A Long Read recorded before the Anthropic-government-Fable blowup, NLW argues that: the American economy now is the AI trade, the labs need ever-more token consumption, enterprises are slamming on cost caps — and the only thing that squares the circle is mass-scale AI training.
Only AI training can square what the labs need with what enterprises will pay.
The American economy is now the AI trade — AI investment drives the bulk of GDP growth, and that capital keeps flowing only as long as token consumption keeps rising fast. But enterprises have moved from assisted-AI budgets to agentic-AI reality and are slapping on spending caps. The only force that can give labs their never-ending token growth and give enterprises enough value to lift those caps is mass-scale, high-quality AI training — moving every knowledge worker from assisted to agentic AI. And right now that training market is an abysmal failure.
AI investment isn't a sector story, it's the growth story. In Q1 2026, GDP grew 2% annualized with AI-driven investment contributing roughly 75% of the increase, and AI data centers, hardware and networking hit 1.4% of US GDP — double the 0.7% prior.
St. Louis Fed data suggests AI investment accounted for 39% of marginal GDP growth over the trailing four quarters — bigger than the tech sector's 28% contribution at the height of the dot-com boom. Strip the investment out, and H1 2025 growth would have been a near-standstill 0.1% annualized.
Big tech's 2026 AI CapEx will pass $800 billion — which David Sacks argues could be a 2.5% GDP tailwind this year and 3% next. The build-out was first justified by belief in AI's future; increasingly it's justified specifically by lab revenue growth. That's the contract: as long as token consumption keeps rising fast enough, the capital keeps flowing.
Last year's AI-bubble narrative wasn't really about Sam Altman quotes or the MIT report claiming 95% of pilots failed — underneath it was math. At $20–$200 a month times addressable knowledge-worker seats, the TAM simply wasn't enough to justify trillions in infrastructure.
The shift we've all lived through is from an assisted, seat-based paradigm to an agentic, usage-based one. Per-person economics move from $20–$200 a month to potentially thousands of dollars — and the revenue evidence is clear.
Anthropic surged to a $30 billion annual run rate, then to $47 billion by late May — driven not by new $20–$200 seats but by an insane amount of Claude Code usage. OpenAI's revenue jumped similarly via Codex. Enterprises spending $1M a year on Anthropic went from 500 to more than 1,000 in under two months.
SemiAnalysis estimates the $200/month Claude plan allowed up to ~$8,000 a month of token value, and the max ChatGPT plan up to $14,000 — huge subsidies. As AI consumption rises while infrastructure capacity lags, basic market pricing is kicking in and the subsidy era is giving way to a scarcity era.
GitHub Copilot was one of the first to shift to usage-based billing for agentic sessions. Google I/O lowered some premium prices while adding usage limits that push you to the API, and Anthropic sparked a developer dust-up when it moved all third-party-harness usage to usage-based billing.
2025 assisted-AI budgets are colliding with 2026 agentic-AI reality. Uber blew through its entire AI budget in four months and moved to a $1,500/month per-employee cap; Walmart did something similar. NLW's argument: for the foreseeable future every AI business is, in some form, a token-efficiency business.
Harness companies are routing routine tasks to cheaper models and saving state-of-the-art models for what matters. After Factory launched a model-routing feature in early June, it reported $13 million saved in the first 30 days of private preview.
DeepSeek became Ramp's top trending SaaS vendor, and startups like Linde shifted off expensive American models. Others post-train their own: Cursor's Composer 2.5 hits Opus- and GPT-5-class levels at a tenth of the cost, and Harvey mixes post-trained open models like Kimi K2.6 with Opus for higher performance at lower cost.
The Citadel Securities note showing the Silicon Data LLM token expenditure index rolling over isn't about demand or volume — it tracks average price paid per million tokens, drawn from third-party routers, so it's biased toward cost-seekers. It does show leading companies hunting for cost advantages for the first time.
Already, the relationship between lab revenue (token consumption) and available infrastructure capital is the defining one. Once Anthropic and OpenAI IPO, public-market pressure to show massive token-consumption growth every quarter will be relentless — NVIDIA already gets punished for merely beating estimates by too little.
OpenAI and Anthropic both launched forward-deployed consulting efforts, but value won't come from a few centrally-planned agents. It will come from many diverse knowledge workers building and using agents well — a bottoms-up experimentation that FDE efforts can't deliver on their own.
Over the next 6–12 months, NLW expects dramatic increases in lab investment in enablement, training, and expanding usage depth. Even labs that don't believe everyone should be building agents will be forced to act as if it's true — because they can't hit quarterly token growth by giving leverage to only a select few.
Caps don't just limit spend — they shape what gets attempted, pushing people toward basic productivity use cases and away from the big, unseemly experiments that create the next generation of value. NLW calls it the known-ROI bias: without permission, sandboxes and encouragement, people just do today's work a little faster.
The single thing that serves both the labs' need for token growth and enterprises' need for value, NLW argues, is AI training at mass scale and high quality — moving people from assisted to agentic AI and helping them uncover the use cases that make input costs seem negligible.
An EY survey found only 28% of organizations have empowered employees to actually change business processes with AI. The World Economic Forum notes the half-life of skills is so short that content decays before a course catalog can ship — and agents make it far more complicated than prompt engineering ever was.
— DataCamp, on enterprise AI training. DataCamp surveyed more than 500 enterprise leaders and found video courses are the most common AI training format — but they produce, in DataCamp's words, awareness without confidence and adoption without judgment.
We're shifting from a paradigm where we do things to one where we oversee synthetic intelligences that do them for us. Prompting was a new skill but not a new primitive; managing agents is — and it's far closer to management training than to software training.
He's released three free self-directed programs — the AIDB New Year's program, Claw Camp, and Agent OS (an agentic operating system for the Claude Code / Codex era) — and teases new initiatives with Superintelligent, 'returning to our roots.' He points to Riley Brown's how-to videos and sponsor Section as bright spots, but says it's nowhere near enough.