AI Is Turning Intelligence Into Infrastructure. Who Owns the Pipes?

What happens to the AI economy if the intelligence gets cheaper but everything needed to deliver it stays expensive?

That question is hiding inside a surprisingly ideological fight over open AI.

In July, Dean Ball, OpenAI’s head of strategic futures, posted on X that open-weight models could deter further AI capital spending and that one possible endpoint of an open-weight-dominated market was what he called “full AI communism”: AI provided by the state as “digital public infrastructure.” TechCrunch subsequently reported the remarks in its coverage of Moonshot AI’s Kimi K3. These were Ball’s arguments, made in his own name; this article does not treat them as an OpenAI corporate position.

Ball was pointing at a genuine economic tension.

If increasingly capable AI models become cheaper and easier to obtain, companies may find it harder to charge a premium simply for access to a model. But building and running advanced AI still depends on chips, data centers, electricity, networking, cloud capacity, engineering, data, and capital.

AI could therefore make one part of intelligence cheaper without making the infrastructure underneath it abundant.

And that raises a more concrete question than capitalism versus communism: If the model layer becomes easier to access, where does the economic scarcity—and therefore the value—move next?

The “AI communism” argument starts with who funds frontier AI

Kimi K3 is a useful example.

Moonshot AI describes K3 as an open-weight model and has released its model weights under the Kimi K3 License.

That terminology matters.

The Open Source Initiative’s definition of open-source AI requires more than downloadable weights. It includes the freedoms to use, study, modify, and share a system, plus access to the code and data information needed to meaningfully modify it. A model can therefore make its weights available without satisfying the full open-source definition.

Even open weights alone can change the market.

Organizations with sufficient computing resources can deploy a model outside the developer’s own API. Other companies can offer inference services around the same weights. Developers may adapt or fine-tune the model within the terms of its license.

That creates more competition over how access to a model is sold and delivered.

Ball’s argument goes further. He argued that open-weight models could deter further AI capital spending and suggested that a market dominated by them could push advanced AI toward state-supported infrastructure. (TechCrunch)

His prediction is debatable. The underlying financing question is not.

Frontier AI development requires enormous investment. If the ability to monetize the model itself falls, investors and companies will look for returns elsewhere—or reconsider how much they are willing to spend.

Open models redistribute some control, not all of it

Making weights available does not make the rest of the AI stack free.

Kimi K3 itself illustrates the distinction. Moonshot describes it as a 2.8-trillion-parameter model. Its weights are available, but operating a system at that scale still demands substantial computing resources. (Moonshot AI)

So openness at the model layer can reduce dependence on the original model provider without eliminating dependence on hardware, hosting, electricity, technical expertise, or other infrastructure.

China’s open-model strategy also has an industrial dimension.

A March 2026 staff paper from the U.S.-China Economic and Security Review Commission, written by senior policy analyst Ngor Luong, argues that widespread adoption of Chinese open models can reinforce China’s broader AI ecosystem. The paper describes two feedback loops: rapid model adoption and iteration on the digital side, and deployment across manufacturing, logistics, and robotics on the physical side.

That is the analysis of a USCC staff paper, not proof that every open Chinese model inevitably produces industrial dependence.

But it illustrates something broader: giving away or commoditizing one layer of technology can be a competitive strategy if value is captured somewhere else.

The American market is also more complicated than a simple closed-U.S.-versus-open-China split.

Anthropic CEO Dario Amodei recently wrote that the company has never advocated a blanket ban on open-weight models and called models without dangerous capabilities a “public good.” At the same time, he argued for stronger safeguards around models that reach sufficiently dangerous capabilities. (Anthropic)

So the emerging argument is less about choosing “open” or “closed” AI once and for all.

It is about which layers remain controlled, by whom, and under what conditions.

If models get cheaper, value can move elsewhere

Suppose model capability becomes increasingly substitutable for some workloads.

A company might have less pricing power simply because it owns a capable language model.

But someone still owns the data center where that model runs. Someone supplies the accelerators. Someone pays for the electricity and networking. Someone may own proprietary data, a cloud platform, a popular application, or the customer relationship through which the AI becomes useful.

Commoditizing one layer does not necessarily commoditize the entire stack.

Authors of a 2026 IMF Note based on a scenario-planning exercise make a related point. They write that economies of scale in frontier AI increase barriers to entry and market power, while large investments in compute and proprietary datasets can concentrate market power among a relatively small number of dominant firms and hyperscalers. The Note also discusses the possibility of “winner-take-most” dynamics as rents accrue disproportionately to AI-intensive firms and countries.

The publication carries an important qualification: the views in IMF Notes belong to their authors and do not necessarily represent the IMF, its Executive Board, or IMF management.

