Nvidia, AMD and Intel now back the same free software for running AI models. None has said what it will spend

Three rival US chipmakers have agreed to help steer one piece of free software. Nvidia, AMD and Intel joined Google, Meta, Cloudflare, Modal and the start-up RadixArk in a new SGLang Consortium. It was announced on November 12 at the first SGLang Summit in San Francisco.

SGLang is an open-source inference engine. Inference is the work of running a trained AI model to answer users, as opposed to building the model. The engine sits between a company's AI models and its chips, and decides how efficiently that hardware is used.

That layer is where AI spending is heading. LMSYS Org, the nonprofit behind SGLang, cites a Deloitte forecast that inference will take about two-thirds of AI computing in 2026, up from about half in 2025. That is a forecast, not a measured figure.

Why chip rivals would share the same software

For buyers, the appeal is choice. A big reason companies default to Nvidia is that most AI software is tuned for its chips first. If a shared, free engine runs well on AMD and Intel hardware too, switching suppliers gets cheaper and less risky.

Intel's side deal points the same way. Italian outlet 01net reported that Intel and RadixArk will tune SGLang for Intel processors and accelerators. They will merge the work into the main code rather than a separate version. The stated goal is fast AI software on Intel hardware that leaves developers free to choose their own tools.

Intel has not published benchmarks for the work, and no financial terms were disclosed. AMD's Ramine Roane said SGLang already runs on AMD's Instinct chips.

Nvidia's reason for joining is different. Nvidia executive Ian Buck said better software gets more performance from Nvidia's computing hardware. The logic: free software that squeezes more work from each chip can make Nvidia's hardware more valuable, not less. 01net noted that SGLang does not replace CUDA, the Nvidia programming system that underpins much of its lead.

The software already has heavy users. LMSYS says it runs on hundreds of thousands of GPUs, the chips that do most AI work. It names Meta, Microsoft, Google, Amazon and Thinking Machines Lab among production users. SiliconANGLE reported in May, citing RadixArk, that clusters running SGLang held more than 400,000 graphics chips in total.

What the group has not said

The announcement is light on commitments. LMSYS says members will contribute engineering help and testing resources, but no member has disclosed money, staffing levels or dues. The group lists broad focus areas but no timeline or performance target. Each member names one technical representative and attends two planning meetings a year.

SGLang's own maintainers keep final say over which code is accepted. That is meant to keep any one sponsor from steering the project. It also means the consortium cannot force the chip-specific work each member wants.

Tech site Byteiota compared the structure to the industry foundation that oversees Kubernetes, widely used data center software that no single vendor controls. It said the real test is whether support for new models lands on AMD and Intel chips on day one. The group has not yet reported progress on that.

Tech site Xenospectrum noted that SGLang already supported Nvidia and AMD chips, Intel processors and Google's custom chips before the group formed. In that sense, the consortium builds on existing work rather than starting it.

The start-up in the middle

The clearest commercial winner may be RadixArk. Its co-founder and CEO is SGLang co-author Ying Sheng, who spoke for the SGLang team at the launch. It sells paid hosting and tools built on the free engine.

RadixArk raised a $100 million seed round in May, led by Accel alongside Spark Capital. Nvidia's venture arm and AMD both invested, SiliconANGLE reported. It cited a Wall Street Journal report valuing RadixArk at $400 million. A consortium seat beside its own investors gives RadixArk a central role in where the free software goes next.

That mirrors a wider pattern in open-source AI. The free project builds the user base, and an affiliated company sells services around it. SGLang's best-known rival, vLLM, came out of the same UC Berkeley lab and has its own venture-backed company. TechCrunch reported in January, citing Forbes, that it was in funding talks at about a $1 billion valuation. A vLLM co-founder called the reported information inaccurate without saying which details were wrong.

Head-to-head comparisons published online are mixed. Byteiota said SGLang leads on some workloads, while vLLM leads on others. Buyers should test both on their own traffic. Vendor and community figures are not independent audits.

For executives, the practical takeaway is modest for now. The group lowers the risk of standardizing on SGLang, because every major chipmaker has a stake in it. Whether it actually loosens Nvidia's grip will depend on work none of the members has yet put a price on.