OpenAI weighs a $1.2 trillion private round before its IPO
The frontier's price keeps rising while its exit keeps slipping. Two stories from the same afternoon, one about money and one about method.
OpenAI has held early conversations with large investors about another private round that would value the company at roughly $1.2 trillion, the Financial Times reported Tuesday, with Reuters and Bloomberg corroborating the reporting. The talks were initiated by investors rather than by OpenAI, the figure is early and could change, and the company declined to comment. The comparison that gives it meaning: OpenAI raised $122 billion in committed capital in March at an $852 billion valuation, and CEO Sam Altman said on Saturday that the company would not go public in 2026 at all, citing safety concerns about the technology. Read those two facts together and the story is not the number — it is where the money is queuing. Investors who were waiting for a listing now get offered a private entry at a much higher mark, days after the company told them the listing they were pricing against is not arriving this year. We covered the tender that set the previous mark — OpenAI closes $7B tender offer at $852B valuation.
The inversion carries a cost, and it is one the labs' own arguments have to absorb. A company asking Washington to slow the frontier down, then raising at a trillion-dollar valuation, is appealing to two audiences for two different things — restraint from regulators, growth from investors — and this week it supplied both in the same news cycle without having to reconcile them. It also puts an option on the public market: private primary capital raised now is IPO demand deferred, at a price that no filing has had to defend. Anthropic, the rival that filed first, is expected to begin marketing its listing in mid-October at the earliest.
A reinforcement-learning paper posted Tuesday argues that the models' repeated failures on hard problems are manufactured by the training loop itself, and that a change in batch composition fixes much of it. In "Learning to Solve Hard Problems in RL for LLMs by Never Giving Up," Michael Noukhovitch and co-authors Hamish Ivison, Nathan Lambert and Aaron Courville, working with compute from Ai2 and Mila, show that standard fine-tuning on verifiable rewards trains only on prompts whose sampled completions contain both passes and failures. Prompts a model gets entirely wrong — the hard ones — are filtered out and never trained on, so competence grows where it already exists: easy tasks get easier, hard ones stay hard. That is the Matthew effect the RL literature has been circling for two years, and the paper locates it in the data pipeline rather than in the model.
Their method, Never Give Up, keeps earlier correct completions and their counts for prompts already solved, so the hardest questions stay in the batch instead of being silently deleted once they look hopeless. On AIME-style math broken out by difficulty, the authors report gains on the hard third without losing the easy two-thirds, beating difficulty-curriculum baselines. They also test the obvious counter-argument — that this is a plasticity problem, meaning the network stops being able to learn new things — and report that models recover strong performance even after long stretches of bad training data. The signal was the bottleneck, not the weights.
What to watch: whether Anthropic's October marketing window pulls public money out of the private rounds now competing with it.
If the hardest problems are the ones a training loop quietly deletes, where else is capability growth stalling because the data got filtered out first? Tell us in the comments.
Sources: never-give-up (GitHub)
