Intel / article

Can AMD break the CUDA Moat? AMD Advancing AI 2026

Source: newsletter.semianalysis.com
Author: Bryan Shan
Date: Published Jul 25, 2026

Context

Reliable access to test hardware is central to AMD's software progress, according to research publisher SemiAnalysis. Its account describes shortages and instability in the GPU clusters engineers use to develop and test software spanning multiple machines. Coding agents increase that demand because their changes also need hardware-based testing. The publisher argues that supplying more coding assistance cannot remove this bottleneck by itself: development teams need dependable clusters and automated checks that must pass before changes are accepted. This is the report's assessment of an engineering constraint on making AI chips useful at scale. Original analysis.

Key takeaways · AI-distilled

SemiAnalysis moved AMD from a 0% chance of closing the gap with Nvidia to a good chance, based mainly on leadership behavior: Lisa Su took the calls and shipped the suggested changes rather than routing them through committees.
AMD's Helios rack is ramping slowly because it kept a cabled compute tray design and its weak SerDes force retiming on up to 85% of the backplane, which means over 550 Broadcom ethernet retimers per rack plus reliability problems.
Gating tests, the ones that block a merge until they pass, are the number that matters; non-gating pass rates are for non-technical audiences. AMD's Kubernetes inferencing CI for its Pollara NIC sits at 0% parity with Nvidia's ConnectX nightly CI.
Agentic coding multiplies the hardware each engineer needs, since every wants GPUs to test against and one human runs dozens of them. AMD's internal GPU shortage is therefore worsening, and clusters were pulled off its own vLLM team mid-push.
AMD's equity rebate gives Meta and OpenAI close to a 105% discount, triggered once the stock reaches $600 and they buy enough compute, which makes Helios racks effectively free with 5% on top.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

Key quotes

“Software was broken, progress was unexciting, and we were the top bug submitter for many months with dozens of AMD engineers triaging our bug reports.”

“The pie is growing rapidly for everyone, and Nvidia will continue to massively grow revenue.”

“This lack of GPU nodes is a problem that has been getting even worse due to the rise of agentic coding.”

“The MI455X is the most advanced chip that comes out of a fab on the silicon front.”

“If AMD can overcome these challenges, we strongly believe that AMD will be well positioned to do well and take market share.”

Read the source newsletter.semianalysis.com