Sean Linehan, writing on X, offers a reframe for the debate over whether OpenAI, Anthropic, and Google can stay dominant as DeepSeek, Qwen, and GLM close the capability gap. His claim: both things can be true at once. Frontier labs can build some of the largest companies on Earth even as most of what their models actually do turns into a commodity.

The mechanism he proposes centers on two thresholds. Minimum viable intelligence is the bar a model must clear before it can do a task at all. Maximum necessary intelligence, which he shortens to MNI, is the point past which more capability stops adding value. Consider customer support email triage: a model that correctly routes a ticket 99 percent of the time is not meaningfully improved by a smarter successor, because the task has no further reward for getting sharper. Between those two thresholds, Linehan argues, more intelligence means more economic value. Above MNI, it means nothing, and buyers switch their attention to price, latency, and reliability of service instead.

That is the load-bearing move in his argument, and it doubles as a direct explanation for a pattern already visible in the market: procurement teams downgrading from frontier models to cheaper ones for tasks that used to require the expensive option. If a task’s ceiling has already been reached, spending more on intelligence is not conservatism or budget pressure. It is buyers correctly pricing a good that has stopped scaling in value for their specific use case.

Linehan draws the analogy to semiconductors, where mature process nodes still power cars and appliances long after cutting-edge transistor density stopped mattering to those customers, and to NVIDIA, which he notes drew roughly $194 billion of its recent fiscal year Data Center revenue against $16 billion from Gaming, evidence that a technology’s biggest market can shift entirely once new capability opens use cases the original product never touched. His broader wager is that frontier labs stay profitable not by defending the commodity tasks but by continually creating new markets: capabilities so far beyond the previous ceiling that no task existed to saturate yet.

The argument is strongest where it explains behavior we can already observe, like enterprise buyers shopping capability against cost line by line rather than defaulting to the newest release. It is weakest on the question Linehan himself flags as unresolved: how much valuable, unsaturated intelligence remains to be discovered. His entire thesis about frontier labs staying enormous depends on that unknown staying large, and he offers reasoning, not evidence, for betting it will.

He also raises a structural tension worth watching: a lab that sells both a model router and its own frontier models has an incentive to steer traffic toward the expensive option, while the buyer wants the opposite. That conflict, if he is right, is what opens space for independent routing layers that pick models on cost and task fit rather than vendor preference.

For teams buying model access, the practical takeaway is to benchmark task by task rather than model by model. A workflow that already clears its accuracy bar on a mid-tier model gains nothing from upgrading, and the budget saved is better spent identifying which of your remaining tasks still sit below their ceiling.

Posted by Sean Linehan on X, August 24, 2026.