The useful economic point is narrower than “AI inevitably creates monopolies.”

As one source of scarcity falls, companies compete to control the remaining scarce assets.

The question is where those assets end up.

AI can transform work without eliminating most jobs

Paul Mason takes the capitalism argument further in a recent Guardian opinion essay, asking what happens if AI increasingly performs work associated not only with employees but with professionals, innovators, and entrepreneurs.

It is a provocative argument about where automation might eventually lead.

Current labor evidence supports a more restrained near-term conclusion.

The International Labour Organization’s 2025 update on generative AI and jobs estimates that one in four workers worldwide is in an occupation with some degree of generative-AI exposure. Because most exposed occupations still require human input, the ILO says transformation rather than replacement is the more likely outcome for most jobs.

That distinction changes the economic debate.

Suppose AI allows an accountant to complete in one hour a task that previously required four.

The technology creates three hours of potential time savings.

It does not determine who captures their value.

The employee might handle more clients. The employer might reduce staffing. Customers might receive lower prices. Workers might get shorter hours or higher pay. Profit margins might rise.

Those outcomes can begin with the same technological improvement.

AI affects the productivity frontier. Labor markets, competition, ownership structures, corporate decisions, and public policy influence how the gains are distributed.

Even today’s measured AI gains are uneven

Researchers are already finding uneven patterns of AI use.

A July 2026 IMF working paper used five waves of the Anthropic Economic Index, each based on one million sampled Claude consumer-web conversations, to measure how AI usage is distributed across occupations. Country-level data were available for the later waves. The dataset excludes enterprise API usage, so it should not be treated as a census of global AI activity.

Under assumptions about Claude’s share of AI use, the researchers estimated an annual labor-cost equivalent of roughly $2.7 trillion across an 86-country sample, or 3.4% of those countries’ combined GDP.

That figure is easy to misread.

It is not $2.7 trillion of measured GDP growth, realized company revenue, or income paid to workers. The researchers describe it as an indicative estimate of the labor cost associated with time saved under their methodology.

Their distributional finding may be more useful: in developing economies, AI usage-based value in the dataset was concentrated heavily among relatively small professional groups, while high-income economies tended to show broader occupational distribution.

And again, the publication type matters. IMF Working Papers describe research in progress, and their authors’ conclusions do not necessarily represent IMF management, its Executive Board, or the institution as a whole.

The evidence is therefore better read as an early signal about diffusion than as a final accounting of AI’s economic impact.

Using AI is different from owning the AI value chain

That distinction becomes particularly important for developing economies.

A country can become a heavy user of AI without controlling much of the infrastructure behind it.

Its companies can buy access to foreign models. Its workers can use foreign AI products. Domestic developers can build applications on foreign clouds or imported hardware.

Those uses can produce real productivity gains.

But adoption alone does not tell us where the resulting revenue, margins, intellectual property, or infrastructure investment will accumulate.

A separate 2025 IMF working paper modeled the potential global effects of AI and estimated that growth gains in advanced economies could be more than twice those in low-income countries. Its model attributes the gap partly to differences in AI preparedness, economic exposure, and access to essential data and technologies.

Those are modeled scenarios, not predictions that poorer countries are destined to fall behind. The authors explicitly find that better preparedness and access can mitigate some of the disparity.

The more defensible distinction is this:

Using AI and capturing a large share of AI’s economic returns are not the same thing.

For countries outside the major AI powers, local applications, skills, energy infrastructure, language resources, data, computing access, and domestic businesses can all influence how much value stays in the local economy.

Open-weight models can lower one barrier.

They do not remove the others.

The AI dividend still has to go somewhere

This is where Mason’s political argument becomes useful even if you don’t share his prescriptions.

Mason proposes universal basic services, redistribution, cooperative ownership, and public-interest AI. Those are political choices, not consequences baked into the technology. (The Guardian)

Other paths exist.

Greater competition among model and cloud providers could push prices down. Open standards can reduce vendor lock-in. Companies can invest productivity gains in growth, wages, lower prices, or profit-sharing. Governments can invest in infrastructure, research, workforce transitions, or shared computing resources. Competition enforcement can target excessive concentration where it emerges.

Different countries and companies will choose different mixes.

What AI itself changes is the economic map underneath those decisions.

If machine intelligence becomes cheaper for a growing range of tasks, some forms of scarcity decline. Other constraints—compute, energy, capital, data, distribution, expertise, trust, or access to customers—can become relatively more important.

That is why the capitalism-versus-communism framing is ultimately too narrow.

AI does not contain an economic system inside its architecture.

It changes the economics of what is abundant and what remains scarce.

And if model-layer intelligence becomes more commodity-like, scarcity does not simply disappear.

The economic question becomes where it moves next.