AMD Made 12.4% of vLLM’s Commits in a Year, 2.6x Nvidia’s

Commits from AMD email addresses made up 1,652 of the 13,360 commits to vLLM, the open-source inference engine, between 30 October 2025 and 30 October 2026, by our count of the project’s public git history. That is 12.4% of all commits and 2.6 times the 628 we could attribute to NVIDIA addresses.

AMD’s share rose from 9.5% in the six months to April to 14.3% in the six months to October. More than half of all commits (54.6%) came from personal or GitHub no-reply addresses that cannot be tied to any company, so every company share here is a floor.

Business Engineer Course · The FDE Playbook

An evergreen Business Engineer course on forward-deployed engineering, how AI gets put to work inside organisations. It teaches the general idea only and does not discuss AMD, vLLM or this commit count.

How We Counted

We read the git log of the vllm-project/vllm repository for the 12 months to 30 October 2026. We attributed a commit to a company when the author email, or an email in the commit’s Signed-off-by line, used that company’s domain, such as amd.com or nvidia.com. Red Hat includes neuralmagic.com, and Intel includes habana.ai.

Of the 13,360 commits, 6,064 could be attributed to a company domain. AMD accounted for 27.2% of those attributed commits. After AMD came Red Hat with 1,124 commits (8.4% of all), Intel with 760 (5.7%), NVIDIA with 628 (4.7%), Inferact with 537 (4.0%), Meta with 306 (2.3%) and IBM with 300 (2.2%). Google addresses accounted for 48 commits (0.4%).

By our count, 157 distinct amd.com author addresses appear in the window, against 90 at nvidia.com and 38 at redhat.com, out of 2,563 distinct author addresses overall. An address is not always one person.

What Was Posted on X

The post on X by @sonnet_xu argues that AMD is trying to buy its way out of the CUDA moat by putting engineering and compute into vLLM, and that the returns are asymmetric because AMD has more to gain from hardware-agnostic inference. The view is the poster’s.

Not Only AMD’s Own Code

We also looked at which files the commits touched. 1,017 commits changed files whose paths contain “rocm” or “aiter”, AMD’s GPU software stack and its kernel library, and 541 of those (53.2%) came from AMD addresses.

But 1,111 of AMD’s 1,652 commits, or 67.3%, did not touch any such path, by our count. A path test like this one does not catch AMD-specific code stored under other file names.

What AMD Says

In a January 2026 blog post, AMD wrote that “AMD ROCm™ is now a first-class platform in the vLLM ecosystem” and credited “months of dedicated engineering and cross-team collaboration”. It said that in mid-November 2025 “only 37% of the vLLM test groups in the AMD CI pipeline were passing”, and that by mid-January 2026 “93% of the vLLM AMD test groups are succeeding”. The post gives no commit or headcount figures.

The Argument on X

On 30 October, @sonnet_xu wrote on X that “AMD is trying to buy its way out of the CUDA moat”. The post says AMD is “pouring eng + compute resources into vLLM, benefiting the entire ecosystem, including NVIDIA”, and that “the returns are asymmetric” because AMD “has much more to gain from making inference hardware-agnostic”.

The chart attached to the post lists vLLM’s cash backers as a16z, Dropbox, Sequoia, Skywork AI and ZhenFund, and its compute backers as including AMD, NVIDIA, AWS, Google Cloud, Intel, IBM and others. Our commit count measures something different from that chart: who writes the code, not who pays for or supplies the compute.

This connects to our coverage of the CUDA moat, today’s chip-stock selloff, Google’s TPU compute growth and open model weights that run on local hardware.

The Structural Read

As we read our own count, the ranking is among company email domains only; 54.6% of commits came from personal or no-reply addresses, so every company share is a floor.

As we read the file paths, 67.3% of AMD’s commits did not touch paths containing rocm or aiter, though a path test does not catch AMD-specific code stored under other names.

As we read AMD’s January blog post, it gives test pass rates rising from 37% to 93% but no commit or headcount figures, which is why we counted the git log ourselves.

Three Implications

FOR THE CUDA DEBATE The post on X by @sonnet_xu argues that AMD is trying to buy its way out of the CUDA moat.

FOR READING OPEN SOURCE Our count measures who writes vLLM’s code; the chart attached to that post lists who funds the project or supplies compute.

FOR THE NUMBERS By our count, AMD addresses made 541 of the 1,017 commits that touched rocm or aiter paths.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Open vs Closed Meta-Framework.

The framework states the pattern this way: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.

The framework is the owner’s general argument about technology stacks, not evidence about vLLM; it does not discuss AMD, NVIDIA or this commit count.

What Is Not Established

Email domains are a proxy. Engineers at any company may commit from personal or no-reply addresses, which covers 54.6% of commits in the window, so the true company shares are unknown. Commit counts do not measure the size or importance of a change. The count covers one repository and one 12-month window.

The Bottom Line

By our count of vLLM’s public git history, AMD addresses made 1,652 of 13,360 commits in the 12 months to 30 October 2026: 12.4% of all commits, 2.6 times NVIDIA’s 628, and rising from 9.5% to 14.3% between the two halves of the year. 67.3% of AMD’s commits touched code outside its own ROCm and AITER paths.