AI Is Not Ending Scarcity, It Is Moving the Bill

Jun 24, 2026

11 min read

2 views

86%

The strange thing about AI is that it does not feel like a productivity revolution

The headline story says AI makes everything cheaper, faster, and easier. But the lived story inside firms is more unsettling: the work expands, the risks multiply, and the costs do not disappear, they relocate. A company that uses AI to draft decks, analyze data, and automate reporting may indeed get more output per employee, but it also buys a new class of bills: token spend, security infrastructure, compliance overhead, model governance, and often, a return to systems everyone assumed were obsolete.

That is the real tension hiding underneath today’s AI boom. AI does not simply reduce friction. It changes where friction lives. Some of it moves from labor to compute. Some of it moves from manual coordination to system design. Some of it moves from the edge of the organization to the center, where trust, confidentiality, and control suddenly matter more than ever.

Once you see that, a lot of apparently unrelated developments snap together. Hard drive shortages, on-prem AI, prediction markets, surveillance, federal debt, minimum wage pressure, and the rise of AI agents are not separate stories. They are all variations on the same question: what happens when intelligence becomes abundant, but trust, compute, and governance become the scarce resources?

The economy is shifting from labor scarcity to infrastructure scarcity

For decades, the dominant economic problem was labor allocation. Could you hire enough people? Could you get them to the right place? Could you pay them enough? Could you keep them productive? AI is changing that equation, but not by removing scarcity. It is moving scarcity down the stack.

A useful way to think about it is the three bottlenecks model:

Labor bottleneck: Can a human do the work?
Compute bottleneck: Can the machine do the work cheaply enough?
Trust bottleneck: Can the organization allow the machine to do the work safely?

AI is reducing the first bottleneck and exposing the second and third.

The shortage of memory, hard drives, and other physical infrastructure is not a side note. It is a signal that intelligence now consumes industrial supply chains. If AI companies are buying up storage and memory, the economy is telling us that language models are not just software. They are heavy users of material infrastructure. The cloud may feel weightless, but its underlying demands are very physical: chips, drives, energy, cooling, and factories.

That matters because it means the AI economy will not be priced like the old software economy. Software used to scale with a near-zero marginal cost model. AI does not. Every query, every agent, every workflow, every retrieval step, every reasoning pass has a bill attached. And once organizations begin deploying agents broadly, they are forced to ask the most old-fashioned question in management: what is the unit economics of this decision?

Jason Calacanis’s observation about token budgets is more profound than it first sounds. A company that once budgeted for salaries now has to budget for inference. The employee is no longer the only unit of cost. The employee plus their agent stack becomes the unit. When that happens, the workplace stops being organized around headcount and starts being organized around throughput per trust boundary.

The future org chart may not be people first. It may be permissions first.

That is a dramatic shift. The limiting factor is no longer just who can work. It is who can be trusted to work through systems that remember, call APIs, touch documents, update databases, and act on behalf of the firm.

Why on-prem is coming back, and why that matters more than cloud hype

For years, the cloud looked like the final answer to enterprise computing. Why own infrastructure when you can rent it? Why maintain servers when the platform can abstract them away? AI is now forcing a partial reversal, and not because companies have become nostalgic for the old days. They are becoming paranoid in a rational way.

The issue is not only cost, though cost already looks brutal. Some agents are generating token bills that approach or exceed the compensation of the humans they assist. That is an astonishing fact. It means the economics of delegation are becoming visible in real time. A firm may save money by automating repetitive work, but only if it can control the infrastructure layer enough to avoid leaking value to external model providers.

The deeper issue is confidentiality as a systems problem. If employees use public endpoints to analyze strategy documents, customer data, legal materials, pricing models, or trade secrets, the organization may be exporting its own edge into someone else’s training and telemetry pipelines. Even if the vendor promises safety, the risk is no longer hypothetical. In high stakes domains, data leakage is not a bug. It is a business model concern.

That is why on-prem, or at least private provisioned AI, is likely to return in serious enterprises. Not every company will do it. But those with sensitive pricing, legal privilege, proprietary research, or regulated information probably will. The striking part is that AI, the most cloud-friendly technology imaginable in theory, may revive the most unfashionable corporate architecture in practice.

This is not regression. It is optimization under constraint. If using a public model means giving away the company’s private reasoning, then the cloud is not free. It is financed by hidden leakage.

The better mental model is not “cloud versus on-prem.” It is convenience versus control. AI makes convenience irresistible, then makes control mandatory.

AI does not reduce work. It intensifies the demand for higher leverage work

There is a comforting story about AI: it will free people from drudgery, shorten the workweek, and let knowledge workers coast. But the evidence points in the opposite direction. When people get better tools, they usually do not do less. They do more, faster, with a wider scope, and under greater expectation.

That is the paradox of productivity. Productivity gains often become ambition gains.

A worker with AI does not just finish the spreadsheet sooner. They are expected to do the spreadsheet, the presentation, the follow-up analysis, the email, the research memo, the competitor scan, and maybe the first draft of the strategy deck too. The result is not less work, but a thicker layer of work across more hours of the day. AI can make work feel more meaningful because it removes the annoying parts, but that also makes it easier for the organization to ask for more.

This is why the emerging skill is not prompt engineering in the narrow sense. It is agent management. The high-value employee is becoming a designer of work systems, not just a performer of tasks. They know how to structure a project so that humans do the judgment, machines do the repetitive passes, and the output loops back through quality control.

Think of it like a small film production set. In the old model, one person had to do everything, from writing to editing to distribution. In the new model, the director does not replace the crew. The director becomes more important because coordination is now the scarce skill. AI turns many workers into mini producers.

