RadixArk

Overview

RadixArk builds infrastructure for running open AI models in production and improving them over time. Its founders created and still maintain SGLang, an open-source inference engine that serves trillions of tokens per day for organizations including Google, Microsoft, NVIDIA and xAI, and the company also maintains Miles, an open-source framework for reinforcement learning. On top of those projects it sells managed infrastructure and tooling to teams that would rather own and operate their AI systems than rent tokens through another company’s API.

Thesis

For most of the period since ChatGPT’s release, training dominated the cost of AI, and inference has now overtaken it. In August 2026 Gartner forecast that spending on inference within AI-optimized cloud infrastructure would reach $23.3 billion in 2026, passing training spending of $19 billion for the first time, with inference rising from 55% of that spending in 2026 to 59% in 2027. On NVIDIA’s February 2026 earnings call, Jensen Huang told investors that inference now equals revenue. The change comes from how models are used. An agent makes dozens of model calls to finish one task, and a reasoning model generates thousands of tokens before it answers. By November 2025, production LLM systems were reportedly processing tens of trillions of tokens per day.

Enterprise budgets point the same way. According to Contrary Research, enterprise LLM budgets grew 75% year over year in 2025 and moved from experimental lines into permanent ones, and 37% of enterprises reportedly run five or more models across experimentation and production. Agentic systems, which plan, call tools and iterate, use five to 30 times more compute per task than a standard chatbot.

At that volume, the software sitting between the GPU and the application starts to matter financially. The inference engine decides how GPU memory is managed, how requests are batched and scheduled, which low-level kernels run and which chips are supported. An engine that serves twice as many requests on the same hardware effectively doubles its operator’s GPU fleet, at a time when both chips and power are scarce. As of 2026 two open-source engines dominate open-model inference, vLLM and SGLang, and both started in 2023 in the UC Berkeley lab of Databricks co-founder Ion Stoica.

Enterprises, meanwhile, have started building whole products on AI, often running multiple models, retrieval pipelines, fine-tuning loops and continuous post-training as one system. Closed-model APIs are easy to adopt, but they limit customization, control and cost management at scale. Open-weight models such as Llama, Qwen, DeepSeek, Kimi and GLM give teams that control, at the price of specialized performance engineering for every new model and every new chip. Most teams lack that expertise, and RadixArk is betting they will pay someone who has it.

Founding Story

Ying Sheng (CEO) and Banghua Zhu founded RadixArk in 2025 together with a group of SGLang’s core developers and contributors to related open-source reinforcement learning projects.

Sheng studied in the ACM Class at Shanghai Jiao Tong University, earned a master’s degree at Columbia and completed a PhD in computer science at Stanford. There she won a best paper award for work in formal verification before turning her research to high-throughput inference. She went on to work as a research scientist at Databricks, which she credits with shaping her commitment to open source, and co-founded LMSYS, the nonprofit research collective best known for Chatbot Arena, started by students from Stanford, Berkeley, Carnegie Mellon and other universities.

SGLang began in the summer of 2023 and was released publicly in January 2024. The name stands for structured generation language, and the original goal was a better way for programs to hold structured interactions with models, but performance is what won it users. Sheng has described several waves of growth after the first release, including long nights in the summer of 2024 spent debugging with a small group of other core developers.

In October 2024 Sheng joined xAI, where she co-led the team that served the Grok models. She left in August 2025, reportedly two months before her first equity cliff, to work on SGLang full time. When she announced the company in December 2025, she wrote that hundreds of people had contributed to the project over two years and that demand had grown to the point where the project needed a dedicated organization. Zhu joined from NVIDIA, where he had worked on AI systems.

HOF Capital, one of the seed investors, has said it met the founders before the company was incorporated and invested mainly because of the team. News of the financing surfaced in January 2026, and RadixArk formally launched on May 5, 2026. At launch Sheng said the company treats “inference, training, and post-training as first-class citizens,” and described the goal as end-to-end infrastructure that gives developers more speed and control. Accel, which led the round, pitched the engine as a way to expand the capacity of hardware customers already own, since demand for compute is growing faster than the supply of chips.

Product

RadixArk’s offering has three parts. Two are open-source projects it maintains and gives away, and the third is a commercial platform built on top of them. The company keeps no private fork of SGLang. The project belongs to the LMSYS community, and all of RadixArk’s contributions to it are public.

