In late 2025, a little-known Silicon Valley startup bet much of its remaining cash on a lofty goal: building the strongest open-weight AI model it could.
The startup, Arcee AI, pulled it off after a 33-day pretraining run, with far less funding than the industry’s biggest labs. The result is a cheap artificial-intelligence system that users can download and customize, an American counterpart to a host of upstart Chinese models that are shaking up the AI race.
The new batch of open-weight companies, which includes Arcee, Reflection AI and Poolside, are betting on an increase in demand for capable open models that offer the efficiency of China’s without the geopolitical complications.
“There is a vast, vast degree of want for an American company producing the most-capable open-source artificial intelligence,” said Jason Warner, co-chief executive and co-founder of Poolside, an AI company that in July released a new version of its family of open-weight models called Laguna S 2.1.
Open-weight models allow anyone to download the numerical values, called weights, for each of the billions of parameters stored in their machine minds. That raw information allows anyone to run a model on specialized hardware and to build on top of the model, altering its weights with additional data in a process called fine-tuning. In the industry, some also refer to “open-source” models, which provide a higher level of transparency, offering users full access to weights, training code and other data.
The U.S. was initially the leader in open-weight AI models, with much of the open-AI sector and tools having grown up around Meta Platforms’ Llama-series models. But China quickly started catching up.
Over the past year, as companies have watched in awe as the size of their AI bills multiplied, many have turned to China’s open-weight models because they are often much cheaper. The shift has brought renewed attention to open-weight AI and the widening gap between the U.S. and China. U.S. companies including OpenAI have recently reduced their prices to make some of their high-end models more competitive.
“It’s the beginning of an awakening that it can happen—that the default model that you use will be an open-source model in the future,” said Michael Stewart, a managing partner at M12, Microsoft’s venture fund.
Nvidia and its CEO, Jensen Huang, have emerged as being among the strongest supporters of open AI. In July, Nvidia signed an open letter urging greater support of open models and asking policymakers to avoid “premature restrictions.”
The leaders of Anthropic and OpenAI have responded with broad support for open models but added that there are trade-offs. There is the possibility of the models’ being misused for cyber or biological attacks, Anthropic Chief Executive Dario Amodei said in a recent blog post.
Nvidia has built several open models, including a family called Nemotron, and earlier this year started a coalition with several AI labs to contribute to open models. It has also become one of the largest backers of open-weight companies. It made big investments in Reflection AI, the artificial-intelligence startup that is expected to debut its first open model later this year after raising more than $2 billion; Poolside; and Thinking Machines Lab, which released its first open-weight model, called Inkling, in July.
Outside Nvidia and the startups it has backed, the U.S. open-weight ecosystem remains small, and China’s open models continue to outperform.
A lack of investment interest by the venture-capital establishment in open-weight AI businesses hasn’t helped. Many investors have questioned whether a company developing an open-weight model, which is free to use, can generate real revenue. The investors are also fearful of backing technology that could erode OpenAI’s and Anthropic’s businesses.
The closed, proprietary AI labs continue to dominate and have forged relationships with virtually every top investor in Silicon Valley. In the first quarter alone, AI startups raised $255.5 billion. Nearly two-thirds of that total came from just three deals: funding rounds for OpenAI, Anthropic and xAI, according to the data firm PitchBook.
“I don’t want this to succeed. I don’t want to invest because it will hurt my investment in Anthropic or OpenAI. Those are the excuses I heard repeatedly,” said Joe Floyd, an early investor in San Francisco-based Arcee and a general partner at Emergence Capital who has helped the startup raise money.
Arcee’s McQuade said he expected more interest from venture capitalists. Instead, he encountered skepticism and disbelief. Recently, there has been increased outspoken support on X for developing U.S. open-weight AI to compete with China.
“Probably every venture fund that you see writing on X, I spoke to,” he said. “I look at it this way: I have a chip on my shoulder now. We are in the streets training and releasing models. We are not just talking about it. We are trying to have the U.S. catch up to and surpass China.”
McQuade is a former employee at the open-source AI platform Hugging Face who previously rose from field technician to Amazon Web Services team lead at Unity Connected Solutions. He started Arcee in 2023 with Jacob Solawetz and Brian Benedict, not intending to build a foundation model. McQuade said he saw an opportunity after Meta seemed to shift its focus away from its Llama models.
Arcee started with a small, 4.5 billion parameter model before scaling up to Trinity Large, which it trained on 2,048 of Nvidia’s Blackwell B300 chips during a tightly planned effort designed to get the most out of its roughly $20 million budget. Arcee previously raised $50 million in total at a valuation of $240 million. It has about 30 people on staff.
“It was more to prove that we could do it with such limited capital and resources,” said McQuade. “We’re just a bunch of grinders—a bunch of misfits that are going after it and want to prove the world wrong.”
Trinity Large is still much smaller than the top AI models and trails models from Anthropic and OpenAI on many leading benchmarks. Arcee plans to build larger, more-capable models with capital from a new funding round expected to close soon.
Despite its fundraising hurdles, the team remains optimistic about the future of the U.S. open-weight ecosystem.
In July, the startup joined with the Energy Department to build an AI model focused on scientific research.
“The tailwinds are behind us now based on everything that is happening in the world,” said McQuade. “It’s just a matter of time before widespread adoption.”
Write to Kate Clark at kate.clark@wsj.com and Sam Schechner at Sam.Schechner@wsj.com