The Physical Stack of AI · Chips and accelerators
You can name the three credible Nvidia challengers in 2026 (AMD, Google TPU, AWS Trainium), explain what each one is built for, and recognize the cautionary tale of Tesla Dojo.
Every conversation about AI hardware eventually arrives at the same question: when will someone catch Nvidia? In 2026 the honest answer is "they haven't, but three players are putting real silicon and real customer dollars on the table — and one publicly tried and failed."
The credible challengers fall into two categories. AMD is the only company building a direct, merchant-market Nvidia competitor: a discrete GPU sold to anyone willing to buy. The MI400 preview, coming in 2026, is AMD's memory-heavy rack-scale bet: up to 432 GB of HBM4 per GPU, a 72-GPU Helios reference rack, and support for the open Ultra Accelerator Link (UAL) standard.
Google TPU and AWS Trainium are different — both are in-house chips, built for one company's use, available to the public only through their respective cloud. Google's TPU v7 "Ironwood" is now general-availability on GCP. AWS Trainium3 went GA in December 2025 and is the backbone of Anthropic's 5 GW capacity commitment.
Chapter contains 4 lessons.