The development demonstrates how major tech companies are increasingly building proprietary solutions rather than relying on third-party vendors for critical infrastructure components.
• This trend could accelerate as AI chip requirements continue to evolve and create new technical challenges for data center operations.