AI Is Getting More Expensive for You and More Expensive for Everyone Building It. Only One of Those Groups Is Celebrating.
The free tier is shrinking. The token bills are growing. And the companies losing billions on every query are currently valued at hundreds of billions of dollars. At least one of those things has to g
You’ve noticed this, even if you haven’t put a number to it. The things that used to be free in AI are becoming paid features. The messages you used to get in bulk now run out faster. Image generation that came included is now metered or gated behind higher tiers. Enterprise companies that enthusiastically pushed their employees toward AI tools are now looking at monthly invoices in the millions and having difficult conversations about whether any of it is worth it.
Wall Street is in a state of excitement about artificial intelligence that is genuinely hard to overstate. Startups are raising not millions but billions, sometimes hundreds of billions. Governments are treating AI sovereignty the way previous generations treated nuclear capability. The largest companies in the world are announcing capital expenditure plans that dwarf anything the technology industry has previously attempted.
AI becoming more expensive for users and AI investment accelerating are not mutually exclusive phenomena. The common root of these two symptoms is a structural problem, and comprehending it necessitates an examination of the operational costs of these systems.
The Teleportation Problem
Let me use an analogy that will become uncomfortable as soon as you recognize what it maps to.
Imagine someone invents teleportation. Not a simulation of teleportation, not a metaphor, but genuine instant physical transport. It works, and the demonstration is real. The implications are revolutionary. The problem is that teleporting one object 100 kilometers currently costs $1,000 in energy and infrastructure. It works better than every alternative; it’s just that it costs many times more than every alternative.
You can still build a business around it. No one would purchase the technology if it were priced at its actual cost. So you charge $150. And every time you move something, you absorb an $850 loss and hope that scale, efficiency, and time will eventually bring the cost down to a point where the gap closes.
Roughly, what running generative AI looks like today.
Running ChatGPT costs OpenAI an estimated $700,000 per day in compute and infrastructure. OpenAI’s audited 2025 financial documents, obtained by technology writer Ed Zitron and independently confirmed by the Financial Times, show the company recorded a net loss of $38.5 billion in 2025. Revenue was $13.07 billion, more than tripling from 2024. But total costs and expenses reached $34 billion. The net loss was amplified to $38.5 billion by a $41.55 billion charge related to the conversion from a nonprofit to a for-profit entity; the operating loss alone was $20.92 billion. Either figure is historically unusual for a company receiving this level of investor confidence.
The standard response to this, the one you’ve probably heard, is that this is normal for transformative technology in its early phase. New technologies are always expensive at first. Costs come down with volume and experience. The same economics that made computing affordable eventually apply to everything.
Here’s the problem with that argument applied to generative AI: the curve is going in the wrong direction. Since the launch of ChatGPT approximately four years ago, running the systems has not gotten cheaper in aggregate. It has gotten more expensive. Every capability improvement requires more compute. Every performance improvement consumes more energy. The race between labs to maintain a frontier position means that the savings from efficiency are immediately consumed by the push to the next capability level. The cost reduction that the argument promises keeps getting deferred by the ambition the market rewards.
The Business Model We Haven’t Solved Yet
The hyperscalers, meaning Google, Microsoft, Amazon, and their peers, have announced aggregate AI-related capital expenditure that exceeds one trillion dollars across 2025 and 2026. These are genuine commitments, already being turned into physical infrastructure: data centers, power agreements, custom chips, network buildouts.
The revenue coming back against those expenditures, across the AI industry, currently represents roughly 15 cents on the dollar spent. The gap isn’t closing. It’s widening because competitive pressure forces companies to deploy capacity before demand justifies it.
This is not a temporary growth-stage challenge with a clear resolution path. It’s a structural mismatch between what it costs to produce AI output and what the market will pay for it. For individual companies, the business logic is to keep investing because the alternative, falling behind in the capability race, is assumed to be worse than the losses. For the industry, no single company’s decision to keep investing changes the underlying arithmetic of what the infrastructure costs and what the services generate.
OpenAI’s relationship with Microsoft makes this concrete in a way that should concern anyone watching closely. In 2025, OpenAI paid Microsoft $17.2 billion for research and development, compute services, and related expenses. Microsoft has been one of OpenAI’s largest investors, providing capital that OpenAI then uses to build on Microsoft’s infrastructure. In the same year, Microsoft paid OpenAI $303 million. This is a documented circular financial relationship: investment flows in, and a large portion flows back out to the investor as service payments, creating revenue for the investor without requiring external customers.
This practice, sometimes called round-tripping in financial analysis, is not unique to this pair. Nvidia, which has arguably benefited more than any company from the AI investment cycle, also invests in AI startups. Those startups use that capital to purchase Nvidia’s chips. Nvidia’s revenue grows while the stock rises.
