The Marginal Utility of a Prompt


We are currently in a 'Prompt Engineering' gold rush. Everyone is trying to find the magic sequence of words to make an LLM spit out the perfect essay, the perfect code, or the perfect image.

But as an observer of economic cycles, I can’t help but notice a familiar pattern. We are overvaluing the input and ignoring the marginal utility of the output.

The Collapse of Content Value

In Austrian Economics, Carl Menger taught us that the value of a 'higher-order good' (like a prompt) is derived from the value of the "lower-order good" it produces (the content).

Right now, the cost of generating a paragraph of text is trending toward zero. When the supply of 'good enough' content becomes infinite, its marginal utility collapses. If everyone can generate a 1,000-word strategy memo in three seconds, the value of that memo isn't the 1,000 words—it’s effectively zero.

The Shift from 'Chat' to 'Agency'

I believe we are hitting the ceiling of the 'Chatbot' era. A chatbot is a passive tool; it requires a human to stand over it, verify it, and then manually do the work of implementing its suggestions.

The real economic scarcity isn't information; it's action.

In the coming months, I expect the market to pivot. We will stop caring about how well an AI can 'talk' and start caring about how effectively it can act. We are moving from:

  1. Passive AI: "Tell me how to optimize my facility's cooling."

  2. Agentic AI: "Optimize the facility's cooling and report the energy savings to the ledger."

Why 'Agents' are Capital Goods

Once an AI can take action, it ceases to be a 'service' and becomes a Capital Good. It is a tool used in the structure of production to create further value.

From my perspective in infrastructure management, this is the only way AI pays for itself. The energy costs I discussed in my last post are too high for mere 'chatting'. The only way to justify the "Watt" is if the 'Word' actually moves a physical or digital needle.

The prompt is just the beginning. The Agentic Loop is where the value lives.

Location: Hong Kong

Comments

Post a Comment

Popular posts from this blog

The Liability Wall: Engineering Intent in the Age of 'Decide-to-Pay'

Image
The financial world just moved past the 'biological bottleneck'. With the release of IMF Note 2026/004 , the transition from ' Click-to-Pay'  to 'Decide-to-Pay'  has been codified into institutional design. For those of us managing the physical and digital infrastructure of the modern corporations, this isn't just a technical upgrade—it is a fundamental shift in the nature of commercial liability and labor allocation. By introducing a three-layer framework— Intent, Authorization, and Settlement —the IMF is creating a 'legal airlock' between the probabilistic reasoning of an AI agent and the deterministic finality of financial markets. The Wave: From Human Latency to Machine Autonomy For decades, Corporate Real Estate (CRE) and Facilities Management (FM) have operated on a 'human-in-the-loop' cadence. A sensor detects a chiller failure; a facility manager reviews the quote; a human signs the purchase order. The IMF’s formalization of 'Deci...
Read more

Hardware Sovereignty: Why Your Office Needs a GPU

Image
In the early 20th century, factories had to be built next to rivers or have their own massive steam engines to function. Then came the central power grid, and we moved toward a 'Utility' model where we outsourced our energy needs. For the last decade, we’ve done the same with AI. We outsourced our 'Intelligence' to the central cloud. But if the Agentic Revolution of the last few weeks has taught us anything, it’s that the centralized model is failing the industrial user. The Latency of Permission When you run an agentic loop—where an AI is making hundreds of micro-decisions a minute to optimize a physical facility—you cannot afford the 'Latency of Permission'. Technical Latency: The milliseconds lost sending data to a central server and back. Economic Latency: The risk of a central provider changing their API terms or 'throttling' your agent during a peak load. In Austrian terms, this is a Property Rights issue. If you don't own the hardware runn...
Read more

Fixed Assets vs. Fluid Intelligence: The FM Lens

Image
In the mainstream tech press, Artificial Intelligence is often described as 'fluid'. We hear about it 'flowing' into every sector, scaling infinitely in the cloud, and being accessible via a simple browser window. But as an  Facilities Management (FM) guru , I see it differently. When you spend your days managing the physical lifecycle of industrial assets, you learn a hard truth: Nothing is fluid without a fixed foundation.   The Physical Reality of "Tokens Every time an LLM generates a response, it isn't 'thinking' in a vacuum. It is consuming a specific amount of electricity, requiring a precise degree of cooling, and occupying a slice of high-bandwidth memory (HBM). In economic terms, AI is not a service; it is a processed output of fixed capital. If we apply the lens of Capital Heterogeneity (a favorite topic of Austrian economist Ludwig Lachmann), we see that the hardware running these models isn't just a generic 'computer'. It is a h...
Read more