Anthropic Can Model AI’s Economy. It Can’t Yet Model Who Gets the Intelligence.

Anthropic has released one of the more serious attempts yet to answer a question everyone seems to have an opinion about: What happens to the economy if AI becomes capable of doing a large portion of human knowledge work?

Its Economic Scenarios for Transformative AI is substantially better than the usual collection of breathless predictions about millions of jobs disappearing—or equally unsupported assurances that technology always creates more jobs than it destroys. Anthropic instead builds an economic model, which is both the strength of the work and its greatest weakness.

The model gives us a useful framework for thinking about AI. But it also risks creating a false sense of precision around a transformation that may change the assumptions underneath the model itself. And there is another question hiding beneath nearly all of these projections: Who will actually be able to afford the most capable AI?

That question may turn out to be just as important as what AI can do.

Jobs Aren't Jobs

Anthropic starts in the right place by treating occupations as collections of tasks rather than assuming that a job is simply automated or not automated. A nurse might draw blood, talk with patients, enter information into a chart, schedule follow-up visits, order supplies and review treatment plans.

AI may automate some of those activities, augment others, leave some essentially untouched and create entirely new tasks. That is far more realistic than asking whether "AI will replace nurses." The same applies to programmers, lawyers, accountants, designers and engineers.

A programmer does not disappear because an AI can write a function. What changes is how much human labor is required to produce a given amount of software.

Anthropic then asks what happens as AI becomes capable of performing progressively larger portions of cognitive work. Its scenarios range from relatively familiar technological change to something unprecedented. In the modest scenario, AI has an effect roughly comparable to the internet. In the substantial scenario, AI can perform about half of knowledge work by 2030, although adoption remains incomplete. In the extreme scenario, AI becomes more productive than humans at most knowledge-work tasks and performs nearly all of them autonomously.

This is a useful way to frame the problem. But the farther the model moves from today's economy, the more heavily its conclusions depend on assumptions that are extremely difficult to estimate.

The Model Knows How AI Replaces Work Better Than It Knows How AI Creates Work

The hardest variable in this entire exercise may be the creation of new tasks.

Economic history is full of jobs that would have been nearly impossible to anticipate before the technologies that created them existed. Someone living in 1900 could have understood that automobiles might reduce demand for some horse-related occupations. Predicting semiconductor engineers, app developers, cybersecurity analysts, cloud architects, drone operators and digital marketers would have been considerably harder.

Those jobs were not simply replacements for existing tasks. They emerged because technology created entirely new products, markets, institutions and behaviors.

Anthropic acknowledges that new tasks matter and explicitly includes them in the model. But there is an unavoidable problem: you cannot count the jobs created by industries that have not been invented yet.

That is not a minor parameter problem. It may be one of the largest uncertainties in the entire exercise.

AI makes that uncertainty even greater because it may not merely automate existing production. It may accelerate the invention of new products and industries. If AI increases the rate of scientific discovery, software development, engineering and entrepreneurship, then it could also increase the rate at which entirely new forms of human work appear.

Anthropic itself acknowledges this limitation. Some of its external reviewers argued that the model may underestimate how much AI accelerates technological progress.

Productivity Is Not Demand

Suppose AI allows one programmer to produce ten times as much software. Does the software industry therefore need 90% fewer programmers?

That conclusion only follows if society wants roughly the same amount of software.

What happens if software becomes so inexpensive to create that we consume fifty times as much of it? The employment outcome becomes much less obvious.

We have seen this repeatedly with computing. When computing became vastly cheaper, we did not simply buy the same amount of computing for less money. We put computers into cars, thermostats, phones, watches, televisions, factories, medical equipment and billions of other devices.

AI-generated software could follow a similar pattern. Businesses that could never justify building custom applications may build dozens of them. Individuals may have software written specifically for them. Small organizations may operate systems that once required enterprise IT departments.

The price of producing something affects how much of it people demand.

That matters because AI is not simply another machine that replaces a production task. It is increasingly becoming a technology for producing other technology, and that feedback loop makes traditional automation comparisons much harder.

AI May Change the Definition of a Company

Anthropic's model largely considers workers moving among occupations inside a recognizable economy. But AI may also alter the structure of the firm itself.

A small company can already use AI for software development, research, design, documentation, marketing, customer service and administrative work. The more interesting question may therefore not be how many employees a large corporation will need, but how many large-corporation capabilities suddenly become available to a three-person company.

That distinction matters enormously.

If AI primarily allows today's corporations to produce the same goods and services with fewer workers, then capital captures much of the benefit. But if AI enables millions of people to create businesses that previously required hundreds of employees and millions of dollars in capital, the outcome looks very different.

A person equipped with powerful AI is not necessarily an unemployed worker. That person may instead become a newly viable economic enterprise.

Traditional labor-versus-capital models have difficulty describing someone who simultaneously becomes worker, manager and owner of the productive machinery.