SGLang

A model generates text one token at a time, and at each step it has to account for everything that came before. To avoid recomputing that work, the engine keeps intermediate results in a structure called the KV cache, which is large, lives in scarce GPU memory and grows with every token. Much of inference performance comes down to how well that cache is managed.

SGLang’s signature technique is RadixAttention. When many requests start with the same text, such as a system prompt, a set of tool descriptions or the earlier turns of a conversation, the shared portion only has to be computed once. SGLang indexes every prefix it has processed in a radix tree and maps each one to memory that already holds its results, so a new request with a matching prefix jumps straight to its new portion. Nearly all production workloads share prefixes, and agentic applications share the most, because each step resends the history of the steps before it.

According to HOF Capital, SGLang has shipped day-of-release support for the Llama, Qwen, DeepSeek, Kimi, GLM, Gemma and Mistral model families, and it runs on NVIDIA GPUs, AMD GPUs, Intel CPUs and Google TPUs. For users that cadence may be the product itself, since it assures them the next important model or chip will work well on the day it arrives.

Miles

Miles is an open-source reinforcement learning framework aimed at enterprises, released in 2026. Reinforcement learning has become the main way to specialize and improve models after initial training. A model attempts a task many times, the attempts are scored, and the model is updated toward the better ones. The most expensive step in that loop is generating the attempts, which is an inference workload, and Sheng has pointed out that a reward engine is essentially an inference engine. A team that leads in efficient inference therefore starts with an advantage in making reinforcement learning affordable, which is why RadixArk pairs the two projects.

The managed platform

On top of the open-source projects, RadixArk sells managed infrastructure for inference and training with enterprise reliability, security and support. A press report in January 2026 said the company had begun charging for managed hosting. Longer term, the company describes a complete foundation for building specialized models, made up of codebases, tools, sandboxes, training environments and intermediate checkpoints that teams assemble around their own data and goals. HOF Capital contrasts this with inference APIs, which rent out a finished product, and says RadixArk instead “gives you the factory.”

Customer

SGLang’s open-source users include some of the largest AI organizations in the world. According to the company, it serves trillions of tokens per day for Google, Microsoft, NVIDIA, Oracle, AMD, Nebius, LinkedIn, xAI, Thinking Machines Lab and humans&, across hundreds of thousands of GPUs. All of them use SGLang as free software, and none has been publicly identified as a paying customer.

The commercial platform targets a different buyer, an enterprise or AI-native company that wants to run and adapt open models on infrastructure it controls without staffing a low-level optimization team. Such a buyer typically serves several models, cares about latency and cost at high volume, has proprietary data it wants to specialize models on, and may face regulatory or contractual limits on sending data to third-party APIs. Its alternatives are a closed-model API, which offers little control, or self-hosting an open engine, which takes scarce expertise.

Market Size

Gartner projected in August 2026 that worldwide spending on AI-optimized infrastructure as a service would grow from $21.5 billion in 2025 to $42.3 billion in 2026 and $66.1 billion in 2027, with inference making up 55% of the 2026 total and 59% of the 2027 total.

Fortune Business Insights estimated the broader AI inference market at $103.7 billion in 2025 and $117.8 billion in 2026, and longer-range estimates point the same way. Figures compiled by Contrary Research project inference growing from $106 billion in 2025 to $255 billion by 2030 and accounting for 75% of all AI compute by then. Enterprise spending on foundation model APIs rose from $3.5 billion in all of 2024 to $8.4 billion in the first half of 2025. Separately, 23% of enterprises have fully integrated agentic AI and 74% expect to use it at least moderately within two years. Those two numbers measure different things and should not be read as points on one curve.

So far usage has grown faster than prices have fallen. Each drop in the cost per token has been followed by a bigger rise in tokens consumed, as agents and long-form reasoning become affordable to run. If that holds, efficiency gains at the engine layer grow the market, and RadixArk’s whole position rests on that assumption.

Competition

RadixArk competes on two levels. At the engine level its rival is vLLM and the company formed around it. At the platform level it faces managed inference providers, the AI services of the large clouds and chip vendors’ own serving software, and many of those competitors also use SGLang.

