The AI Token Shortage Begins

June 1, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
June 01, 2026

MAIN STORY

The AI Token Shortage Begins (May Monthly Recap)

May 2026 was one of the most consequential months in AI's short history — not because of any single product launch, but because of a fundamental shift in the economic reality underpinning the entire industry. We are exiting the AI subsidy era, where flat-rate max subscriptions quietly transferred enormous value to users at the labs' expense, and entering a token scarcity era defined by structural compute constraints, rising costs, and a scramble to figure out who pays and how. That shift touched everything: business models, enterprise deployment strategies, geopolitics, infrastructure, and even narrative.

AI SHORTAGE SET IN

From Scarcity to Shortage: The New Economic Reality
GPU prices up 2x+ since January. The toilet paper is gone.
The month opened with a growing recognition that the explosion in agentic AI usage — code pushed to production, multi-hour autonomous sessions, entire agents running at scale — had completely blown past the assumptions baked into enterprise AI budgets. Uber became the capstone story twice: first when its CTO revealed the company had burned through its entire 2026 AI budget in four months, then when its COO expressed serious skepticism about whether they'd gotten ROI on it. The Axios piece "AI Sticker Shock Hits Corporate America" captured the mood. Meanwhile the revenue numbers from the labs tell a different story: OpenAI hit $30B ARR, and Anthropic — starting from just $3B at the beginning of 2025 — reached $47B in annualized run rate, with its first profitable quarter anticipated. The Atlantic's "So About That AI Bubble" served as a quiet mea culpa for a media cycle that had badly underestimated what token-based revenue could look like.

The End of the Subsidy Era: Business Model Shifts
The current model is no longer sustainable.
The clearest sign of the transition was the wave of companies moving from flat-seat pricing to usage-based billing. GitHub Copilot led the way, announcing the shift in late April with an unusually candid post explaining that a quick chat and a multi-hour autonomous coding session can no longer cost the user the same amount. Google IO brought nominal price cuts to Gemini plans alongside new usage limits and overage billing — meaning power users will in many cases pay more. Anthropic moved similarly, targeting billing for usage through third-party harnesses specifically. Token maxing experiments — internal leaderboards that encouraged maximum AI consumption as a proxy for adoption — are now being scrapped. Amazon is the most recent example. The incentive structure has reversed.

Implementation Needs: The Labs Move Into Consulting
Closing the AI adoption gap is a ~$100B opportunity.
The agentic capability overhang — the gap between what models can do and what most enterprises are actually extracting — has exploded. Both OpenAI and Anthropic responded this month by moving more directly into enterprise deployment support. OpenAI launched a majority-owned deployment company to embed forward-deployed engineers inside large clients. Anthropic went a different route, partnering with Blackstone, Hellman & Friedman, and Goldman Sachs to back a separate enterprise AI consulting firm (built on the bones of Fractional). The thesis is the same: deploying agents is categorically harder than deploying software, and companies need a lot of help. The forward-deployed engineer is becoming one of the hottest roles in tech, with job postings up over 700% in the past year.

Market Responses: Cursor, Google, DeepSeek
The market is already moving to bring token costs down.
Not all the responses to token scarcity are business model changes — some are genuinely technical. Cursor shipped Composer 2.5, which is performing comparably to Opus 4.7 and GPT-5.5 at meaningfully lower cost, showing that model efficiency innovation can outpace raw compute constraints. Google's Gemma small model series is seeing fast adoption — outpacing comparable Chinese models — as companies look for affordable alternatives. DeepSeek made its 75% price cut on V4 permanent; that's not a cost breakthrough, it's a deliberate land-grab play targeting companies priced out of the frontier models. Price wars, in other words, are coming.

