Nomic's CEO alleges startups resell frontier traces generated on lab credits

A trace is a record of a model working a task: the prompt, the tool calls, the step-by-step reasoning and the grade someone put on the answer ^[10]. Distillation runs on exactly that, taking a stronger model's outputs to train or improve another one ^[10]. Companies that build reinforcement-learning environments hold the largest supply, because they write the tasks, the simulations and the scoring systems ^[11]. Epoch AI's survey of that market described those environments as a central input for frontier-model training, spanning coding, computer use and enterprise workflows ^[12]. The prompts, responses, tool calls and reward signals they generate become training material once someone retains them and transfers them ^[12].

Anthropic's February disclosure describes how it found the traffic it found. Proxy services assembled the account networks, mixed distillation requests into ordinary customer traffic and swapped in new accounts as old ones were banned ^[14]. Anthropic said it attributed the activity using request metadata, infrastructure indicators and information from industry partners ^[16]. The volumes it published work out to roughly 670 exchanges per account ^[18]. Moonshot AI's share was more than 3.4 million exchanges, about 21 percent of the total, including attempts to reconstruct Claude's reasoning traces ^[15][19].

Each of those signals depends on many accounts, thin volume in each, an intermediary in the request path ^[14][16]. A startup in one of the lab programs holds a single account and a rate-limit upgrade the lab granted it ^[8][9]. Anthropic could ban a fraudulent account as it found it ^[14]. It cannot do that with an account it upgraded on purpose.

Andriy Mulyar, founder and CEO of Nomic, wrote on September 14th that he had heard of "so many companies" using elevated model privileges to generate and resell traces during the past month ^[2]. The thread invoked Kimi, the model family from Moonshot AI, and Z.ai's GLM models ^[7]. Mulyar did not publish transaction records, datasets or buyer identities, and the thread does not establish that any OpenAI-, Anthropic- or Y Combinator-backed startup sold model output ^[4]. The screenshots he attached identified the accused founders and startups clearly enough that RuntimeWire cropped the names ^[5]. RuntimeWire reports that the three institutions were cited for their alleged connections to the startups and were not accused of authorizing any sale ^[6].

What a lab controls in this situation is the contract. OpenAI for Startups advertises free API credits, rate-limit upgrades and technical support for eligible companies in its investor network ^[8]. Anthropic's program offers credits and priority rate limits, with additional benefits for startups backed by partner investors and accelerators ^[9]. Both exist to recruit developers and help young companies absorb the cost of building on frontier models, and both create accounts able to produce larger volumes of output at lower effective prices ^[20]. The report does not describe any change to the terms of either program ^[21]. RuntimeWire's framing is that resale by recipients would turn those programs into a subsidy for rival-model training ^[22].