Wafer’s write-up (blog) reports GLM-5.2 served on AMD’s Instinct MI355X at 2,626 tokens per second per node for less than half the cost of Nvidia’s Blackwell, and credits AI agents writing the kernels for closing the software gap that has protected the incumbent. The newsletter reads it as ‘the silicon is cooperating’ with the cheap post-training economics of the same open model. In the recursion trajectory it carries the kernel strand, GPT-5.5 topping KernelBench in April and Fable 5’s megakernel the day before, from benchmark scores to production serving on a rival vendor’s chips, a month before self-improving agents rebuild an inference stack on B200s.
Development
A challenger GPU serves an open model at half the cost as agents write kernels
The kernel-writing loop leaves the leaderboard and lands on market structure: if agents write the kernels, the software moat that kept a challenger GPU out of frontier serving stops being a moat.
Claim
Wafer served GLM-5.2 on AMD's MI355X at 2,626 tokens per second per node at less than half Blackwell's cost, arguing AMD's software gap is closing because AI agents now write the kernels.
Reported metric: 2,626 tok/s/node
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to July 4, 2026 (theinnermostloop.substack.com)
- Performance per dollar is getting faster and cheaper (www.wafer.ai)
Verification record
These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.
- Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.