Steve Eisman is looking for the thing that could break the AI trade. He thinks he has found it in OpenAI and Anthropic.

I think he has identified the right companies but perhaps the wrong failure.

Eisman, whose bet against the housing market was chronicled in The Big Short, recently called OpenAI and Anthropic the “Achilles’ heel” of the AI trade. He estimates that approximately 70% of the AI revenue at Microsoft, Amazon, Google and Oracle comes from those two labs. If something bad happens to either one, he argues, the consequences could spread across the companies supplying their compute, hosting their models and building businesses around them.

That is a legitimate concentration risk. But it assumes the principal danger is that OpenAI or Anthropic might fail.

There is another possibility. AI could succeed beyond our expectations while OpenAI and Anthropic fail to capture the greatest fortunes it creates. In fact, the more successful AI becomes, the less dependent the world is likely to be on either company—and the smaller their share of the total profits generated by AI may become.

That is the indispensability trap.

OpenAI and Anthropic may become this era’s railroad or fiber providers. They are building foundational infrastructure that makes an enormous new economy possible. They may remain important, valuable and even indispensable in the abstract. But that does not mean they will remain essential to every transaction conducted across the economy they helped create.

The railroads connected markets, moved people and goods, and reorganized national commerce. Fiber providers supplied the capacity beneath the commercial internet. Both created extraordinary economic value. Yet the greatest fortunes did not necessarily remain with the companies that laid the tracks or buried the fiber. They migrated to the businesses that used those networks to create new products, control customer relationships and reorganize entire industries.

AI is following a similar path.

OpenAI and Anthropic made advanced machine intelligence accessible and commercially useful. They showed businesses what was possible and set off one of the largest infrastructure investment cycles in history. Their models helped turn AI from a research discipline into a general-purpose business platform.

But the breakthrough and the fortune are not the same thing.

The current AI economy is heavily dependent on two expensive, proprietary laboratories with enormous capital requirements. That dependence may support extraordinary valuations for OpenAI and Anthropic in the short term, but it may also be one of the factors holding AI back from achieving its full potential.

Businesses do not want to build critical operations around a technology they can obtain from only one or two providers. They want model choice, price competition, portability and the ability to move their workloads. They want to choose the least expensive model capable of performing a particular task. They want to combine frontier models with smaller specialized models and, where appropriate, models they can operate themselves.

As AI becomes more important, those requirements become stronger, not weaker. No enterprise wants its intelligence layer to become a new form of proprietary lock-in. No government wants its economy or national security to depend entirely on two private American companies. No cloud provider wants the bulk of its AI future permanently tied to models it does not fully control.

The more indispensable AI becomes, the more pressure there will be to make intelligence abundant, affordable and interchangeable.

That is where Kimi K3 and the broader rise of Chinese and open-weight models matter. Kimi is not merely another entrant trying to beat OpenAI or Anthropic on a benchmark. It represents a different economic threat. A model that is close enough, cheap enough and open enough does not need to be the world’s absolute best model to alter the market. It needs only to be good enough for an enterprise orchestration system to choose it over a more expensive alternative.

The release of Kimi K3 rattled semiconductor stocks because investors saw another potential DeepSeek moment: A Chinese open-weight model reportedly approaching the capabilities of leading American systems while requiring fewer advanced chips. The immediate market losses were largely recovered, as they were after DeepSeek, but the strategic signal should not be dismissed merely because stock prices rebounded.

China may not be trying to win the same economic contest as the American labs. OpenAI and Anthropic are attempting to preserve scarcity at the model layer and monetize access to frontier intelligence. China can benefit by commoditizing that layer, encouraging widespread adoption and allowing value to accumulate elsewhere—in cloud infrastructure, applications, devices, robotics, manufacturing and industrial systems.

Kimi does not have to destroy OpenAI or Anthropic to spring the trap. It only has to give customers a credible alternative.

Nor must one Chinese model become the global winner. Kimi, Qwen, DeepSeek, Zhipu and the open-model ecosystem collectively weaken the assumption that advanced intelligence will remain scarce enough for one or two American laboratories to control its price. Every capable alternative gives enterprises more leverage, encourages portability and reduces the strategic importance of any single model provider.

Model routing turns that competitive pressure into an operating model.

An enterprise application does not always need the most intelligent model available. It needs the least expensive adequate model for the task at hand. A frontier model might be appropriate for difficult reasoning, while a smaller specialized model handles routine classification, summarization or retrieval. An orchestration layer can make that decision invisibly and route each request accordingly.

Once that happens, the durable power moves away from the individual model and toward the layer deciding which model gets called. The company that controls the workflow, proprietary data, customer interface and resulting action owns the more valuable relationship. OpenAI or Anthropic may provide the intelligence without controlling the product in which that intelligence is used or capturing most of the profit it generates.

The model becomes a component rather than the product.

None of this means OpenAI or Anthropic will disappear. The indispensability trap is not a prediction of bankruptcy. Both companies could continue growing rapidly. Their models could process extraordinary volumes. They could remain technology leaders and rank among the most valuable enterprises in the world.

But technological importance does not guarantee proportional economic returns.

Cheaper intelligence could cause AI usage to explode. Companies may apply it to far more workflows, products and decisions than they would at today’s prices. Total demand for compute could continue growing even as the cost of performing an individual task falls. AI could create trillions of dollars in new economic activity while model pricing declines and profit margins migrate into the applications and companies using it.

OpenAI and Anthropic can become more important to the economy while capturing a smaller percentage of the value the economy creates. They can succeed by almost every conventional measure and still fail to justify the fortunes investors expect from them.

This is where Eisman’s Achilles’ heel analogy falls short. OpenAI and Anthropic are not the Achilles’ heel of AI. The belief that AI must remain dependent on them is.

If one of the labs suddenly collapsed, the short-term disruption could be substantial. Cloud providers could lose revenue, infrastructure commitments could be reconsidered and confidence in AI spending could decline. But AI itself would not disappear. The incentives to replace the failed provider would be overwhelming. Capital, talent and customers would move to competing models and alternative platforms.

The more important transition will probably be less dramatic. Instead of a Lehman-style collapse, we may see the steady commoditization of model intelligence. Prices will fall. Performance differences will narrow. Enterprises will adopt multi-model architectures. Open models will improve. Orchestration platforms will shift workloads among providers. AI usage will grow while the laboratories’ share of the resulting profits declines.

That would not represent the failure of AI. It would represent AI becoming mature enough to escape dependence on its pioneers.

The railroads did not fail because commerce grew larger than the railroad companies. Fiber did not fail because the internet’s greatest fortunes went to companies built on top of it. Those networks fulfilled their economic purpose by becoming widely available infrastructure on which others could build.

OpenAI and Anthropic may be moving toward the same destiny. They are laying the tracks and burying the fiber for an intelligence economy. Their work may ultimately touch nearly every business, profession and product. But the companies that use abundant intelligence to reshape health care, manufacturing, finance, cybersecurity, software development and other industries may capture far more value than the companies supplying the underlying models.

AI will succeed beyond our expectations. OpenAI and Anthropic may be remembered as the companies that made that success possible. But the more completely AI succeeds, the less the world will depend on either one—and the greater the fortunes built on top of their breakthrough will become.

That is not AI’s Achilles’ heel.

That is the indispensability trap.

The Indispensability Trap: How Becoming Essential Caps the Fortune—From the Railroads to AI is Alan Shimel’s forthcoming book examining why the technologies that become essential to an economy often create their greatest fortunes for the companies built on top of them rather than for the pioneers who supplied the foundational infrastructure. The book is scheduled for publication in September 2026.