Amazon: Workloads, Not Benchmarks, Shaped A Decade Of Its Chips
Amazon published a recap of a fireside chat that Peter DeSantis, its senior vice president of foundational AI models and custom silicon, gave at AI Infra Summit on September 16, 2026, with SemiAnalysis founder Dylan Patel.
Transcript
Amazon recapped a decade of its chip design. Customer workloads guided each generation.
Peter DeSantis leads its custom silicon work. He said understanding the workload deeply is a huge asset.
Some early customers asked for bare metal performance. That produced the Nitro System, built with Annapurna Labs.
Graviton came from years of real workload data, not microbenchmarks. Amazon says the vast majority of AWS workloads run faster and cost less on it.
The same approach shaped the AI chips. DeSantis expects most AI compute to go toward inference. He also expects more chips better suited to specific workloads.
Inboxsmith builds AI phone and email agents for small business. Please like and subscribe for more news.
Sources
Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:
- At AI Infra Summit, Amazon SVP Peter DeSantis explained how real customer workloads guided every generation of Amazon's chips, from Nitro to Trainium.(Amazon (aboutamazon.com), first party article titled "Inside a decade of Amazon chip design, built workload by workload", published Sep 22, 2026, written by Amazon Staff, recapping SVP Peter DeSantis at AI Infra Summit on Sep 16, 2026; read from the alert-time snapshot captured 2026-09-22T18:47:11Z.)
- said DeSantis, Amazon's senior vice president of foundational AI models and custom silicon, in a fireside chat with SemiAnalysis founder Dylan Patel.(Amazon (aboutamazon.com), first party article titled "Inside a decade of Amazon chip design, built workload by workload", published Sep 22, 2026, written by Amazon Staff, recapping SVP Peter DeSantis at AI Infra Summit on Sep 16, 2026; read from the alert-time snapshot captured 2026-09-22T18:47:11Z.)
- "When you're building hardware and building complex systems, understanding the workload deeply is a huge asset," said DeSantis(Amazon (aboutamazon.com), first party article titled "Inside a decade of Amazon chip design, built workload by workload", published Sep 22, 2026, written by Amazon Staff, recapping SVP Peter DeSantis at AI Infra Summit on Sep 16, 2026; read from the alert-time snapshot captured 2026-09-22T18:47:11Z.)
- Most workloads ran well on early EC2, but some customers kept asking for bare metal performance.(Amazon (aboutamazon.com), first party article titled "Inside a decade of Amazon chip design, built workload by workload", published Sep 22, 2026, written by Amazon Staff, recapping SVP Peter DeSantis at AI Infra Summit on Sep 16, 2026; read from the alert-time snapshot captured 2026-09-22T18:47:11Z.)
- The result was the Nitro System, which moves the virtualization stack onto dedicated hardware. That work brought Amazon together with Annapurna Labs(Amazon (aboutamazon.com), first party article titled "Inside a decade of Amazon chip design, built workload by workload", published Sep 22, 2026, written by Amazon Staff, recapping SVP Peter DeSantis at AI Infra Summit on Sep 16, 2026; read from the alert-time snapshot captured 2026-09-22T18:47:11Z.)
- Amazon designs its chips using data from real workloads, its own and its customers', rather than benchmarks.(Amazon (aboutamazon.com), first party article titled "Inside a decade of Amazon chip design, built workload by workload", published Sep 22, 2026, written by Amazon Staff, recapping SVP Peter DeSantis at AI Infra Summit on Sep 16, 2026; read from the alert-time snapshot captured 2026-09-22T18:47:11Z.)
- that's the right way to optimize a chip and a system, not microbenchmarks," DeSantis said.(Amazon (aboutamazon.com), first party article titled "Inside a decade of Amazon chip design, built workload by workload", published Sep 22, 2026, written by Amazon Staff, recapping SVP Peter DeSantis at AI Infra Summit on Sep 16, 2026; read from the alert-time snapshot captured 2026-09-22T18:47:11Z.)
- That approach produced Graviton. Today, DeSantis noted, the vast majority of workloads on AWS run both faster and more cost-effectively on it.(Amazon (aboutamazon.com), first party article titled "Inside a decade of Amazon chip design, built workload by workload", published Sep 22, 2026, written by Amazon Staff, recapping SVP Peter DeSantis at AI Infra Summit on Sep 16, 2026; read from the alert-time snapshot captured 2026-09-22T18:47:11Z.)
- The same principle has guided Amazon's AI chips. DeSantis expects the majority of AI compute to go toward inference, whether that's serving end users or generating the rollouts used in reinforcement learning.(Amazon (aboutamazon.com), first party article titled "Inside a decade of Amazon chip design, built workload by workload", published Sep 22, 2026, written by Amazon Staff, recapping SVP Peter DeSantis at AI Infra Summit on Sep 16, 2026; read from the alert-time snapshot captured 2026-09-22T18:47:11Z.)
- but I think we're going to see more chips that are better suited to specific workloads.(Amazon (aboutamazon.com), first party article titled "Inside a decade of Amazon chip design, built workload by workload", published Sep 22, 2026, written by Amazon Staff, recapping SVP Peter DeSantis at AI Infra Summit on Sep 16, 2026; read from the alert-time snapshot captured 2026-09-22T18:47:11Z.)