Does the Frontier Advantage Shrink?

A company spends ten times more compute. Its model improves. Both statements can be true while the improvement becomes progressively smaller. The commercial question is how much that smaller gain matters.

Scaling laws need not predict a sudden cliff. Diminishing marginal returns mean that another unit of resource buys less improvement on the chosen measure. Change the measure and the apparent shape of progress can change too.

The case for convergence

In Meek Models Shall Inherit the Earth, Hans Gundlach, Jayson Lynch and Neil Thompson argue that diminishing returns can narrow the capability difference between highly resourced developers and those with limited budgets. Their model examines a fixed-distribution next-token objective. As raw compute grows, marginal gains shrink, potentially weakening the advantage of scaling faster than competitors. Paper.

This is an important challenge to the idea that a widening spending gap must produce an ever-widening capability gap. It is not proof that all models converge on every commercially important task. The relationship between prediction loss and the outcome being purchased carries much of the argument.

Capability on a chosen task

Compute / cost →

Conceptual, not measured. A smaller slope is not a universal ceiling; changing the metric or algorithm can change the picture.

The 2026 qualification matters

The July 2026 paper Two AI Metrics Diverged: Will it Make All the Difference? revisits the question explicitly. It finds that the answer depends on how capability is valued and measured: a shrinking gap in validation loss can coexist with a growing frontier lead on other measures. Paper.

That is not a minor footnote. Imagine two systems that are close on average prediction quality, but only one can reliably complete a particularly difficult engineering task. A buyer of that task may experience a substantial difference. Conversely, a customer with routine work may gain little from paying for the leading model.

The example is our interpretation, not an empirical result from the paper. It illustrates why an aggregate score cannot settle every purchasing decision.

Evidence that still favours scale

MIT FutureTech’s March 2026 account of a separate study of 809 models released from 2022 to 2025 reports that scale explains much of the observed frontier performance variation, while proprietary techniques and shared algorithmic progress matter differently away from the frontier. Its regression and attribution methods offer evidence about that sample, not a controlled experiment proving that expenditure alone causes every gain. Is there a “secret sauce” in AI development?.

Taken together, these studies do not support a simple obituary for large models. They support more careful questions about the objective, the comparison group and the metric.

What would test the hypothesis?

Compare systems on a fixed workload over time. Record the task success rate, cost per accepted result, latency, human correction time and severe failures. Include both a fixed budget and a frontier budget. Keep the evaluation difficult enough to avoid a saturated test making all models appear equal.

Then repeat when tools, retrieval and verification are allowed. If the smaller model catches up only with an expensive supporting system, count that system. If the frontier model also benefits from better orchestration, give it the same opportunity.

Evidence for convergence would be a narrowing useful performance gap at a durable cost advantage. Evidence against it would be new valuable tasks that remain inaccessible to smaller budgets, even after efficient orchestration. Both can occur in different markets at once.

The economic implication is conditional: lower reproduction costs may erode scarcity in some capabilities while the frontier opens new scarce ones. That is a more demanding thesis than “bigger is finished”, and a better basis for strategy.

References

• Gundlach, Lynch and Thompson (2025), Meek Models Shall Inherit the Earth, research paper.
• Two AI Metrics Diverged (July 2026), workshop research paper; metric-dependent conclusions.
• MIT FutureTech (March 2026), Is there a “secret sauce” in AI development?, institutional account of empirical research.