Product1 publisher3 min readPublished
Reflection AI wants companies to build their own models instead of paying OpenAI and Anthropic
Reflection AI, backed by about $800 million from Nvidia, is building software that lets companies make their own AI models. The pitch aims at the enterprise spend frontier labs now depend on, though the leverage it creates sits mostly with those labs' largest customers.
The Product Desk · Product desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- Anthropic said in February that more than 500 business customers each spent over $1 million a year on its API or Claude subscriptions.
- A leaked Anthropic IPO prospectus showed that two unnamed customers supplied a quarter of the company's revenue last year.
- Axios reports Reflection is preparing an open-weight model, expected soon, with capabilities on par with the most advanced Chinese open-weight models.
- Reflection signed a multiyear deal in June to pay SpaceX $150 million a month for compute from the Colossus data center in Tennessee.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- exposure Anthropic's dependence on two accounts gives those buyers leverage at renewal that a typical product team buying from the same price list does not have.
- capability A competitive US-built open-weight model would give companies that ruled out Chinese models over data concerns a model they can modify without a China-linked developer.
- cost A company that moves to Reflection buys from a vendor committed to about $1.8 billion a year in compute, a cost Reflection's own enterprise pricing will eventually have to cover.
The person renewing an AI contract this quarter is working from pricing pages that Gizmodo describes as "a confusing jumble of pricing plans that can be difficult to track" [20]. Gizmodo also writes that the "token-maxxing" period, when companies pushed staff to use AI tools as much as possible, has given way to "sober frugality" [21]. Both OpenAI and Anthropic make most of their money selling to that buyer through APIs and enterprise subscriptions, according to Gizmodo [4]. OpenAI CFO Sarah Friar reportedly told investors in August that revenue, historically split about 60/40 in favor of individual customers, had reversed [5]. "The majority of our revenue is now enterprise," Friar said, according to CNBC [6].
Reflection AI, founded in 2024 by two former Google DeepMind researchers, calls its product an "AI factory": software that lets businesses build their own models instead of subscribing to closed systems such as ChatGPT [9]. The public record behind that pitch is thin. The Axios report on its first open-weight model was, in Gizmodo's words, "scant on specifics" [17]. The reporting does not include a price or a benchmark, and Reflection did not immediately respond to Gizmodo's request for comment [18].
The most concrete part of the pitch is what Misha Laskin, Reflection's cofounder and CEO, says buyers already do. "Today the best open models are coming out of China," Laskin told CNBC in July. He said that as a result, global AI adoption is being built on top of Chinese models, including by American companies [1]. Those models are cheaper than American frontier models and can be modified for each customer [3]. Most major US companies still hesitate to hand sensitive customer data to a developer with direct ties to the Chinese Communist Party, according to Gizmodo [2].
Bargaining power is harder to place. Enterprise spend at the frontier labs is concentrated. Anthropic's 500-plus customers above $1 million a year spend more than $500 million between them [13]. The two customers named only in aggregate in the leaked prospectus averaged about 12.5 percent of revenue each [14].
Reflection carries a large bill of its own. Its SpaceX compute deal works out to $1.8 billion a year [15], more than twice what Nvidia had invested as of March, according to the Wall Street Journal [10][16].
Gizmodo forecasts that open models will push OpenAI and Anthropic to spend less effort on new systems and more on making existing products affordable and accessible as both near stock market debuts [22]. It also says it is too soon to tell whether Reflection's model will be good enough to pull businesses off ChatGPT or Claude in large numbers [12].
The decision splits on two questions: whether a team's spend is large enough to register in a lab's revenue concentration, and whether its data can go to a hosted API.
- Large spend, hosted API acceptable: Chinese open-weight models already give the renewal a cheaper reference point [3], and a US-built model would add one without the data objection. - Large spend, data must stay in-house: Reflection's factory is built for this buyer, and the case rests on a model still being prepared for launch [17]. - Small spend, hosted API acceptable: the published price list is what this team pays, so it gains from open-model competition only through the affordability push Gizmodo forecasts [22]. - Small spend, data must stay in-house: the other Western open-weight models Axios said are due this month are the nearest new option [19].
What to watch
- Independent benchmark results for Reflection's first open-weight model against the leading Chinese open-weight models.
- Whether the other Western open-weight models Axios expects this month arrive, and from whom.
- Whether OpenAI or Anthropic cut or simplify enterprise pricing as they move toward stock market listings.