On the first day of June, NLW recaps one of the most consequential AI months in a very long time. May opened with the bubble narrative collapsing under $47 billion run rates and closed with corporate America in sticker shock — the hinge month between the AI subsidy era and a world defined by a structural shortage of tokens.
The AI subsidy era is ending. The token scarcity era has begun.
May was the month the agent era's economics caught up with everyone. Usage shifted from seats to tokens, revenue exploded past the bubble doubters, and then the bill arrived: repriced plans, burned budgets, sticker shock across corporate America. The anchor fact of the period ahead is a structural shortage of AI tokens — there simply isn't enough compute to produce all the AI people want to consume — and everything from cheaper models to SpaceX becoming a neocloud is the world's first round of responses.
The first arguably began in late 2025, when Claude Code, Codex, and the Opus 4.5 / GPT 5.2 generation of models came together to unleash the true agent era at the start of the year. The second is the one May made undeniable — and it's about economics, not capability.
Once engineers started pushing agent-written code into production and knowledge workers graduated from vibe coding to building full agentic systems, lab revenue stopped being capped by paid-seat conversion and became a function of raw token consumption. NLW's own personal context portfolio builder racked up a $5,000 API bill in six weeks — more than two years' worth of the $200-a-month Claude Max seat he'd been paying for forever.
OpenAI surged to $30 billion in ARR while Anthropic hit $47 billion annualized — up from $3 billion at the start of 2025. The Atlantic's "So About That AI Bubble" served as the media's mea culpa for Q4's bubble obsession: token-based revenue has no seat-shaped cap, and a lot of people readjusted their priors accordingly.
The New York Times asked "How Anthropic Got So Big So Fast" as the company closed the month with a fundraising round valuing it just under $1 trillion, raced ahead of OpenAI on business adoption per Ramp's statistics, and forecast what would be the first profitable quarter of any big foundation model lab.
The CTO's April admission became the capstone story of the new constraint era. Those token budgets were set before anyone understood Opus 4.5-class agentic usage — nobody planned for what agents would actually consume once they came online.
Yes, Goodhart's law applies, and yes, tokens are an input metric when outputs are all that will matter in the long run. But in a period where no one knows the best way to use these tools, the experimentation those internal leaderboards incentivized was a necessity — even if the consequences were about to rear their ugly head.
An interview full of skepticism about how much value Uber actually got from its burned-through AI budget was flattened into The Information's uncharacteristically oversimplified headline "Uber COO says AI lacks ROI" — and Axios's framing of the moment, "AI sticker shock hits corporate America," stuck.
The $100–$300 max plans were often delivering 10 to 20 times their price in actual token value to power users — $2,000 to $5,000 a month of subsidized usage for $200. Providers can only subsidize that for so long, and once they stop, many companies paying per usage can only afford it for so long. That double bind is the secular shift — and underneath it sits a structural shortage: there simply is not enough compute to produce all the AI people want to consume.
— GitHub, announcing Copilot's shift away from flat-seat billing. Copilot, GitHub wrote, "has evolved from an in-editor assistant into an agentic platform" where a quick chat question and a multi-hour autonomous coding session can cost the user the same amount — and GitHub had been absorbing much of the escalating inference cost behind it.
Google I/O's headline was cheaper plans — Gemini Ultra down to $200 plus a new $100 tier — but usage limits with usage-based billing on top mean a big cost increase for heavy users. Anthropic kept the subsidy inside its own harnesses like Claude Code while moving third-party environments to per-token billing, with financial consequences that caused a bit of an uproar all month.
OpenAI announced a majority-owned deployment company putting forward-deployed engineers inside big clients; Anthropic took a smaller stake in an enterprise AI services firm launched with Blackstone, Hellman & Friedman, and Goldman Sachs, built on the team at Fractional. Same instinct behind both: the agentic capabilities overhang has completely exploded, and companies are going to need a lot of help.
Companies that ran token-maxing experiments are shutting down their internal leaderboards, with Amazon the most recently announced example — not just over leaderboard-gaming concerns, but because the business-model shifts simply made token maxing too expensive.
The next-generation Composer is performing well at a much lower cost than the frontier Opus and GPT models — early evidence that part of the response to token scarcity will be market-based innovation that brings the cost of tokens down without sacrificing performance.
Google gave main-stage lip service at I/O to Gemini 3.5 Flash as the way for enterprises to cut costs, but Artificial Analysis finds it costs about five times as much as Gemini 3 Flash, on both higher token prices and higher token usage. Google's better shot may be Gemma: its smallest, cheapest models are seeing adoption that's outpacing similar Chinese models.
It's China's tried-and-true playbook of artificially keeping prices low for competitive advantage, not a breakthrough in serving costs. In a token shortage, companies around the world will be forced to look past the state-of-the-art OpenAI and Anthropic models toward affordable alternatives — and DeepSeek wants to be right there scooping up that business.
Inference provider Base10 is raising $1 billion at an $11 billion valuation — more than doubling in a single quarter — and OpenRouter, which lets developers automatically toggle between models on cost, efficiency, and performance trade-offs, raised a $113 million Series B to become an AI unicorn.
After his lawsuit against OpenAI was thrown out on statute-of-limitations grounds, Elon teamed up with Anthropic instead: SpaceX AI is letting severely compute-constrained Claude run on Colossus-1 — and then, at least temporarily, on Colossus-2 as well. The realignment carries massive implications for the upcoming SpaceX IPO.
Moving from Grok cheerleader and Altman antagonist to builder of big, ungodly physical infrastructure puts Elon exactly where he's better than just about anyone — and a clear pathway from SpaceX-as-neocloud to future orbital data centers makes the SpaceX IPO make far more sense than investing in an also-ran model.
AI memory stocks surged, with SK Hynix and Micron becoming trillion-dollar companies, and even Meta is talking about becoming a cloud business — if it can sell back its roughly $130 billion of compute at a premium, the big CapEx spend gets significantly de-risked. For the first time in a long time, Meta's AI narrative wasn't freaking investors out.
Six weeks ago, Elon's orbital data center talk was getting sci-fi blank stares. Now Jeff Bezos isn't debating whether they'll happen — he's quibbling with the timeline.
— Riley Brown, on the Claude Opus 4.8 release. Opus 4.8 landed at the end of the month to a telling reaction: the emphasis has shifted from models alone to the harnesses they sit in. Claude Code shipped dynamic workflows the same week, and /goal jumped from Codex into Claude Code to become a real primitive.
— Greg Eisenberg, on why he skipped covering Claude Opus 4.8. 4.6 to 4.7 to 4.8: "Nobody can agree if it's better or worse. The benchmarks say one thing, the vibes say another." The thing that actually matters right now, he argues, is what's happening around the models.
May might go down as the month both CEOs backed off the everyone-loses-their-jobs narrative — Sam with actual effort to articulate why (he'd overestimated how the transformation would happen, much to his delight), Dario more nascently. The reversal opens narrative space for a more nuanced AI policy conversation.
The Bernie and AOC wing is calling for data center moratoriums, while Elizabeth Warren's Time op-ed "Why We Need to Tax AI" argues the opposite: don't stop AI — get our cut. Expect novel taxation structures like token taxes to become a more prominent theme in the months to come.
Beyond cybersecurity concerns around Anthropic's Mythos, part of the administration's opposition to expanding access was reportedly that it knew there was a token shortage and didn't want other people using up tokens it might want for itself. Anthropic says some version of Mythos arrives in the coming weeks.