There Is Another Kind of Capital Now

Anthropic's most important conclusion may still be one of its most convincing: the distribution of AI's gains could matter more than the absolute amount of growth.

Anthropic begins with roughly 60% of economic output flowing to labor and 40% to capital. In its substantial scenario, labor's share falls to about 56%. In the extreme scenario, it falls to approximately 45%, while capital receives almost 55%.

That would represent an enormous shift.

But AI introduces an unusual complication. Historically, capital meant factories, industrial machinery, land, buildings and financial assets. A powerful AI system may instead be productive capital that someone accesses from a laptop.

That sounds enormously democratizing. A consultant could possess capabilities once available only to a corporation. A five-person company could potentially compete with an organization employing hundreds.

But there is a condition attached: the small company has to be able to afford the intelligence.

That is where my own thinking about this changed.

My Wake-Up Call: How Much Intelligence Can I Afford?

I recently experienced a small version of this problem myself.

My individual OpenAI account was offered access to a newer model, Astra. I assumed, almost automatically, that the next generation of AI would simply mean greater capability. What I had not seriously considered was what the next generation of AI might mean to my pocketbook.

In the work I was doing, the newer model consumed my available usage close to an order of magnitude faster than what I had become accustomed to. Suddenly the decision was no longer simply, Which model works best?

It became: How much intelligence can I afford?

I could use the newer model and spend substantially more. I could reserve it only for problems where I believed the additional capability justified the cost. Or I could continue using an older model.

That was a wake-up call because I had unconsciously assumed that AI would follow the familiar trajectory of personal computing: the latest technology arrives expensive, becomes commonplace, and eventually gives everyone access to enormous capability for relatively little money.

That may still happen over the long term. But it does not mean everyone will have affordable access to the frontier at every stage along the way.

If the capability difference between generations of AI becomes economically meaningful, then we may be creating another kind of digital divide. The important question is no longer simply who has AI, but who can afford to continuously use the best AI.

Technical Access Is Not Economic Access

Imagine two companies that both say they use AI.

One is a small firm whose employees carefully decide when to invoke an expensive frontier model because every difficult problem consumes a meaningful portion of their monthly budget. The other is a billion-dollar corporation running thousands of frontier agents continuously across software development, research, sales, operations and product design.

Technically, both have access to AI. Economically, they do not have remotely the same thing.

A large corporation can justify spending millions on AI if that expenditure eliminates tens of millions in labor costs or produces hundreds of millions in new revenue. An independent developer or small business has a much harder ceiling.

That difference can create a powerful feedback loop:

• Large companies can afford more capable AI.
• More capable AI increases productivity.
• Greater productivity produces additional resources.
• Those resources buy still more compute and intelligence.
• Smaller competitors risk falling farther behind.

If that occurs, AI does not merely automate labor. It amplifies existing differences in access to capital.

Intelligence May Become a Metered Resource

We have traditionally thought about software as something we purchase. You bought a copy of a program or paid a relatively predictable monthly subscription.

AI increasingly behaves differently. We are beginning to consume intelligence.

More difficult reasoning requires more computation. More capable models may require more resources. Agents can operate for extended periods while performing hundreds or thousands of intermediate steps. The economics start to resemble a utility more than a conventional software license.

That leads to a strange possible future in which everyone technically has access to extraordinary intelligence, but the quantity and quality of intelligence they can afford differs dramatically.

A large corporation might purchase thousands of hours of frontier reasoning every day. A small company might carefully budget a few. An individual may decide that a cheaper model is good enough.

All three technically have "AI access." But that phrase would conceal the economically important distinction between them.

The AI Companies Have Their Own Economics

There is another side to this equation that deserves more attention: the companies building the frontier models are not public utilities.

They require staggering amounts of capital. Training and operating these systems requires chips, data centers, power, cooling, networks, engineering talent and continuous infrastructure investment.

The money required to build that infrastructure has to come from somewhere, and the organizations supplying it generally expect a return.

Today, frontier AI companies have major private investors, strategic partners and conventional economic stakeholders. Over time, some of these companies may also enter the public markets. If and when they do, pressure to produce returns will become even more visible, but the underlying economic incentive already exists.

Investors do not provide enormous amounts of capital simply because they hope advanced intelligence eventually becomes free. They expect companies to produce revenue, margins, competitive advantage and returns on invested capital.

That does not mean AI companies will deliberately restrict useful technology. Competition can push prices downward. Open and open-weight models may create alternatives. Hardware and algorithmic improvements may make today's expensive capabilities dramatically cheaper.

But it would be equally unrealistic to assume that the economic objective of frontier AI providers is to make every future generation of intelligence equally affordable to everyone.

Their incentives are more complicated than that.

The Best AI May Become a Premium Economic Input

Suppose a future model makes a software engineer five times more productive than someone using the previous generation. What should access to that model cost?

If the model creates enormous economic value for enterprise customers, basic economics suggests that the provider may be able to capture some portion of that value. Pricing does not necessarily fall immediately toward the cost of electricity, chips and data-center time.

