How an AI bust can make Big AI bigger

The transformational technology of the 21st century is beginning to resemble the transformational technology of the 19th century: the railroad. As artificial intelligence (AI) overcomes constraints on human cognition to redefine commerce and culture, it echoes how the railroad overcame the constraints of geography with similar effects. Both transformations have been accompanied by similarly massive investments in the infrastructure that makes them possible.

Ultimately, the railroad boom produced the railroad bust. Both the causes and consequences of that bust offer lessons for today. Before the collapse, railroad companies and investors overextended themselves in the belief that future demand would justify enormous capital expenditures. After the financial reckoning, those assets remained while their ownership consolidated. J.P. Morgan and other financiers reorganized the financially weakened and bankrupt railroads in a process that came to be known as “Morganization.” By 1906, seven companies controlled roughly two-thirds of U.S. railroad mileage.  

Could AI be heading toward a “digital Morganization”? As the market for AI becomes more competitive, control of its infrastructure could get more concentrated.

The AI economy comprises three highly interconnected groups. The leading model developers—Anthropic, OpenAI, Google, and Meta—boast either hefty balance sheets or impressive fundraising capabilities. The infrastructure necessary for AI models to work is similarly concentrated with almost two-thirds of worldwide cloud infrastructure revenue generated by three hyperscalers with powerful balance sheets: Amazon, Microsoft, and Google. The immediate financial concern is about the stability of a third group of independent companies. These so-called neocloud companies such as CoreWeave, Nebius, and Crusoe have borrowed billions of dollars to build competitive computing infrastructure. They may be corporately independent, but the financial stability of these companies is tied to the business they receive from Big AI.

This independent computing capacity is important to Big AI. Perhaps its greatest importance, however, is the alternative it provides for those choosing to develop AI applications outside the dominant companies. Just as the railroad tracks mattered less than what ran on them, so is it that while AI models are the base, it is the applications that run on them that will be transformational. If digital Morganization comes to pass, it could mean the loss of these competitive alternatives and the further concentration of AI power.

According to the Financial Times, “Investors are losing confidence that revenues from the AI boom will be sufficient to support the vast debt piles financing data centres.” One institutional investor said independent data centers are facing a “Darwin moment” in which the weakest could struggle to survive.

These financial issues could restructure the entire sector to further increase the power of Big AI. As we saw with the railroads, when a financially troubled infrastructure company goes bankrupt, its essential assets remain. In the 21st century, this means that data centers with their microchips, power contracts, fiber connections, and customers become available to those with the financial strength to absorb them.

The conditions that could produce such a reprise are now visible. The open question is whether the rumblings in today’s financial markets are a lead-up to a railroad-like outcome for the 21st century’s transformative technology. If the infrastructure remains valuable while some of its independent owners do not survive, the result could be the modern equivalent of railroad consolidation. Anticipating the potential for such an outcome should begin now by identifying what is acceptable and the policy decisions necessary to protect against digital Morganization.

As the railroad experience demonstrated, demand can be booming while the companies financing the capacity to meet that demand are failing. The demand for a transformational technology and the financial viability of those building its infrastructure are two different realities linked by financial physics. The trillions of dollars invested in computing infrastructure must be supported by the cash flow that infrastructure makes possible.

Whether anticipated revenues can support these financial commitments is the central concern. “OpenAI’s decision to delay its initial public offering until next year and a slowdown in Anthropic’s quarterly sales growth have triggered doubts about the sustainability of the AI boom,” the Financial Times reported. A Gartner study warned that a continuation of the current investment trajectory risks exposing more than $3.5 trillion of AI infrastructure to write-down by the end of 2030. Data center owners will need to generate about $8.5 trillion through the end of the decade to cover their spending and capital outlays, yet current forecasts suggest no more than $3.6 trillion, the Gartner study found.

