Featherless AI is betting that open models win, and that AMD is how they get there

In the winter of 2023, a small team of open-source researchers in Singapore, Toronto and San Francisco ran an experiment. They had built an inference platform for RWKV, the attention-free language model architecture they had created, and the business was called Recursal. The experiment was a differently priced product under a different name, spun up to test whether flat-rate access to open models would sell. Within days it was generating more revenue than the parent platform.

That experiment was Featherless and 2.5 years later, it serves a catalogue that has grown past 40,000 open models across language, vision, audio and multimodal use cases, and in April 2026 it closed a $20 million Series A co-led by AMD Ventures and Airbus Ventures, with participation from BMW i Ventures, Kickstart Ventures, Panache Ventures and Wavemaker Ventures. The investor list reads less like a venture syndicate and more like a statement of intent: a chipmaker, an aerospace group and a carmaker, each with a structural interest in avoiding dependency on a single American AI vendor.

The thesis Featherless sells is straightforward. “Most companies run their AI on a handful of closed APIs, so cost and control sit with the vendor, not them,” said Joe Fleming, Chief Revenue Officer. “Bills scale unpredictably, models change or get deprecated underneath you, and access itself isn’t guaranteed.”

The company that emerged from an architecture argument, not a market gap

Featherless was founded in 2023 by Eugene Cheah, Harrison Vanderbyl and Wesley George, the core team behind RWKV. The architecture rejected the transformer paradigm in favour of a recurrent design, producing linear-time inference, constant memory consumption without a key-value cache, and theoretically unlimited context length. In September 2023, it became the first foundation-model project accepted by the Linux Foundation under its Generative AI Commons umbrella.

The commercial platform outgrew the research identity, and in February 2025 the company formally retired the Recursal.AI name in favour of Featherless.AI, by which point it had become the largest provider of Hugging Face language models. A $5 million seed round followed in March 2025, backed by Airbus Ventures, 500 Global, Kickstart Ventures, HF0, Oakseed Ventures and Panache Ventures. Research work continues under Recursal Labs, which now operates alongside the commercial business rather than above it.

That lineage explains an unusual quality in the company’s engineering. Most inference providers optimised for a small number of high-demand models. Featherless started from the opposite constraint.

Serving 40,000 models is a categorically different engineering problem to serving 10

“Each has different requirements, so hand-tuning doesn’t scale,” Fleming said of the catalogue. The response was an optimisation stack combining smart scheduling, caching, hot-loading, speculative decoding and custom kernels, designed to serve any model quickly and cheaply without human intervention. Underneath sits a GPU orchestration layer that swaps models into active memory on demand rather than keeping dedicated capacity warm for each one.

The economics of that design invert the usual infrastructure curve. Utilisation improves as the subscriber base grows, because more models stay resident in cache tiers and cold loads become rarer. Revenue therefore does not scale linearly with compute consumption in the way it does for a conventional GPU cloud.

The commercial model reflects the breadth. Individual chat plans run at $25 a month for unlimited tokens, developer API keys at $50 on a pay-per-use basis, and dedicated enterprise instances from $7,500 a month with SLAs and MLOps support. An on-premises licensed inference engine sits at the top of the range for organisations that cannot send workloads off their own infrastructure.

AMD is the supply advantage the market has been slow to price

The second engineering problem was hardware, and it is the one that attracted the company’s lead investor. “AMD GPUs are more available and cheaper, but the market defaults to NVIDIA because AMD’s software stack is hard to unlock,” Fleming said. “We did the work to run major open models natively on AMD and ROCm alongside NVIDIA, giving customers a real second source and better economics.”

