The Curve Always Bends: Why AI Won't Deliver Infinite Exponential Growth

Everyone selling you the AI future draws the same chart: a line shooting straight up, forever.

AGI arrives. Then ASI. Then exponential growth. Then prosperity so ridiculous it barely fits on the slide. The story is emotionally satisfying. It flatters our sense of living through something unprecedented. And it completely ignores how the real world actually behaves.

Here's a different take, grounded in something unfashionable: history.

That "exponential" curve everyone keeps pointing to? It's not what they think it is. What we're actually looking at is the steep middle portion of an S-curve. The early adopters have bought in. Growth is accelerating. Everything feels limitless.

But the curve always bends.

The more accurate model isn't the exponential. It's the logistic curve: slow start, rapid takeoff, then saturation as constraints kick in. Or as someone once put it more bluntly: things continue until they can't.

History Is Littered With "Exponentials" That Weren't

If you want to evaluate the AI prosperity narrative honestly, don't argue ideology. Argue shape. Look at what actually happened to technologies that once seemed destined to grow forever.

Railroads seemed utterly unstoppable in the late 1800s. Track mileage was doubling. Fortunes were being made. Entire industries were restructuring around this new infrastructure. The railroad men of 1870 flat-out couldn't conceive of saturation. Then they hit it. Profitable routes got exhausted. Overbuilding led to financial crashes. By 1916, U.S. rail mileage peaked and began declining. The revolution became a utility.

Electrification followed the same arc. Factory productivity gains from electric motors were extraordinary in the early 20th century. But by mid-century, everyone who could be electrified was. The infrastructure revolution was complete. Further growth required entirely new applications, not more of the same underlying transformation.

Moore's Law is perhaps the most relevant example, since it's the foundation of the current AI boom. For fifty years, transistor density doubled roughly every two years. This trained an entire generation to think the future is always faster, forever. Then physical limits intervened. Dennard scaling ended around 2006. Clock speeds plateaued. We're now approaching atomic scales where quantum tunneling and heat dissipation create massive resistance. The cost per transistor is no longer falling as fast. The industry has shifted tactics to multicore processors, specialized accelerators, and clever architectures. The curve is visibly bending.

The Green Revolution in agriculture offers another sobering case. High-yield crops, synthetic fertilizers, and irrigation caused food output to surge from the 1940s through the 1980s, preventing predicted famines and enabling global population to more than double. But yields have since plateaued in many regions. Soil degradation, water limits, and diminishing returns on fertilizers have created new ceilings. Biology has hard limits on how fast a plant can metabolize sunlight.

Global population itself followed this pattern. Post-industrial medicine and nutrition caused human numbers to spike vertically. Yet we didn't keep breeding exponentially. Cultural and economic pressures acted as braking mechanisms. We're now heading toward a peak around 10-11 billion, followed by likely decline. The "population explosion" turned out to be a population S-curve.

The pattern is overwhelming. Every claimed exponential in history, upon completion, has been revealed as logistic. The question isn't whether there's a ceiling. It's where the ceiling sits and what causes it.

The Counter-Argument, Fairly Stated

To be fair, the techno-optimist response to all this has some force. The argument goes like this: ASI could be recursively self-improving. It could solve the constraints themselves. Invent new energy sources. Design novel materials. Discover new physics. In this view, intelligence is the ultimate resource, and sufficient intelligence can escape any carrying capacity.

Sure, that's conceivable. But it requires strong assumptions. It assumes that all constraints are technical rather than fundamental. It assumes solutions can be implemented faster than constraints bind. It assumes the AI remains aligned and controllable throughout the process. Each of these assumptions deserves skepticism.

More importantly, even if ASI generates brilliant solutions instantly, physical construction has irreducible time constants. Semiconductor fabs take three to five years to build. Power plants take five to fifteen years from permitting through grid integration. Moving to space takes years as well. Workforce retraining operates on five-ten year scales. Supply chains reconfigure over years, not months.

ASI can optimize within these constraints. It can improve upon them. It cannot abolish them. The digital-physical interface is a fundamental bottleneck.

The Walls That Will Bend the Curve

Even granting that superintelligent AI arrives and is transformatively capable, multiple constraint categories could flatten the growth trajectory. None of these require exotic assumptions. They're normal-world limits that apply regardless of how smart the machine is.

Energy and thermodynamics slam into the first wall. Intelligence is physical. It requires energy to process information and mechanisms to dissipate the resulting heat. Current data centers are already stressing regional power grids. Scaling to superintelligence implies computational needs rivaling small nations. Even with efficiency gains, capability requires scale. Unless fusion or some other breakthrough arrives alongside ASI, we'll be energy-constrained. You cannot compute faster than you can cool the hardware. There's even a theoretical floor: Landauer's principle sets a thermodynamic minimum for computation. We're not close to it yet, but it's a hard wall.

