The Innovation Plateau: Why AI Adoption Is About to Slow Down

Part 1 of a three-post series. Part 2: Coding Was Easy Mode · Part 3: The Token Subsidy Is Ending

Every technology that changed the world looked unstoppable right before it slowed down. AI’s last three years — explosive capability gains, aggressive market penetration, trillion-dollar promises — have convinced a lot of smart people that exponential progress is a permanent feature of the technology. I think the opposite: the mechanics of AI development, inference economics, and physical infrastructure all point toward a period of friction and plateauing adoption. Not collapse — a plateau. And two old frameworks from innovation theory predicted the shape of it.

Two Curves Everyone Forgets

Richard Foster’s S-Curve (1986) models how a technology’s performance responds to invested effort: slow ferment, explosive takeoff, then maturity, where ever-larger amounts of capital buy ever-smaller gains. The transformer era from 2017 to early 2024 was a textbook takeoff. The question is whether we’re now watching the top of the S start to bend — and the early signs say yes. (Note this is not Gartner’s Hype Cycle, which measures sentiment. A technology can be hyped, un-hyped, and re-hyped while its actual performance curve quietly flattens.)

Carlota Perez’s Technological Revolutions and Financial Capital (2002) models the money. Across 250 years and five technology surges — from railways to the internet — every revolution follows the same two-act structure, split by a crisis she calls the Turning Point:

The railway mania of the 1840s and the dot-com bubble both sat exactly where AI sits now: deep in Installation, accelerating toward the Turning Point. Subsidized tokens are the classic Installation-phase tactic — speculative capital forcing adoption before profit becomes urgent.

The Two Load-Bearing Arguments

If the frameworks give the shape, two structural arguments give the mechanism — and each gets its own post in this series.

The software engineering exception. AI’s one frictionless win happened in the single domain that has peer-reviewed training data at scale, telemetry on every keystroke, and a compiler to catch hallucinations deterministically. Those advantages don’t travel to law, finance, or strategy — which means the flagship success story is a poor forecast for everything else. → Part 2: Coding Was Easy Mode

The token economics. Flat-rate pricing was the Uber playbook, and it’s being unwound just as agents multiply token consumption 5–1,000x per task. Unit prices are falling 10x a year and enterprise bills are exploding anyway, because stateless LLMs re-read their entire context on every call and GPU memory limits break the caching that was supposed to fix it. → Part 3: The Token Subsidy Is Ending

Meanwhile, the Scaling Party Winds Down

All of this would be tolerable if models kept getting dramatically better. But the scaling laws that funded the boom are power laws, and power laws on a linear scale look like exponential decay: huge early gains, then diminishing returns. The data wall makes it concrete — a Chinchilla-optimal 1T-parameter model wants ~20 trillion tokens of high-quality text, and the entire high-quality internet is somewhere between 10 and 50 trillion. Synthetic data risks model collapse. And you can feel it in the releases: the GPT-4-to-GPT-5 jump landed with far less shock than the ones before it. If capability gains keep shrinking while agentic costs keep compounding, enterprises will stop paying — that’s the plateau mechanism in one sentence.

Lawyers and Neighbors

Two more brakes, both from outside the industry. The courts: the fair-use question for training data is heading toward dispositive rulings, and Anthropic’s tentative $1.5B settlement with authors (~$3,000 per book) hints at the price of losing. Adverse rulings mean multi-billion-dollar licensing regimes, passed straight into token prices.

And the neighbors: in Q1 2026 alone, US communities blocked or delayed over $130 billion in planned data center construction over water, grid strain, and noise — from Indiana to Arizona, Memphis to Ireland. Capital is fleeing to pre-permitted Nordic sites, but relocation is slow, expensive, and adds latency. Compute cannot scale exponentially if the buildings it lives in can’t get permits.

Where This Leaves Us

The plateau isn’t pessimism; under the current brute-force paradigm it’s arithmetic. Coding’s frictionless adoption was a one-off, built on telemetry and compilers other domains don’t have. The subsidy that made AI feel cheap is being withdrawn exactly as agents multiply consumption. Scaling is hitting data and physics walls, and courts and city councils are adding hard constraints on both ends of the pipeline.

Perez’s framework offers the consolation prize: after the Turning Point comes Deployment — the phase where the technology stops being a speculative frenzy and becomes durable infrastructure. Getting there won’t be about bigger parameter counts. It will be about architectural efficiency, ruthless memory orchestration, small local models, and the unglamorous institutional work of wiring probabilistic systems into deterministic workflows. The interesting part, in other words, is just beginning — it’s the exponential part that’s ending.

Sources & Further Reading

Works cited

• Foster, R. (1986). Innovation: The Attacker’s Advantage. The original S-Curve argument.
• Perez, C. (2002). Technological Revolutions and Financial Capital. Installation, the Turning Point, and Deployment, across five surges of history.
• Hoffmann, J. et al. (2022). “Training Compute-Optimal Large Language Models” — the Chinchilla scaling laws.
• Shumailov, I. et al. (2024). “AI models collapse when trained on recursively generated data”, Nature — the model-collapse result behind the synthetic-data worry.
• Epoch AI. “Will we run out of data?” — the data-wall estimates.

People who cover this beat well

• Stratechery — Ben Thompson on the business strategy of the AI buildout.
• One Useful Thing — Ethan Mollick on what AI adoption actually looks like inside organizations.
• Marcus on AI — Gary Marcus has been arguing the scaling-limits thesis longer than anyone.
• Simon Willison — the best running notebook on what agents and LLM tooling can really do.
• Don’t Worry About the Vase — Zvi Mowshowitz’s exhaustive weekly synthesis of AI developments.