The Efficiency Paradox:

Why ‘Infinite Compute’ is a Legacy Dream

Levie argues compute budgets will climb forever. But brute-forcing old workflows with new power is paving the cow path.

In a recent post, Aaron Levie argued that as AI agents become more effective, worker compute budgets will “monotonically go up over time.” The vision is one of scale: engineers running parallel agents overnight, lawyers generating thousands of drafts, and sales reps flooding the zone with automated outreach.

It’s a compelling vision of the future, but it’s built on a fundamental misunderstanding of how technology evolves. Levie is describing “paving the cow path” — using new power to brute-force old workflows. The real future of AI isn’t about doing more tasks; it’s about Targeted Inference and the collapse of the task chain.

The Core Problem

If every company 10x’s their outreach compute, the value of an individual email drops to zero.

When volume explodes, the market pivots to precision.

The Problem with More

Levie’s argument assumes that productivity is token-dependent. In this world, a sales rep’s value is tied to how many AI-driven touchpoints they can generate. But we are already hitting “peak noise.”

We don’t need agents to “review more drafts” (an output-based mindset); we need agents to ensure a contract is “unbreakable” (an outcome-based mindset). The former requires a mountain of tokens; the latter requires one perfect realisation.

The former requires a mountain of tokens; the latter requires one perfect realisation.

Collapsing the Task Chain

Most modern workflows are “work about work.” They are linear chains — Research → Draft → Edit → Review — designed to manage human limitations. Brute-forcing these chains with agents is inefficient.

True AI maturity involves Targeted Inference. Instead of running 1,000 agents to pick up every piece of straw in a haystack, a high-reasoning model acts as a magnet, pulling the needle out instantly. As models gain “reasoning density,” the number of hops required to reach an outcome shrinks. We move from “Infinite Tasks” to “Zero-Step Outcomes.”

The Efficiency Principle

The goal of a business isn’t to maximise its compute budget; it’s to minimise the distance between a problem and its solution.

As agents get smarter, they will actually do less work to get better results.

The idea that compute budgets will climb forever ignores the inevitable drive toward efficiency. Just as we moved from massive mainframes to efficient cloud microservices, AI will shift from massive, “chatty” token consumption to surgical, high-value strikes.

The Brute Force Path

Scale everything. More agents, more tokens, more output. Run parallel workflows overnight. Flood the zone. Measure success by volume.

The Precision Path

Collapse the task chain. Use fewer, smarter inferences to reach decisive outcomes faster. Measure success by the gap between problem and solution.

The winners won’t be the companies burning the most tokens — they’ll be the ones using the least amount of compute to achieve the most decisive outcomes.

The winners won’t burn the most tokens. They’ll use the fewest to achieve the most decisive outcomes.

That is the efficiency paradox.