Solow among the thinking machines: the illusion of the silicon moat

Writing on a boat from New Zealand to England in 1863, Butler looked at the fast pace of early industrial machinery and argued that engines were evolving at a pace that, following Darwin, would result in machines becoming the superior race. He saw, over 160 years ago, that humans were already sacrificing a growing portion of our labor and capital just to keep the hardware clean, organized, and lubricated.

Many read his work as an early warning of an existential, science fiction apocalypse. I have always looked at it differently. As an economist, an asset allocator, and an entrepreneur, I look at the hundreds of billions of dollars currently being funneled into artificial intelligence infrastructure and see that we have walked right into the trap Butler anticipated. We are sacrificing our energy grids, our capital, and our natural resources to sustain a brute force infrastructure that misaligns with basic production economics. We are treating raw compute as a permanent competitive moat, and we are about to hit an inevitable economic wall.

How do we measure the productivity increase from hundreds of billions of dollars in AI infrastructure? In a knowledge economy, reasoning for both low and complex tasks is what we pay people for. For the sake of this argument, a unit of reasoning is equivalent to a unit of production.

To trace where the financial alpha will be captured, strip away the marketing and look at the foundation of modern growth theory: Robert Solow's total factor productivity residual. Output (Y) represents reasoning quality and quantity, Capital (K) represents the hardware stack, and Labor (L) represents the raw token workload. The residual, A, is everything that turns the same capital and the same labor into more output. For the past three years, venture funds and technology monopolies have disproportionately allocated capital to a single linear bet: if you want more cognitive output, you buy more GPU.

That brute force thesis is already showing diminishing marginal return. Blackstone's joint venture with Google puts up an initial five billion dollars in equity to fund computing power and custom TPUs directly, intentionally decoupling the silicon from traditional data center real estate. Traditional asset managers have financialized compute into its own sovereign asset class, treating it like a standard utility.

The first crack is the accelerating depreciation schedule. By 2023 and 2024, hyperscalers had quietly stretched the assumed useful life of their servers from three or four years out to six, a move that collectively pulled an estimated $18B out of annual depreciation expense and flattered earnings accordingly. Then the schedules started reversing. In early 2025, Amazon shortened the life of a subset of its servers back to five years, citing the accelerating pace of AI hardware development, and booked a corresponding hit to net income. In November 2025, Michael Burry went public with the claim that the sector was understating depreciation by as much as $176B through 2028. When the hardware itself depreciates faster every quarter, owning a large pile of it stops looking like a moat and starts looking like a melting asset.

The second crack sits on the labor side of the ledger, and it cuts against the narrative that AI replaces headcount. As of mid 2025, AI native firms were posting revenue per employee that dwarfs the incumbents: OpenAI near $2.8M per head and Anthropic near $2.5M, against a tech industry average closer to $450,000. That is the prize everyone is chasing. But the chase is turning out to be expensive. By 2026, an Nvidia research executive was saying plainly that for his team the cost of compute now runs higher than the cost of the employees. Microsoft pulled thousands of engineers off Claude Code licenses once the bill became hard to justify, and by April 2026 Uber's CTO reported burning through the company's entire 2026 budget for AI coding tools in only four months. The promise was that a token would be cheaper than a salary. For a growing number of firms, it is not.

Read the matrix down the right hand column and the conclusion writes itself. If the capital is commoditizing, and the labor it was supposed to replace is now more expensive than the compute, the only column left with a defensible margin is the residual: the orchestration layer.

This stops being an abstract infrastructure debate the moment it touches the operating budget. A CFO needs to forecast and control expenses. Headcount is a known cost, budgeted every year with precision. Compute is a wild card. The cost of an agentic workflow is not a function of users and it is not a function of revenue. It is a function of how much compute the engine decides to spend, and a static engine spends the same compute on a comma as it does on a hard deduction.

Instead of treating the token stream as a uniform block of labor, an intelligent runtime layer reads token signals in real time. It transforms the inference engine from a blunt instrument into a dynamic router, and in doing so it breaks the linear lockstep between hardware spend and reasoning output.

Think of how junior employees tend to do the simpler tasks, while more senior contributors plan, oversee, and handle the most complex parts of a task. This is how AI models should behave, and this is how PrizmalSwitch treats workflows.