Map

AI Is Too Expensive

Wiki summaryai-bubbleeconomicspredictions ↳ show in map Markdown
title
AI Is Too Expensive
type
summary
summary
Zitron's May 2026 follow-up to the subprime-AI thesis — Microsoft sunk $87B into OpenAI, hyperscalers need $3-6T to break even, enterprises blow through token budgets, Zillow is AI Chernobyl
tags
ai-bubble, economics, predictions
sources
zitron-ai-too-expensive
created
2026-05-20
updated
2026-05-20

Edward Zitron's May 2026 piece updates the "subprime AI" thesis with the numbers that came out in the intervening two months: the Musk-OpenAI trial testimony revealing Microsoft's $87B sunk into OpenAI infrastructure, Anthropic's March affidavit showing $3 spent for every $1 of compute revenue, RPOs jumping by hundreds of billions but with almost all of the jump coming from Anthropic and OpenAI themselves, and a new section of evidence: enterprises blowing through their entire annual token budgets in months. The thesis hasn't changed; the supporting numbers got worse.

The hyperscaler arithmetic

Zitron stacks the capex numbers:

  • Microsoft has spent $293.8B in capex since FY23 (back half of 2022). Bloomberg, via Musk-OpenAI trial testimony from executive Michael Wetter, reveals that ~$87B of that — about 30% — went specifically into OpenAI infrastructure. Per Zitron's Azure sources, this is the vast majority of Microsoft's operational capacity.
  • 2026 hyperscaler capex plans: Microsoft $190B, Google $185B, Meta $125B+, Amazon $200B+. Roughly $700B in 2026, and another ~$1T projected for 2027.
  • Combined revenue of Microsoft, Meta, Amazon, Google FY2025: $1.599T for every product. They need at least $3T in AI-specific revenue to break even, $6T to come out ahead.
  • Microsoft 365 Copilot has 20M paid subscribers — absolute maximum revenue $7.2B/year if every one paid $30/month (they don't; discounting is multi-year). Total Microsoft AI revenue FY2025 Zitron estimates at ~$17.9B against $88.2B capex.

The pull-through math: at $42-44M per megawatt, OpenAI's claimed 1.9GW of capacity (end of 2025) is ~$79.8B of infrastructure cost. Microsoft has invested nearly four times that. Most of what Microsoft has spent is for OpenAI alone.

The four-conditions test

For any of this to make sense as an investment, all four must be true:

  1. AI revenues have to explode.
  2. Capex has to stop being invested.
  3. GPUs need to be margin-positive, including debt service.
  4. AI revenue has to stay consistent before and after capex stops.

Zitron walks through why each condition fails. If condition 1 fails alone, the best case is break-even. If condition 2 fails (capex keeps growing), revenues have to roughly double Microsoft/Meta/Google's entire businesses and triple AWS. Condition 3 is unproven and contested; the on-the-record claims ("Anthropic is profitable on inference") rely on misquoted Altman remarks or hedged Amodei statements about "stylized facts." Condition 4 is the worst: even if revenue is good now, the question is whether it survives the capex stop.

RPOs — the demand illusion

Remaining performance obligations (RPOs) are the multi-year contracted revenue commitments on hyperscaler balance sheets. They jumped dramatically in Q1 2026:

  • Microsoft RPOs: $392B → $625B, driven by OpenAI's $250B "incremental Azure spend" lockup (October 2025) and the $30B from Anthropic (November 2025). Without those two, RPOs would have been flat.
  • Amazon RPOs: $244B → $364B, driven by Amazon's $100B expansion of its OpenAI compute deal and the Anthropic 5GW capacity partnership.
  • Google RPOs: $242.8B → $467.6B, driven mostly by Anthropic's $200B TPU spend commitment. Without it, the jump is $24.8B (and a chunk of that is the Meta TPU rental deal).

Zitron's reading: outside of Anthropic and OpenAI, these three companies aren't significantly increasing their AI revenues. They are extending each other credit through two intermediary companies. Per The Information, more than 50% of hyperscalers' revenue backlog comes from Anthropic and OpenAI alone.

Anthropic and OpenAI's four-year hole

What Anthropic appears to owe over the next four years:

  • $200B to Google
  • $100B to Amazon
  • $30B to Microsoft
  • $30B to CoreWeave
  • $20B to xAI ($5B/year × 4)
  • ≥$20B in non-compute operating expenses

Total: ~$380B over four years. Anthropic per its March 2026 affidavit had spent $10B on training and inference against ~$5B in total lifetime revenue — $3 spent per $1 earned. The Information reports its gross margins were 40% in 2025, 10 points below the optimistic projection, with inference costs 23% higher than expected.

OpenAI per The Information plans to burn $852B through 2030, with gross margins falling from 40% in 2024 to 33% in 2025 — 13 points below projection because of last-minute compute purchases.

