Divya van Mahajan

The AI Token Shortage Begins [AI Monthly Recap]

ai-daily-brief-podcast

AI Monthly Recap: The AI Token Shortage Begins (May 2026)

Overview

This episode of The AI Daily Brief (published June 1, 2026) serves as a retrospective on May 2026, which the host characterizes as one of the most consequential months in recent AI history. The central thesis is that the AI industry is undergoing a structural transition from an AI subsidy era — in which foundation model companies effectively subsidized heavy usage through flat-rate subscriptions — to a token scarcity era, in which demand for AI compute fundamentally outstrips available supply. The host argues this shift has profound implications for business models, enterprise adoption, infrastructure investment, and policy. The speaker is the host of the AI Daily Brief podcast and video channel; no full name is explicitly stated in the transcript.

Source video URL: not provided.

Prerequisites

  • Basic familiarity with large language model (LLM) foundation model companies (OpenAI, Anthropic, Google DeepMind)
  • Understanding of AI API pricing models (seat-based vs. usage/token-based billing)
  • Awareness of agentic AI coding tools (Claude Code, Codex, GitHub Copilot, Cursor)
  • General knowledge of AI infrastructure concepts (data centers, inference compute, neoclouds)
  • Familiarity with key industry figures: Sam Altman (OpenAI CEO), Dario Amodei (Anthropic CEO), Elon Musk (xAI/SpaceX)

Main Points

1. The Second Big AI Transitional Moment of 2026

  • The first transitional moment began in late 2025 (November–December) with the rise of Claude Code, Codex, and next-generation models (Opus 4.5, GPT-5.2), catalyzing the “true agent era” at the start of 2026.
  • Engineers shifted from prototyping and “vibe coding” to deploying agent-created code in production environments.
  • Non-technical users previously using tools like Lovable and Replit migrated to more advanced harnesses such as Claude Code, Codex, OpenClaw, and Hermes.
  • May 2026 represents a second transitional moment: the recognition that agentic AI usage creates real, binding constraints on compute supply.

2. The Revenue Explosion and Collapse of the “AI Bubble” Narrative

  • The most relevant economic unit for AI companies shifted from the seat (monthly subscription) to the token (per-unit API consumption).
  • Personal anecdote illustrates the scale shift: a six-week side project at contextportfolio.ai generated ~$5,000 in API costs, equivalent to more than two years of a $200/month Claude Max subscription.
  • OpenAI surged to $30 billion ARR; Anthropic reached $47 billion annualized run rate, up from $3 billion ARR at the start of 2025.
  • Anthropic closed May with a $65 billion fundraising round valuing the company just under $1 trillion, and anticipated its first profitable quarter — also the first profitable quarter expected for any major foundation model lab.
  • A widely-cited Atlantic article (“So, About That AI Bubble”) effectively served as a media mea culpa, as revenue figures challenged the prior bubble narrative.
  • The bubble argument had centered on whether token sellers could monetize fast enough to justify infrastructure costs; the growth numbers caused many analysts to revise their priors upward.

3. Token Maxing and the Emerging Cost Crisis

  • A trend of internal corporate token maxing — companies creating leaderboards incentivizing maximum AI token consumption — reflected the experimental, “letter rip” culture of the subsidy era.
  • The host acknowledged the tension: token maxing measures an input (tokens consumed) rather than an output (business value), invoking Goodhart’s Law.
  • Uber became a dual case study:
    • Its CTO revealed in April that Uber burned through its entire 2026 AI budget in four months.
    • Its COO subsequently expressed skepticism about the ROI actually delivered by that spend.
  • An Axios article (“AI Sticker Shock Hits Corporate America”) captured the broader corporate reaction.
  • Amazon was among companies scrapping internal AI leaderboards, partly due to gaming concerns and partly due to rising token costs.