Infrastructure: Everything Going Vertical
The entire AI supply chain is getting repriced.
In a token shortage, compute infrastructure becomes the most valuable thing in the world — and the market is behaving accordingly. Inference provider Base10 raised $1B at an $11B valuation, more than doubling its value in a single quarter. OpenRouter, which lets developers toggle automatically between models based on cost and performance trade-offs, raised a $113M Series B to become a unicorn. AI memory stocks are surging, with SK Hynix and Micron becoming trillion-dollar companies. Even Meta is floating the idea of becoming a cloud business — if they can sell back some of their $130B compute investment at a premium, the CapEx math looks very different.

Elon's Realignment: From Antagonist to Infrastructure Czar
Elon doesn't need the best model. He needs to control the best hardware.
The single most structurally important development of the month may be Elon's repositioning in the AI landscape. His lawsuit against OpenAI was thrown out. More importantly, Grok has never caught up to the leading models, and rather than keep trying, Elon pivoted to something he does better than almost anyone: building massive physical infrastructure. SpaceX announced it would allow Anthropic to use Colossus-1 for additional Claude capacity — significant given how compute-constrained Anthropic has been all year. Then weeks later, Anthropic was also getting access to Colossus-2. In the span of a month, SpaceX became a neocloud, with Elon self-appointing as a kind of compute czar. This reframes the SpaceX IPO entirely: it's not just a rocket company, it's an AI infrastructure play. When Elon started talking about orbital data centers six weeks ago, he got sci-fi blank stares. Now Jeff Bezos is talking about them too — just quibbling about whether 2-3 years is an ambitious timeline.

Government and Mythos: The White House Gets Involved
The US government is now explicitly thinking about token allocation.
The White House's involvement in Anthropic's Mythos release turned out to be about more than cybersecurity. A key part of the opposition to expanding access was that, in a world of token shortage, the government didn't want others consuming tokens it might want to use itself. Trump ultimately postponed signing the AI executive order after last-minute objections — reportedly from David Sacks among others — and separately approved $9 billion for spy agencies to catch up on AI. On the Democratic side, the battle lines are still being drawn: Bernie and AOC are calling for data center moratoriums while Elizabeth Warren is pivoting to taxation, publishing an op-ed in Time making the case for token taxes specifically. That framing is going to be a growing theme.

Public Backlash and Internal Politics
48 data center projects worth $156 billion were blocked last year.
The infrastructure build-out that the token shortage demands is running headlong into a growing public revolt. A Gallup poll found Americans broadly oppose AI data centers in their area. Data center project cancellations are accelerating. Energy bills are rising, water tables near server farms are dropping, and the anti-tech backlash is attracting a law enforcement response — with the FBI warning about "anti-tech extremism" following escalating incidents including an attack on Sam Altman's home.

Harnesses Over Models: The New Priority Stack
Model releases are starting to feel like iPhone releases.
It was a relatively quiet month for new model releases. Claude Opus 4.8 arrived at the end of May, but the reaction was telling: both Riley Brown and Greg Eisenberg essentially said they didn't care about the model unless it came with a Claude Code update. Eisenberg compared it to iPhone releases — incremental improvements where benchmarks say one thing and vibes say another. What actually mattered this month was the harness layer: Claude Code shipping dynamic workflows, and /goal becoming a real primitive that moved from Codex into Claude Code. The shift from "what model is this running?" to "what can this harness do?" is now firmly underway.

Narrative Competition: Sam and Dario's Reversal
OpenAI has pivoted to "AI makes humans more powerful." Anthropic is still catching up.
May may go down as the month both Sam Altman and Dario Amodei figured out it was probably bad for business to keep telling everyone that the thing making them rich was going to destroy jobs and livelihoods. Sam went further, publishing a piece articulating why the evidence had shifted his view — he now believes he overestimated the pace of economic disruption. Dario's reversal is more nascent. Noah Smith put it bluntly: OpenAI has pivoted to augmentation, Anthropic is still messaging replacement. The narrative shift opens up more nuanced policy conversation — which Nathaniel says is welcome. Karpathy joining Anthropic this month also turned heads.