One path to the necessary cash flow is from proprietary AI services such as those offered by OpenAI, Anthropic, and Google. These companies generate revenue by bundling use of their “closed” models with the necessary computing power. The revenue from that bundled service helps support the commitments the companies have made for computing capacity. These commitments form a major backstop to funding the necessary compute infrastructure being built.

Challenging that plan are unbundled “open-weight” models. Typically less expensive, these models make trained weights available under varying licenses that allow the user to download, modify, and run them. The weights are often free to download, so that the buyer pays only for the compute to run them. The result creates pricing pressure on Big AI while also threatening their bundled compute revenue and thus the companies’ long-term commitments to purchase computing capacity.

As the AI frontier pushed forward and model competition increased, Big AI responded by lowering the price of older models much like a baker marks down day-old bread. Cutting-edge models continue to demand a premium. For example, OpenAI cut prices of ⁠its low- and mid-tier AI models earlier this year, which Reuters reported “may intensify competition ⁠in the industry as U.S. companies battle cheaper Chinese rivals for customers increasingly wary ‌of the technology’s ballooning costs.”   

The availability of cheaper intelligence can trigger what economists call the Jevons Paradox: that greater efficiency lowers the cost of using a resource, stimulating demand that can increase overall consumption. William Stanley Jevons developed the concept while studying coal use in Victorian Britain. A report from institutional investor Apollo demonstrated its applicability to AI, finding that while total spending on AI models almost doubled since late 2025, it occurred while the price per token fell more than 90% since 2023. Greater demand does not necessarily mean greater financial returns.

At the frontier model level, the economic question becomes whether the revenues from selling access to increasingly powerful models can support the cost of their development and operation. This, in turn, has a ripple effect on independent infrastructure providers. As of June 30, CoreWeave’s $104 billion order backlog, on which it relies to secure financing, is largely anchored by contracts from OpenAI, Meta, and Microsoft. The Gartner study predicts that more than 80% of the independent compute providers will “cease independent operations” by 2030.

If falling token prices prevent model developers from generating enough to support their long-term compute commitments, even as use soars, the availability of that capability could come into question. The consequences would then extend to the independent compute providers that built capacity on the strength of those expectations.

The irony is that open-weight models, which helped to drive prices down, could be among the biggest losers. Open-weight models may be less expensive, but they still require computing power. If financial distress reduces the number of independent infrastructure providers, application developers could find fewer compute alternatives. What would remain are hyperscalers with the built-in self-preferencing conflict of interest to protect their own models and those they back.

The paradox is that competition in models makes intelligence more affordable while making the infrastructure behind it less competitive.

That customers of frontier models are acting rationally in the face of high frontier prices is exemplified by communications giant AT&T. In July, the company announced that it was consuming an average of 45 billion AI tokens per day. A token is the unit of management, measurement, and billing of an AI model. In large language models, a token represents approximately three-fourths of a word. Each model developer creates and counts tokens differently, and their price varies depending on whether they are used to input a query to the model or output an answer. As of Oct. 17, OpenAI (GPT-6 Astra) and Anthropic (Claude Fable 5.1) charge $10 per million input tokens and $50 for a similar unit of output tokens for their frontier models.

While the above prices are illustrative, they identify how maximizing token use can quickly become a budget-buster. AT&T addressed that challenge by behaving like a rational consumer. The company says it reduced its AI costs by as much as 90% by developing its own routing and caching gateway that matches tasks to the most cost-effective closed or open-weight model. Not all AI-assisted activities require the horsepower of a frontier model. The decision of a user revolves around whether a particular model is capable of performing a specific task.  To take advantage of such decisions, there are multiple open-weight models from Big AI including Meta’s Llama, Google’s Gemma, Nvidia’s Nemotron, and OpenAI’s gpt-oss. Cloud companies have also begun their own lower cost-bundling with open-weight models at competitive prices.

