In 2025, Mark McQuade took a gamble wholly specific to the AI era.
Arcee AI, the startup he founded in 2023, was focusing on post-training—the process by which an existing model is honed for humans. But McQuade saw an opening: Meta had just backed off its push into open-weight models, a vital (and tricky) middle ground in AI. Open-weight models allow companies (and people) to securely run state-of-the-art AI without turning their data over to, say, OpenAI or Anthropic.
But if McQuade wanted to fill the gap, he and Arcee would have to build their own model completely from scratch, a technical and financial mountain to climb.
“We saw an opportunity, and we had $30 million in the bank,” said McQuade. “I said ‘let’s do it’ and I bet the company on it. ‘Let’s spend 65% to 70% of our capital.’”
The startup over the coming months burned about $20 million training four open‑weight models, including a 400‑billion‑parameter model called Trinity Large, released in early 2026. $20 million, in AI, is a shockingly low number—conventional wisdom states that you need billions to train a new model, a belief challenged dramatically when China-based DeepSeek surfaced in 2025, with reports that the top-notch model was trained for under $6 million.
In AI, open-weight models are definitionally geopolitical, and China, so far, has dominated the game. Arcee—whose models have beat Meta’s Llama 3, and have benchmarked on par with Mistral and Chinese models—has been relatively quiet, but is now stepping forward: the startup has raised its Series B at a $1 billion pre-money valuation, Fortune has exclusively learned. The round was led by Vista Equity Partners, Cambium Capital, and Emergence Capital, with participation from Microsoft’s M12, AI10 Ventures, Hitachi, IAG, P7, and Wipro. Arcee declined to disclose the amount raised in this round, but a source familiar with the matter told Fortune it was at least $150 million.
The cash will funnel towards new open-weight models and products, along with growing Arcee’s partnership with the U.S. Department of Energy. (Arcee will also be working extensively with Vista’s portfolio companies.) However, McQuade is very clear: The ultimate goal is to catch China.
“Everyone talks about China versus the U.S.,” said McQuade (who was previously an early employee at Hugging Face, just acquired by Nvidia for almost $13 billion). “My stance is: just do something great. Catch up to them. Do the work… In order for the U.S. to take the absolute lead in the AI race, you have to be in the lead when it comes to closed and open. The U.S. is far ahead in closed source, but kind of dropped the ball on open source.”
McQuade says it’s time to pick up the ball: “We’re not chasing [Poolside’s] Laguna,” he says. “We’re chasing [Beijing-based Z.ai’s] GLM Flash.”
To get there, efficiency is non-negotiable, McQuade said. And, for Arcee, efficiency’s also the game plan: “We can comfortably say we’re the most efficient lab in the world based on what we’ve done.”
Building in AI is a high-wire act on its own, but the game at the model layer is that much more viscerally competitive, in the U.S. and abroad. McQuade didn’t just pivot, he dove into the deep end, and I asked him: Why cannonball, when you could’ve dipped a toe in?
“Gosh, adrenaline? I don’t know, I like to push the envelope,” he said. “In AI, things change so fast, so what’s the point of dipping your toe in. If you fail, you fail, and at least you tried. If you’re going to do something, go all in.”
See you tomorrow,
Allie Garfinkle X: @agarfinks Email: [email protected]
Submit a deal for the Term Sheet newsletter here.
Joey Abrams curated the deals section of today’s newsletter. Subscribe here.
The post Arcee AI trained four models for $20 million. Now, it’s worth $1 billion. appeared first on Fortune.