The size of that gap is documented. NVIDIA still controls roughly three-quarters of AI GPU revenue in 2026, down from a peak near 87% in 2024, while AMD’s Instinct line generated an estimated $7 billion to $8 billion in 2025, or 5% to 7% of the market. ROCm has closed considerable ground, with version 7.0 delivering up to 3.5 times the inference performance of version 6.0, PyTorch listing it as a first-class option, and seven of the 10 largest AI model builders now running production workloads on AMD Instinct GPUs. On memory-bound inference, the hardware case is stronger still, since the MI355X carries 288GB of HBM3e against 180GB on NVIDIA’s B200, which lets a 70 billion-parameter model in FP16 sit on a single accelerator.

Featherless has taken the AMD position further than most. The original QRWKV6 32B conversion was carried out on 16 AMD MI300X GPUs donated by TensorWave, and the company now tunes dedicated enterprise instances for MI325X hardware.

The research arm functions as a hedge against the transformer consensus

Recursal Labs gives Featherless something its better-funded competitors do not have, which is a credible claim to changing the cost structure of a model rather than the cost structure of running one. “We believe the future isn’t transformers,” Fleming said. “Hybrid and linear architectures are already winning, and through Recursal Labs we’ve done the work to cheaply convert existing models and experiment with new ones.”

The published research supports the economics. The RADLADS protocol converts softmax attention transformers into linear attention decoders using only 350 to 700 million tokens, less than 0.005% of the token count used to train the original teacher models. Converting a 72 billion-parameter model costs under $2,000 at current prices, with inference quality remaining close to the original transformer. Featherless completed that 72 Billion-parameter conversion on eight GPUs and is now scoping client pilots to run the same process on customer models.

The client evidence Fleming cites is where the two threads meet. A major telecommunications operator cut a sustained agent workload from approximately $150,000 a month to $7,600, a 95% reduction, by moving to a dedicated GLM 5.2 instance tuned for AMD MI325X hardware. Drop-in Qwen VL replacements at 30 billion and 205 billion parameters cut costs by around 50% versus closed alternatives for computer-use agents and image verification, while Chatterbox replaced ElevenLabs voice at roughly a tenth of the cost inside the enterprise voice suite of a large European IT solutions company. With BMW, the company works as an open-source AI consultant, assessing model maturity, putting cost observability in place for finance and engineering teams, and shaping long-term open-model strategy.

The market has moved to open weights faster than enterprise procurement has

The macro case for Featherless has strengthened considerably in the past year. On OpenRouter, open-weight models grew from a negligible share in late 2024 to 33% by May 2025 and crossed 50% by mid-2026, with the five highest-volume models on the platform all open weights. Vercel’s AI Gateway Production Index recorded open-weight models at 29% of all tokens in June 2026, up from 11% in April. Research and Markets values the open-source AI model market at $23.08 billion in 2026, growing to $50.03 billion by 2030.

Capital has followed this market size - Fireworks AI raised $1.505 billion at a $17.5 billion valuation in July 2026, up from $4 billion seven months earlier; Baseten closed a $1.5 billion Series F at $13 billion in June 2026, and Together AI raised $800 million at an $8.3 billion valuation on 1 July 2026 with annual bookings above $1.15 billion. Cloudflare absorbed Replicate, NVIDIA licensed Groq’s LPU technology for approximately $20 Billion, and DeepInfra raised a $107 Million Series B.

Against that backdrop, a $20 Million Series A is a modest cheque, and Featherless is competing on a different axis. Its rivals concentrate on depth, with fine-tuning sophistication at Fireworks, serving reliability at Baseten and price competitiveness at Together. Featherless competes on surface area, hardware neutrality and architectural research, a combination that suits customers whose primary anxiety is dependency rather than latency. The company describes itself as unaligned with any hyperscaler or proprietary ecosystem, with infrastructure across the EU and the United States and teams in San Francisco, Toronto, Singapore and Europe.

The constraint Fleming identifies is neither capital nor engineering. It is that most buyers do not yet know the alternatives exist. “The remaining challenge is awareness,” he said. “Most customers don’t yet know these innovations exist, so a big part of our job is showing and teaching them the real MLOps and cost benefits.” For a company whose entire proposition rests on customers understanding what they currently pay for, that may prove the harder problem to solve.