A conversation with Mark McQuade, CEO, and Lucas Atkins, CTO of Arcee AI, with Emergence General Partner Joe Floyd.
There is no shortage of opinions right now on whether AI's future runs on open or closed models. What Emergence backs is choice: a market where builders and enterprises aren't locked into a single architecture or a handful of labs deciding what capabilities exist and who gets them.
Arcee AI, meanwhile, has made a much more specific bet than "choice matters." The company moved from post-training other people's open models to pre-training its own foundation models from scratch, a capital-intensive pivot that most people around them thought was a mistake.
We sat down with Mark and Lucas right around their Series B about why they made the bet anyway, and why the "mistake" is now leading an open revolution.
Joe Floyd: We just closed our Series B investment in Arcee, and I'm glad to be having this conversation now. Let's start with what's driving your commitment to open-weight models.
Mark: My background is early Hugging Face, so a lot of this traces back to that mission: democratizing AI, putting it in more people's hands, making it customizable in ways that closed models typically don't allow. It gives people ownership of their own AI as an asset, instead of renting it from a closed model lab.
Joe Floyd: A year ago Arcee made a capital-intensive pivot from customizing other people's models to pre-training your own. Most people thought that was a mistake. Why do it?
Mark: A little over a year ago, when we decided to move from customizing other people's open models to training our own from scratch, people thought it was a bad idea. There was an assumption that a couple of big players would just keep releasing competitive open models indefinitely, and relying on that single point of failure didn't sit right with us. A year later, we haven't talked to anyone at the enterprise, government, consumer, or venture level who thinks the bet we made was wrong.
Look at what OpenAI, Anthropic, and others actually ship. It's rarely just a UI on top of a general model. It's a custom variant, trained and tuned alongside a harness built specifically for that product. If the rest of the industry doesn't have access to that same capability, you end up with two or three labs holding disproportionate control.
— Lucas Atkins
Joe Floyd: You mentioned customizability and ownership. Is that the core advantage of open over closed, or is there more to it?
Lucas: Choice is the umbrella that motivates it, for us and for the broader ecosystem. If you're only relying on a company making the product decisions behind its API, deciding what capabilities matter and what they're optimizing for, every product built on that model eventually starts to look the same, because they're all running on the same backbone. Markets don't tolerate that for long.
Look at what OpenAI, Anthropic, and others actually ship. It's rarely just a UI on top of a general model. It's a custom variant, trained and tuned alongside a harness built specifically for that product. If the rest of the industry doesn't have access to that same capability, you end up with two or three labs holding disproportionate control, not just over profit, but over political, economic, and safety decisions. If AI is genuinely as transformational as people say, that kind of concentration in a small number of hands at a small number of companies is a real risk.
Joe Floyd: From the builder's side, what are the practical advantages of working in open weights?
Lucas: It lets us serve customers at a much lower price point than the larger labs can. Companies like Google, Anthropic, and OpenAI have to raise enormous amounts of capital, and most of it goes to compute for serving government, enterprise, and consumer traffic at massive scale. Very little of it goes to training. We get to put most of our compute toward training and toward the work we do directly with customers.
Joe Floyd: Arcee was one of a handful of signatories on the open-weight letter alongside Microsoft, Nvidia, and others. Why did that letter matter enough to put Arcee's name on it?
Mark: Signing alongside the giants of the industry, all vouching for open-weight support, mattered because it's our entire business. It made the point that the U.S. can catch up to and eventually surpass China in open weights.
Competitive Chinese models started arriving not months behind the West but ahead of it in places, and that sent a real shock through the ecosystem, especially in government. The fear wasn't just about Chinese models specifically. It was that the government's response might be to restrict open models generally. This letter was a statement, from massive enterprises, startups, and the labs themselves, that restricting open models isn't the answer. The answer is making the American and Western open ecosystem competitive. And the vibe shift since has been real.
Joe Floyd: Why does U.S. leadership in open weights matter specifically?
Mark: To lead in AI overall, you need to lead in both closed and open. The U.S. is ahead on closed, with Anthropic, OpenAI, Google, and xAI. On open, there's a real gap. China has taken a different approach: state-supported, mutually reinforcing labs pushing open weights out as a deliberate strategy, not a side project. The U.S. needs to treat open weights as something that can be genuinely lucrative, on the business side and on the adoption side, not just a hobby.
We're one of very few companies solely focused on building open-weight models in the U.S., and culturally we're different from most of the other labs. I call it blue-collar AI. There isn't much of that in this industry, which tends to run elitist. We think of ourselves as the every-person AI company.
— Mark McQuade
Joe Floyd: Shifting to Arcee specifically. What makes the company distinct in this space?
Mark: We're one of very few companies solely focused on building open-weight models in the U.S., and culturally we're different from most of the other labs. We're not staffed with a team of Anthropic or OpenAI alumni with the traditional pedigreed resumes. We're skilled, but we're scrappy, thirty people total, grounded in outcomes over credentials. I call it blue-collar AI. There isn't much of that in this industry, which tends to run elitist. We think of ourselves as the every-person AI company.
Lucas: What makes us able to capitalize on the open-weight wave is that we're hungry. A lot of the team comes from nontraditional backgrounds for this industry, mostly engineering rather than research. We treat model development as a product discipline: a well-defined series of steps you can execute in a structured way, not some dark art only a handful of people in the world understand. There's real complexity in the architectures and pipelines, but the process itself doesn't require having been on the original OpenAI team.
We also don't isolate researchers into narrow slices of the pipeline the way larger labs often do. We have about sixteen researchers who move across the entire lifecycle, from data curation and pre-training through mid-training, context extension, SFT, and RL. Everyone on the team understands the full process, which lets us move fast and develop each phase with more nuance than a much larger, more siloed organization typically can.
Joe Floyd: What are Arcee's ambitions post-Series B?
Mark: In the next eight months, I want Arcee mentioned alongside OpenAI, Anthropic, DeepSeek, and Moonshot as one of the companies driving the frontier. It's achievable. We're not tiptoeing anymore.
