Rival Chipmakers Join SGLang Consortium for Open AI Inference

LMSYS Org launched the SGLang Consortium on November 12, 2026, in the opening keynote of the inaugural SGLang Summit in San Francisco. The eight founding members are NVIDIA, AMD, Intel, Google, Meta, Cloudflare, Modal and RadixArk. They aren't merging anything. They're taking seats around one open-source project, the inference engine that serves trained AI models to users, and agreeing to build its roadmap together instead of in eight separate forks.

The structure is deliberately light. Each founding member designates a technical representative and joins two roadmap meetings a year with SGLang's maintainers. The consortium's stated work splits into five areas: support for frontier models, reasoning workloads, multimodal inference, distributed serving and hardware optimization. LMSYS is explicit that its maintainers keep final authority over technical decisions and code acceptance, so the group steers priorities rather than taking over the project.

The membership is the story. NVIDIA and AMD sell the accelerators that dominate AI data centers, and they compete for the same sockets. Intel is trying to win some of those sockets back. Google runs its own chips and its own cloud. Meta operates AI at consumer scale, Cloudflare serves models at the network edge, Modal sells serverless GPU capacity, and RadixArk is the infrastructure company founded by Ying Sheng and Banghua Zhu, whose team includes creators and core maintainers of SGLang itself.

SGLang's footprint explains why rivals bothered. The project says its version 1.0 milestone counts more than 330 adopting organizations and day-zero support for 217 models, and companies including Meta, Microsoft, Google and Amazon run production workloads on it. Inference is where AI bills accumulate: training happens once, but every question a user asks costs compute, and a faster, cheaper serving stack cuts the price of every answer a company ships.

Two concrete pieces of engineering arrived with the announcement. Intel and RadixArk said they'll optimize SGLang for Intel CPUs, GPUs and accelerators, including support for Intel Advanced Matrix Extensions. Separately, SGLang-Neuron version 0.1.0, co-developed by RadixArk and AWS Annapurna Labs, brings the serving stack to AWS Trainium chips through a native PyTorch plugin. AWS lists that plugin as a public preview, not a general release, and readers shouldn't treat it as finished software.

What the launch doesn't include matters as much as what it does. No new feature shipped with the announcement, no member disclosed funding or engineering headcount, and nobody published a cross-vendor benchmark. A consortium is a governance instrument, and governance only pays off in code: faster model support on launch day, serving costs that fall in public tests, and hardware from competing vendors that works without private patches. LMSYS says it plans to publish updates on the group's direction and progress.

The summit around the launch ran November 12 and 13 at the Fort Mason Center for Arts & Culture, with more than 100 speakers covering reinforcement learning, multimodal models and hardware. That's the audience the consortium has to convince. The engineers in those sessions don't grade press releases, and they'll know within a few release cycles whether eight rivals sharing a roadmap produces better inference software or just better keynote slides.