That is why AI adopters can appear superhuman. Not because they are working harder in the old sense, but because they are operating a higher-leverage assembly of cognition. Their advantage is not raw intelligence. It is orchestration.

This also explains why organizations are splitting into two classes of employees:

AI natives, who can compound their output with agents.
AI draggers, who use the tools occasionally but remain manually bound.

The gap between these groups can become enormous, quickly. In that sense, AI is not flattening the workforce. It is steepening it.

Prediction markets and the return of information asymmetry

Prediction markets sound like a different subject, but they are actually part of the same structural shift. They expose a timeless truth: markets are not morally neutral information machines. They are asymmetry engines.

When people trade with unequal information, prices move faster toward truth. That can be socially useful. It can surface corruption, fraud, or hidden events more quickly than official channels. In that sense, prediction markets can behave like crowdsourced surveillance of reality. They can incentivize people to reveal what they know.

But asymmetry has a dark side. If a market is dominated by insiders, sharps, or people with privileged access, it can become a machine that burns through ordinary participants. The platform gets liquidity, but the users without an edge get churned out. The marketplace may remain active, but it becomes structurally extractive.

This is the same tension as enterprise AI. In both cases, the promise is openness, but the actual advantage comes from control over information. In both cases, the system can be socially useful while also being exploitative. And in both cases, the question is not whether asymmetry exists. It always will. The question is who owns it, who benefits from it, and whether the system has guardrails.

Here is the deeper connection: AI increases the speed at which asymmetry can be discovered, deployed, and monetized. Prediction markets monetize informational edges. AI agents help individuals and firms create, exploit, and defend those edges faster than before. This is why the same technology can look emancipatory in one context and predatory in another.

A good rule is this: whenever a new tool improves the speed of inference, ask whether it also improves the speed of exploitation. If the answer is yes, regulation, governance, and architecture will matter more, not less.

The fiscal state is the macro version of the same problem

If AI is the micro story of control, then sovereign debt is the macro story. A government can behave like a company that keeps adding new obligations while assuming future growth will cover the bill. But when the bill compounds faster than growth, you enter a debt death spiral: more debt requires more future servicing, which narrows the room for policy, which increases pressure for bailouts and transfers, which adds more debt.

This is not just a numbers problem. It is a governance problem. And it mirrors the AI dilemma almost perfectly.

At the enterprise level, the question is whether the firm can afford to let its data and workflows leak into open systems. At the public finance level, the question is whether the state can afford to let its liabilities leak upward into the federal balance sheet. In both cases, short-term relief tempts institutions to push costs into a larger shared pool. In both cases, the larger pool eventually notices.

There is a deep symmetry here:

Enterprises face a choice between convenience and confidentiality.
Governments face a choice between political ease and fiscal sustainability.

When private pensions, state obligations, and federal deficits all interact, the system becomes a giant version of the token budget problem. Someone, somewhere, must pay for the compute. Someone must pay for the debt. Someone must absorb the hidden cost. The system can only defer the question for so long.

The reason this matters to the AI conversation is that both stories are about the same thing: the return of metering. In a low-friction world, people pretend scarcity is gone. In reality, scarcity returns as a bill. Whether it is tokens, pensions, memory chips, or federal interest expense, the pattern is the same. A new abundance creates a new metering system.

The real revolution is not automation. It is accountability

The most common mistake is to think AI is mainly about doing more with less. Sometimes it is. But the deeper transformation is that AI forces organizations to make costs visible again.

A human assistant could absorb ambiguity. An AI agent emits logs, traces, calls, invoices, errors, and side effects. A cloud system can be convenient, but it also leaves footprints. A prediction market can reveal hidden truth, but it also exposes who knew what and when. A government can roll liabilities forward, but eventually the arithmetic arrives.

This is why AI may end up being less like a labor-saving device and more like an accounting device. It makes the firm count again.

That has consequences for strategy:

Firms will need token governance the way they once needed headcount planning.
Executives will need to decide which work can live in public models and which must remain private.
Employees will need to become operators of systems, not just performers of tasks.
Regulators will need to decide when information asymmetry creates discovery and when it creates abuse.
Governments will need to recognize that debt is just another form of deferred metering.

The world is not becoming less constrained. It is becoming more precisely constrained.

And that is why the old categories keep failing. Cloud versus on-prem, labor versus automation, insider versus public, austerity versus growth, productivity versus burnout: these are not either or choices anymore. They are design choices in a world where every gain comes with a new boundary to manage.

AI does not abolish scarcity. It reveals where scarcity was hiding.

That is the frame worth keeping.

Key Takeaways

Track the whole bill, not just the obvious savings. AI can reduce labor costs while increasing token spend, security needs, and infrastructure obligations.
Treat confidentiality as an architecture problem. If your most valuable edge lives in sensitive data, public AI endpoints may be too expensive in hidden leakage.
Invest in agent management, not just tool usage. The biggest gains go to people who can orchestrate humans and AI systems together.
Assume productivity gains will increase expectations. Faster work usually means more work, broader scope, and new performance standards.
Look for asymmetry wherever a new market or system emerges. Whether it is prediction markets, AI platforms, or public finance, the key question is who holds the edge and who absorbs the cost.

Conclusion

The conventional story says AI is making intelligence cheap. The better story is that AI is making intelligence measurable. That difference matters. Cheap things disappear into the background. Measurable things become governable, taxable, securable, and politically contentious.

So the true impact of AI may not be a world of effortless abundance. It may be a world where every institution is forced to confront what it has been externalizing for years: labor, trust, risk, and debt. In that sense, AI is less a machine for replacing people than a machine for making hidden costs impossible to ignore.

That is a much bigger revolution than automation.

It is the return of responsibility.