The rational response of AT&T illustrates the distinction between demand and revenue. Its demand for AI is enormous and growing, yet the amount needed to pay for each unit of intelligence is declining. Multiply that behavior across all AI users, and it becomes apparent that rapidly increasing AI usage may not necessarily produce the revenues assumed when the capacity was first financed.

China is also behaving rationally. While Chinese models may lag the most capable American frontier models, they are sufficiently capable for many productivity-enhancing tasks. As a result, U.S. companies have turned to Chinese models to avoid rising costs, or as the Washington Post framed it, “The hottest AI models in Silicon Valley face a powerful source of competition.”

It is another development that could have a significant impact on the cash flow of American model developers and thus alternative infrastructure providers.

Typically, these Chinese models are open weight. In fact, “China has opted to go all in on an open-source approach to AI,” according to the U.S.-China Economic and Security Review Commission. While AT&T made a national security decision not to use Chinese models, companies such as Airbnb, Pinterest, and DoorDash have reportedly used open-weight models from China in at least some part of their business.   

These highly capable but lower-cost open-weight models have great strategic value for China. They serve to redirect the discussion away from the American lead in frontier models while establishing Chinese AI as a credible price and performance alternative. They also threaten American Big AI with competition that feeds financial fears.

There is no box score that ranks American and Chinese open-weight models. Any such comparison would depend on variables such as task, model size, and cost of deployment. “The Token State of Play,” published Sept. 22 by OpenRouter, a marketplace for accessing both open and closed models, points to some interesting trends in activities on that platform. Closed-weight models garnered 73% of the spending on models yet represented only 28% of total token volume. According to the report, Chinese tokens commanded 69% of total token usage during the week monitored. These statistics do not establish the profitability of either open- or closed-weight models, nor do they demonstrate that Chinese models are necessarily taking revenue from American companies. They do, however, demonstrate how open-weight models can command a disproportionately large share of token consumption while generating a relatively small share of spending. It is this difference between usage and revenue that could complicate the financial assumptions underlying the infrastructure boom.

All AI relies on computing capacity. The headlines regarding financing AI infrastructure have redirected attention from the gee-whiz capabilities of frontier models and their eye-popping valuations to the bits and bytes that underlie everything. Such reorientation reinforces how competition at the model layer of the AI stack could have the perverse effect of driving consolidation at the essential compute layer.

Whether prices fall because of open-weight models, Chinese rivals, or a price war among the AI labs themselves, the frontier labs take in less for each token they sell. While this could threaten the viability of their compute commitments, for the major model labs, their balance sheets and ability to raise capital provide a degree of insulation. Similar protection, however, may not be available for independent infrastructure providers. If those companies cannot raise the necessary capital to honor their debt obligations, their compute assets could become available for acquisition. Among the most likely potential acquirers are the companies that already dominate AI and/or allied financial interests. This creates a reality that necessitates antitrust policy that goes beyond the traditional failing firm defense as well as open access and interconnection requirements for the surviving entities.

This brings us back to the lessons learned from the railroads. The problem with Morganization was that the owners of essential infrastructure could determine the competitive opportunities of businesses dependent on that infrastructure. AI infrastructure represents a similar danger that controlling computing capacity can affect who gets to deploy the applications that will shape the economy.

The prospect of digital Morganization calls for an anticipatory awareness that addresses both ownership and conduct. Antitrust authorities need to guard against the dominant firms and their allies using financial distress as a convenient justification for acquisitions that eliminate independent computing alternatives. Such scrutiny should not be limited to outright acquisitions; the industry has amply demonstrated the ability to create interlocking relationships with the potential for similar concentration of control. Equally important, the owners of essential AI infrastructure should face enforceable expectations of nondiscriminatory access and interconnection to ensure that control over computing capacity does not become control over who can compete in AI.

Meaningful railroad oversight came only after the rail map had largely been drawn. AI policy still has time to act before the AI map is set. American capitalism has once again created tremendous new opportunity. Defining the outcome of that opportunity is too important to be left to a handful of Big AI CEOs and their financiers.