The letters AI, which stands for “artificial intelligence,” stand at the Amazon Web Services booth at the Hannover Messe industrial trade fair in Hannover, Germany, on March 31, 2025.
Julian Stratenschulte | Picture Alliance | Getty Images
Amazon said Wednesday that its cloud division has developed hardware to cool down next-generation Nvidia graphics processing units that are used for artificial intelligence workloads.
Nvidia’s GPUs, which have powered the generative AI boom, require massive amounts of energy. That means companies using the processors need additional equipment to cool them down.
Amazon considered erecting data centers that could accommodate widespread liquid cooling to make the most of these power-hungry Nvidia GPUs. But that process would have taken too long, and commercially available equipment wouldn’t have worked, Dave Brown, vice president of compute and machine learning services at Amazon Web Services, said in a video posted to YouTube.
“They would take up too much data center floor space or increase water usage substantially,” Brown said. “And while some of these solutions could work for lower volumes at other providers, they simply wouldn’t be enough liquid-cooling capacity to support our scale.”
Rather, Amazon engineers conceived of the In-Row Heat Exchanger, or IRHX, that can be plugged into existing and new data centers. More traditional air cooling was sufficient for previous generations of Nvidia chips.
Customers can now access the AWS service as computing instances that go by the name P6e, Brown wrote in a blog post. The new systems accompany Nvidia’s design for dense computing power. Nvidia’s GB200 NVL72 packs a single rack with 72 Nvidia Blackwell GPUs that are wired together to train and run large AI models.
Computing clusters based on Nvidia’s GB200 NVL72 have previously been available through Microsoft or CoreWeave. AWS is the world’s largest supplier of cloud infrastructure.
Amazon has rolled out its own infrastructure hardware in the past. The company has custom chips for general-purpose computing and for AI, and designed its own storage servers and networking routers. In running homegrown hardware, Amazon depends less on third-party suppliers, which can benefit the company’s bottom line. In the first quarter, AWS delivered the widest operating margin since at least 2014, and the unit is responsible for most of Amazon’s net income.
Microsoft, the second largest cloud provider, has followed Amazon’s lead and made strides in chip development. In 2023, the company designed its own systems called Sidekicks to cool the Maia AI chips it developed.
WATCH: AWS announces latest CPU chip, will deliver record networking speed