Inferact

Inferact was founded by the creators and core maintainers of vLLM, which, like SGLang, came out of Ion Stoica’s Berkeley lab in 2023. vLLM’s founding contribution was PagedAttention, which manages the KV cache in pages the way an operating system manages memory. Its chief executive has put its deployment at more than 400,000 GPUs, the same figure SGLang claims, so public statements cannot settle which installed base is larger. In January 2026 Inferact raised $150 million in seed funding at an $800 million valuation, co-led by Andreessen Horowitz and Lightspeed with participation from Sequoia, Altimeter, Redpoint and ZhenFund. It plans a commercial product described as a universal inference layer that works with existing providers.

Fireworks AI

Fireworks AI was founded in 2022 by former Meta engineers who worked on PyTorch, and it runs a managed platform for open and enterprise-owned models. In July 2026 it raised $1.5 billion at a $17.5 billion valuation and announced more than $1 billion in annualized revenue and 40 trillion tokens processed per day. It competes head-on with RadixArk’s managed platform, at far greater scale.

Together AI

Together AI, founded in 2022, provides cloud infrastructure for training and deploying open models. Its $800 million Series C in July 2026 brought total funding to $1.3 billion. It runs its own large-scale GPU capacity and chases the same enterprises, and it could also become a user of open engines such as SGLang.

Baseten

Baseten, founded in 2019, provides infrastructure for deploying and serving models in production and competes on latency, throughput and reliability. Its $1.5 billion Series F in June 2026 brought total funding to $2.1 billion. It already has a mature commercial platform and an established enterprise sales motion, which is the capability RadixArk has yet to build.

The clouds

Amazon Bedrock reached a multi-billion-dollar annualized run rate by the fourth quarter of 2025, with customer spend up 60% quarter over quarter. Google and Microsoft offer comparable services, and both appear on SGLang’s user list. These platforms plug into the security, identity and procurement systems enterprises already use, can host the open engine without paying a license fee, and sell the hosting themselves. They win on convenience and contracts already in place, where RadixArk’s case rests on specialization.

Chip vendors

NVIDIA and AMD both ship their own inference software, and both invested in RadixArk’s seed round. An engine that runs well on their hardware helps them, and either one could fold its techniques into their own software.

Business Model

RadixArk runs an unusually strict version of open core. The engine is free, owned by its community and has no proprietary version, so revenue has to come from managed hosting, enterprise support and tooling around reinforcement learning. As of September 2026 the company had disclosed no pricing, revenue or gross margin.

The revenue mix will shape how the business is valued. Managed hosting means reselling compute, and comparable businesses report gross margins near 50%. Fireworks AI’s was about 50% as of July 2025, with a management target of about 60%. Software and support typically earn 80% or more.

Among companies built on open infrastructure, the ones that grew large built a product clearly broader than the open project, and several later tightened their licenses to stop cloud providers from reselling their work. RadixArk’s governance rules out that second route entirely.

Traction

As of May 2026 RadixArk reported that SGLang ran on hundreds of thousands of GPUs and served trillions of tokens per day, with named users including Google, Microsoft, NVIDIA, Oracle, AMD, Nebius, LinkedIn, xAI, Thinking Machines Lab and humans&. Sheng has said hundreds of people have contributed since 2023. vLLM claims deployment of the same order, more than 400,000 GPUs, so neither project’s public numbers show which is larger.

Release cadence is the indicator to watch. SGLang has supported the Llama, Qwen, DeepSeek, Kimi, GLM, Gemma and Mistral families on the day of their release, across four classes of accelerator, several of whose vendors are investors. That cadence makes SGLang the default for a team standing up a new model, and it is expensive to keep up.

Commercial traction is not public. A January 2026 press report, citing a person familiar with the company, said RadixArk had begun charging for managed hosting, and as of September 2026 no revenue figures, customer counts or named paying customers had been disclosed. In May 2026 the company was hiring across research, engineering, product, sales and operations.

Valuation

In May 2026 RadixArk announced $100 million in seed funding at a $400 million post-money valuation. Accel led the round, first reported in January 2026, and Spark Capital co-led it. The seed proceeds equal a quarter of the post-money valuation.

The cap table’s most unusual feature is that NVIDIA, AMD and MediaTek invested alongside the chief executives of Intel and Broadcom, even though their companies compete with one another. The likely common interest is an inference engine that is open, well funded and outside the control of any one chip vendor.

For comparison, Inferact was valued at $800 million at its seed round in the same month, and Fireworks AI at $17.5 billion in July 2026 on more than $1 billion of annualized revenue. Acquirers have also started paying heavily for this layer. In August 2026 Stripe reportedly agreed to acquire OpenRouter, which routes inference requests across models and providers, for $7.5 billion, three months after a Series B valued it at $1.3 billion, according to Contrary Research.

