Dario Amodei warns that frontier AI may advance faster than humanity can safely control it. The less discussed risk runs the other way: demand for machine intelligence may grow faster than inference capacity, turning tokens from cheap software output into a scarce economic resource.
Dario Amodei is worried that artificial intelligence may soon become too powerful to control safely. His September proposal to “pace the frontier” is not a call to stop AI. It is a request to slow capability gains enough for evaluators, security work and alignment research to keep up. He argues that within 6–12 months, a sufficiently capable but misaligned swarm could plausibly create a persistent botnet with enormous economic consequences.
Take the warning seriously. Then invert it.
The AI economy may have another problem before intelligence becomes unlimited: we may not be able to manufacture enough of it.
Intelligence Has a Factory Floor
A token looks like software. Economically, it is closer to industrial output. Every useful unit of inference requires accelerators, memory bandwidth, networking, cooling and electricity. Long reasoning chains consume more of all of them: a 2026 Joule study estimated that a roughly 5,000-output-token reasoning query used about thirteen times the energy of a standard query.
That physicality is easy to miss because the product appears instantly on a screen. But power and cooling already cap AI data-center scale. A model can be copied cheaply; serving it millions of times cannot.
Epoch AI estimates that global inference capacity is growing roughly 3.4 times per year. Several imperfect demand proxies suggest token consumption at current prices could be growing closer to tenfold annually. Epoch’s range is deliberately wide, and efficiency gains may relieve the pressure. Still, its conclusion is uncomfortable: a compute crunch is plausible, particularly for the long-context workloads used by coding agents and other autonomous systems.
Agents Turn One Request Into an Economy of Requests
The old demand model was one person asking one chatbot one question. Agents break it.
A single instruction can now trigger planning, search, coding, testing, criticism and revision across multiple model calls. One visible answer may conceal millions of tokens of machine work. That is why the software team is becoming an economic artifact: organizations increasingly substitute computation for coordination.
The consequence is recursive demand. Humans do not merely consume inference; software begins consuming inference on their behalf. Better models may reduce the tokens required for some jobs, but better models also make entirely new jobs economical. Jevons would recognize the problem.
The first political economy of artificial intelligence may not be abundance. It may be allocation.
If frontier inference becomes scarce, the important question changes from what can AI do? to who gets to use the best AI, and for what? A pharmaceutical lab, a cybersecurity team, a hedge fund and a public school would all be bidding—directly or indirectly—for the same underlying stack of silicon, power and bandwidth.
AI Safety Has an Allocation Problem Too
This complicates the familiar argument for slowing capability development.
Amodei’s own proposal contains the complication. He warns that democratic countries cannot slow so much that unpaced rivals overtake them; he explicitly treats advanced chips and compute access as determinants of geopolitical power. A global pause, he writes, would be difficult to verify and tempting to defect from.
The same asymmetry exists between attackers and defenders. If autonomous systems find vulnerabilities and coordinate attacks at machine speed, human-speed defense is not a stable equilibrium. Advanced offensive AI increases the value of advanced defensive AI.
That does not mean “more AI automatically makes us safe.” It means capability, deployment and access are different control surfaces. A powerful model inside a constrained environment is a different risk from a weaker model with credentials, network access and permission to act. Regulation that treats intelligence itself as the only dangerous variable can miss the permissions around it.
Delay Is Not the Neutral Baseline
There is also a cost that safety debates struggle to price: the discoveries not made while development is slowed.
Amodei himself argues that AI could help cure major diseases within five to ten years and accelerate economic growth. That matters because a delay is not merely additional safety research. It is also delayed drug discovery, materials research, tutoring, defensive cybersecurity and productivity—benefits whose magnitude is uncertain, just as catastrophic-risk estimates are uncertain.
The AI utility debate usually asks whether applications can justify enormous infrastructure spending. Scarcity reverses the question. If useful applications proliferate faster than the infrastructure, the issue will not be empty data centers. It will be queues.
That is why the AI boom is already moving beyond Nvidia toward custom silicon, memory, power and the machinery required to serve inference cheaply. The next advantage may belong less to whoever owns the smartest model than to whoever can run it continuously at scale.
The Risk Is Not Either-Or
Amodei could be right about rapidly improving agents and still be wrong about which constraint society encounters most often.
A world of powerful AI can simultaneously face misalignment risk, misuse risk and an acute shortage of affordable inference. Indeed, scarcity could make the safety problem nastier by concentrating the strongest systems among governments and corporations able to secure chips, power and capital.
The useful test is empirical. Watch token prices, quotas, long-context availability, accelerator utilization and the spread between frontier and smaller-model inference. If capacity keeps compounding faster than demand, the scarcity thesis fades. If demand outruns it, intelligence begins behaving like electricity: abundant in theory, rationed at the point where infrastructure runs out.
The AI debate has spent years preparing for limitless machine thought. We should also prepare for intelligence that is valuable, physically constrained and unevenly distributed.
If the existential-risk camp is right, the defining problem will be controlling intelligence.
If the scarcity case is right, it will be deciding who gets it.
Sources
- Dario Amodei, “We Must Pace the Frontier”
- Epoch AI, “Is a compute crunch coming?”
- Epoch AI, AI scaling and compute datasets
- Joule, “Energy use of AI inference, efficiency pathways, and test-time scaling,” 2026.
More in AI & Compute
Sources
Dario Amodei's September 2026 pacing proposal; Epoch AI inference-capacity analysis; Joule research on inference energy; Epoch compute and usage datasets.
