Uber burned through its entire 2026 AI budget by April. Tesla is capping employees at $200 per week starting July 6 — with a carve-out for Grok that its own engineers didn’t ask for. Meta is fighting a practice employees dubbed “tokenmaxxing,” which its CTO told staff to stop. Amazon scrapped its internal AI adoption leaderboard after engineers started gaming it. Walmart is rationing tokens for Code Puppy.

Five companies. Same problem. The bill for “just let people experiment” came due.

This isn’t a story about enterprise reluctance. These companies spent aggressively. The story is about what happened when per-token billing replaced flat subscriptions — and what it means for you if you sell AI tools to enterprises.

The Five Companies and What Actually Happened

Tesla — $200/Week with a Grok Carve-Out

Tesla’s AI spending cap took effect July 6, 2026. The number is $200 per week per employee — roughly $800 a month. The rationale is straightforward: before the cap, some software engineers were spending “thousands of dollars’ worth of tokens each week”.

The notable detail is what’s exempt: beta versions of xAI products, including Grok, aren’t subject to the limit. That’s a carve-out Tesla leadership built in, not one engineers requested. Electrek reports Tesla engineers largely prefer Anthropic’s Claude.

The competitive lesson here isn’t flattering to builders: at the enterprise level, policy exemptions can override model quality in determining which tool your customers use. If your tool isn’t named in the carve-out, the cap applies to you.

Uber — $1,500/Month After Burning the Annual Budget in Four Months

Uber is the most dramatic case because it was first and most public. CTO Praveen Neppalli Naga confirmed the company had used its entire 2026 AI budget by April — four months into the year, saying the company was “back to the drawing board” on its assumptions. The cap that followed was $1,500 per month per employee, per tool, applied to agentic coding tools including Claude Code and Cursor.

Uber’s public admission of AI budget loss of control carries its own lesson: the companies that moved fastest got the biggest bill.

Meta — Approaching Billions, Coined “Tokenmaxxing”

Meta employees built an internal leaderboard called “Claudeonomics” that ranked the company’s roughly 85,000 workers by token consumption, with some staff instructing agents to run multiple tasks at once purely to climb the rankings — a practice that came to be called “tokenmaxxing.” Meta CTO Andrew Bosworth pushed back on the practice in an internal memo: “Nobody should be using AI tools just for the sake of using them. All motion is not progress and token usage alone is not a measure of impact of any kind.” An internal memo sent to roughly 6,000 employees flagged an “exponential increase” in AI usage and warned the company faced billions of dollars in internal AI costs. Meta’s response was spending controls on employee use of outside AI tools alongside heavier internal investment in its own infrastructure.

The leaderboard detail is important for builders: any tool with a usage dashboard can create the same incentive. Design for value delivery, not raw usage volume.

Amazon — The Leaderboard That Backfired

Amazon rolled out an internal adoption dashboard called “KiroRank” for its Kiro AI coding tool and watched engineers begin gaming it — feeding trivial or fabricated tasks into the tool primarily to climb the ranking rather than to do work. Amazon SVP Dave Treadwell told staff, “Please don’t use AI just for the sake of using AI,” and the company scrapped the leaderboard, saying it “was not a formal or approved tool.” The lesson is the same as Meta’s: when you instrument adoption and tie recognition to it, you measure gaming, not value.

Walmart — Capping Code Puppy Tokens

Walmart built its own internal AI assistant called Code Puppy, and in June 2026 moved it from unrestricted access to a fixed per-employee token allocation after employee demand outpaced budget projections. Even fully internal tools face the same cost structure once they’re running at scale. Walmart’s approach — building the tool, deploying it, then capping it when costs ran ahead of projections — is a preview of the governance cycle that any large organization deploying AI will go through.

Why This Is Happening Now

The root cause is a pricing structure shift. Through late 2025 and into 2026, Anthropic restructured its enterprise plans to unbundle Claude, Claude Code, and Cowork from flat seat fees, moving customers to per-token billing at standard API rates, and OpenAI shifted Codex from flat-message pricing to token metering in early April 2026. Per-token pricing was good for providers (it scales with actual usage) and initially felt fair to buyers (you only pay for what you use). But it made cost invisible at the point of use — until it wasn’t.

