THE ESSENTIALS IN 30 SECONDS

  • LMSYS Org announced the SGLang Consortium on November 12, 2026, at the first SGLang Summit in San Francisco.
  • Founding members are NVIDIA, AMD, Intel, Google, Meta, Cloudflare, Modal and RadixArk.
  • Members name technical representatives and join two roadmap meetings a year; maintainers keep the final say on code.

LMSYS Org, the non-profit behind the SGLang inference engine, has set up an industry consortium to steer the project's development. The SGLang Consortium was announced on November 12, 2026, during a keynote at the first SGLang Summit in San Francisco, with eight founding members: NVIDIA, AMD, Intel, Google, Meta, Cloudflare, Modal and RadixArk.

How the consortium will work

According to the announcement, each founding member appoints a technical representative and takes part in two roadmap meetings per year. Contributions can take several forms, from engineering expertise and deployment feedback to testing resources and jointly developed features. Working groups may continue between meetings, and LMSYS says it will publish updates on roadmap direction and progress.

The structure stops short of handing control to the companies involved. SGLang's maintainers remain responsible for technical decisions and for deciding which code is accepted, and the project stays open to contributors outside the consortium. The stated focus areas are support for frontier models, reasoning and multimodal workloads, distributed inference and hardware optimisation.

Why the hardware makers are at the table

SGLang is software for serving trained models in production, handling the requests that hit a model after deployment. It is known for techniques such as prefix caching and for spreading work across many GPUs, and it already supports NVIDIA and AMD GPUs, Intel Xeon processors and Google TPUs. LMSYS says the engine runs on hundreds of thousands of GPUs, and the announcement names Microsoft, Amazon, Thinking Machines Lab and Reflection among companies using it in production, although they are not listed as consortium members.

For chip vendors, an open engine that performs well on their hardware is a route to customers who would rather not be tied to one supplier. The announcement cites a Deloitte estimate that inference will account for roughly two-thirds of AI compute in 2026, up from about half in 2025, which helps explain the interest. Ying Sheng, a co-author of the original SGLang paper and chief executive of RadixArk, said the largest operators "see what inference needs next" before others do.

Open questions

No release schedule, feature list or performance targets were announced for the consortium's work. Its influence will depend on how much engineering time members commit and on how the maintainers balance competing hardware priorities. SGLang also competes with other open-source serving engines such as vLLM, and some vendor tools, including AMD's Hyperloom, already support both.