Fal, Fireworks, Baseten, Modal Valuations: The Chinese Open-Model Risk in the Inference Boom
Fal, Fireworks, Baseten and Modal are in talks at a combined valuation of close to $90 billion. Their revenue comes from serving open-weight models more cheaply than closed labs charge, and the best of those models are Chinese. The valuations do not price that political risk.
In a single week, four companies that run other people's AI models have gone looking for close to $90 billion of combined valuation. Fireworks AI is considering a round at up to $30 billion. Baseten is talking at $26 billion, and Fal and Modal at about $15 billion each. Investors are paying software multiples for the fastest-growing layer of the AI stack. They are underpricing a political input: a large share of what these companies sell is Chinese open-weight models served to American companies, and Washington has started to restrict that trade.
The Information reports that Fal's revenue is running at $800 million a year and that it wants a valuation above $15 billion. Its last priced round, led by Sequoia in December, valued it at $4.5 billion. Fireworks raised at $17.5 billion in July, with more than $1 billion of annualised revenue, and now has a $30 billion target. Bloomberg reported that Baseten's valuation could double from the $13 billion it reached in June, and that Modal's could roughly triple from the $4.65 billion it was valued at in May. At $15 billion, Fal would be valued at about 19 times its run rate. At $30 billion, Fireworks would be valued at close to 30 times the revenue it disclosed in July.
What these companies sell is a price gap. A business that wants frontier-level results without frontier-lab prices rents an open-weight model from one of them. The provider earns the difference between what it pays for GPUs and what it charges per token. Fireworks' own price list shows which models customers want. Most of the model families at the top of the list are Chinese: Moonshot's Kimi, DeepSeek, Zhipu's GLM, Alibaba's Qwen and MiniMax. OpenAI's GPT OSS and Nvidia's Nemotron are the American exceptions. Fal, which works on images and video, is in a similar position. Its catalogue depends heavily on Kling, ByteDance's Seedance and MiniMax, along with Google's Veo.
The reason is price for performance. Larridin, which measures enterprise AI use, told Fortune that Chinese models such as GLM 5.2 and Kimi handle about 75% of the engineering tasks it tracks reasonably well, at a fifth of the cost of American frontier models. DoorDash has said Kimi is cheaper and better. Cursor used Kimi to help build a coding agent. Kimi K3 costs $3 per million input tokens on Fireworks, and DeepSeek's V4 Flash costs $0.22. Those prices are low enough to reprice a coding agent or a customer-service bot. They also leave room for a margin that a startup can call a moat.
That margin is where the political exposure sits, and it is building up. This month the FBI, the NSA and CISA said DeepSeek, Moonshot and four other Chinese labs had extracted "capabilities worth billions" by training on the output of American models. Representatives Josh Gottheimer and Nick LaLota introduced the China FIREWALL Act. It would ban Chinese open-weight models from government devices and stop federal agencies buying software that relies on them. Beijing is adding its own pressure: its internet regulator is examining DeepSeek and Moonshot over data security. Fireworks has already priced in part of the problem. Since September 1, models served only from US infrastructure cost 1.5 times the standard rate. That premium shows buyers already care where a model runs. They will soon care who built it.
The platforms say they do not depend on any one model. If Chinese weights become hard to sell, they can serve American open models, and Nvidia is putting money into Nemotron partly to fill that gap. That is true about how the business works, but it does not solve the profit problem. The gap they sell exists because the best cheap models are Chinese. Replace Kimi and GLM with weaker American models and the difference between the platform's price and a closed lab's price narrows. At that point the customer's alternative is Anthropic or OpenAI through AWS Bedrock or Azure, and those clouds can sell that service at a loss. A model-agnostic platform is still exposed to the best model it is able to serve.
Rising costs make this worse. Fal lists H100 time at $1.90 an hour for committed customers, against $4.50 at list price. Fireworks charges $8 an hour for an on-demand H100. The distance between those numbers is the business, and the company with the most capital to secure GPUs keeps more of it. That is why these rounds are this large. Fal has raised $337.5 million in total, less than a quarter of what Fireworks or Baseten each raised in a single round this year. Most of the new money will pay for chips, and chips do not become a software asset.
The companies can still be good businesses. Demand is real, revenue grows several times over each year, and the shift of computing from training to inference favours whoever runs models most efficiently. The point is narrower. A valuation of 20 to 30 times revenue assumes gross margins like software and a customer base that is safe from regulation. A business that resells mainly Chinese models on rented GPUs has neither.
Investors in these rounds should ask for one number the companies do not yet publish: the share of tokens served from Chinese-developed models. We expect all four rounds to close this quarter at or near the reported targets, because demand from GPU-constrained customers is strong enough to carry them. The repricing will come from procurement. Federal contractors should start requiring certification of model origin in 2027. When that happens, the companies that have moved their token mix toward models they can defend in Washington will keep their multiples, and the others will be valued as neoclouds.