When the frontier becomes harder to defend

The lead is becoming expensive to create and increasingly difficult to keep

The original economics of frontier AI seemed fairly intuitive.

Building the best model required extraordinary amounts of compute, engineering talent, data and capital. Only a handful of companies could afford to compete. If the gap between the frontier and everyone else remained sufficiently large, those costs could ultimately be supported by a durable technological advantage.

What looks less obvious today is the durability of that advantage.

The leading proprietary models are still ahead. On the latest Artificial Analysis index, Claude Fable 5.1 and GPT-6 Astra score 53, compared with 42 for GLM-5.3 Flash, one of the strongest open-weight models. That is still a meaningful gap.

But the economics underneath it look different. Artificial Analysis estimates a cost per task of $7.63 for Fable 5.1 at maximum effort versus around $0.25 for GLM-5.3 Flash. The cheaper model is not equivalent, but it is already performing surprisingly close to the frontier on some individual tasks.

And capabilities are not simply being recreated independently.

Anthropic says DeepSeek, Moonshot and MiniMax used more than 16 million Claude exchanges across roughly 24,000 fraudulent accounts to extract specific capabilities for their own models. The targets were not random: coding, agentic reasoning, tool use, computer use and other areas where Claude was differentiated. Anthropic says distillation can allow competitors to acquire capabilities at a fraction of the time and cost required to develop them independently.

I think that changes the economic question.

It is no longer only:

Can you build the best model?

It is also:

How long can you remain the only company that has what you just spent billions building?

If that interval keeps shortening, the moat is not really the model. It is the time advantage before the capability diffuses.

And that makes the current discussion around pacing much more interesting.

Pacing does something very useful economically

Dario is quite explicit that pacing does not mean stopping AI research.

His proposal is to continue technical progress while putting limits around the rate of unchecked capability advancement, introduce embedded external evaluators and coordinate standards between frontier laboratories.

One sentence in the essay stood out to me. He writes that coordinated pacing could give frontier developers more time to work on safety “without sacrificing commercial advantage or the United States’ lead in AI.”

That is obviously compatible with the safety argument.

But it is also economically convenient.

If Anthropic independently decides that its next model is too capable to release, while OpenAI releases an equivalent model, Anthropic bears the commercial cost of restraint.

If both companies slow releases or agree on comparable capability thresholds, neither necessarily does.

This is why the antitrust part of Dario’s proposal matters.

He explicitly argues that frontier companies should coordinate around standards and limits on unchecked progress, and notes that government mediation or a narrow antitrust waiver may be required for some of those discussions.

I don’t think this proves anything about motive.

But it solves a very specific economic problem: how do competing labs exercise restraint without allowing the other lab to take the market?

Nic Carter describes this as a prisoner’s dilemma. I think that framing is useful.

The other asymmetry is China

The proposal becomes even more interesting when looking at what should and should not be slowed.

Dario does not argue that the US frontier should simply accept a smaller technological lead.

Quite the opposite.

The same essay calls for restricting advanced chips, preventing model-weight theft and cracking down on unauthorized distillation because it allows lagging competitors to narrow the gap much more cheaply. His stated objective is to preserve enough of a US lead to make pacing possible and, if these measures work, potentially widen that lead over the next several years.

So the structure is roughly:

continue developing the frontier internally

while

making uncontrolled diffusion of the frontier harder.

Again, there are coherent safety and geopolitical arguments for doing this.

What I find interesting is that it is also almost exactly what I would expect a company to want if the economic value of its technological lead were eroding faster than expected.

The safety incentive and the commercial incentive do not need to contradict each other.

They may simply have started pointing in the same direction.

We can already see a small version of this

Anthropic’s Mythos architecture makes the distinction less theoretical.

Mythos 5.1 is Anthropic’s most capable model for cybersecurity and biology research, but access remains limited to vetted organisations. Fable 5.1 uses the same underlying model with additional safeguards that restrict some of those capabilities for broader availability.

I would not overstate what this proves.

Cybersecurity and biology are obvious areas where restricted access may be justified independently of economics.

But it establishes an important precedent:

the capability a frontier laboratory possesses does not have to be identical to the capability available to everyone else.

There can be a private frontier and a public frontier.

If that difference remains narrow and specific to genuinely dangerous capabilities, it is mostly a safety architecture.

If the gap begins to widen systematically, it starts becoming economically much more consequential.

That is one of the things I would watch.

There may also be a better customer for frontier intelligence than the public

The second part of Carter’s argument initially sounded more speculative to me, but Anthropic’s own data makes it harder to ignore.

As of August, Claude was already classified as “leading” 26% of Anthropic’s measured AI R&D work, up from less than 1% in February. More than 90% of that work involved Claude at least collaboratively, and Anthropic says around 30,000 agents were simultaneously doing research and engineering work on its main internal platform.