Physical-world throughput creates the second bottleneck. Most GDP is still stubbornly tied to atoms: housing, transportation, healthcare delivery, food systems, energy infrastructure. ASI might design a perfect fusion reactor in three seconds. We still have to mine the materials, pour the concrete, smelt the steel, and deal with local zoning permits. You can accelerate blueprints far faster than you can accelerate construction. That's precisely how an exponential becomes an S-curve: deployment hits throughput limits.

Diminishing returns on intelligence present a subtler but equally important constraint. There's a prevailing assumption that more intelligence always equals proportional increases in capability. This is likely false. Scientific discovery often gets harder as you progress, not easier. Newton could discover calculus with pen and paper. Discovering the Higgs boson required the Large Hadron Collider, billions of dollars, and thousands of researchers. Some problems are measurement-limited, requiring experiments that run in real time regardless of how smart you are. Some are complexity-limited: biology and medicine aren't just puzzles of intelligence but problems of noisy data and intricate feedback loops. An ASI might be a thousand times smarter than any human, but if the experiment still takes six months to grow cells in a petri dish, progress is capped by biology, not IQ.

Resource scarcity will bite hard eventually. Advanced technology requires rare earths, platinum-group metals, ultra-pure silicon, specialized chemicals. Mining throughput has physical limits. Supply chains are geographically concentrated and geopolitically fragile. Even perfect recycling can't overcome stock limitations for rapid expansion.

Institutional drag is perhaps the most underappreciated constraint. Deployment is where the real friction lives. Organizations, laws, and social norms evolve slowly. Rapid technical capability bumps against slow institutional adaptation. If ASI raises systemic risks around finance, cybersecurity, or critical infrastructure, societies will impose speed limits. Sometimes those limits will be smart. Sometimes they'll be bureaucratic. Either way, the more powerful the capability, the stronger the incentive to regulate it.

Market saturation provides another ceiling. If ASI enables radical abundance in material goods, demand growth slows. You can only use so many cars, houses, or appliances. Growth shifts to positional goods, experiences, and status competition, which are zero-sum by definition. Even in a post-scarcity world, additional productivity may not translate to proportional economic growth if measured in consumption. People have limited time, attention, and biological needs.

Geopolitical competition could eat a substantial fraction of the gains. A serious ASI capability isn't just a productivity tool. It's an intelligence tool, an offense-defense tool, a strategic advantage. That invites arms-race dynamics, restrictive export controls, security-driven fragmentation. Resources get diverted to defense, redundancy, and resilience. This is one of the most under-discussed prosperity killers: huge potential output, but a growing slice burned on security competition.

The Shape of Things to Come

So what does the realistic trajectory look like?

Look, the best model isn't "ASI flips a switch and GDP goes vertical." The better model is that ASI triggers one or more giant S-curves. Those S-curves overlap enough to look exponential for a while. Then they bend as constraints become binding.

This can still be a big boom. Perhaps the biggest economic transition in human history. It just won't be a frictionless singularity of prosperity.

Think of it as stacked S-curves rather than runaway exponential growth. Individual technologies diffuse along logistic arcs, bounded by adoption limits. The macroeconomy can look exponential for long stretches because it's often a stack of these curves: one wave saturating as the next ramps up. But "exponential forever" requires something much stronger. It requires no binding constraints, no regime changes, no saturation effects, no institutional drag, no physical-world bottlenecks.

History doesn't do "no binding constraints."

The strategic question isn't whether ASI will be powerful. It almost certainly will be. The question is which bottleneck bites first, and how hard. That determines whether the next era brings broad abundance, uneven abundance, stagnation alongside high capability, or a security-dominated world where gains get consumed by the costs of keeping systems stable.

A More Grounded Optimism

None of this is an argument for pessimism. The logistic curve can still reach impressive heights. ASI could produce transformative gains, possibly larger than any previous technological transition. The ceiling may be dramatically higher than what we've seen before.

But the existence of a ceiling is more likely than its absence.

The intellectually honest position is something like this: ASI could deliver extraordinary benefits. Growth will likely accelerate substantially for years or decades. And then the curve will bend as physical, economic, and social constraints assert themselves. We'll experience a powerful boom that eventually matures into something more stable.

That's still a remarkable future. It's just not the infinite exponential on the slide deck.

The burden of proof lies with those claiming this time is different. They've rarely specified why fundamental constraints wouldn't apply. Energy still requires generation. Atoms still need moving. Institutions still adapt slowly. Markets still saturate. Competition still consumes resources.

"Things continue until they can't" is folk wisdom that happens to encode a deep truth about dynamical systems. It applied to railroads. It applied to electrification. It's applying to Moore's Law. There's no particular reason to believe it won't apply to artificial superintelligence.

The curve will bend. It always does.

The real debate isn't about whether to be excited about AI. It's about understanding which bottlenecks will shape the transition, preparing for the constraints that will eventually bind, and building systems robust enough to navigate the inevitable flattening of the curve.

From my vantage in physics and engineering, that's less thrilling than "exponential prosperity forever." But it has the advantage of being grounded in how the world actually works.