Is the AI Race Slowing Down, or Is This a Cold War?

What the Leaders Actually Said

Amodei's proposal has three layers. Anthropic says it will give embedded third-party evaluators ongoing access to inspect safety practices. He also wants frontier companies in democratic countries to coordinate on standards and limits, with governments making that coordination legally possible. The hardest layer is international coordination, where every participant would need confidence that rivals were not using the pause to move ahead in secret.

Altman said he agreed that the frontier needed pacing and that pacing did not mean stopping. Hassabis described the proposal as pointing in the right direction while noting the difficulty of implementation. Musk's endorsement was shorter. None of these statements creates an enforceable agreement, a shared threshold, or a timetable. Agreement on the word is easier than agreement on who measures progress and who decides when a model is safe enough to continue.

OpenAI has already described one version of pacing in practice. It said a large frontier reinforcement-learning run remained on hold while the company used smaller runs and evaluations to gather evidence about model behaviour and safeguards. That is closer to engineering discipline than surrender: stop one risky step, study what happened, improve the controls, and continue when the evidence supports it.

Capability Has Run Ahead of Ordinary Use

I keep thinking about smartphones. The processors, cameras, displays, and radios have improved enormously. Yet the basic reasons most people carry one have changed much less. We call, message, browse, watch video, take photographs, use maps, and play games. A small group pushes the hardware hard; everybody else lives comfortably inside a fraction of what it can do.

New phone capabilities keep appearing because the hardware permits them, but software has not always become lighter or clearer. Faster chips often allow applications to consume more memory, run more background work, and add interfaces that the average owner never asked for. Some of the improvement is real. Some of it is absorbed by less efficient software.

AI now has a similar gap. Models can reason for longer, operate tools, write software, inspect documents, and coordinate multi-step tasks. That is remarkable. It is also well ahead of how most people use them. A large share of everyday demand is still summarising, drafting, searching, translating, asking questions, and tidying information. Those jobs do not always need the most expensive frontier model available.

This does not make frontier research pointless. Advanced models reveal new risks, create scientific and engineering capabilities, and eventually make smaller systems better. But another capability jump is not automatically the most urgent product improvement for the person who wants a dependable answer on an ordinary device at an affordable price.

The Workload Is Growing Longer

A chatbot generally receives a prompt and produces a response. An agent may plan, search, call a tool, inspect the result, revise its plan, call another model, retry a failed step, and continue until it reaches a stopping condition. One visible request can therefore contain many inference calls and a great deal of generated text. If several agents collaborate, the multiplication becomes easy to miss because the user still sees one task in the interface.

The International Energy Agency now treats reasoning and agentic tasks as a distinct infrastructure concern. Its 2026 analysis says energy-intensive uses such as video, reasoning, and agentic work can consume hundreds or thousands of times more energy per query than simple text generation. That is a range across very different workloads, not a claim that every agent burns through power at the top end. The direction is enough to matter: longer work requires more compute, and widespread adoption multiplies it.

Efficiency gains complicate the picture. A cheaper token or more efficient chip reduces the cost of one action, but it also makes more actions economical. An agent that used to stop after five steps may run fifty. A company that could afford one automated workflow may deploy hundreds. Better efficiency is necessary; it does not guarantee lower total electricity use.

This is where safety and infrastructure stop being separate stories. Long-running agents have more opportunities to make a bad decision, and every additional evaluation, monitor, retry, or safety model also uses compute. We need those controls. We also need agent designs that know when to stop, use smaller models for routine steps, cache stable results, and ask a person rather than spending another hour guessing.

Adoption Is Not the Same as Revenue

OpenAI told the Associated Press in April that ChatGPT had more than 900 million weekly users and that about 95 percent did not pay. One company and one product cannot stand in for the whole market, but the ratio exposes the commercial problem neatly. The world is interested in AI. Interest does not mean most people are ready to pay even for the entry subscription, let alone the true cost of a long frontier-model task.

The free tier is not wasted capacity. It teaches people what the product can do, creates habits, widens access, and supplies a path towards paid individual and business use. But frontier capability becomes awkward when it is too expensive to give away generously and too costly for smaller subscribers to use often. A model can be technically superior and still fail to become part of ordinary life because people ration it.

This is why I am sceptical of treating every slower release cycle as evidence that AI progress has stalled. The better commercial move may be to make existing intelligence cheaper, faster, safer, and less demanding to operate. A model that finishes the common task on modest hardware, with predictable latency and fewer retries, may create more value than a spectacular model that most customers can barely touch.

Safety and Economics Can Point Together

A genuine safety slowdown and a commercial efficiency push are not mutually exclusive. The same pause that gives evaluators time to understand a frontier system gives engineers time to improve inference, reduce failure rates, simplify serving, and bring capable models to more customers. The responsible choice can also be the economically sensible one.

That is the optimistic reading. The colder reading is that every laboratory wants coordination because none can slow down alone. Each leader fears a dangerous capability jump, but each also fears that a rival will make it first. Shared evaluations, visible checkpoints, and outside scrutiny become the verification machinery of a technological cold war. Everyone says they want restraint; nobody can afford to rely on goodwill.

For businesses using AI, the distinction matters less than it may seem. We should not build operations around an assumption that every new frontier release will be cheaper, unlimited, or automatically better for our work. We should measure the task, choose the smallest model that performs it reliably, restrict how long agents can run, preserve human approval for consequential actions, and track cost per accepted result rather than cost per token alone.