Sulat.com / Writing on AISaturday, November 21, 2026
Business & Industry/21 November 2026

Chart of five open-source inference companies draws pushback from Inferact

A chart from Sonnet Xu maps five inference companies to the open-source work behind them. An Inferact co-founder disputes his company's row, and the funding runs from US$800M to none announced.

Five glossy 3D plaques bearing the Inferact, RadixArk, FriendliAI, Novita and together.ai logos on one pedestal, fed by purple roots from a glowing code cube

At a glance

  • Sonnet Xu's 20 November chart pairs Inferact with vLLM, RadixArk with SGLang, FriendliAI with Orca, Novita with speculative decoding and Together AI with FlashAttention, and says each monetises that work differently.
  • Inferact co-founder Roger Wang replied that production-grade vLLM is already shipped upstream and isn't how Inferact differentiates. He named control, continuous optimisations and efficiency instead.
  • The roots aren't one pattern: two are engines turned into companies, one is a research paper, one is a method the company published itself, and one is a researcher who is now chief scientist.
  • The latest announced rounds run from Together AI's US$800M Series C to FriendliAI's US$20M seed extension. No round has been announced for Novita.

Sonnet Xu, whose bio reads “cs @ stanford”, posted a chart on 20 November that pairs five AI infrastructure companies with the open-source work behind each. Xu’s point is that the companies monetise that expertise very differently, and an Inferact co-founder replied within the hour to dispute how the chart describes his company’s offer.

Sonnet Xu on X, 20 November 2026: the chart and the thread that started the argument.

Roger Wang, one of Inferact’s founders, wrote that production-grade vLLM is what Inferact already ships upstream and not how it differentiates. He said the difference is control, continuous optimisations and efficiency, and that he wasn’t sure about the other four.

Five rows pairing Inferact with vLLM, RadixArk with SGLang, FriendliAI with Orca, Novita with DSpark and Together AI with FlashAttention, plus what each sells
Five open projects, five different things to sell.

What the chart claims

The chart has four rows: open-source roots, core technology, how each company monetises and what differentiates it. The table keeps the first and third in the chart’s own words. The framing text says open-source success buys credibility, distribution and a head start on the technology, but that the customers and business model look different.

CompanyOpen-source rootsHow they monetise
InferactvLLM, inference engineEnterprise inference software and support
RadixArkSGLang, inference engineManaged inference and RL infrastructure
FriendliAIOrca, continuous batchingHigh-performance inference platform and deployment
NovitaSpeculative decoding, inference optimizationOptimized inference and model serving (inc. speculative decoding)
Together AIOpen kernels and systems research (e.g. FlashAttention)Compiler, kernels, and full-stack inference platform
Source: Sonnet Xu on X, 20 November 2026.

Five roots, not one pattern

The chart sets the five roots side by side as if they were the same kind of thing. They aren’t.

Two are open-source engines that their maintainers turned into companies. Inferact is led by maintainers of vLLM, and RadixArk launched on 5 May around SGLang. TechCrunch reports that both engines were incubated in 2023 at the UC Berkeley lab of Ion Stoica, a Databricks co-founder.

FriendliAI’s root is a paper. Its site says CEO Byung-Gon Chun led the Seoul National University team behind continuous batching, introduced in the Orca paper at OSDI 2022, and that continuous batching has become the standard across modern inference engines. Orca is a research paper describing a serving system, so the root is an idea other engines adopted rather than an engine with its own community.

Novita runs the other way. Its blog describes DSpark as a speculative decoding method that Novita published, and says it released DSpark draft models for two Kimi models and integrated them with vLLM. Here the research comes from the company rather than leading to it.

Together AI’s link is a person. Its chief scientist is Tri Dao, who wrote FlashAttention, and Together introduced him alongside the release of FlashAttention-2. The chart’s row for Together covers a body of work, with FlashAttention as the example.

Five different cheques

The funding behind the five is just as uneven. Together AI raised US$800M on 1 July at a US$8.3B valuation, in a round Reuters says Aramco Ventures led. Together also said its new investors had secured commitments for more than 500 MW of compute capacity, to be financed separately from the equity.

US$800M

Together AI, Series C

US$150M

Inferact, seed

US$100M

RadixArk, seed

US$20M

FriendliAI, seed extension

The other three announced rounds are seeds. Inferact raised US$150M at a US$800M valuation on 22 January, co-led by Andreessen Horowitz and Lightspeed. RadixArk raised US$100M at a US$400M post-money valuation on 5 May, led by Accel with Spark Capital co-leading. FriendliAI announced a US$20M seed extension on 28 August 2025, led by Capstone Partners.

Bar chart of latest funding rounds: Together AI US$800M, Inferact US$150M, RadixArk US$100M, FriendliAI US$20M, Novita none announced
Together’s Series C is the only round of the five that isn’t a seed.

Novita is the gap. Tracxn lists no funding raised, and no announced round turned up in a web search. That doesn’t prove there isn’t one, only that it hasn’t been announced.

So the chart’s columns hide a bigger split. Together is a Series C company with a compute-backed round, and the other four are seed-stage or unfunded as far as announcements show.

Where Inferact disagrees

Roger Wang of Inferact on X, 21 November 2026 NZ time: his reply to the chart.

production grade vLLM is what we already ship in upstream and not how we differentiate. It’s more about control, continuous optimizations and efficiency

Roger Wang, Inferact, on X
Side by side: the chart says production-grade vLLM, enterprise features and support; Inferact says control, continuous optimisations and efficiency
The chart sells vLLM support; Inferact says it sells control and efficiency.

That matters most for the Inferact column, where the chart’s differentiation row reads as a pitch for supported vLLM. Inferact’s own site headline is “Frontier Inference You Control“, which lines up with Wang’s reply on control. No other company has replied in the thread.

What the chart can’t tell you

The chart is a sketch. It gives no revenue, customer or pricing figures, so it can’t show which of the five models earns more. Its open-source roots row mixes a codebase, a paper, a technique and a research area, and its differentiation row is Xu’s reading, which one of the five has already pushed back on.

If you’re choosing among these vendors, the question the chart raises is the right one: what do you pay for on top of the free software? Inferact’s answer, in Wang’s words, is control and efficiency rather than supported vLLM. Ask each of the others directly, because the chart’s answers for them are one observer’s, not theirs.

Terms in this post

Inference engine
Software that runs a trained model to answer requests. vLLM and SGLang are two open-source examples.
Continuous batching
Scheduling that adds and removes requests from a running batch as they arrive and finish, instead of waiting for a whole batch. The Orca paper introduced it.
Speculative decoding
A small draft model proposes several tokens and the main model checks them in one pass, which speeds up generation without changing the output.
FlashAttention
A faster, more memory-efficient way to compute the attention step in transformer models.
Seed extension
An additional round raised on top of an earlier seed round.

Sources


Discover more from Sulat AI & Tech News

Subscribe to get the latest posts sent to your email.