Featherless AI

$3M+ ARR From the Models Nobody Else Would Host

September 6, 2026

$3M+ ARR

Revenue

10,000+

Customers

2023

Founded

~30

Team

$25M raised

Funding

What Featherless AI Does

Featherless AI runs open-source AI models so you don't have to.

One API key, 40,000+ open models
Chat, coding, reasoning and vision
Any model in the catalogue loads in about five seconds
No GPUs to rent, size, or babysit

Think Vercel or Heroku, but for AI models. You pick the model. They own the hardware.

Who actually pays for it

Three kinds of customer, three ways to charge them:

Enthusiasts and prosumers — the $25/month Chat plan. Unlimited tokens for hands-on use, up to 32K context.
Developers shipping apps — $50/month of Developer credits. Bigger context, agent environments, unused credits roll over.
Companies in production — custom Business contracts. Dedicated H100, MI325 or B200 GPUs, fine-tuning, failover, annual volume pricing.

The cheap tier is the funnel. People arrive to try one model, then upgrade once they know what their product needs. The money concentrates at the top: the largest single customer pays $1–2M a year. Logos on the site include Ubisoft, Dropbox, Cisco, VMware and Meta.

“The average customer does not know what a B200 is. That's essentially what our job is. We abstract away all the complexity.”

The Problem They Solve

There are two markets in AI inference. Everyone fights over one of them.

Look at any provider. Most host fewer than 100 models — the famous ones. Eight to ten rivals compete for each.

Featherless measured what that actually covers:

Top ~100 models: 50% of their inference workload
Everything else: the other 50%

That other half is the fine-tunes: models tuned to give farming advice in Asia, tuned for one language, or trained on one company's own data.

Nobody hosted them. So nobody could run them.

And the pile keeps growing. Cheah's argument is blunt:

“In a world where various companies are fine-tuning their own model, you can't go to a provider who can only support 100 models. There are more than 100 companies on earth.”

The bottom 50% — how Featherless AI found a market nobody was fighting for

Watch: Eugene Cheah on the growth story behind Featherless AI

The Growth Story

2023: it started as research, not a business

Three people, three continents, one open-source project.

Eugene Cheah (CEO, Singapore), Harrison Vanderbyl (CTO, Australia), Wesley George (COO, Toronto)
Together they co-created RWKV — an attention-free AI architecture
RWKV became the first foundation-model project under the Linux Foundation
The company was called Recursal.AI and sold inference for its own models

They weren't building a hosting company. They needed to run thousands of open models for their own research, so they built the tool for themselves.

2024: a name test that took over the company

Featherless started as a pricing experiment. New name, new price, same team.

Then this happened:

“Within the first few days it became more profitable and more revenue than the original company platform. I guess we are Featherless now.”

The first users didn't come from ads. Community members posted it to AI forums on Reddit. That was the whole launch.

A Product Hunt launch followed on 24 June 2024: #6 Product of the Day, 279 upvotes — pitched on $10 and $25 monthly plans with unlimited usage.

Then came two years of firefighting.

“We get more users, the servers are on fire, we add more servers, we get more users, we add more servers.”

2025: Hugging Face turned coverage into distribution

The catalogue was the product. Hugging Face was how people found it.

February: Recursal.AI officially renamed Featherless.AI
March: $5M seed from Airbus Ventures, 500 Global, Kickstart, HF0, Panache and Oakseed — 4,000+ models live
June: 6,700+ models, making it the largest LLM inference provider on Hugging Face
Year end: around $400K ARR

Hugging Face is where developers go to pick an open model. Every model page carries a widget listing the providers who will run it — Featherless among them, alongside Groq, Together and Fireworks.

Two details turn that listing from a badge into a channel:

It runs itself. Any model that passes 100 downloads gets onboarded automatically. The catalogue grows without a salesperson touching it.
On the long tail, they are often the only name in the dropdown. On a famous model you are one of ten choices. On a fine-tune, you are the choice.

That is the flywheel. Hosting the unfought models is what makes Hugging Face's traffic convert — and the stated goal is to cover 100% of public Hugging Face models.

“Featherless AI is doing for inference what Hugging Face did for open-source model hosting.”

2026: $250K a month, then a $20M round

By April 2026 the scoreboard read:

10,000+ paying customers
More than $250K a month in revenue, less than $500K
Largest single customer paying $1–2M a year
27 staff — 12 on infrastructure — heading past 30

On 30 April 2026 they raised a $20M Series A, co-led by AMD Ventures and Airbus Ventures, with BMW i Ventures, Kickstart, Panache and Wavemaker. Total raised: about $25M.

That is roughly 7x ARR growth in twelve months. It also bought something the team hadn't had in two years: enough servers.

Featherless AI growth journey — from a 2024 soft launch to $3M+ ARR in April 2026

Key Growth Tactics

Three plays did most of the work.

Own the supply. Every fine-tune they add is a model with no competing supplier. One example: a 200-billion-parameter model that needs four H100s — roughly $20 an hour of hardware — shipping billions of tokens a day, with Featherless the only provider listed for it.
Let the channel do the selling. Being the most-listed provider on Hugging Face puts them in front of a developer at the exact moment that developer decides to try a model. That is a buying moment, not an impression.
Sell the savings, not the tech. For bigger accounts, sales opens with the customer's AI bill, then names the open model that does half those jobs for a fraction of the price.

Cheah backs that last one with a guarantee:

“If we don't help you save, we'll refund you your bill.”

What they have not done:

No paid acquisition. Growth has run on Reddit, Hugging Face and word of mouth.
No fight for the top-10 models, where eight to ten providers already slug it out.
No frontier model of their own. Not yet.

The Featherless flywheel — own the long-tail supply, ride Hugging Face distribution, and monetise with a three-tier price ladder

Key Takeaways for Builders

✓Serve the customers nobody else will. Featherless measured it: the top ~100 models are only half the workload. The other half had almost no suppliers — and almost no competition.
✓Go where your buyers already shop. Every Hugging Face model page lists the providers who will run it. Featherless got onto more of those pages than anyone, so it appears at the moment someone decides to try a model — no ad budget required.
✓Build a price ladder, not a price. A $25 hobbyist plan feeds $50 developer credits, which feed annual contracts worth $1–2M. The cheap tier is the funnel, not the business.
✓Ship the experiment under its own name. Featherless was a pricing test inside another startup. In days it out-earned the real product, so the founders followed the money and renamed the company.
✓Lead sales with the customer’s invoice, not your tech. "Here is your AI bill, here is the open model that does the same job for less" — backed by a refund guarantee — is what closes the large contracts.