A little-known American AI startup, Reflection AI, has released its first open-weight model, Beam: 501 billion parameters, trained from scratch, text-only, under an Apache 2.0 license that lets developers modify it and build businesses on it without special commercial restrictions. The company says it will publish Beam's weights and full technical details this month, distributed through hyperscalers and neoclouds with integrations across open-source libraries at launch. The pitch is "AI factories": enterprises and sovereign nations can keep training on their own proprietary data and run customized, local AI systems instead of sending sensitive data to an outside API. Reflection says Beam outscores Inkling on the four coding tests where both report results, and uses roughly a quarter of GLM-5.3's inference compute in its reasoning-efficiency comparison.
Reflection was founded in 2024 by two former Google DeepMind researchers: CEO Misha Laskin, who led reward modeling for Gemini, and CTO Ioannis Antonoglou, who helped create AlphaGo and AlphaZero and later led Gemini's reinforcement learning from human feedback effort. The startup has raised roughly $4.7 billion, per PitchBook, from backers including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, at a latest valuation of about $25 billion pre-money; it raised $2 billion last October at an $8 billion valuation in a round led by Nvidia. This summer it signed compute deals collectively worth more than $7 billion with SpaceX and Nebius, securing access to Nvidia's GB300 chips through 2029. Laskin likens closed models to "renting an apartment," arguing that "an ownership market" for AI is starting to form.
That is exactly why OpenAI and Anthropic should be nervous: both companies' business models now revolve around business customers, via APIs and enterprise subscriptions. OpenAI CFO Sarah Friar told investors at an August closed-door meeting that revenue had historically split about 60/40 in favor of individual customers, but that the ratio had reversed: "The majority of our revenue is now enterprise," she said, according to CNBC. In February, Anthropic said it had more than 500 business customers spending over $1 million a year on API usage or Claude subscriptions; a leaked IPO prospectus last month revealed that one-quarter of last year's revenue came from just two unnamed customers, almost certainly businesses. Enterprise contracts are the two giants' bread and butter.
Beam first threatens the turf of Chinese open-weight models: DeepSeek and Alibaba's Qwen have long dominated open-weight leaderboards on coding, math, reasoning, and price-performance, and are cheap and customizable. But most large U.S. companies are reluctant to hand sensitive customer data to developers with direct ties to the Chinese Communist Party — the gap Reflection wants to fill. Axios reports that hedge funds and trading firms are already lining up to build their own localized systems, and Reflection has begun testing a sovereign AI factory partnership with South Korea's Shinsegae Group. Nvidia CEO Jensen Huang has long championed the "AI factory" idea — systems that would run, naturally, on Nvidia GPUs.
Once Beam's weights land this month, the real test is commercial: if the model is free, how does the company make money? The answer is customization and deployment services. The South Korea pilot and the queue of hedge funds will be the first proof points. If enterprises truly shift en masse to building their own AI, the API-subscription model of the frontier labs will face its first serious diversion of demand from a U.S. rival.