Featherless.ai, a serverless inference platform based in 2023 by Eugene Cheah, Harrison Vanderbyl, and Wesley George, has secured $20 million in Sequence A funding co-led by AMD Ventures and Airbus Ventures.
Extra buyers embrace BMW i Ventures, Kickstart Ventures, Panache Ventures, and Wavemaker Ventures.
Featherless addresses a structural problem that’s usually ignored outdoors the developer group. Whereas Hugging Face hosts over 30,000 open-weight AI fashions, many tailor-made for particular languages, domains, and duties that flagship fashions from OpenAI and Anthropic don’t deal with nicely, accessing these fashions in manufacturing stays tough.
“Usually, the fashions obtainable from suppliers are solely the preferred ones. Accessing fashions skilled on extra area of interest areas may be very tough. Making these obtainable constantly on-line, at a worth the place you don’t must lease 1000’s of {dollars} of compute to have a dialog with a chatbot that may communicate your language — that’s the genesis of Featherless,” says George to Tech Funding Information.
Together AI, Replicate, and Baseten present API entry to open fashions, whereas Groq provides high-speed inference on customized silicon. Featherless contends that none are really impartial, as every has {hardware} preferences, mannequin partnerships, or ecosystem alignments that recreate the lock-in the platform seeks to keep away from.
The startup stands out with a hot-swapping method that masses fashions into GPU reminiscence on demand in underneath 5 seconds and releases them when idle. This allows a flat-rate pricing mannequin, providing fastened month-to-month capability as an alternative of per-token billing.
“Most inference suppliers have 50 to 100 fashions obtainable of their public cloud. We now have the whole catalogue of 30,000 fashions obtainable on-line. You’ll be able to’t run 30,000 fashions by dedicating $2,000 of {hardware} to every one. That’s what our opponents do. That’s the differentiation,” explains George.
Enterprises acquire price predictability not obtainable from main cloud suppliers. For area of interest and bespoke fashions, this makes deployment economically viable for the primary time.
The investor lineup displays the corporate’s global-first orientation. AMD’s involvement is strategically particular: Featherless ensures the preferred open fashions run natively on AMD’s ROCm platform, offering a aggressive different to Nvidia’s ecosystem.
“AMD is aware of we will do nice issues with their {hardware}. They’re very dedicated to open supply. There’s a really pure match,” says George. Airbus, which backed the corporate on the seed stage, is targeted on deploying open-weight fashions in enterprise settings.
That sovereignty angle is more and more the corporate’s greatest progress alternative.
“A 12 months in the past, there was nonetheless a query of whether or not open fashions could be clever sufficient to do productive work. Right now, that’s now not the case. The main target is now shifting to who controls the AI, particularly in markets outdoors the US, the place there’s a huge push to manage your fashions, your infrastructure, and the liberty to take no matter you’ve constructed wherever you need,” George concludes.
The funding will assist infrastructure growth into new areas, improvement of a market for specialised open fashions, and enhanced integration with {hardware} architectures past Nvidia’s platform.
