2025-11-19
In the history of technology, there is a "Golden Ratio" for revolutions: High Fixed Costs, Zero Marginal Costs.
This was the magic of the Compact Disc. The first CD cost millions to engineer and master. The millionth CD cost pennies of polycarbonate plastic.
This was the magic of Windows 95. The code took years to write. Pressing the "copy" button cost nothing.
Silicon Valley has spent the last two decades addicted to this math. They believe that if you capture the user base with subsidies today, the "Zero Marginal Cost" future will inevitably make you profitable tomorrow.
But in the AI cycle of 2025, this math is a lie.
We are not witnessing a software boom. We are witnessing a manufacturing crisis disguised as a software boom. And the unit economics are fundamentally broken.
1. The Category Error: "Compute" is Not a Copy
The fatal flaw in the current valuation models is a category error. Investors are valuing AI companies like SaaS (Software as a Service) businesses, which scale infinitely with minimal friction.
But Generative AI is not SaaS. It is EaaS (Energy as a Service).
When a user plays a song on Spotify, the server sends a static file. It costs micro-pennies.
When a user asks a model to "write a python script" or "generate an image," the server doesn't send a file. It spins up a factory. It engages thousands of GPU cores, burns massive wattage, and performs billions of matrix calculations from scratch.
Every single interaction is a new manufacturing event.
The industry is currently selling these "manufacturing events" for $20 a month. But the factory (the H100/Blackwell clusters) costs billions to build and millions to run. They are selling jet fuel for the price of tap water, hoping that if they sell enough of it, the cost will drop to zero.
Physics says no.
2. The "Uber" Trap: Subsidizing a Commodity
The only reason to burn cash on subsidies is to build a "Network Effect" or a "Moat" that allows you to raise prices later. Uber subsidized rides to kill taxis, then raised prices.
But the AI incumbents have trapped themselves. They cannot raise prices because they have no moat.
If OpenAI or Google tries to charge the real cost of inference (let’s say $1,500/month per user to actually break even on capital expenditures), the user base will vanish. Why? Because the "Wikipedia Moment" is already happening in the shadows.
Open-source models, running on optimized, event-driven architectures, are providing "good enough" intelligence for free on local hardware. The incumbents are subsidizing a product that is rapidly becoming a commodity. You cannot charge a premium for tap water when it’s raining outside.
3. The Inverse Deflation Curve
In the CD era, the technology got cheaper. The pressing machines became commodities.
In the AI era of 2025, the technology is getting more expensive.
To keep up with the "State of the Art," companies are trapped in a Red Queen Race—running faster just to stay in place. They are forced to buy larger, hotter, more complex clusters (from H100 to Blackwell to Rubin). The capital expenditure required to stay relevant is compounding, not shrinking.
They are running up a down escalator. They have to spend $50 billion this quarter just to maintain the market share they bought with $40 billion last quarter.
4. Conclusion: The "Encarta" Bankruptcy
We are watching the financial equivalent of Microsoft Encarta in 1999. Encarta was a brilliant product, but it was built on a heavy, expensive, centralized production model (CD-ROMs, paid editors).
It didn't die because it wasn't good. It died because a decentralized, lightweight, zero-cost alternative (Wikipedia) made its business model obsolete.
The current AI giants are building "GPU Encartas." They are gorgeous, expensive cathedrals of compute. But they are mathematically doomed because they are trying to fight a war against efficiency.
5. The "Kingdom of Dirt" Fallacy
The final defense of the bulls is the "Battle Royale" theory: “Google, Meta, and OpenAI are fighting to the death because the Last Man Standing gets a monopoly on God. The winner takes the entire global GDP, so burning $500 billion now is rational.”
This assumes the prize is a Monopoly. It is not. The prize is a Commodity.
(Yes, this is the old Microsoft playbook: "Embrace, Extend, Extinguish." It worked for operating systems. It worked for browsers. It worked for office suites. That track record is exactly why they are so confident it will work again. But they are wrong. Intelligence is not an application suite to be bundled; it is a pervasive utility like electricity. Here is why.)
A. Intelligence Has No Network Effect
This is the fatal flaw in the comparison to Facebook or Amazon.
- Social Media: I use Facebook because you use Facebook. (Winner Takes All).
- Intelligence: I use an AI model because it gives me the right answer. If a local, open-source model on my laptop gives me the right answer for free, I do not care if 3 billion people use ChatGPT. Intelligence is a utility, like electricity. There is no "Winner Takes All" in electricity; there is just the race to the lowest price per kilowatt.
B. Winning the Ice Trade in 1920
Fighting a "Battle Royale" for AI dominance right now is like fighting a bloody war to monopolize the Ice Block Trade right before the electric refrigerator was invented.
- The Hyperscalers are fighting over who owns the massive, centralized "Ice Factory" (The H100 Cluster).
- Meanwhile, the "Wikipedia Moment" (Open Source/Local Compute) is inventing the personal refrigerator.
- The "Winner" of the Battle Royale will find themselves ruling over an empire of melting ice while everyone else just makes their own cold air at home.
C. The Pyrrhic Victory
Even if one company manages to "win" and kill the others, look at the cost. If you spend $1 Trillion to win a market that generates $50 Billion a year in profit (because open-source pressure keeps prices down), you haven't won. You’ve committed financial suicide. The winner of this Battle Royale inherits a throne made of ash.
6. The Verdict: The Host and The Parasite
We are witnessing a Multi-Trillion Dollar Symbiosis of Doom.
Group A: The Hosts (Hyperscalers) They are running a business model that loses money on every unit sold. They have no path to raising prices because open source is driving their product value to zero. They are burning cash to build factories they cannot afford to run.
Group B: The Parasite (The Shovel Seller) Nvidia is priced at $4.5 Trillion based on the assumption that Group A can keep buying $3 million racks forever. Nvidia is booking record profits by selling tools to an industry that has yet to prove it can make a profit using them.
This isn't sustainable commerce; it is Vendor Financing at a global scale.
When Group A finally admits the math doesn't work and turns off the capex spigot, Group B's backlog evaporates overnight. Nvidia isn't the bubble; Nvidia is the beneficiary of the bubble.
And when the host dies, the parasite starves and then... y'know.