Key Opportunities

Reinforcement learning as a second platform

Enterprises are early in adopting reinforcement learning, and no standard tooling has emerged. In its July 2026 funding announcement, Fireworks AI said more than 95% of the tokens it processes come from models specialized on customers’ proprietary data, which suggests heavy demand for customization. Reinforcement learning workloads are mostly inference, so RadixArk’s engine gives it a cost advantage there. If Miles becomes a default way to run reinforcement learning on open models, the company would own a product that is harder to commoditize than raw throughput. It would sell as a platform and a workflow instead of by the token, in a market where vLLM’s installed base counts for little.

A neutral runtime across hardware

Enterprises want to avoid depending on one cloud or one accelerator, because prices change and supply is uneven. An engine that performs well on NVIDIA, AMD, Intel and Google hardware lets customers move workloads as conditions change, and an investor base spanning rival chip makers supports that pitch. Contrary Research has observed that investment from both NVIDIA and AMD may support positioning as a neutral runtime, while noting that it does not prove demand will go that way. The opportunity depends on comparable performance and reliability across hardware types, an engineering burden that grows with every new chip.

Agentic workloads favor prefix reuse

Agents and multi-turn applications resend large amounts of shared context with every call, and RadixAttention was designed for that pattern before agents became a commercial category. As agents take a larger share of inference, an engine that reuses prefixes well pulls further ahead of one that does not.

Key Risks

Monetizing software it does not own

RadixArk’s best-known users pay nothing for SGLang, and the company has given up the tools other open-source companies used to make people pay. No proprietary version runs faster than the open one, and because the project belongs to the LMSYS community the license cannot change. HOF Capital describes the plan as open-sourcing the core engine and then selling the operational complexity of running it at production scale. That pattern produced Databricks. It also produced widely used projects with modest businesses attached, and the closest comparison is Anyscale, which commercialized Ray from the same Berkeley lab and agreed in July 2026 to be acquired by Nscale for $1.65 billion. Public information cannot yet say which outcome RadixArk is headed for.

Competition for the default engine

Open infrastructure markets have usually settled on a default and an alternative, with most of the value going to the default. Inferact has 50% more capital, investors with deep enterprise networks and a claimed installed base of the same order. Sheng has argued that inference is an incremental market where everyone grows together, and the market’s growth rate backs her up for now. Over time, though, enterprise buyers standardize, and the engine with the broader ecosystem of integrations, documentation and trained engineers usually wins.

Value capture by clouds and chip vendors

Cloud providers can host SGLang themselves and bundle it with services their customers already buy, and chip vendors can fold its techniques into their own software. Both groups benefit from the engine staying free, and several are investors, which gives them a stake in the project’s health and none in the company’s revenue. Contrary Research lists bundled cloud platforms winning by default as a major risk even for Fireworks AI, which has more than $1 billion in annualized revenue. A seed-stage company with no disclosed revenue is more exposed still.

Talent retention and commercial execution

The core asset is a small group of systems engineers whom frontier labs recruit heavily. The founding team is mostly researchers, and the company has announced no experienced leaders in enterprise sales or product. A large share of SGLang’s contributor community, and of the open models it supports, comes from China, which could complicate sales to the US government and regulated enterprises.

Summary

RadixArk is the company formed around SGLang, one of the two open-source inference engines that serve most open-model workloads. Inference spending passed training spending in 2026, agentic workloads are its fastest-growing segment, and RadixAttention was built for exactly those workloads. The cap table adds its own evidence. NVIDIA, AMD, MediaTek and the chief executives of Intel and Broadcom rarely agree on anything, and their joint investment suggests they agree that no single one of them should control the layer above the chip.

Little of that turns into revenue by itself. Google, Microsoft, NVIDIA and xAI run SGLang for free, the license cannot change to stop them, and a faster private version would abandon the openness that made the project the default. The paying customer is a different company entirely, an enterprise that wants to own its models without owning a kernel team, and reaching it takes the kind of enterprise sales operation that Fireworks, Together and Baseten have already built with several times the capital.

Miles is the most interesting thing RadixArk owns. Reinforcement learning is an inference workload wearing different clothes, it sells as a platform instead of by the token, and vLLM’s installed base gives no advantage there. Whether RadixArk ends up like Databricks or like Anyscale probably depends more on Miles than on the engine everyone already uses.