Inferact

Scroll sideways for more →

Inferact is an artificial intelligence infrastructure company founded by creators and core maintainers of vLLM, the open-source inference engine that emerged from the University of California, Berkeley in 2023. The company came out of stealth on January 22, 2026 with a $150 million seed round at an $800 million valuation, led by Andreessen Horowitz and Lightspeed Venture Partners.^[1][2][3] Inferact states that its mission is "to grow vLLM as the world's AI inference engine and accelerate AI progress by making inference cheaper and faster"; alongside stewardship of the open-source project, it plans a commercial product its backers describe as a "universal inference layer".^[1][2][14]

Background

vLLM began in 2023 as a research project at UC Berkeley, in the Sky Computing Lab directed by Ion Stoica, the computer science professor who also co-founded Databricks.^[3][4] The project's central contribution was PagedAttention, a block-based memory manager for the KV cache retained during large language model generation, published at SOSP 2023 with Woosuk Kwon as first author.^[11] The team announced the engine publicly on June 20, 2023.^[15] It became one of the most widely used pieces of open-source AI inference software: UC Berkeley contributed it to the Linux Foundation in July 2024, and the PyTorch Foundation announced vLLM as a hosted project under its governance in May 2025.^[12]

By the time of Inferact's launch, the vLLM contributor pool exceeded 2,000 developers. Andreessen Horowitz, which said it had supported the project since 2023, wrote in its investment announcement that vLLM was "running on 400k+ GPUs concurrently around the world (that we know of)", was maintained by a dedicated team of more than 50 core developers, and was "used in production by companies like Meta, Google, Character.ai, and many others".^[2] These figures are claims by the project's maintainers and investors rather than independently audited measurements.

Launch and funding

Inferact announced itself on January 22, 2026, and Andreessen Horowitz said the same day that it was leading a $150 million seed round for the company.^[6] TechCrunch reported that the round was co-led by Andreessen Horowitz and Lightspeed Venture Partners and valued the company at $800 million.^[3] The company's own launch page lists the round as "led by a16z and Lightspeed" with a broader syndicate.^[1]

Scroll sideways for more →

Both lead investors framed the deal as a bet on inference demand. Andreessen Horowitz's announcement, written by Matt Bornstein, Jason Cui, and Raghu Raghuram, argued that the AI industry was moving from a phase bottlenecked by training to one bottlenecked by inference, and that serving many models across diverse hardware is an "{m*n}" problem that inference engines exist to solve.^[2]

Founding team

Inferact's launch statement is signed by "Simon Mo, Woosuk Kwon, Kaichao You, Roger Wang, Joseph Gonzalez, Ion Stoica, and the rest of founding members of Inferact".^[1] Andreessen Horowitz separately described the startup as "led by the maintainers of the vLLM project, including Simon Mo, Woosuk Kwon, Kaichao You, and Roger Wang", and its welcome message also named Zhuohan, a reference consistent with vLLM co-creator Zhuohan Li, though Inferact's own page does not list him among the named founding members.^[2]

Scroll sideways for more →

Strategy and products

Andreessen Horowitz described Inferact's goal as twofold. First, supporting the open-source vLLM project with dedicated financial and developer resources, which the firm called "explicitly the main goal of the company for the foreseeable future". Second, building what the team sees as a next-generation commercial inference engine, a "universal inference layer" intended to work with existing inference providers rather than compete against them.^[2]

At launch the company had not announced a shipping commercial product. SiliconANGLE reported that Inferact's launch blog post hints at a paid serverless version of vLLM, and that a job posting indicated plans for observability, troubleshooting, and disaster recovery features running on Kubernetes.^[4] Inferact's own statement sketches the direction rather than a product: "Today, deploying a frontier model at scale requires a dedicated infrastructure team. Tomorrow, it should be as simple as spinning up a serverless database."^[1]

On the company's relationship to the open-source project, Mo said at launch: "I fundamentally believe that open source, especially how vLLM itself is structured, is critical to the AI infrastructure in the world. And what we want to do with Inferact is to support, maintain, steward and push forward the open source ecosystem."^[5] The launch page makes the same commitment ("vLLM was built in the open. That's not changing.") and says optimizations the company develops will flow back to the community.^[1] vLLM itself remains an Apache-licensed project hosted by the PyTorch Foundation.^[12][13]

Scale claims

Inferact and its investors have published several figures about vLLM's footprint, all of which should be read as maintainer or investor statements:

