The Race In Anthropic’s Own Words
Anthropic CEO Dario Amodei wrote this month that AI companies “must slow the pace at which we improve the capabilities of AI models.” Reuters reported Monday that Anthropic’s IPO prospectus describes a continuous, overlapping release schedule as essential to staying at the frontier. Customer usage and revenue, the filing says, are driven by new models.
Those statements expose a real tension. A lab can believe its technology needs stronger safeguards while its business depends on producing new versions of it. Anthropic’s proposed stock-market debut puts that tension in front of prospective investors.
Harvard psychologist and author Steven Pinker posed this question in an open letter to Scott Alexander, a prominent rationalist writer and supporter of effective altruism who has examined the case for slowing AI: “Why are the same people who warn that a technology will kill us all building it as fast as they can?” Anthropic’s prospectus offers part of the answer: new models help bring in the revenue that funds further development.
Whose Values Set The Pace?
Effective altruism’s moral calculus can reach striking conclusions. In 2019, philosopher Peter Singer and researcher Michael Plant questioned the donations pledged to rebuild Notre-Dame, asking how many lives the money might save elsewhere. An EA Forum writer argues that insects’ aggregate suffering dwarfs human suffering. These are moral priorities many people wouldn’t bring to decisions about medical research and technological progress.
An EA exponent also helped finance the AI race. The SEC described Sam Bankman-Fried as a prominent supporter of effective altruism; Anthropic says he led its $580 million Series B in 2022. At FTX and Alameda, his professional and sexual relationships overlapped: he had an on-and-off relationship with Alameda CEO Caroline Ellison, and they shared a bedroom in the Bahamas. Ellison later testified against him in his fraud trial.
The social ties are visible around Alexander, too. He wrote that he met his wife at one of Aella’s parties. Aella, a former cam model and OnlyFans creator who uses that one-word pseudonym, documented her 2024 birthday gangbang and said its organizers excluded applicants affiliated with the pro-AI-acceleration movement because, in her view, they were “hastening our doom.” She co-founded an AI-risk creator bootcamp funded by the Machine Intelligence Research Institute, which Eliezer Yudkowsky co-founded. The public should know whose values and relationships stand behind demands to slow a technology that will affect everyone.
What Amodei Is Asking For
Amodei’s pacing plan calls for three steps. Anthropic says it’ll invite independent evaluators into its development process. It then wants frontier labs in democratic countries to coordinate on safety standards and limits on unchecked capability gains, with some forms of coordination requiring government support. A third step would seek international coordination, including with geopolitical rivals.
He explicitly says pacing won’t halt model training or technical progress. Anthropic released Claude Opus 5.5 after his essay and said outside groups had tested it before release. That kind of testing is consistent with the plan, though Amodei’s proposed evaluators would have ongoing access inside the lab. A meaningful industry-wide slowdown still depends on competitors and governments.
That’s the hard part. Amodei argues that a coordinated slowdown could buy safety researchers time without costing American labs their commercial position or the United States its lead. Until such coordination exists, each lab faces the same incentive to keep improving and shipping its models.
The Prospectus Puts A Price On Time
The economics make the dilemma sharper. According to the prospectus reviewed by Reuters, Anthropic generated nearly $4.6 billion in revenue in 2025 and recorded an $8.06 billion operating loss. It spent $7.33 billion on compute and infrastructure that year and disclosed $518 billion in cloud, computing and infrastructure obligations over the coming years.
Anthropic also warns in its filing that advanced AI could pose catastrophic or existential risks. Reuters says risk factors occupy roughly 80 pages of the 261-page prospectus, compared with 48 pages describing the business. The company is telling investors that its product can be dangerous and that remaining competitive requires a steady stream of new versions of it.
There is a counterweight to the pressure from public shareholders. Reuters reports that Anthropic’s seven co-founders plan to control 50.1% of voting power through a Founder LLC, while the company will remain a public-benefit corporation. That gives its leaders room to make safety decisions that hurt near-term financial results. It doesn’t make the cost of falling behind Google, OpenAI or other rivals disappear.
Pinker takes risks such as cyberattacks and unconstrained agents seriously. In a follow-up to Alexander, he argued that power-seeking and self-preservation are design-contingent behavior, not automatic consequences of intelligence. His question lands because even people who assign extreme stakes to AI safety continue to lead the race. Amodei’s answer is that stopping one lab could hand the field to a less cautious rival. The prospectus shows how expensive it would be for Anthropic to test that proposition alone.
What This Means For The Buildout
For investors, the immediate signal is continued competition for model capability, customers and compute. Safety work may change how models are tested and released. A broad cap on the pace of advancement would require coordination that Amodei himself says Anthropic can’t accomplish unilaterally. We’ll watch whether policy turns that proposal into enforceable limits, and whether demand for chips, power and data centers changes before it does.
We also have indirect exposure to Anthropic’s prospective IPO through options on Zoom Communications (ZM 0.00%↑) and SK Telecom (SKM 0.00%↑), two publicly traded companies with pre-IPO equity stakes in Anthropic. We explained those positions last week. A successful listing at a higher valuation could prompt investors to assign more value to both stakes.
Our view has been that the AI buildout will continue, and we’ve positioned our trades accordingly. So far, that’s been the right way to bet. These are three examples from our record of exited trades.
Credo Technology Group (CRDO 0.00%↑) makes high-speed interconnect products for AI data centers. This is how our first completed trade in Credo turned out:
Micron Technology (MU 0.00%↑) supplies high-bandwidth memory for AI training and inference. Here is the result from the trade we opened in March:
Advanced Micro Devices (AMD 0.00%↑) supplies AI accelerators. Our December trade in it delivered the strongest return on maximum risk of these three:
We’ll keep screening the AI stack for compelling trade setups. For a heads-up on our next AI-related trade, subscribe or upgrade below.
New to Portfolio Armor? Our Start Here page explains how we select trades, size their maximum risk, and manage exits:
