Every few weeks, the financial press publishes an apocalyptic column warning that hundreds of billions of dollars poured into artificial intelligence infrastructure will inevitably trigger a systemic economic crisis. The investment banks, for their part, are having a marvellous time underwriting $30 billion data centre debt packages and preparing prospectus filings for $1 trillion initial public offerings, pausing only occasionally to publish polite research notes wondering where the enterprise software revenue has gone. Reading these frantic debates over a civilisation-ending capital expenditure bubble, one feels an overwhelming urge to walk the commentators to an office window and point quietly at the nearest mobile phone mast.
Between 2019 and 2024, the global telecommunications sector deployed roughly $1 trillion dollars rolling out fifth-generation mobile networks. The accompanying marketing campaign was an operatic triumph. Promoters promised remote surgeons performing delicate procedures from the back of speeding ambulances, autonomous lorries platooning down motorways in millimetre synchronisation, and tactile smart cities where municipal streetlights conversed politely with passing refuse collection vehicles. Telecommunications executives assured their boards that subscribers would happily pay an extra $10 a month for the privilege of loading a website half a second faster.
The subscribers did nothing of the sort. They flatly refused to pay a single additional dollar for faster connection speeds, and carriers generally ended up giving away the service for no incremental fees. Carrier returns on invested capital drifted quietly below their cost of funding. Share prices settled into a multi-year slumber.
Yet the most instructive aspect of that historic capital misallocation was how little anyone seemed to mind. There was no systemic banking crisis. There were no emergency parliamentary inquiries into the collapse of Western capitalism. Pension funds quietly absorbed the modest dividend reductions, the physical antennas remained bolted to the masts, and the global economy kept ticking along with perfectly serviceable mobile bandwidth. The total capital poured into that entire five-year telecommunications cycle matches the scale of capital currently being committed to the foundation model tier. It was a massive, mediocre commercial investment, yet society absorbed it without breaking stride.
The contemporary panic over an artificial intelligence bubble makes an elementary accounting mistake. It aggregates every dollar spent on high-voltage electrical substations, every dollar spent on experimental model training runs, and every dollar spent on enterprise workflow integration, dumps them into a single ledger labelled capital expenditure, and declares that the entire sector must either deliver immediate transformation or collapse into insolvency.
Capital behaves in three completely different ways depending on which layer of the stack it occupies.
At the physical foundation sit the hyperscalers and utility infrastructure providers. Their balance sheets carry high-voltage grid connections, industrial cooling towers, optical transceivers, and sprawling warehouses filled with specialised silicon.
These represent tangible industrial assets supported by multi-year take-or-pay leases with corporate treasuries. If an experimental model architecture becomes obsolete overnight, the underlying electrical substation remains intact. The kilowatt-hours and matrix multiplication engines retain durable commercial utility across the broader economy. They can run closed-weight frontier models, they can host open-weight alternatives, they can simulate complex protein folding, or they can generate photorealistic cat videos at planetary scale.
The compute landlord sells access to physical scarcity. The landlord collects cash rent regardless of which research laboratory wins the cognitive arms race.
In the middle tier sit the foundation model companies, where capital undergoes a far more combustible chemical reaction.
Here, hundreds of billions of dollars in equity are converted into electricity, run through GPU clusters for ninety days, and distilled into a static file of floating-point weights. The output is an intangible asset with an operational half-life measured in months.
The moment a training run concludes, open-source researchers in Hangzhou or university laboratories begin reverse-engineering and distilling its capabilities for a tiny fraction of the original discovery cost. The pioneer absorbs the expensive exploratory research bill. The fast follower captures the functional equivalent for cents on the dollar.
Model weights behave like unpasteurised milk, depreciating in economic value the instant they encounter open market competition.
At the top of the stack sit the enterprise software platforms, vertical agents, and developer environments that hold actual commercial relationships.
These applications command genuine operational switching costs. They store corporate permission hierarchies, maintain compliance audit trails, integrate with messy legacy databases, and manage the bespoke business logic that institutions take decades to negotiate.
For these software businesses, every collapse in the unit cost of underlying intelligence expands their gross margin. They treat the intelligence beneath them as a cheap commodity input, extracting durable economic rent by owning the human interface and operational context.
