I use AI every day. And new models often feel a bit smarter. But what’s the ROI on that marginal gain? Thinking about that has let me to develop a new interpretation of the latest calls to slow frontier AI development.
The usual argument is regulatory capture. OpenAI and Anthropic have already spent billions reaching the frontier. Expensive testing, audits, and regulatory approval could protect that lead by making it harder for smaller competitors to participate.
But there may be another economic incentive:
What if the frontier labs would genuinely prefer to stop spending so much money racing to train the next model?
That does not mean their safety concerns are insincere.
Dario Amodei has warned about catastrophic AI risks for years. In his recent proposal to “pace the frontier,” he cites rapidly improving capabilities, recursive self-improvement, and troubling examples of agent behavior.
“We must slow the pace at which we improve the capabilities of AI models,” he writes.
Sam Altman quickly agreed:
“I agree with Dario that we need to pace the frontier.”
Altman said this had become a primary topic of discussion inside OpenAI and endorsed Amodei’s proposal for independent evaluators with employee-like access. OpenAI committed to doing the same.
It is entirely possible that both men genuinely believe the systems they are building could become dangerous. But sincerity and economic self-interest can point in the same direction.
Escaping the training race
The economics of frontier development increasingly resemble a prisoner’s dilemma.
The existing models are already capable enough to support years of product development. The labs could focus on improving reliability, lowering inference costs, building better agents, and converting their technical capabilities into profitable products.
There are also growing questions about diminishing returns. Training runs keep getting more expensive, while each generation does not necessarily deliver a correspondingly dramatic improvement for users.
But no lab can slow down by itself.
Even if OpenAI would prefer to monetize its current models, it cannot risk Anthropic or Google spending tens of billions to leap ahead. The same is true in reverse. The return on the next training run may be getting worse, but conducting it remains rational if a competitor might do the same.
Regulation or government-enabled coordination could offer an escape. It would allow the labs to slow down together without requiring one company to surrender its position.
Amodei effectively acknowledges this advantage. He argues that coordinated pacing could give developers more time for safety work “without sacrificing commercial advantage.”
That is a safety argument. It is also a solution to a very expensive competitive problem.
Ceding the model as the moat
The open-model question complicates this theory.
Amodei argues that unauthorized distillation allows lagging companies to narrow the gap at a fraction of the cost of developing frontier capabilities independently. His solution combines pacing with chip controls, better protection of model weights, and efforts to restrict unauthorized distillation.
But distillation would not stop simply because frontier development slowed. Open models would also have more time to catch up to the current frontier.
That means slowing down could amount to partially conceding that the underlying LLM will become commoditized.
Perhaps the labs are increasingly comfortable with that.
OpenAI and Anthropic do not necessarily need permanent model-level differentiation if they can build defensible businesses around the model. Their moat could shift toward the complete system: the agent harness, user experience, proprietary data, enterprise controls, integrations, distribution, and the infrastructure required to deliver reliable AI at scale.
A longer model cycle would give them more time to make that transition. Instead of repeatedly spending enormous sums to preserve a temporary capability lead, they could use the intelligence they already created to build products customers depend on.
The bet would be that an open model can eventually reproduce much of the raw intelligence, but reproducing the entire product and operating stack will be harder.
That is a much more conventional software strategy, and potentially a much more profitable one.
Why Nvidia wants acceleration
Nvidia has a different economic interest.
The labs need AI to become useful and profitable. Nvidia benefits when producing that usefulness requires ever-growing quantities of computing infrastructure.
Jensen Huang’s regulatory position reflects that difference:
“Don’t regulate hypothetical theoretical harm, regulate actual and pragmatic harm.”
He has also urged countries to accelerate AI adoption and build their own infrastructure, arguing that being left behind would be the worst outcome. That position makes strategic and economic sense for Nvidia.
For a frontier lab, slowing the training race could reduce costs and extend the useful life of its models.
For the company selling the infrastructure required to conduct that race, a slowdown is considerably less attractive.
AI can win while infrastructure loses
This creates a counterintuitive way for the AI investment bubble to unwind.
OpenAI and Anthropic could become more profitable. Existing models could support better agents and applications. AI adoption and inference usage could continue growing.
At the same time, frontier-training growth could slow sharply, leaving some planned GPU, data-center, and power investments unnecessary.
In that scenario, AI does not disappoint. It succeeds with less capital than the infrastructure market expected.
That matters because today’s buildout assumes that training runs will continue becoming dramatically larger and more expensive. If progress instead shifts toward inference efficiency, product design, workflow integration, and better use of existing intelligence, AI could create enormous economic value without validating every infrastructure investment being made around it.
None of this proves that safety is a pretext. The risks may be genuine, and pacing may be responsible.
But pacing would also help frontier labs escape an expensive arms race and redirect investment toward products with more durable moats.
The usual cynical interpretation is that AI incumbents want regulation so they can pull up the ladder behind them.
The larger possibility is that they want permission to stop building a more expensive ladder every year.
If they get it, the labs may become better businesses even as the infrastructure boom built around them begins to deflate.