Amazon Unveils Powerful New AI Servers To Support Nvidia's Most Advanced Chips

Amazon.com AMZN and Amazon Web Services (AWS) announced on Wednesday that they have designed the P6e-GB200 UltraServers to meet the skyrocketing compute demands driven by the latest advancements in generative AI, including trillion-parameter foundation models, reasoning systems, and agentic AI.

AWS has engineered custom hardware to cool next-generation Nvidia NVDA GPUs used in AI workloads.

Recognizing the high energy demands of Nvidia’s latest chips, AWS decided not to wait and build new liquid-cooled data centers.

Also Read: Amazon Expands Global Data Centers, Boosts Access To Nvidia AI Chips For Cloud Customers

Instead, its engineers developed the In-Row Heat Exchanger (IRHX)—a solution that integrates with existing and new data centers to manage heat from dense GPU configurations, CNBC reported Wednesday.

These UltraServers feature up to 72 Nvidia Blackwell GPUs, all interconnected using fifth-generation Nvidia NVLink, allowing the GPUs to operate as a single, unified compute system.

Dave Brown, AWS VP of Compute and Machine Learning Services explained that off-the-shelf cooling options couldn’t meet the demands of Nvidia’s GB200 NVL72 systems, which pack 72 Blackwell GPUs into a single rack. Traditional air cooling was sufficient for previous GPU generations, but AWS had to innovate to keep pace with the increased compute intensity of today’s systems.

AWS now offers this infrastructure through its P6e instances, providing customers with the performance required to train and deploy large-scale AI models with advanced thermal support.

Microsoft MSFT and CoreWeave CRWV have previously offered computing clusters built on Nvidia’s GB200 NVL72 architecture.

In March, Intel INTC positioned its SuperFluid cooling technology as a strong contender for managing the heat generated by Nvidia’s high-powered GB300 AI chips.

Launched in 2023, SuperFluid has reportedly passed verification tests showing it can handle thermal loads up to 1,500 watts, surpassing the 1,400-watt consumption expected from Nvidia’s Blackwell Ultra GB300 chips.

In June, AI server company Super Micro Computer SMCI expanded its liquid-cooled AI server solutions for Nvidia’s Blackwell architecture into the European market.

The company introduced over 30 solution stacks for Nvidia HGX B200, GB200 NVL72 and RTX PRO 6000 Blackwell deployments, helping enterprises speed up AI factory rollouts.

Supermicro is scaling its portfolio with new liquid-cooled systems, including a 4U front I/O HGX B200 server powered by its DLC-2 technology. These systems enable customers to run high-density AI workloads while maintaining thermally efficient systems.

By working closely with Nvidia, Supermicro is also preparing to support next-generation Blackwell Ultra chips, such as the GB300 NVL72 and HGX B300, later this year.

Price Action: AMZN stock was down 0.34% at $221.79 premarket as of the last check on Thursday. NVDA is up 0.85%.

Read Next:

• Supermicro’s AI Boom, Liquid Cooling Edge Spark Analyst Confidence In Growth Potential

Photo: Shutterstock