The price may partly reflect the economic value of the intelligence being sold.

That could produce several layers of AI capability:

• inexpensive commodity models,
• capable mass-market models,
• expensive frontier models,
• specialized enterprise systems,
• proprietary systems available only inside particular organizations.

Today's frontier model may eventually become tomorrow's cheap model. But if the frontier keeps moving, the largest organizations may continually operate with tomorrow's intelligence while smaller organizations work with yesterday's.

Competition may prevent that outcome. Open models may prevent it. Falling inference costs may prevent it.

But we should not simply assume it away.

AI Could Democratize Capital — Or Concentrate It

This leaves us with two very different possible futures.

In one, AI becomes inexpensive, abundant and widely distributed. A small company gains essentially the same cognitive machinery available to a global corporation. Individuals can start sophisticated businesses with very little money, and small firms gain capabilities previously limited to massive enterprises.

That would be one of the greatest democratizations of productive capital in history.

In the other future, frontier capability remains expensive. The largest companies buy enormous quantities of the best models, use them to become still more productive, and then use the resulting financial advantage to buy even more advanced intelligence.

In that world, AI does not democratize capital. It becomes one of the most powerful mechanisms for concentrating it.

The distinction between those futures may depend on something that sounds mundane compared with discussions about artificial general intelligence:

the price of inference.

The Model Stops Just Before Things Get Really Weird

Anthropic's horizon is 2030, which is understandable because forecasting farther becomes increasingly speculative.

But its extreme scenario contains an assumption that makes even a four-year economic model difficult to interpret: recursively self-improving AI.

Once that possibility is introduced, conventional economic extrapolation becomes much more questionable. If AI materially accelerates materials science, drug discovery, robotics, energy technology, manufacturing, transportation, software development and scientific research, then we are no longer merely automating today's task list.

We are creating tomorrow's task list faster.

That creates a fundamental modeling problem. The model is trying to predict the future economy while simultaneously introducing a technology capable of changing the rate at which that economy invents itself.

That is extraordinarily difficult to put into an economic model.

And Anthropic Knows This

To Anthropic's credit, the researchers are unusually explicit about the limitations of their work.

They say the model is not a forecast. It leaves out policy responses, business cycles, financial disruptions, some aggregate-demand effects, advanced robotics and other forces that could materially change the result.

Reviewers also questioned whether heavily AI-exposed occupations would necessarily shrink at all. Some considered the extreme scenario closer to a thought experiment than a conventional economic scenario, while others argued that the model may actually underestimate AI's ability to accelerate technological progress.

That intellectual honesty makes the work much more valuable.

The danger comes when the charts escape their assumptions.

"AI could produce massive knowledge-worker unemployment" is easy to repeat. The more accurate statement is considerably less dramatic: under a particular combination of assumptions about capability, autonomy, adoption, productivity, worker mobility and task creation, the model produces a particular unemployment outcome.

That difference matters.

The Missing Variable May Be Access

Anthropic asks us to consider AI capability, adoption, autonomy, productivity and how quickly displaced workers adjust.

I would add another variable:

Economic access to frontier intelligence.

The question is not merely whether AI exists or whether a model technically has a consumer interface. The question is whether an individual, a five-person company, a mid-sized business and a multinational corporation can afford to deploy roughly comparable intelligence at the scale their work requires.

That variable could dramatically alter the economic outcome.

If frontier intelligence becomes inexpensive and abundant, AI may distribute productive capital to billions of people. If it remains expensive and the best systems are disproportionately consumed by organizations with enormous financial resources, AI may accelerate the shift toward capital that Anthropic's own model already predicts.

The companies with the most capital can purchase the most intelligence. The companies producing that intelligence have investors expecting returns. The price of advanced intelligence therefore becomes part of the mechanism determining who captures the gains from AI.

That deserves to be inside the economic model, not outside it.

The Biggest Unknown Isn't AI

The central question may ultimately be larger than How intelligent will AI become?

It may be:

Who gets to use that intelligence, how much can they afford, and what will they do with it?

If AI is broadly accessible, inexpensive and competitive, small businesses may gain capabilities once reserved for enormous corporations. If frontier intelligence becomes an expensive economic input, the largest organizations could gain a compounding advantage.

If humans use cheap intelligence to create millions of businesses, products, professions and markets that do not yet exist, the labor market may prove far more adaptable than automation models suggest. If corporations primarily use expensive frontier systems to substitute capital for labor, Anthropic's more disruptive distributional scenarios become easier to imagine.

Both forces may occur simultaneously.

That may be the deepest limitation of any attempt to model the economics of transformative AI. We are trying to forecast a technology that may simultaneously change productivity, demand, the creation of new industries, the structure of corporations, the definition of capital, the ownership of productive capability and the market price of intelligence itself.

Anthropic has built a valuable map.

But there is one feature I would add prominently to it:

Not everyone traveling through this new economy may be able to afford the same vehicle.