This is what Zitron has called the knife-catching problem: order compute years ahead and you can bankrupt yourself if revenue doesn't materialise (Amodei: "there's no force on earth, there's no hedge on earth that could stop me from going bankrupt"); buy at the last minute and you pay spiked spot prices that destroy whatever margin you might have had. Both companies are in the bad version of this trap.

Enterprises blowing through token budgets

The new evidence in this piece is enterprise token-budget overruns. Anthropic moved enterprise customers to token-based billing a month or two before publication, and the stress test is producing predictable results:

  • ServiceNow CIO Kellie Romack (talking to the Information's Laura Bratton): the company is unsure whether it can keep employees on Claude Enterprise for the rest of 2026 without containing costs.
  • Salesforce CEO Marc Benioff: $300M planned Anthropic token spend in 2026.
  • Stripe (per sources to Zitron): ~5,000 technical staff burning an average of $94K/day in tokens — $2.8M/month, $33.6M/year. Against ~$765M in technical headcount, this is ~4.4% of headcount cost. Goldman reports AI costs approaching 10% of total headcount across the industry, "on track to be on par with headcount costs in the next several quarters."
  • Uber and other organisations are exhausting yearly token budgets in months.

Zillow is AI Chernobyl

The case study Zitron leans on:

  • Zillow spent $1M on AI services in Q1 2026, then $749K in April alone (Cursor, Anthropic, AWS Bedrock).
  • By mid-May, total AI spend was over $3M; Cursor budget consumption hit 85% of the annual $1.1M.
  • 2026 trajectory: $7-10M total AI spend, against $23M FY2025 net income.
  • Internal slide deck "AI-Native Engineering" promises a "path to an agentic Zillow." Talks about engineers moving from "soloist" → "conductor" → "composer."
  • The "2027: A Tuesday" slide describes a hypothetical engineer who never opens a codebase, just specs and eval dashboards.
  • Reality, per Zitron's sources: engineers still open IDEs, code is "slowly becoming AI slop," human review work increased by ~50% even with engineering headcount flat. Software reviewer load up 29,000 hours per month — about 19 extra review hours per engineer per month, "just looking at extra code written by LLMs."
  • Blind anonymous posts from Zillow engineers: "the slop is job security"; "people are burning tokens just to hit internal AI adoption targets."

This is the rare named case where the "corporate AI mandate" dynamic is being documented from the inside in near-real-time.

Anthropic's transparency problem

Per Bratton's Information reporting, Anthropic does not offer:

  • Telemetry data of the kind ServiceNow, SAP, Microsoft, Workday provide (which user, which tool, how much, how)
  • Service-level agreements
  • Granular billing decomposition

Zitron's reading: with multi-million-dollar contracts, no SLA, no telemetry, and a model whose token consumption is non-deterministic per task, the conditions are set up for either deliberate or accidental cost amplification with no customer-side audit path. He stops short of accusing Anthropic of anything specific — but he's clear that the structure of the relationship is unusual for enterprise software.

The unmeasurable ROI

Zitron's structural argument: AI token budgets are inherently bullshit because tokens-per-task isn't a fixed quantity. Tokens vary across models, across users, across runs, even with identical prompts. Sample-of-one measurements are useless. KPI replacements all gameable:

  • "Burn as many tokens as possible" — Amazon and Meta employees already gaming token-leaderboards.
  • "Use AI every day" — no success criterion.
  • "Ship more software" — emphasises velocity over quality.
  • "More pull requests" — any number an engineer can manipulate, they will.

Without per-task token measurability, every AI budget is a dart thrown blindfolded. Without a usable KPI, "productivity" claims can't be falsified — which is why the Business Idiot class loves the technology.

The Business Idiot turn

Zitron's closing argument is structural rather than financial: AI is grift-shaped because the modern executive is grift-receptive. LLMs do an "impression of work" the same way most executives do an impression of work. They'll say yes to anything, never push back on timelines, produce "convincing-sounding responses" instead of asking why the engineer just said no. See business-idiot for the standalone concept.

The connection to the rest of his analysis: a Business-Idiot–run economy is what makes the demand side of the bubble look real. Enterprise spend looks like demand. It's not — it's executives "eating the cost while employees experiment" (Workato CIO) with no return-on-investment measurement and no exit plan.

Where this leaves the bubble thesis

Zitron's prediction in subprime-ai-crisis was that the collapse sequence runs through (1) lab cost escalation → (2) startup margin compression → (3) consumer rate-limit revolt → (4) lab compute over-commitment. This piece adds a fifth: (5) enterprise token-budget revolt, where CIOs and CFOs are forced into the cost-containment conversation in 2026, not 2027.

The first CFO to publicly cut AI spend, Zitron predicts, sends the rest of them running for the doors.

Cross-references

Linked from