Nvidia’s ability to invest in more startups increases. The FTC and several other regulatory bodies are reported to be examining these investment structures for exactly this reason. An unwinding of those relationships, or a regulatory intervention that forces cleaner separation between them, could be the trigger that breaks the cycle.
Why This Looks Like 2000, and Why It May Be Worse
In March 2000, the NASDAQ reached its peak before collapsing. The same technology that genuinely changed the world, the internet, had attracted more capital than any existing business model could justify. The mechanism was simple: companies with no credible path to profitability were valued on the assumption that someone, somewhere, would eventually figure out how to make the underlying technology generate returns commensurate with the investment.
The dot-com bubble had four specific characteristics that the AI situation shares almost exactly. New technology with genuine utility. Euphoric investor behavior that disconnected valuations from fundamentals. A cultural pressure, particularly in Silicon Valley, that reframed cash burn as a sign of ambition rather than recklessness. And a pervasive belief that being first or dominant in the new paradigm would eventually justify any amount of capital consumed getting there.
After the crash in 2000, the S&P 500’s technology sector lost 80% of its value. Cisco, one of the most foundational companies in internet infrastructure, did not return to its year-2000 valuation for 25 years. The damage spread well beyond technology stocks into the broader economy, and the full recovery took the better part of two decades.
The initial consideration is the investment scale; AI infrastructure funding is significantly larger in proportion to the economy compared to the internet companies of the 1990s. That means the exposure is more concentrated, the leverage positions are more significant, and the systemic implications of a sharp correction are correspondingly larger. Forced selling ensues when investment funds borrow to engage in a perceived unmissable technological cycle, causing prices to drop, which then accelerates this decline and necessitates further forced selling.
The second is the circular investment structure described above. The dot-com bubble involved genuine third-party investors losing their money on companies that turned out not to be viable. This AI situation involves companies whose reported revenues are partly funded by the same investors who provide the capital, creating an appearance of commercial traction that doesn’t fully reflect external demand. When those circular flows eventually stop, the reported numbers change significantly.
The concentration of AI company valuations in the top tier of the S&P 500 has reached a level, with the top technology and AI holdings now representing over a third of the index, that would make any prior market cycle look modest by comparison.
Nobody Has Proven It Will Be Worth It
The most honest thing you can say about the current AI investment cycle is this: no one has demonstrated, with evidence rather than projections, that generative AI will eventually generate returns proportional to what’s being spent on it. The argument for those returns is a belief, backed by analogies to other transformative technologies, that the pattern of enormous early-stage investment followed by enormous long-term returns will hold. It might. But it is a bet, and the size of the bet being placed globally right now is extraordinary relative to the evidence available.
This isn’t a novel observation. It’s been made consistently since at least 2024. The reason the investment cycle hasn’t slowed is a combination of genuine belief and competitive necessity. If you’re a company like Google or Microsoft and you believe that AI will eventually be as important as electricity or the internet, the cost of not investing is potentially catastrophic even if the cost of investing is currently enormous. The fear of being left out of a transformation of that magnitude is rationally more powerful than the evidence that profitability is distant.
That same fear operates at the fund manager level. The funds that missed the first wave of internet companies, the ones that didn’t invest in Amazon or Google early enough, spent decades explaining their absence. No fund manager wants to be the one who sat out the AI era because the fundamentals were unclear.
What Happens When It Corrects
This is where I want to be clear about what I am and am not saying.
Generative AI won’t disappear. The internet didn’t disappear in 2000. Amazon was worth $6 a share in 2001 and is worth several thousand dollars today. The correction of a speculative bubble around a technology is not an argument that the technology isn’t real or isn’t transformative. It’s an argument that the price being paid for exposure to that transformation has gotten significantly ahead of what the transformation has actually produced.
When the correction comes, and the signs, including cancelled data center projects, investors showing early nervousness, and companies beginning to apply genuine scrutiny to their AI expenditures, suggest it’s no longer a question of whether but when, the most likely outcome is a significant consolidation. The companies that provide genuine, measurable, documented value will survive. These companies that exist primarily because AI is a word that generates investment will not. The business model will shift from “grow the user base and worry about monetization later” to “demonstrate that the revenue justifies the infrastructure cost.”
That shift, when it happens, will probably benefit users over the long term. The companies that survive it will be the ones that figured out how to make the technology genuinely cost-efficient, not just impressive. And the conversation will move, finally, from what AI can theoretically do to what it actually produces at a sustainable cost.
It’s also worth noting something that rarely gets said in coverage of the AI investment cycle: the companies that have been most measured about their AI spending, the ones that have been gently mocked for not joining the capital arms race at full speed, may eventually look a great deal less foolish than they do today.
Is this bubble concern overblown, or are you already seeing the cracks in your own industry’s AI spending? Curious what the view looks like from your seat.