Factory Tripled in Five Months. Arcee Says It Spent $20M.

On September 15, Factory raised $200 million at a $5 billion valuation for its enterprise "Droid" coding agents — up from $1.5 billion in April 2026, a 3.3x step-up in five months. Investors: Blackstone, Khosla Ventures, Sequoia Capital, NEA, with Marc Benioff as an angel. Total announced funding now exceeds $400 million.

On September 16, open-weight model developer Arcee AI raised a $150 million+ Series B at a valuation above $1 billion, led by Vista Equity Partners, Cambium Capital and Emergence Capital, with AI10 Ventures, Hitachi, IAG, M12, P7 and Wipro participating.

The two rounds are a day apart and argue opposite theses about where the money in coding agents ends up.

Factory's bet: the buyer is procurement, not the developer

CEO Matan Grinberg founded Factory in 2023 and plans to reach 300 employees by year-end. The named customer list is the pitch: Nvidia, Blackstone, Royal Bank of Canada, Palo Alto Networks, Adobe, T-Mobile. The company says hundreds of thousands of developers use the platform.

Published pricing runs $20/month for Pro, $100 for Plus, $200 for Max, and $60/month plus $40/seat for Teams up to ten seats, with Business and Enterprise quoted. The product includes Droid Computers — cloud or on-premises virtual environments for agents — plus AutoWiki auto-documentation and a readiness-report command that assesses repo compatibility.

Read the list again. Royal Bank of Canada, Palo Alto Networks, Blackstone. Those are not accounts you win through developer word-of-mouth and a credit card. They are procurement cycles with security review, on-prem requirements, and a compliance function that has to sign.

That is the tell. The on-prem Droid Computers offering exists because regulated enterprises will not send their source code to someone else's cloud, and building that is expensive and unglamorous and precisely the kind of moat that individual-seat competitors skip. Factory is not competing on model quality. It is competing on being deployable inside a bank.

A 3.3x step-up in five months on enterprise logos rather than seat growth says the coding-agent market has moved from IDE subscriptions to platform procurement — the motion that made Palantir and Databricks defensible, and the one that makes a $5 billion mark defensible on revenue quality rather than revenue size.

Arcee's bet: the capital moat is smaller than advertised

Arcee was founded in 2023 by CEO Mark McQuade, formerly of Hugging Face, with CTO Lucas Atkins. Its model family is Trinity Nano, Trinity Mini and Trinity Large — the last a 400-billion-parameter sparse mixture-of-experts with 13 billion active parameters per token.

The company says it built its entire 2025 model lineup, Trinity Large included, for roughly $20 million.

That number is the actual news, and it deserves scrutiny rather than acceptance. Twenty million dollars is a Series A. The conventional account of frontier training economics puts a 400B-class model an order of magnitude above that, and the gap is usually explained by pretraining compute. If Arcee's figure is accurate and refers to what it implies, the capital moat around closed frontier labs is considerably thinner than the financing rounds of the past eighteen months assume.

The caveats are real. Sparse MoE with 13B active parameters is dramatically cheaper to train and serve than a dense model of the same nominal size — comparing a 400B sparse model to a 400B dense one is not a comparison. "Entire 2025 lineup for ~$20M" is a company-stated figure with no disclosed accounting of what it includes. And frontier-class is a claim about benchmarks that third parties have not yet adjudicated.

Still: even discounted heavily, it is a datapoint in the direction everyone has been arguing about.

The strategic asset is quieter. Arcee works with the US Department of Energy and its 17 national laboratories on Genesis-Science-1 under the Genesis Open Models Initiative. That is a procurement channel for US-sovereign open weights that Chinese open-weight releases — which have dominated the open leaderboards for two years — structurally cannot access. Open weights with a federal customer is a different business than open weights with a Hugging Face download count.

One correction worth making: some aggregator sites reported a "$10 billion pre-money" valuation for Arcee. The company's own release and independent trade coverage put it above $1 billion. Treat the larger figure as erroneous.

The same week, the same thesis from two directions

Factory says value accrues to whoever can deploy agents inside a regulated enterprise. Arcee says value accrues to whoever can produce competitive weights cheaply enough that model access stops being scarce.

Both can be right, and if they are, the squeezed party is the middle: companies selling model access as the product, with neither an enterprise deployment moat nor a cost structure that survives open weights getting good. For context on where the week's capital actually went, the largest US round was neither of these — Temporal Technologies raised a $550 million Series E at $12.55 billion from Lightspeed, Wellington, Goldman Sachs Alternatives and Tiger Global. Durable execution infrastructure, not agents.

What to watch

Whether Arcee's $20M figure survives contact. An itemized breakdown, or a third-party benchmark placing Trinity Large where the company does, converts this from a claim into the most important number in open-weight economics.

Factory's net revenue retention. Enterprise logos are easy to name and hard to keep. Expansion inside RBC and Adobe is the metric that justifies $5 billion.

Whether DOE work becomes a template. If national-lab partnerships become the standard channel for US open weights, sovereign AI procurement is a market, not a slogan.

Consolidation. Two coding-agent rounds in two days at these marks means the shakeout starts within a year.