This creates another question.

If I have the most capable intelligence available, is selling it by the token necessarily its highest-value use?

Perhaps not.

If the model can help me build the next model faster, using it internally may produce more value than selling those tokens externally.

The logic could eventually extend further. If frontier intelligence becomes capable of generating genuinely valuable scientific discoveries, new materials, drug candidates, chip architectures or industrial designs, selling the inference that produced those outputs may capture only a small fraction of the value created.

I think this part remains much less proven.

Anthropic using Claude heavily to improve Claude is observable today.

Anthropic turning itself into some kind of vertically integrated owner of AI-generated intellectual property is not.

But the incentive is easy to understand.

And it gives another reason why the distinction between the private frontier and the public product might matter.

Where I currently land

My working thesis is therefore a little different from simply saying that LLMs are commoditising.

I think the more interesting possibility is this:

Frontier AI companies may be discovering that the technological lead they spent enormous amounts of capital creating is less durable than expected once the resulting model is exposed to the market. Distillation, open weights and cheaper competitors reduce the time during which frontier capability remains genuinely scarce. “Pacing the frontier” may therefore address a real safety problem while also helping preserve that scarcity: continue progressing internally, slow the diffusion of the strongest capabilities, coordinate restraint among domestic competitors and make replication by external competitors more difficult.

I don't need Dario's safety argument to be false for this to be true.

In some ways the thesis is more interesting if he genuinely believes it.

It would mean that the emerging safety architecture and the commercial interests of frontier laboratories are becoming aligned.

The result could be a market in which intelligence bifurcates.

There is a public frontier, where models become cheaper, increasingly substitutable and progressively commoditised.

And there is a private frontier, where a much smaller number of organisations retain capabilities that are not immediately released and increasingly use them internally.

Whether that gap actually widens is probably the most important empirical question in this thesis.

Then there is the capital sitting underneath all of this

There is another part of the story that I think is easy to miss when the discussion stays focused on models.

Even as the cost of individual AI tasks falls, the physical capital being deployed around AI is rising extraordinarily quickly.

The IEA estimates that capital expenditure by five large technology companies exceeded $400 billion in 2025 and could increase another 75% in 2026. It notes that those five companies alone are now investing more than the entire global oil and gas production sector.

There is an interesting apparent contradiction here.

The IEA estimates that energy consumption per AI task has been falling by at least an order of magnitude annually in recent years.

At the same time, electricity consumption from AI-focused data centres grew 50% in 2025, as cheaper and better models enabled much more intensive use cases such as reasoning, video and agents. Data-centre electricity demand is now projected to rise from roughly 485 TWh in 2025 to around 950 TWh by 2030.

I think this may be the second half of the same thesis.

Intelligence can become cheaper without making the AI economy less capital intensive.

Potentially the opposite happens.

Cheaper intelligence creates more economically viable uses for intelligence.

Those uses require more inference.

More inference requires compute.

Compute requires power, networking, memory, cooling, transformers, grid connections and data centres.

The IEA is already describing shortages or constraints across high-bandwidth memory, electrical equipment, grid capacity and other parts of that physical stack.

So even if the model layer commoditises, economic scarcity does not disappear.

It moves.

Two different ways this could develop

I currently see two broad paths.

In the first, frontier enclosure works.

The best laboratories maintain a meaningful private lead, access to frontier intelligence remains scarce and part of the economic rent stays concentrated inside those organisations.

In the second, it fails.

Open weights, distillation and competition keep compressing the gap until general-purpose model intelligence becomes much more interchangeable.

In that world, the rent should migrate further toward whatever remains difficult to reproduce: compute, electricity, infrastructure, proprietary data, distribution and eventually the physical systems through which intelligence acts on the economy.

The two outcomes are not completely exclusive.

We could easily end up with very cheap and abundant public intelligence alongside a much smaller private frontier, while both still consume enormous amounts of physical capital.

That may actually be the more interesting possibility.

What would change my mind

There are several things that would weaken this thesis.

If frontier labs continue releasing their strongest capabilities rapidly and broadly, the idea of a widening private frontier becomes much less convincing.

If the gap between open and closed models stops compressing, the economic argument for restricting frontier access also weakens.

If distillation proves much less useful than Anthropic currently suggests, then the durability of the existing model moat may be greater than it appears.

And if pacing remains narrowly focused on very specific dangerous capabilities rather than expanding into more general restrictions on frontier access, Carter’s interpretation would look overstated.

For now, I think the evidence supports a narrower conclusion.

The frontier is still valuable.

But keeping the frontier valuable is becoming a different problem from building it.

And that may have consequences well beyond the labs themselves.

If intelligence continues to become easier to reproduce, the important capital-allocation question will increasingly be less about who has the smartest public model and more about where scarcity survives around it.

That is the part I am trying to understand.