Open-source AI projects are becoming companies en masse, but business and code are two different things
A new class of AI infrastructure companies is growing out of open-source systems research. Open-source success brings credibility and distribution, but it does not directly translate into customers and a business model. Using Inferact, founded by the vLLM team, as an example, take a look at how these companies turn technology into business, and what people online are saying.
What’s the next step after an open-source project gets big? A group of AI infrastructure companies is emerging from open-source systems research.
Sonnet Xu posted a list on X:
- vLLM → Inferact
- SGLang → RadixArk
- Orca / continuous batching → FriendliAI
- speculative decoding optimizations → Novita AI
- open kernels and systems research → Together AI
These names are familiar to people who know AI development, but ordinary people may have no idea what they are.
vLLM is one of the most popular LLM inference engines today, SGLang is another inference framework, and Orca is representative work in scheduling optimization. They determine the speed and cost of AI services.
The key observation from the list is: these companies monetize technical expertise in very different ways. Open-source success gives you credibility, distribution, and a first-mover advantage in technology, but customers and a business model are a different matter. When you do open source, you serve the community; when you run a company, you serve customers who are willing to pay.
Inferact is the most typical example. It is a company founded by the vLLM founding team, and its website says “Frontier inference you control.” You can deploy the platform on your own cloud or data center, or use the GPUs they provide. The selling points are controlling cost, latency, and throughput, as well as Day-0 support on the day a new model is released. Partners include AMD, MiniMax, Moonshot AI, and others, covering everything from chip makers to model labs.

Someone pointed out in the replies that production-grade vLLM is actually already in the upstream open source, so that is not their differentiation; the real differentiation lies in control, continuous optimization, and efficiency. Inferact is positioned to go beyond open-source vLLM by offering services such as embedding maintainers into teams, custom kernels, and training speculators.
This tweet sparked quite a bit of discussion. Some said that building open source successfully does not mean business will go smoothly, and customers and a business model have to be figured out separately. Some said maintainers can finally get paid, and it’s about time. Others said it’s interesting that vLLM has turned into a commercial company.
A user named Victor raised a more practical issue: the energy consumption of GPU inference. For the same cluster, costs are completely different between UK industrial electricity prices and the Gulf region. Profit margins follow electricity bills. No matter how strong the technology is, it cannot withstand the electricity bill.
When commercializing open-source projects, technical leadership is only a ticket to entry. Who pays, and why they pay, is the real question. The answers from Inferact and similar companies differ, but at least maintainers are starting to be able to make a living.
发布时间: 2026-11-21 04:50