# AI Token Budget Explosion

The "AI token budget explosion" is the dynamic [[ai-too-expensive|Zitron documents]] when enterprises move from flat-fee subscriptions to token-based billing: budgets that were set for a year are exhausted in months, but spend keeps growing, because nobody — at either side of the transaction — can measure tokens-per-task accurately enough to budget against.

## What it looks like in the field

The May 2026 examples:

- **Stripe**: ~5,000 technical staff burning $94K/day in tokens — $2.8M/month, $33.6M/year — primarily on Anthropic coding models. Against ~$765M in technical headcount cost, that's 4.4%.
- **Salesforce**: Marc Benioff publicly committed to $300M Anthropic token spend in 2026.
- **ServiceNow**: CIO Kellie Romack openly unsure (to The Information's Laura Bratton) whether Claude Enterprise can be sustained for the rest of 2026 without containment.
- **Uber**: blew through its annual token budget in months without slowing spend.
- **Zillow**: $1M Q1 + $749K April + on pace for $7-10M annual against $23M FY2025 net income. See the [[ai-too-expensive#Zillow is AI Chernobyl|"AI Chernobyl"]] case study.
- **Goldman Sachs**: industry-wide, AI costs approaching 10% of total headcount, "on track to be on par with headcount costs in the next several quarters."

## Why budgets miss

The structural reasons enterprise token budgets are non-functional:

- **Tokens-per-task isn't fixed.** The same prompt, same model, same configuration can produce wildly different token counts on different runs. Reasoning models can think for ten seconds or three minutes on the same input.
- **Models change underneath you.** Anthropic and OpenAI push silent updates that change behaviour. Last quarter's "this task uses X tokens" can be 2× different next quarter without any user-visible signal.
- **No telemetry granularity.** Per Bratton, Anthropic does not provide the per-user / per-tool / per-task telemetry that SAP, ServiceNow, Microsoft, Workday all consider standard. CFOs can see total spend but not what drove it.
- **No SLAs.** Per Romack and Mehta, Anthropic does not offer service-level agreements. There's no commitment about performance or cost-stability.
- **Adoption mandates inflate demand.** Executives — see [[business-idiot]] — have given engineers free rein to burn tokens with no measurement. Engineers at Amazon and Meta have been documented gaming token-leaderboards to hit adoption KPIs. Spend grows without producing output.

## Why ROI can't be measured

Zitron walks through the gameable-metric problem:

- "Burn more tokens" → employees write scripts to burn tokens.
- "Use AI every day" → no success criterion attached.
- "Ship more software" → optimises for velocity over quality.
- "More pull requests" → engineers game any number you set.

The deeper problem: without a stable cost-per-task, ROI denominator is missing. Without a measurable output-per-task, ROI numerator is missing. Both pieces are unstable. The aggregate "we spent $X, we shipped Y" comparison can't separate AI's contribution from headcount or process changes.

## What it predicts

Zitron's prediction in [[ai-too-expensive]] is that 2026 is when CFOs and CIOs are forced into the cost-containment conversation, two or three quarters earlier than the "deep into 2027" timeline industry boosters expect. The first publicly named enterprise to cut its AI budget triggers the herd. The mechanism is the same one that started the explosion: enterprise AI spend was social-proof-driven on the way in, and will be social-proof-driven on the way out.

## Related concepts

- [[ai-too-expensive]] — the full case
- [[subprime-ai-crisis]] — the broader bubble thesis
- [[knife-catching-compute]] — the supply-side bind that feeds back into customer pricing
- [[business-idiot]] — the executive class that bought without measurement
- [[ai-great-leap-forward]] — the corporate mandate side of the same story
- [[ai-bubble-pale-horses]] — checklist
- [[ai-subsidy-economics]] — the upstream mechanism
- [[tokenmaxxing]] — the employee-side gaming that inflates the budgets further