4. The End of the AI Subsidy Era

  • Under flat-rate “Max” subscription plans ($100–$300/month), the most active power users could theoretically consume $2,000–$10,000 worth of tokens for $200/month — a 10–20x subsidy by the provider.
  • This subsidy was sustainable during the seat-based growth phase but becomes untenable as agentic usage (long, multi-step sessions) becomes the default.
  • GitHub Copilot (announced late April) shifted to usage-based billing, explicitly stating: “The current premium request model is no longer sustainable.”
  • Google (at Google I/O) nominally reduced plan prices (Gemini Ultra to $200, new $100 tier) but introduced usage caps and usage-based overage billing, representing a net cost increase for power users.
  • Anthropic introduced per-token billing for usage in third-party harnesses (outside Anthropic-owned environments like Claude Code), causing significant user backlash.

5. The Capabilities Overhang and Enterprise Support Initiatives

  • A capabilities overhang exists between what AI models can do and what most enterprises are actually extracting from them; agentic AI has dramatically widened this gap.
  • Both OpenAI and Anthropic launched separate enterprise services ventures:
    • OpenAI announced a majority-owned deployment company placing forward-deployed engineers inside large clients.
    • Anthropic partnered with Blackstone, Hellman & Friedman, and Goldman Sachs to launch a separate enterprise AI consulting firm (built on Fractional).
  • Both initiatives reflect a shared recognition that enterprises need substantial implementation support to navigate the agentic transition.

6. Structural Token Shortage and Infrastructure Response

  • The host’s central thesis: there is a structural shortage of AI compute/tokens — demand exceeds available supply, driving costs up.
  • Market-based responses already emerging:
    • Cursor released Composer 2.5, delivering competitive performance at significantly lower cost than frontier models (Opus 4.7, GPT-5.5).
    • Gemma (Google’s smallest model series) is seeing fast adoption outpacing comparable Chinese models, indicating enterprises are adapting to cost constraints.
    • DeepSeek made a temporary 75% price cut on its V4 model permanent, interpreted as a deliberate strategy to capture cost-sensitive enterprise customers priced out of frontier Western models.
  • Infrastructure valuations surging:
    • Inference provider Base10 raised $1 billion at an $11 billion valuation, more than doubling its valuation in a single quarter.
    • OpenRouter raised a $113 million Series B, achieving unicorn status.
    • AI memory chip companies SK Hynix and Micron became trillion-dollar companies.
    • Meta raised the possibility of monetizing its compute infrastructure as a cloud business, partially de-risking its $130 billion+ CapEx commitment.

7. Elon Musk, SpaceX, and the Compute Realignment

  • Elon Musk shifted from his prior role (Grok cheerleader / OpenAI antagonist) to a new role as a compute infrastructure provider.
  • SpaceX AI (SpaceX’s internal AI division) agreed to provide Anthropic access to Colossus 1, addressing Anthropic’s severe compute constraints.
  • Weeks later, Anthropic was also granted access to Colossus 2, effectively making SpaceX a neocloud provider.
  • The host argues this reframes the SpaceX IPO narrative: rather than an investment in an also-ran LLM (Grok), investors are buying into a critical AI infrastructure play at a moment of structural compute scarcity.
  • Musk’s discussion of orbital data centers — previously dismissed as science fiction — gained credibility; Jeff Bezos publicly engaged with the timeline rather than the premise, suggesting two to three years may be slightly ambitious.

8. Model Releases and the Shift Toward Harnesses

  • May was relatively quiet on model releases; Claude Opus 4.8 was the primary launch.
  • Industry sentiment increasingly reflects that marginal model improvements matter less than improvements to the harnesses (interfaces and environments) surrounding models.
    • Greg Eisenberg: “We’re entering the era where model releases start to feel like iPhone releases… The thing that actually matters right now is what’s happening around the models.”
    • Riley Brown: more excited about super-app updates to Codex and Claude Desktop than model version bumps.
  • Claude Code shipped dynamic workflows alongside the Opus 4.8 release, which the host identifies as more significant than the model itself.
  • Slash Goal emerged as a new primitive, moving from Codex to Claude Code and enabling more structured agentic task execution.

9. Narrative and Policy Shifts

  • Both Sam Altman and Dario Amodei moved away from public messaging about AI displacing jobs and transforming society in destabilizing ways.
    • Altman articulated a view that he had overestimated the pace and nature of workforce transformation.
    • The host characterizes this as opening space for more nuanced AI policy conversation.
  • On the political side, divergent Democratic approaches emerged:
    • Bernie Sanders / AOC wing: calling for data center moratoriums.
    • Elizabeth Warren: opposed moratoriums; proposed taxing AI in a Time op-ed (“Why We Need to Tax AI”), with token taxes as a potential mechanism.
  • The White House became directly involved in model access decisions around Anthropic’s Mythos model, reportedly opposing expanded access partly to preserve token availability for U.S. government use — framing compute as a strategic national resource.

