On November 12, 2026, LMSYS Org, the nonprofit that supports SGLang, launched the SGLang Consortium with eight companies, including NVIDIA and AMD. SGLang is an engine for "inference," the process of running a trained AI model to respond to user input. Under the new body, each company will send a technical representative and contribute development challenges and resources, while the project's maintainers will continue to make technical decisions and decide which code to adopt. The question now is how companies can jointly develop software that already supports multiple types of hardware.
The eight founding members are NVIDIA, AMD, Intel, Google, Meta, Cloudflare, Modal, and RadixArk. LMSYS announced the launch in a keynote at the first SGLang Summit in San Francisco. The event runs November 12-13 at the Fort Mason Center for Arts & Culture, with more than 100 speakers covering topics such as reinforcement learning, multimodal models, and hardware.
Twice-yearly meetings, with maintainers making technical decisions
Each founding member will name a technical representative and take part in a roadmap meeting held twice a year. Members will bring specialized expertise and feedback from production operations to those meetings, and will also work on providing testing resources and co-developing features. Between meetings, working groups may collaborate as needed, according to the announcement.
The areas of focus include support for frontier models, so that new high-performance models can run; processing for reasoning models, which work toward an answer step by step, and for multimodal models; distributed inference, which splits workloads across multiple compute resources; and hardware optimization. The structure allows model developers, service operators, and compute providers to bring their challenges to the same table.
| Party | Role described in the announcement |
|---|---|
| Founding members | Name technical representatives; take part in twice-yearly meetings and co-development |
| Working groups between meetings | Collaborate as needed |
| Maintainers | Handle technical decisions and code adoption |
Note: Roles by party, as described in LMSYS's November 12, 2026 announcement.
In August, before the consortium was formed, AMD said its GPU optimization tool Hyperloom would support SGLang. Under the new setup, companies bring challenges and resources, while maintainers handle technical decisions and code adoption. That makes it possible to distinguish existing support from future results. LMSYS also said it will publish and update information on the consortium's direction and progress, and that it plans to open the project to the broader community.
AMD mentioned SGLang support in August materials
SGLang is an inference platform for running large language models and multimodal models in production. Its official documentation says it aims for low latency and high throughput using techniques such as prefix caching, which reuses computation results for the beginning of an input, and parallel processing across multiple GPUs. The design addresses the burden of repeatedly computing the same content and the limits of handling a large workload on a single GPU.
Supported hardware includes NVIDIA and AMD products, as well as Intel Xeon and Google TPUs. The participation of multiple companies does not mean this hardware is being supported for the first time now.
AMD introduced Hyperloom at its August 27 ROCm 10 announcement. It optimizes for AMD Instinct GPUs and also supports vLLM. According to the company, Hyperloom analyzes workloads to identify bottlenecks, searches for optimization options, and implements changes. It then runs benchmarks to verify performance and correctness. It performs optimizations using HIP, Triton, and FlyDSL, and reports proposed changes along with measured or expected improvements.
Building on that August support, the November launch adds a venue where companies can continuously discuss direction. AMD's Newsroom also posted an announcement of the launch dated November 12 and linked to the LMSYS post. Optimizations already under way individually, and changes that emerge from future meetings and joint development, can be evaluated on separate timelines.
Improving inference efficiency through software
According to LMSYS, SGLang runs on hundreds of thousands of GPUs. LMSYS named Meta and Google as users, along with Microsoft, Amazon, and Thinking Machines Lab, among others. The list of users partly overlaps with the founding members. In the consortium, founding members plan to bring specialized expertise and operational feedback.
Ying Sheng, a co-author of the original SGLang paper and co-founder and CEO of RadixArk, cited the difficulty of building new data centers and said, "software approaches that improve AI efficiency are extremely important." This reflects expectations for efficiency gains. NVIDIA's Ian Buck also said software improvements can unlock more performance from the company's compute resources, and suggested that lower inference costs would make AI easier to use.
In a forecast published in November 2025, Deloitte projected that inference's share of total AI compute would rise from about half in 2025 to roughly two-thirds in 2026. That is a forecast rather than a tally of 2026 actuals, but it illustrates the background: repeatedly running trained models also requires compute resources. The aim of bringing together model, operations, and hardware companies is to advance improvements that span their different conditions.
How to read the published plans and implementations
LMSYS plans to publish and update information on the consortium's direction and progress. However, the launch announcement does not give release timing or priorities for individual features, or specific performance improvement figures. For users and developers, the key questions are which features are actually adopted and under what conditions, for their own models and hardware configurations, the benefits can be confirmed.
The open question is whether the operational challenges companies bring can be turned into implementations that maintainers can adopt. If published changes come with consistent measurement conditions, users will be better placed to identify the improvements their own environments need.