Wall Street is currently obsessed with a 19th-century ghost.
Recently, KKR sounded the alarm on a trend they view as "extreme" and historic. They pointed to the late 1800s—the era of railroad overbuilding—to warn that we are barreling toward a massive capital expenditure bubble in artificial intelligence. The consensus narrative is now set: tech giants are overinvesting, infrastructure is getting ahead of demand, and a painful market correction is inevitable.
It is a neat, tidy historical parallel. It is also entirely wrong.
The danger facing the market right now is not a 19th-century style infrastructure glut. The danger is a fundamental misinterpretation of what this technology actually does to corporate bottom lines. While private equity giants wring their hands over the supply side—the datacenters, the chips, the power grids—they are completely missing the rot on the demand side.
We do not have an overbuilding problem. We have an execution crisis.
The False Equivalence of Railroads and Large Language Models
To compare the current buildout of computing power to the expansion of the American railroad system is a fundamental misunderstanding of asset economics.
When a 19th-century railroad company laid tracks from Chicago to Omaha, that asset possessed physical exclusivity. If the company went bankrupt, the tracks remained. Another syndicate could buy them out of receivership for pennies on the dollar and run trains on them. The underlying utility of the physical network did not degrade to zero just because the financing failed.
Silicon is different.
The data centers being built right now are tailored for specific compute architectures. More importantly, the software stack running on top of them is depreciating at an unprecedented velocity. A state-of-the-art cluster optimized for today's dense transformer models may look like an expensive, energy-hogging paperweight if the industry shifts toward sparse architectures or neuromorphic computing.
The Reality Check: You cannot buy a bankrupt AI startup’s compute cluster out of receivership and expect it to hold its value for thirty years. The half-life of this technology is measured in months, not decades.
By focusing on the macro-level spend, analysts are treating AI infrastructure like utility networks. It isn't a utility network. It is a hyper-depreciating asset class masquerading as infrastructure. The risk isn't that we build too much capacity for the future; it's that we are building the wrong kind of capacity for right now.
Dismantling the "Productivity Boom" Illusion
Every earnings call features a CEO boasting about a 30% reduction in customer service resolution times or a 40% speedup in software engineering output. These metrics are vanity metrics. They measure activity, not profitability.
I have watched enterprise software companies spend millions deploying generative AI tools across their engineering organizations. They track lines of code written per hour and declare victory. But when you audit the actual product delivery, a grim reality emerges.
Yes, engineers are writing code faster. But they are also generating more bugs, requiring more regression testing, and creating technical debt at double the historical rate. The "productivity boom" is actually just a velocity boom. You are moving faster, but you are also heading toward a cliff.
The Real Math of Enterprise AI Efficiency
Let's break down the actual unit economics of replacing human labor with automated systems in a standard corporate environment.
$$Margin_Impact = (L_s \times W) - (C_{inf} + C_{audit} + C_{liability})$$
Where:
- $L_s$ = Labor hours saved
- $W$ = Fully burdened hourly wage of the human worker
- $C_{inf}$ = Total cost of inference and software licensing
- $C_{audit}$ = Cost of human-in-the-loop verification to prevent hallucinations
- $C_{liability}$ = The priced risk of systematic errors, data leaks, or regulatory non-compliance
When companies run this calculation honestly, the margin expansion vanishes. The cost of running high-context window queries across an enterprise data silo is shockingly high. When you add the human audit layer required to ensure the output doesn't violate regulatory frameworks, the net savings often fall to single digits.
The industry is cheering for top-line efficiency gains while ignoring the exploding operational overhead required to keep these systems stable. It is a shell game.
Why the "People Also Ask" About AI ROI are Flawed
Look at the standard questions dominating corporate boardrooms today. The premises themselves are broken.
"When will AI investments show up in GDP growth?"
This is the wrong question. It assumes AI is an additive force that expands market sizes. In reality, AI is predominantly a deflationary, cannibalistic technology.
If an agency uses automated tools to produce five times as many marketing assets for a client, the client does not pay five times as much money. The client demands a price cut because the commodity value of an asset has plummeted. AI shifts the consumer surplus entirely to the buyer, squeezing the margins of the provider. It compresses industry revenues rather than expanding them. GDP figures will likely remain flat even as computational throughput explodes.
"Which industries will be disrupted first?"
The consensus says white-collar knowledge work: law, finance, coding. This misses the defensive capability of institutional inertia.
The most heavily regulated industries possess structural defense mechanisms against automation. A law firm doesn't just sell legal advice; it sells accountability. A large corporation pays a premium for a top-tier law firm because that firm possesses malpractice insurance and a reputation to defend. An LLM cannot be sued for malpractice. It cannot stand in front of a regulatory body and take the blame.
The disruption is not happening at the top of the food chain. It is happening at the bottom, wiping out the entry-level roles that serve as the training ground for the next generation of experts. We are eating our own seed corn.
The Capital Allocation Trap: The Pivot to Nowhere
Corporate boards are terrified of being labeled laggards. This fear has driven a massive misallocation of capital from core product development into speculative internal AI projects.
Imagine a scenario where an established logistics company diverts 20% of its R&D budget away from optimizing its physical supply chain network to build a proprietary internal AI assistant. They train a model on their internal handbooks. They build a custom interface.
The result? Their employees get answers to HR questions slightly faster, while their core physical distribution network loses market share to a nimbler competitor that focused entirely on asset utilization and route density.
This is the hidden cost of the current trend. The opportunity cost of chasing general-purpose intelligence is the starvation of specialized, proprietary capabilities that actually create competitive moats.
[Standard Corporate R&D Budget]
|
+--> 40% Core Product Optimization (Starved)
|
+--> 60% Speculative AI Frameworks (Overfunded)
|
+--> Result: Infinite prototypes, zero production margin.
The Brutal Truth About the Moat
There is no sustainable competitive advantage in plugging into a third-party API.
If your core business strategy relies on building a wrapper around a foundational model developed by a handful of tech giants, you do not own a business. You own a temporary feature. The moment those foundational providers update their model capability or adjust their inference pricing, your margin disappears.
True differentiation requires proprietary, non-public data loops that are impossible to scrape from the open internet. Yet, most enterprises are actively feeding their unique operational data into public or semi-private models, effectively subsidizing the training costs of the platforms that will eventually replace them.
It is a corporate tragedy played out in real-time: paying for the privilege of training your own executioner.
Stop looking at the CapEx charts of the hyperscalers and worrying about a 19th-century bubble. The real crisis is inside the enterprise balance sheets of the companies buying those services. They are trading capital for computational speed, only to find that speed without strategy simply accelerates their descent into commoditization.
