Former Anthropic Researcher Says IPO Roadshow Slide Pushed Him Out
Key Takeaways
- •Jacob Coxon, a pretraining researcher who spent roughly four months at Anthropic before resigning on September 8, 2026, cited an IPO roadshow slide on compute efficiency as a crucial factor in his departure.
- •The disputed slide charted superexponential growth in pretraining compute efficiency on a logarithmic scale, positioning Anthropic's Mythos model as part of an accelerating trend rather than an isolated breakthrough.
- •Anthropic's roadshow targets approximately $64 billion in proceeds, a share price of $70 to $78, a Nasdaq listing under the ticker ANT, and a valuation approaching $2 trillion.
- •Coxon warns that if efficiency gains continue compounding at that pace, self-improving AI systems could arrive earlier than anticipated and potentially pose existential threats.
- •By resigning early, Coxon gave up unvested equity in a company nearing a $2 trillion valuation, and his safety concerns predate the slide, including an X post on AI safety that drew more than 100 million views.

Former Anthropic pretraining researcher Jacob Coxon said a slide from the company's IPO roadshow was a crucial factor in his resignation. In remarks made public on November 13, 2026, Coxon pointed to a chart showing superexponential gains in pretraining compute efficiency — a data point he described as a selling point for investors and, for him, a warning sign.
The timing is notable. Anthropic is in the middle of a roadshow for an initial public offering that aims to raise approximately $64 billion, and few companies want their pitch deck quoted back at them by a departed employee. A roadshow is the pre-listing phase in which a company's leadership presents its investment case to institutional investors before a final offer price is set.
What the slide showed, and why it concerned him
Some vocabulary helps frame the dispute. Pretraining is the stage in which an AI model learns from enormous amounts of data, while compute efficiency measures how much capability a lab gets out of each unit of computing power.
The slide plotted those efficiency gains on a logarithmic scale. On a standard chart, exponential growth looks like a hockey stick; on a log scale, the same growth flattens into a straight line. When a line on a log chart bends upward, growth is outpacing even exponential rates. That is what "superexponential" means here: the gains are not just compounding, but accelerating as they compound.
According to the research findings, the slide supported a specific argument: Anthropic's Mythos model was presented as part of a trend rather than an isolated breakthrough.
Coxon read the same chart through a different lens. If efficiency gains keep stacking at that rate, self-improving AI systems could arrive sooner than expected. He has publicly warned that such systems could pose existential threats.
A short tenure with long-running concerns
Coxon's time at Anthropic was brief. He joined in May 2026 and resigned on September 8, 2026, after roughly four months in the role. Before that, he worked at OpenAI from 2023 until May 2026. He is 27 years old and holds a mathematics degree from Cambridge University.
His safety concerns predated the IPO slide. He voiced worries about rapid progress toward self-improving AI during his time at the company, and a post on X about AI safety drew more than 100 million views. His arguments also overlapped with views from inside Anthropic itself: the research findings note that his assertions echoed sentiments from Evan Hubinger, who heads alignment science at the company.
Leaving came at a cost. By departing early, Coxon forfeited unvested equity — shares that only become an employee's property after a set period of continued service — a meaningful sacrifice while a former employer prepares to go public. He has emphasized that safety, not money, drove the decision. Walking away from pre-IPO equity at a company targeting a valuation near $2 trillion lends that claim some weight.
What the offering looks like
The IPO itself is shaping up to be enormous. Anthropic aims to raise approximately $64 billion, with a targeted share price between $70 and $78. The company plans to list on Nasdaq under the ticker "ANT," and its valuation target is approaching $2 trillion.
The roadshow pitch leans on Anthropic's computational advantages, and the efficiency slide fits squarely into that story: better results per dollar of compute, at an accelerating pace. That is precisely the claim Coxon has seized on. The roadshow's strongest argument and his central worry rest on the same data. For readers tracking the sector, it is a compact illustration of how a single capability metric can anchor both an investment case and a safety warning. With the roadshow underway, the remaining steps before trading are the close of investor presentations and the setting of a final offer price.
Source: CryptoBriefing. The original article was edited by Diego Almada Lopez.