← All Notes

When the levee breaks, mama, you got to move. — Led Zeppelin, 1971

Companies are capping AI token spend. The story is framed as cost control — a sensible CFO response to runaway AI experimentation budgets. Goldman Sachs has limits. KPMG has governance policies. Fortune 500 legal departments are rationing Claude access. The trade press calls it the "tokenpocalypse."

They're not wrong about what's happening. They're wrong about what it means.

Token caps are not cost control. Token caps are the dam.

The water behind it is the most capable cognitive technology in human history, building pressure against a structure built from a mixture of social responsibility, institutional inertia, and a quiet hope that AI won't quite measure up. The dam is holding. For now. The question worth asking — the one nobody in the trade press is asking — is what we know from history about how long it holds, and what the water looks like when it finally moves.

The Shape of the Wave Is Different

Every major technology disruption since the industrial revolution has followed the same S-curve: slow adoption, then an inflection, then rapid saturation. Electricity. The personal computer. The internet. Mobile. Cloud computing.

Technology adoption S-curves across five historical waves

The curves look similar from a distance. Up close, the dams were different every time.

For electrification, the dam was sunk cost. Factory owners had steam infrastructure. Rewiring a factory floor was expensive and disruptive and required the owner to admit that the thing they'd built their operation around was obsolete. The dam held for roughly 38 years — from Edison's Pearl Street Station in 1882 to majority factory electrification around 1920. It didn't break when electric motors became capable. It broke when the arithmetic was undeniable: rewiring cost less than two years of steam savings.

For the personal computer, the dam was the IT department. Mainframe culture meant that computing was centralized, controlled, procured through formal channels. IBM's PC launched in 1981. The dam broke not when the PC became more capable — it was always capable enough — but when VisiCalc and then Lotus 1-2-3 made the ROI argument at the department level so obvious that managers could buy hardware on discretionary budget and bypass IT entirely. The dam broke in eight years.

For the internet, the dam was infrastructure and trust. The web existed. E-commerce worked technically. But the broadband penetration wasn't there, and consumers didn't trust credit cards online. The dot-com bust in 2000 actually reinforced the dam — it gave every skeptic in every boardroom a reason to wait. The true inflection came when broadband crossed 50% household penetration (around 2004) and mobile became ubiquitous. From 0.6% of retail in 1999 to 20% today. The dam held for a decade, then it didn't.

The pattern across all of them: the dam doesn't break when the technology is good enough. It breaks when one of three things happens.

First: the price crossover — when the cost of not adopting exceeds the cost of adopting. For factories, this was the steam savings calculation. For enterprises now, this is starting to look like a 40-55% productivity premium for AI-augmented knowledge workers. That number is in the research. GitHub Copilot: 55% faster code completion. BCG and Harvard Business School study of AI-augmented consultants: 40% better output, 25% faster. These are VisiCalc numbers. The dam registers this as pressure.

Second: the proof-point event — one company goes all-in, eliminates a layer of workers, reports the P&L impact publicly, and doesn't suffer reputational death. Netflix moving its entire infrastructure to AWS in 2016 broke the "you can't trust cloud at scale" objection overnight. That event hasn't happened yet for AI knowledge work displacement. When it does — and it will — every other enterprise will have quarters, not years, to respond.

Third: infrastructure commitment — the sunk cost that locks in the technology. For the internet, it was broadband rollout. For cloud, it was the data centers AWS built that made walking away economically irrational. For AI, this is the moment enterprises start running their own models. Not because running your own models is necessarily better — it's because the capital investment commits you.

The Tokenpocalypse Is the Metering Moment

Here is what's actually happening with token caps, stripped of the cost-control framing: companies are metering a resource they haven't yet decided to commit to.

This is what electricity looked like in 1895. Early electric companies charged per lamp, then per hour. The metering fight — how do you charge for something that flows invisibly and whose value is distributed across everything that depends on it — preceded mass adoption by decades. The resolution wasn't technical. It was economic and political and took longer than anyone expected.

AI token costs are following a trajectory that makes the metering problem temporary.

AI inference cost trajectory: API vs. self-hosted models

Frontier API costs have dropped roughly 45% per year since GPT-4's launch. The cost of running equivalent capability via self-hosted open-source models (Llama-class, at volume) is already an order of magnitude lower than frontier APIs for organizations generating more than 500 million tokens per month. At current trajectories, today's Claude Sonnet pricing reaches parity with today's open-source self-host costs somewhere around 2028.

