Anthropic, which is racing toward a Nasdaq IPO, disclosed in an updated prospectus that a self-accelerating AI loop is already underway. There have been previous instances in which AI R&D accelerated and humans only recognized it after the fact, and the self-acceleration red line—"achieving two years' worth of R&D progress in a single year"—will soon be crossed. Because the capabilities of large models are difficult to measure in real time, humans can often only observe AI's evolution as it stood several months earlier. Currently, more than 90% of the code inside Anthropic is generated by Claude, and the internal cost of calling Claude has exceeded the monthly cash payroll of all employees combined. The company has drawn a clear line: it will never allow Claude to fully autonomously design the next generation of frontier models without human oversight, in order to prevent hidden risks from accumulating across generations and spiraling out of control. For this reason, there have already been cases of new models being shelved and pretraining being delayed. At the same time, Anthropic has signed long-term computing contracts totaling nearly $900 billion. Caught in the prisoner's dilemma of the AI race—where no leading player dares to slow down—the company has no choice but to keep handing more R&D work to Claude to absorb the massive computing power and catch up on progress.