We backed open-weight AI before consensus

In 2024, Panache Ventures joined the seed round of Featherless AI, a serverless inference platform for open-weight models. We were the only Canadian VC in that round, and betting on open-weight AI was a contrarian position to take at the time.

The largest US tech companies were pouring capital into closed models and proprietary infrastructure, and the working assumption was that AI would belong to whoever could buy the next generation chips. Featherless bet the other way. Its founders built a unique orchestration system that runs open-weight models on both the older-generation and new- generation GPUs. Today, Featherless runs infrastructure in the EU, North America, and APAC, serving tens of thousands of open-weight models as serverless endpoints, dedicated single-tenant capacity, or on-demand GPUs.

Conviction to back early

Panache had previously backed Wesley George, Co-founder & COO of Featherless.  A founder coming back to work with us a second time is the best endorsement a firm can get. Partnering with CEO Eugene Cheah was a no-brainer. Eugene was a core contributor to RWKV, an open-source model architecture built as an alternative to transformers, so Featherless started with real research depth in the open models it set out to serve. Deep trust in the Featherless team made underwriting their thesis an easier decision. 

In 2024, self-hosting a model could take days and cost more than $20,000 a month. Our belief was that open-weight models would keep getting better and cheaper, and that Featherless could be the infrastructure to easily host them. Prashant Matta, our Managing Partner, framed the stakes this way: "The risk of the AI industry ultimately becoming dominated by a few players from one country is very real. Solutions such as Featherless.ai provide a crucial alternative pathway for development."

The market caught up

In January 2025, three months after our investment, DeepSeek caused a deep freak in US markets. The Chinese AI upstart released its R1 model and challenged the idea that AI companies need billions of dollars in the newest chips to train a model. What stood out to us was how it got there: R1 was built on an open-weight approach and trained on older chips, despite US restrictions on chip exports. This was our signal that perhaps open-weight was the secret to global AI adoption.

In April 2026, Featherless closed a $20 million Series A co-led by AMD Ventures and Airbus Ventures, with participation from BMW i Ventures, Kickstart Ventures, Wavemaker Ventures and  Panache Ventures. Featherless is now the fastest-growing inference provider on Hugging Face and its catalog has grown from roughly 3,000 models in early 2025 to more than 40,000 today.

What we’ve learned

Open-weight models are improving quickly, and companies are moving production workloads to them. At a recent webinar we hosted with Featherless, the team walked through why customers are making the switch:

• They cost less at scale. Some engineering teams were spending around $1,000 per developer per month on closed models for coding, and a team of 20 sharing a dedicated open-weight node could bring that to roughly $300.

• The quality holds up. A hiring and classification platform replaced GPT-4o with GPT-OSS-20B and got comparable quality at half the cost.

• Companies keep control of their code and data. Open-weight models can run securely inside a company's private cloud and are cheaply finetuned on proprietary data.

• Teams aren't locked in. As better models come out, Featherless helps them switch without being tied to one provider's API or a specific model.

  
  

We wrote the seed check when open-weight AI was a contrarian bet. Today, it’s becoming the consensus. Panache Ventures backs the next generation of AI pioneers, investing in seed-stage, contrarian ideas that define the future.

Featherless AI is a Panache Ventures Fund II portfolio company. Learn more at featherless.ai.