Arcee, Poolside, Reflection Court VCs Cool on Open-Weight AI

The pitch from Silicon Valley's small open-weight labs is easy to summarize: someone in the US has to build the model American enterprises and Western governments would rather use than a Chinese one. The pitch is not, according to The Wall Street Journal, landing with the venture funds that would normally back it.

Arcee AI, a roughly 30-person startup, bet its remaining cash on a 33-day training sprint and produced Trinity Large, a downloadable and customizable model reportedly built for about $20 million, a fraction of what the industry giants spend. CEO Mark McQuade told the paper that 'every tier-one VC pretty much said no.' Poolside co-CEO Jason Warner, whose company sits alongside Arcee and Reflection AI in the small cluster the WSJ profiles, argues the demand story is there anyway: 'There is a vast, vast degree of want for an American company producing the most-capable open-source artificial intelligence.'

The tension the reporting draws out is a business-model one. Investors have already committed billions to closed-source giants like OpenAI and Anthropic, and an open-weight model that any customer can download, fine-tune and run themselves is harder to underwrite as a subscription business. Michael Stewart of M12 tells the WSJ this is '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,' but even that framing is about eventual belief, not present-day check-writing.

The honest caveat is that the reporting, as retrieved, is thin on the numbers that would settle this either way. It does not detail current revenue at these startups, does not quantify how much Chinese-model usage has actually moved into US enterprise accounts, and does not lay out how Arcee, Poolside or Reflection expect to earn back the compute bill on models they give away. The Chinese systems being cited (Kimi, Qwen, DeepSeek) are named as the competitive pressure, but the specifics of that pressure are asserted more than shown.

What is worth watching is not the next benchmark leaderboard but whether one of these labs lands a signature US enterprise or federal customer that names an open-weight American model as its default. That is the moment the VC cost-of-capital problem changes shape, and the moment the 'root for the underdog' framing turns into a category the tier-one funds can no longer politely pass on.