AWS EC2 AI workloads push AMD instances deeper into cloud compute

AWS EC2 AI workloads are moving beyond conventional high-performance computing as customers use cloud infrastructure for agentic AI, physical AI and inference, according to Art Baudo, principal product marketing manager and head of EC2 product marketing at Amazon Web Services. In an interview with SiliconANGLE’s theCUBE at AMD Advancing AI 2026, Baudo said EC2 demand now spans AMD-based instances, Graviton, Trainium and Inferentia, with the AWS Nitro System providing the common security and performance layer.

The shift is relevant for infrastructure buyers because EC2 is now 20 years old and still absorbing workload types that were not part of its original design center. AWS did not disclose customer counts, revenue contribution, utilization rates or AI-specific EC2 growth figures. Baudo’s comments instead frame how AWS wants customers to think about compute selection as AI inference, bursty CPU jobs and memory-heavy workloads put new pressure on cloud economics.

How is AWS EC2 changing for AI workloads?

Baudo said customers are using AMD-powered EC2 instances for AI inference in addition to established workloads such as high-performance computing and electronic design automation. AWS first added AMD EPYC processors to EC2 in 2018 and has since supported each new generation, including Turin-based instances introduced in 2026.

According to Baudo, the move from multithreaded to single-threaded instances between the sixth and seventh generations produced performance gains that benefited AI and EDA users. AWS has also introduced high-frequency instances with 5GHz clock speeds and larger memory configurations for jobs that need both peak compute and faster access to data.

Baudo said AWS has kept price performance as a priority for AMD instances, including recent instance families. He did not provide specific benchmark results, price cuts or workload-by-workload comparisons, which leaves buyers to validate the economics against their own usage patterns rather than rely on broad vendor positioning.

Nitro remains the control point

The AWS Nitro System sits under EC2 as the architecture AWS uses to offload hypervisor functions to dedicated hardware and software. Baudo said that design lets AWS roll out instance types faster, run them at scale and enforce what he described as zero operator access.

For AI infrastructure teams, the point is consistency across a fragmented compute portfolio. Baudo said Nitro applies across AMD and Graviton instances as well as AWS’s AI chips, Trainium and Inferentia, giving customers the same baseline security model regardless of the compute tier they choose.

That positioning matters as AI workloads split across CPUs, GPUs and purpose-built accelerators. Agentic AI typically increases inference activity as software agents call models repeatedly to complete tasks. Physical AI refers to AI used in systems that interact with the physical world, such as robotics or industrial automation, where compute demand can combine model inference, simulation and data processing.

Cost pressure is part of the pitch

Baudo said AWS is applying the same cost-optimization discipline to AI that it used for broader cloud spending in 2022. His argument was that customers will consume more compute if AWS can make the unit economics work.

One tool AWS is emphasizing is Spot Instances, which let customers use spare EC2 capacity at lower prices when workloads can tolerate interruption. Baudo said AWS is trying to give customers more information about where Spot capacity is available so they can use it more effectively.

The comments reflect a practical constraint for AI infrastructure buyers: model demand may be rising, but budget discipline has returned to cloud procurement. AWS’s answer is a familiar one, more instance options, more specialized chips and more pricing mechanisms. The company did not claim that those tools remove the need for workload-specific benchmarking, which remains the buyer’s problem.

This story draws on original reporting from SiliconANGLE.