• Inferact's site says vLLM supports more than 500 model architectures, runs on more than 200 accelerator types, and was built with more than 2,000 contributors.^[1]
• Andreessen Horowitz's January 2026 announcement put concurrent usage at "400k+ GPUs" worldwide, with the caveat "that we know of".^[2]
• In August 2026, a16z's promotional post for a podcast episode with Mo stated that "vLLM runs on half a million GPUs at any given moment", and vLLM's own account described the discussion as covering "half-a-million-GPU scale".^[9][10]
• TechCrunch reported that Mo told Bloomberg that existing vLLM users include Amazon's cloud service.^[3]

Inferact's site also says that model vendors work with the vLLM team to ensure day-zero support for new architectures and that hardware vendors integrate their silicon with vLLM directly,^[1] a practice a16z echoed by noting that top open-source AI labs and hardware companies "contribute to vLLM directly to ensure compatibility on day 1".^[2]

Competitive landscape

Inferact's debut was part of a broader wave of commercialization around open-source inference engines. TechCrunch noted that it mirrored the transformation of the SGLang project into the startup RadixArk, which sources told the publication had raised capital at a $400 million valuation in a round led by Accel; both vLLM and SGLang were incubated in 2023 in Stoica's Berkeley lab.^[3] Commercial and open alternatives in the serving layer include NVIDIA's TensorRT-LLM and the managed inference services offered by cloud and GPU providers, many of which, according to a16z, use vLLM under the hood.^[2]

Public commentary

Inferact's founders have used Andreessen Horowitz's media channels for their most extensive public statements. On the day of the launch, Mo and Kwon appeared on The a16z Show episode "Inferact: Building the Infrastructure That Runs Modern AI" with a16z partner Matt Bornstein.^[7] On August 6, 2026, the show released a second episode, "The Engine Powering Open-Source AI", in which Mo spoke with Bornstein and a16z's Elena Burger about how open-source AI became critical infrastructure; the episode description lists model licensing, the economics of open-weight AI, Kimi K3, distillation, and Mo's view that the gap between open and closed models is rapidly disappearing.^[8][9]

See also

• vLLM
• SGLang
• AI inference
• PagedAttention
• Andreessen Horowitz
• Ion Stoica

References

1. ^Inferact. "Announcing Inferact." Company launch page, January 2026. inferact.ai
2. ^Matt Bornstein, Jason Cui, and Raghu Raghuram. "Investing in Inferact." Andreessen Horowitz, January 22, 2026. a16z.com/...investing-in-inferact
3. ^Marina Temkin. "Inference startup Inferact lands $150M to commercialize vLLM." TechCrunch, January 22, 2026. techcrunch.com/...lands-150m-to-commercialize-vllm
4. ^Maria Deutscher. "Inferact launches with $150M in funding to commercialize vLLM." SiliconANGLE, January 22, 2026. siliconangle.com/...50m-funding-commercialize-vllm
5. ^a16z (@a16z). Post quoting Inferact cofounder Simon Mo. X, January 22, 2026. x.com/...2014402750331945275
6. ^a16z (@a16z). "We're excited to announce that we're leading a $150M seed round for Inferact." X, January 22, 2026. x.com/...2014387286788698122
7. ^The a16z Show. "Inferact: Building the Infrastructure That Runs Modern AI." Episode 1021, January 22, 2026. podcasts.apple.com/...id842818711
8. ^The a16z Show. "The Engine Powering Open-Source AI." August 6, 2026. a16z.simplecast.com/...cal-infrastructure-3yYJbsEz
9. ^a16z (@a16z). "vLLM runs on half a million GPUs at any given moment." X, August 6, 2026. x.com/...2085429677007736872
10. ^vLLM (@vllm_project). Post announcing the a16z podcast episode with Simon Mo. X, August 6, 2026. x.com/...2085439406069141962
11. ^Woosuk Kwon et al. "Efficient Memory Management for Large Language Model Serving with PagedAttention." SOSP 2023. arxiv.org/...2309.06180
12. ^PyTorch Foundation. "PyTorch Foundation Welcomes vLLM as a Hosted Project." May 7, 2025. pytorch.org/...pytorch-foundation-welcomes-vllm
13. ^vLLM Project. vLLM repository. GitHub. github.com/...vllm
14. ^Inferact. Inferact GitHub organization. github.com/Inferact
15. ^vLLM Project. "vLLM: Easy, Fast, and Cheap LLM Serving with PagedAttention." June 20, 2023. vllm.ai/...2023-06-20-vllm