Key Concepts

  • Token Scarcity Era: The current emerging period in which demand for AI compute (tokens) structurally exceeds supply, making AI costs a binding constraint for enterprises and providers alike.
  • AI Subsidy Era: The preceding period in which foundation model labs effectively subsidized heavy usage by offering flat-rate subscriptions whose value to power users far exceeded the subscription price.
  • Token Maxing: A corporate practice of creating internal leaderboards and incentives to maximize AI token consumption, intended to drive AI experimentation but critiqued for measuring inputs over outputs.
  • Capabilities Overhang: The gap between what AI systems are technically capable of and what enterprises are actually deploying or extracting value from.
  • Agentic AI: AI systems that autonomously execute multi-step tasks, browse resources, write and run code, and iterate across entire workflows without continuous human input — distinguished from single-turn chat interactions.
  • Seat-Based Billing: A subscription model in which customers pay a flat per-user monthly fee regardless of actual AI usage volume.
  • Usage-Based / Token-Based Billing: A consumption model in which customers pay proportionally to the number of tokens (units of AI compute) they consume.
  • Goodhart’s Law: The principle that once a measure becomes a target, it ceases to be a good measure, because behavior optimizes for the metric rather than the underlying goal it was intended to represent.
  • Neocloud: A new class of large-scale compute infrastructure provider that supplies AI inference and training capacity to foundation model labs and enterprises (e.g., SpaceX/xAI Colossus in this context).
  • Dynamic Workflows (Claude Code): A Claude Code feature shipping alongside Opus 4.8 that enables more flexible, context-sensitive agentic task orchestration.
  • Slash Goal: An emerging agentic primitive (originating in Codex, ported to Claude Code) that structures how AI agents interpret and pursue user-defined objectives.
  • Colossus 1 / Colossus 2: Large-scale data centers built by xAI/SpaceX, made available to Anthropic for inference compute, transforming SpaceX into an AI infrastructure provider.
  • OpenRouter: A developer platform that allows automatic routing between different AI models based on cost, performance, and efficiency trade-offs.
  • Forward-Deployed Engineers: Engineers embedded directly within client organizations to assist with AI deployment and integration, a model used by both OpenAI’s and Anthropic’s new enterprise services ventures.

Summary

May 2026 marked a pivotal inflection point in the AI industry’s evolution: the end of the AI subsidy era and the beginning of a structural token scarcity era. Foundation model companies — led by Anthropic’s extraordinary growth to $47 billion in annualized revenue and anticipated first-ever profitability — demonstrated that token-based API consumption had superseded seat-based subscriptions as the primary revenue engine, effectively dismantling the AI bubble narrative. However, this same revenue surge revealed that the economics of unlimited, flat-rate AI usage were unsustainable: GitHub Copilot, Google, and Anthropic all moved toward usage-based billing, while enterprises from Uber to Amazon encountered serious sticker shock and ROI skepticism. The host argues the defining feature of the coming period is a genuine compute shortage — there is simply not enough infrastructure to serve all the AI demand being generated — which is driving a cascade of responses: market innovation in cheaper, performant models (Cursor Composer 2.5, Gemma); infrastructure investment surges (Base10, OpenRouter, SK Hynix, Micron); and a major geopolitical realignment embodied by Elon Musk repositioning SpaceX as a neocloud compute provider for Anthropic, reframing the forthcoming SpaceX IPO as an AI infrastructure play. Simultaneously, the industry’s cultural center of gravity shifted from model capability announcements toward harness improvements (Claude Code dynamic workflows, Slash Goal), policy debates intensified around data center moratoriums and token taxation, and both AI lab CEOs began softening their most alarming public rhetoric about labor displacement. The overarching message is that enterprises and investors who rapidly develop fluency in managing AI costs and deploying agentic capabilities efficiently will hold significant strategic advantages in the token-scarce era ahead.