Again and again, general-purpose technologies have performed the same three-act magic trick on the economy: the pledge, the turn, and the prestige. From steam and electricity to the combustion engine and the railroads, the pattern is eerily similar.
It all starts with an early display of breakthrough capabilities. Enthusiasm is high, and hype follows as everyone rushes to extrapolate from the initial applications to a future in which the technology and its first movers will dominate everything. Next, greed sends massive amounts of capital flooding into infrastructure and capacity to conquer the market. Winners are prematurely declared and celebrated as inevitable.
On the coattails of the initial momentum, the entrepreneurs at the helm become arrogant, and investors stop checking assumptions and fundamentals: “This time is different.”
But there is usually a turn, as a crash, a shakeout, or delayed productivity gains bring valuations and expectations back down to earth. Many investors, hurt and bruised by the end of the euphoria, then retreat from the technology completely.
Once the tourists have moved on, the stage is set for the actual prestige: the same irrational buildout that drove early winners into bankruptcy has also left behind the abundant infrastructure needed for the architectural transformation to finally take place. The technology delivers on every promise, but most of the value accrues to the organizations that quietly owned or built the scarce complementary assets that were actually needed to scale it.
The Inevitable Hubris of First Movers
AI is somewhere between the “pledge” and the “turn” phase. Overconfidence is at an all-time high. The pitch from the labs, roughly: our models may be conscious; our lab is the only one that can be trusted with such powerful technology; please regulate us, or terrible, terrible things will happen; we may one day be powerful enough to become the only private company left in the world. Gavin Baker attributed that last claim to Dario Amodei on All-In. Anthropic researcher Sholto Douglas flatly denied it, but Amodei did not directly address it in his response.
Regardless of whether Amodei believes Anthropic could end up controlling most of the economy, his hubris mirrors that of the entrepreneurs who meaningfully shaped general-purpose technologies before him, many of whom mistakenly confused early control over a new technology’s instantiation with the ability to shape its trajectory and capture the value it creates.
Take steam. Boulton & Watt tightly controlled their design under a patent monopoly for decades. But enormous value traveled downstream. In cotton production, yarn prices collapsed and cotton goods accounted for roughly half the value of British domestic exports by 1830. Or electricity, which was so successful that it became a regulated utility, undifferentiated and metered by the kilowatt-hour. Factories captured the upside as electric motors’ share of U.S. factory drive went from under 5% in 1899 to almost 80% in three decades. Over roughly the same period, manufacturing labor productivity growth jumped from 1.5% a year to 5.1%.
The combustion engine? Similar story. More than 700 firms entered the automaking race in the United States, but the industry contracted to the Big Three and a few independents by 1941. The value migrated to its complements: gasoline went from a near-worthless byproduct to 44% of the crude processed by U.S. refineries in 1931, while each additional highway passing through a city reduced that city’s population by an estimated 18%, benefiting the suburbs.
Railways? The similarity is almost too perfect. In 1876, Tom Scott led what was one of America’s most powerful corporations and the world’s largest freight carrier: the Pennsylvania Railroad. The press compared him to Napoleon, and he acted as though owning the rails meant owning the market. But Standard Oil accounted for the majority of the Pennsylvania’s oil traffic.
Tensions began when the Pennsylvania’s affiliated carrier, Empire Transportation, expanded into refining. Rockefeller demanded that Empire withdraw from the new business, but Scott refused, pushing Rockefeller to move his freight to rival railroads. The Pennsylvania countered with a ruinous price war in a desperate effort to attract other shippers to its lines. In the words of one of Scott’s lieutenants: “We paid them large rebates. … In some cases we paid out in rebates more than the whole freight.” Then the Great Railroad Strike of 1877 depleted the Pennsylvania’s treasury, and Empire’s assets were sold to Standard Oil. Within months, Standard was collecting 20 cents a barrel from the Pennsylvania even on crude the railroad carried for Standard Oil’s rivals.
But the Pennsylvania’s story was not unique. After years of overbuilding, financial leverage, and price wars, the Panic of 1893 pushed more than 100 railway companies into receivership. Yet the tracks survived, and the network thrived: estimates for 1890 put the railroads’ annual private return at just 3.5%, against a social return of 48%. Without the rails, U.S. agricultural land would have been worth 60% less.
Investors financed the buildout, but the value accrued to landowners, manufacturers, shippers, and, of course, consumers. Sears is a great example of the era’s “app” layer: Richard Sears started his career as a railroad station agent and built a massive retail business on top of mail catalogs and rails. By 1906, his three-million-square-foot Chicago fulfillment center boasted rail connections to every major trunk line out of the city. Sears’s sales grew from $750,000 in 1895 to $50 million in 1907.
The Value of Complementary Assets
So if the AI labs are the railway companies, who plays Standard Oil and who plays Sears? Standard Oil is any enterprise that controls valuable demand: AI workloads that benefit from proprietary data and ground truth that can be routed elsewhere. Sears represents the application layer: players whose customer relationships, distribution, and automated labor become more valuable as intelligence is commoditized. NVIDIA is currently a scarce upstream supplier to both categories.
In a recent deep dive into the economics of closed versus open AI models, I argue that because AI models are difficult to protect and easy to learn from, the space exhibits a weak appropriability regime. While the main worry to date has been about open weights undermining future investment in AI, the reality is that they would only affect its direction: what gets built, who builds it, and who appropriates the returns.
