In a falling market, AI capital concentrates in model builders

Anthropic raised $580 million and Hugging Face reached a $2 billion valuation as tech shares fell. Labs and model platforms now draw the largest cheques.
Anthropic closed a $580 million Series B on 29 April, little more than a year after its founding in early 2021. The AI safety and research company will use it to build large-scale experimental infrastructure for studying and improving the safety of computationally intensive models. On 9 May Hugging Face followed with a $100 million Series C at a $2 billion valuation.
Both rounds landed in the weeks when the Federal Reserve raised its policy rate by half a point, on 4 May, and listed technology shares kept falling. While the rest of the market reprices, capital is concentrating in two kinds of AI asset: the lab that trains large models and the platform where models are shared and put to work. Funding for AI is getting more selective, not smaller.
The lab is the expensive half. Compute, a scarce input with one dominant supplier, is its main cost, and a single round now runs to hundreds of millions of dollars, often before any product revenue. A lab needs repeated large rounds and long cloud commitments, so its outcomes sit close to binary.
Hugging Face shows the other half and a different risk. Its hub hosts 100,000 pre-trained models, more than 10,000 companies use it, and it leads BigScience, a research collaboration now training a large open multilingual language model. Investors value it like developer infrastructure, because every new model and every new user makes the hub more useful. A platform of this kind compounds through usage and can take its time to monetise.
Between those two poles the bar has risen. Large cheques go to teams with deep research credentials and a credible compute plan, or to products a large community already uses every day. Application companies have to show revenue and a gross margin that survives the cost of compute. Building on open models from a public hub keeps that cost down and preserves the choice of provider. Tools for evaluation, interpretability and safety now have paying buyers, since well-funded labs treat this work as part of the product.
Higher rates could still break the pattern. If other model labs cannot raise rounds above $100 million by the autumn, the window that opened this spring has narrowed, and the lab side of the split becomes harder to fund. Hugging Face has yet to show how it turns usage into revenue through paid hosting and enterprise plans. The open multilingual model BigScience is training will also show how close open work can come to the best closed models, and with it how much a hub is worth.
Training large models is turning into an industry with infrastructure-like capital needs, while the layers around it stay light and fast to grow. Fanaura, which invests its own and LP capital at any stage over a horizon of years, prices compute risk explicitly for that reason and favours founders whose product gets better as models get cheaper.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building evaluation, interpretability or safety tools for model labs? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.