The primary defence mounted by foundation model venture capitalists relies on Jevons Paradox. If intelligence becomes ninety-five percent cheaper, they argue, demand will explode by several thousand percent. Autonomous agents will run continuously in the background, consuming trillions of tokens per hour, generating an ocean of revenue that easily compensates for declining unit prices.
This argument is mathematically sound for two of the three layers, while quietly executing the third.
Consider the consequences across the stack when token consumption increases tenfold while the market price of a token falls by ninety percent.
The compute landlords in Layer 1 celebrate. Delivering ten times the volume of tokens requires ten times the electricity, ten times the cooling capacity, and ten times the silicon throughput. The physical utility meter spins ten times faster, collecting massive cash flows on non-negotiable physical constraints.
The application providers in Layer 3 celebrate equally hard. They acquire ten times the cognitive horsepower for the exact same operating expenditure. Their products become dramatically more capable, their user workflows become stickier, and their software margins widen into deep moats.
The model labs in Layer 2, meanwhile, find themselves caught in the classic telecommunications curse. They run their clusters at ninety-nine percent capacity to serve an astronomical volume of queries, while their economic margin per query compresses to the marginal cost of compute. They move ten times the traffic, collect roughly the same gross revenue, and face the non-negotiable requirement to spend $3 billion dollars on the next frontier training run just to match the latest open-weight release.
Jevons Paradox creates enormous fortunes for the physical utility and the workflow application; it leaves the wholesale pipe running red-hot with zero economic profit.
In traditional enterprise software, the primary goal of executive leadership is to construct baroque, impenetrable moats that make leaving the platform an operational nightmare. If a corporate customer wants to migrate away from an established ERP system, it must hire a small army of management consultants, spend eighteen months reconciling conflicting database columns, and risk catastrophic disruption to monthly payroll.
The foundation model labs, in an act of breathtaking technical charity, engineered the exact opposite outcome. They universally agreed upon a single, standardized HTTP interface:
POST /v1/chat/completionsEvery developer framework, enterprise proxy, and orchestration library was built around this universal format. As a consequence, migrating an enterprise workload from an $80 billion proprietary model to an open-weight alternative hosted on a regional cloud requires neither board approval nor architectural trauma. It requires altering two environment variables in a configuration file: the base URL and the API key.
The model labs raised hundreds of billions of dollars on the promise of building unassailable monopolies, while simultaneously popularizing the most frictionless commodity interface in the history of commercial computing. A corporate customer can swap out its primary intelligence provider during a morning coffee break.
This structural reality explains the extraordinary haste currently visible in Silicon Valley. Both OpenAI and Anthropic are actively preparing public market debuts at valuations a $1 trillion dollars (or perhaps even $2trn, they’re basically similar numbers, right?), accompanied by a deafening chorus of narrative hype.
To command the revenue multiples traditionally reserved for high-margin software monopolies, these labs cannot afford to be viewed by institutional investors as wholesale utility distributors locked in a brutal price war. They must present themselves to Wall Street as permanent digital sovereigns whose intellectual property will dominate human commerce for the next generation.
The rush toward the public markets represents an urgent liquidity sprint. The early backers, having funded the heroic exploration phase, understand the underlying economics. They can see that the recurring capital expenditure required to stay at the frontier behaves like an endless subscription to an astronomical utility bill.
Taking these companies public allows the private backers to transfer that relentless capital treadmill to public pension funds, sovereign wealth managers, and passive index trackers before the reality of ongoing depreciation arrives on a GAAP income statement.
When the narrative dust settles, the outcome will look remarkably familiar to anyone who paid attention to the telecommunications cycle.
The physical compute buildout will endure as a valuable industrial asset, providing the power grids, data centres, and specialised hardware that will underpin computational research for decades. The software application layer will capture durable corporate profits by embedding cheap intelligence into daily human work.
The equity write-downs will be concentrated almost entirely within the middle tier. The frontier model labs will discover that spending billions of dollars to deliver technological miracles at wholesale commodity prices is a magnificent gift to human civilisation, and an exceptionally poor way to compound shareholder capital.