When appropriability is weak, value does not flow to whoever contributes the most to invention. It flows to whatever constitutes a bottleneck to large-scale deployment. The obvious scarce complements to AI are compute infrastructure and chips, the two Amodei identified as natural sources of concentration if open-weight models were to dominate at the labs’ expense. But control over those assets is not static: labs, hyperscalers, chipmakers, and even sovereign states are moving upstream to secure alternative sources of supply.
In the short run, as supply lags far behind surging inference demand, scarcity may increase concentration and reward whoever can deliver the necessary inputs, on land or in space. Over the long run, however, the same demand signal attracts investment and innovation to disrupt today’s market leaders.
What Amodei conveniently does not mention is that hardware and infrastructure do not follow the same extreme winner-take-all dynamics as software. They can stay concentrated, but their bottlenecks are physical, visible, and investable: high prices invite new global capacity, even if that capacity takes time to build.
On the inference side, the labs will not struggle to monetize SOTA capabilities for as long as they can sustain enough of a lead. But beyond that, they will need either to lock in their customers through more traditional software bundling or to rapidly secure complementary assets of their own. This explains their interest in chips, as well as their push into cyber, where small advantages translate into major differences in customers’ willingness to pay. The same is true for biology and medicine, where a patent can convert a tiny lead into a defensible right to exclude others.
Yet the domains with the steepest demand curves for intelligence also impose a ceiling on private power. There, the marginal buyer is a nation-state, and governments can bypass ordinary markets through procurement, compulsory access, or even nationalization. Ironically, to retain their freedom, the frontier labs need to be successful, but not too successful.
The viability of open weights does not undermine the profit motive; it simply accelerates it in many different directions. As yesterday’s frontier is commoditized, previously uneconomical applications of AI suddenly become viable, giving chipmakers, enterprises, application companies, and governments additional reasons to fund and improve more distributed intelligence “commons.”
But if you zoom out and take the complementary-assets story seriously, you immediately realize that while the labs have been able to collect broad public data cheaply under fair-use arguments, future gains in specialized, currently non-verifiable domains will depend increasingly on proprietary traces, tacit knowledge, systems of record, and unique sources of ground truth and expert decision-making.
Enterprises are increasingly wary of surrendering that information, no matter how capable the labs’ computer-use, memory, and communication tools become. As firms with valuable domain knowledge shift toward owning and controlling their “weights of production,” market leaders in many industries will be able to match or exceed frontier-model performance at a lower cost on the tasks that matter to them by combining open-weight models with the unique information they generate.
The key remaining moats are physical infrastructure, proprietary data and talent, distribution, and the ability to rapidly generate more precise and relevant ground truth through continual friction with the real world. As intelligence diffuses, what stays defensible is not a historical dataset, which can be easily bought to unlock new capabilities, but actual control over a valuable real-world learning loop: the ability not only to deploy AI but also to collect the telemetry needed to verify that the automation ran correctly and improve the system until it can be trusted to support further automation.
Firms that retain the information these loops generate will not only send a decreasing share of tokens to the frontier labs but also protect more of their business from AI disintermediation. The labs are aware of this and are expanding downstream into workflows across finance, health, law, design, and accounting to secure more business. But of course, the further they expand, the more they clash with some of their own customers, generating additional demand for open weights as the escape hatch.
This will likely split the demand for machine intelligence between commodity tokens and value-added ones. The former will be cheap and metered like kilowatt-hours. The latter will secure defensible markups and be increasingly powered by proprietary ground truth.
A World With Many Competing, Specialized Intelligences
For much of the economy, once a model is good enough, additional capabilities will matter far less than deeper integration into a firm’s workflows, lower costs, or greater control. So as open models continuously commoditize the frontier, they will steal meaningful demand, pushing the labs further into whatever can still command a premium.
Rather than centralization, both the underlying economics of AI and the history of general-purpose technology push towards a much more balanced outcome for society. The only force that could drastically reverse or delay that is exactly the type of regulation Amodei is fiercely campaigning for. If regulators set a very high bar for compliance, they will favor the labs with sufficient scale to support a large legal, policy, and safety spend. They may also hurt the thriving open alternative to closed models, for example, by placing a cap on the capabilities open models are allowed to reach. This would turn regulation into the missing complementary asset the technology does not offer on its own. Of course, that does not mean the safety concerns are insincere or unimportant.
If the complementary-assets logic holds as it has in the past, the most likely outcome is still one in which multiple competing intelligences coexist and compete across markets and applications. That diversity would leave society better prepared, more resilient, and, in many domains, safer. Rather than nurturing a few monocultures that could fail in correlated ways, a diverse ecosystem of intelligences would provide us with a variety of tools to deal with the inevitable unknown unknowns of the next few years. In cyber, it is already clear that wider diffusion through open models benefits defense over attack. Other domains such as biology might well be different, but we may be able to more effectively control them by focusing on the complementary assets that stand between an idea and its implementation, such as synthesizers, biomaterials, and related supply chains.
AI, by making intelligence cheap, is possibly the most general-purpose technology we have ever created. If that’s the case, then the chances are high that it will hand the holders of complementary assets their biggest win yet. The labs have laid the tracks on which the machine-intelligence railways can operate and scale, and they are likely to hold the market for frontier intelligence for quite some time. Regardless, value will flow, as in previous cycles, to whoever controls what stays scarce.
Anthropic may well become one of the world’s most valuable companies. But it will never be the only one.
