The Cost Of Going Slower
“Pace the frontier” is an awkward slogan. It sounds like someone walking back and forth along a border. The intended image is a pace car slowing the racers behind it.
Anthropic CEO Dario Amodei puts the point more plainly in his new essay: “We must slow the pace at which we improve the capabilities of AI models.”
Amodei also predicts AI could cure most major diseases within five to ten years. That forecast and his demand for slower progress are in direct tension: delaying better models could mean delaying those cures.
A credible case for going slower has to explain why the reduced risk outweighs that cost to patients. Josiah Lippincott brought the cost into focus with a post about the children’s hospital he passes on his way to work.
The Case For Going Slower
AI extinction forecasts ask us to reason through a chain of possibilities: rapidly increasing intelligence, the ability to act independently, access to resources, and an inability to contain whatever emerges. Each link deserves scrutiny. Putting a frightening outcome at the end of the chain doesn’t establish the probability of reaching it.
Hunter Ash made a useful point about the limits of intelligence: progress requires interaction with the physical world. A brilliant hypothesis still needs experiments, equipment, energy, and time. Medical research illustrates both sides of that constraint. Better AI can accelerate discovery; scientists still have to test treatments.
Our position is that speculative extinction scenarios don’t justify handing incumbent AI companies collective authority to slow technological progress. Demonstrated failures—agents exceeding their permissions, security breaches, unsafe deployments—give us specific problems to fix. Access controls, independent testing, and accountability for damage can be judged against those problems.
A general slowdown therefore has a broad burden of proof. Its advocates have to weigh the risks of development against the patients who would benefit from faster discoveries and treatments. The more credible Amodei’s medical forecast is, the greater the cost of delaying progress.
Who Gets To Set The Speed?
Amodei’s proposal contains several different things. Anthropic is committing to embedded third-party evaluators with extensive internal access. Industry-wide pacing would require coordination, for which he seeks government mediation or a narrow antitrust waiver. Possible limits extend to capabilities, training compute, and AI-assisted AI development. International coordination is a further ambition.
The immediate commitment covers outside scrutiny. A binding industry-wide speed limit remains a proposal. Investors should keep that distinction in mind when reading headlines about an AI pause.
The antitrust request is especially revealing. Alvaro Bedoya pointed to an existing distinction between sharing information about threats and coordinating commercial behavior.
Bedoya’s opening claim is broader than the evidence supports. His antitrust point is stronger. A 2014 DOJ–FTC statement says properly designed cyber-threat information sharing is unlikely to raise antitrust concerns. It distinguishes that activity from exchanging competitively sensitive information about prices, output, or business plans.
That distinction goes to the heart of this debate. Companies can help one another defend against a cyberattack. An agreement governing how quickly competitors improve their products raises a different competition question.
Established labs would have a strong interest in shaping the standards that determine which challengers can enter their market. Sincere concern about dangerous AI can coexist with an incentive to protect a lucrative business. Giving the incumbents a say over competitors’ development schedules would institutionalize that conflict.
A sound test for a proposed rule is whether a new entrant can satisfy it through demonstrable safety performance, or whether compliance effectively requires the money, institutional relationships, and infrastructure of an incumbent.
The Buildout Has Its Own Momentum
On September 6th, The Information reported that Anthropic had struck compute deals over 11 months that could cost $517 billion.
Those commitments predate the current pacing debate. They establish the scale of the commercial plans now intersecting with it. The next useful evidence will be changes to contracts, construction schedules, and equipment orders.
Anthropic’s IPO preparations point in the same direction. Reuters reported on September 11th that Nvidia was discussing an investment of up to $10 billion in the offering. The talks remain provisional, but the capital-raising effort is continuing.
The Financial Times reported Sunday that Anthropic’s gross margins exceed 80% before deducting distribution partners’ revenue shares and model-training costs. It also reported that the company expects positive adjusted operating income for a second consecutive quarter.
A restriction on the pace of frontier-model improvement could coexist with growing demand to run models already developed. Serving more customers, handling more queries, and deploying AI across more businesses all consume compute. A slowdown aimed at training and one aimed at total computing capacity would have different consequences for suppliers.
Competition also makes a collective slowdown difficult to sustain. A company that holds back risks losing customers to one that keeps improving. At the national level, governments face a similar problem with rivals they can’t reliably monitor.
President Trump reinforced that obstacle on Sunday. In remarks reported by Reuters, he rejected what he considered exaggerated AI fears and emphasized maintaining America’s lead over China, while allowing for some guardrails.
Kalshi Finance pushed the contrast even further Sunday:
The post supplies no source or context; even whether Zuckerberg meant AI broadly or Meta’s own work is unclear.
We expect the administration to remain broadly supportive of American AI development through the rest of Trump’s term. A comprehensive, enforceable slowdown would have to overcome that political resistance as well as the industry’s commercial incentives.
Our Base Case: The Buildout Continues
We expect AI infrastructure spending to keep growing over the period relevant to our trades. Compute commitments, Anthropic’s IPO preparations, demand to run existing models, and the administration’s stance all point toward continued expansion. The pacing campaign has yet to produce an enforceable industry-wide spending limit.
Our options trades generally run for much shorter periods than a presidential term. We’ll focus on supplier orders, spending guidance, and binding rules that could change demand during a trade’s life. Broad capex cuts or cancelled orders would challenge this outlook; a new slogan by itself doesn’t change it.
A pace car only sets the speed if the other drivers agree—or are compelled—to stay behind it. That’s the part of this campaign investors should watch.
Trading The Buildout
We’ve had success trading companies that supply the AI buildout.
For example, Credo Technology Group (CRDO) supplies the high-speed connections that move data through AI clusters.
And Advanced Micro Devices (AMD) designs accelerators used to train and run AI models.
We’ll continue to look for promising AI-related names and structure trades when we find compelling setups.
If you’d like to see how we identify these trades and get a heads-up when we place the next one, our Start Here page explains the approach.
“Pace the frontier” sounds like the watchword of people that are politically aligned with China.