The "plug in a server" analogy JP raises is instructive but imprecise. The electrical bill for running a GPU cluster is not the limiting factor — a high-end H100 draws about 700 watts, costing roughly $1,800 per year in power. The limiting factor is the hardware itself ($30,000–$80,000 per H100) and the operational expertise to run it. At current prices, self-hosting wins economically at scale. At current capability levels, the open-source models that run cost-effectively on owned hardware lag the frontier by 12–18 months.

That 12–18 month gap is narrowing. Chinese distillation of US frontier models — Alibaba's Qwen so thoroughly trained on Claude outputs that it occasionally identified itself as Claude — is accelerating open-source capability growth regardless of what Anthropic or OpenAI want. The gap between frontier and self-hostable is the dam's structural weakness.

What Makes This Dam Different

In every prior wave, the dam was made of technology.

The internet wasn't ready for mass retail until broadband existed. The smartphone wasn't useful for enterprise until 4G made mobile data reliable. The dam was physical, and it eroded gradually as infrastructure was built.

This dam is made of management decisions.

The employers capping tokens are not doing so because AI doesn't work. It works. They are doing so because of a deliberate — or at least half-deliberate — choice not to let it work at full speed. The motives, as JP correctly identifies, are mixed. There's genuine social responsibility: the executives signing off on token caps are the same executives who will eventually sign off on the headcount reductions, and they're not eager to be first. There's genuine uncertainty: AI reliability in high-stakes professional contexts is good but not perfect, and the liability for a wrong answer in legal or medical or financial work is real.

Both of those are rational. Neither of them is permanent.

Management-built dams are more brittle than technology-built ones. A broadband infrastructure barrier erodes at the speed of capital investment — slowly, predictably, over years. A management decision reverses in quarters when the competitive pressure changes. And the competitive pressure changes the moment the first proof-point event occurs: the first company to go all-in, report the financial results, and demonstrate that the reputational risk of being seen as an AI-displacement pioneer is outweighed by the financial benefit of moving first.

That company does not yet exist publicly. It is being built right now, inside the budget and strategy sessions of organizations you've heard of. When it surfaces, the dam doesn't crack. It fails.

The Indicators to Watch

If you want to know when the levee breaks, these are the signals:

The productivity premium crossing 50%. The research is in the 40–55% range for knowledge work augmentation. When it crosses 50% consistently and is visible in earnings — when companies that use AI intensively are demonstrably outperforming companies that don't on productivity metrics — the price crossover moment has arrived. The arithmetic works in every spreadsheet in every boardroom simultaneously.

The first public all-in event. Watch for a well-capitalized company announcing significant headcount reduction in a knowledge-work category — legal, analytics, content, customer service at the professional tier — attributing it explicitly to AI substitution, and reporting the margin impact in a way that forces analyst coverage. This is the Netflix moment. It could happen this year. It could be two years. When it happens, the response time for competitors is measured in quarters.

Enterprise self-hosting adoption. When major organizations start building GPU infrastructure for model inference rather than buying API access, they're making an irreversible capital commitment. This is the "rewiring the factory floor" moment — the sunk cost that tells you the organization has decided. Watch procurement announcements, data center buildouts, and the open-source fine-tuning ecosystem for enterprise-scale adoption signals.

The generation entering the workforce. The cohort now in college learned to write with AI. They don't have a non-AI workflow. When they reach mid-career — five to seven years from now — they set the productivity baseline for their generation, and every organization competing for talent has to meet it. This is a slow signal but an inexorable one.

Regulatory calendar. Governments regulate disruption after it happens, not before. When US federal legislation on AI-driven employment displacement starts moving through committee — not just being introduced, but actually moving — the dam has already broken and the government is trying to manage the flood. The EU AI Act is already live. Watch what happens when it meets its first high-profile enforcement case.

The Water

The thing about the Zeppelin lyric is what comes after the levee breaks.

When the levee breaks, mama, you got to move.

The people downstream of the dam aren't warned in advance. The dam holds, and holds, and holds — and then the pressure finds the crack, and the water moves faster than anyone prepared for. The towns downstream had time to see it coming if they'd been watching the right indicators. Most weren't.

The workers in the knowledge economy are the towns downstream. The dam is holding. The water — the productivity premium, the cost trajectory, the open-source capability catch-up, the management decisions reversing one by one — is building.

The token cap is not the story. The token cap is the sound of the dam creaking.

— J.P. Howlett

The Econolypse series: Part 2 · Part 3 · Part 4: Tell Me I'm Wrong · Transitional Tech

Related: You Ain't Seen Nothing Yet — the two doors: post-scarcity or neo-feudalism. The tidal wave is the same wave.

Sources