When engineers ran Claude Code all day on a $20/month subscription, the company had predictable costs. When the same usage became per-token, usage growth showed up as a line item in someone’s budget report. Every developer who spun up an agent was now spending company money in a way that could be measured.

That’s when finance teams got involved. And once finance is involved, governance follows.

What Changes for Builders Targeting Enterprises

Cost Governance Is Now a Feature, Not a Plus

Before 2026, “cost transparency” was a selling point. Now it’s a gate. Enterprise buyers are asking about spend dashboards, per-user attribution, and budget controls before they ask about capabilities. If your tool doesn’t surface cost data at the team and user level, you won’t get past procurement at companies that have already been through a spending episode.

Minimum viable governance features for enterprise AI tools today:

  • Real-time per-user token spend, visible to admins
  • Team or project cost attribution
  • Configurable spend limits with alerts before enforcement
  • Approval workflow for limit overrides (not just hard cutoffs — teams need flexibility with oversight)

Per-Seat Pricing Is Back in Demand

Token-based billing is rational at small scale. At enterprise scale, where FP&A teams need to budget quarters in advance, unpredictable per-token costs are a problem. The companies capping spending are doing it because they can’t forecast token consumption the way they forecasted seat licenses.

If you’re building an enterprise AI product, offering a per-seat tier with predictable monthly costs alongside a consumption tier for variable workloads is increasingly a differentiator. Uber didn’t hit its budget limit on a flat subscription. It hit it because usage scaled in ways no one projected, and the billing model made every increment visible.

Task Routing by Cost Is Becoming a Competitive Differentiator

AWS’s Amazon Bedrock Intelligent Prompt Routing and a growing set of third-party model-router products are built on this idea: send tasks to the cheapest model that can actually handle them. A simple code completion doesn’t need Opus. A complex multi-step agent task might. AWS says its own router — which picks between models in the same family, like Claude 3.5 Sonnet versus Claude 3 Haiku, based on prompt complexity — can cut costs by up to 30% without compromising accuracy; more aggressive routing strategies across a wider range of models can save more, though the exact number depends heavily on the task mix.

If you’re building an agent-based product, model routing by cost-complexity tradeoff is now closer to a requirement than a feature. Enterprises will ask for it. If your product doesn’t do it, theirs will — through a layer they put on top of yours.

The Metric That Closes Enterprise Deals Has Shifted

Before the spend caps, enterprise AI sales conversations focused on adoption rates, activation, and usage volume. “Employees love it” and “usage grew 3x” were wins. That’s now table stakes or, worse, a signal that a spending problem is forming.

The conversation has shifted to value per dollar. What’s the productivity gain? What’s the cost per outcome? Can you attribute a reduction in cycle time to the tool? Can you model what the tool saves versus what it costs?

Builders who haven’t built ROI measurement into their products — even rough signals like time saved per task, completions per dollar, or cost per merged PR — will struggle in late-2026 enterprise sales cycles. The budget-holder question now is: “What did we get for what we spent?”

The Grok Exemption Is a Warning About Partnership Dependence

Tesla’s decision to exempt xAI tools from its AI spending cap while capping Claude, Cursor, and others is a preview of something that will happen more broadly. Large companies have strategic relationships with AI providers that are separate from — and sometimes contrary to — what their engineers would choose based on quality. Those relationships affect which tools get budget, which ones get exempted from caps, and which ones get recommended in procurement.

If your enterprise distribution strategy runs through a large model provider relationship, understand the strategic dependency. If the provider’s relationship with the enterprise changes, so does yours.

The Shift in Brief

The enterprise AI spending crisis of mid-2026 is a predictable consequence of moving large organizations from flat billing to token billing before they had governance tooling to manage it. The companies that moved fastest got the biggest surprise bills. Now the whole market is correcting.

For builders: the enterprises that cut their teeth on this cycle are going to be the buyers of the next generation of AI tools. They’ll know what questions to ask. They’ll have budget processes with AI line items. They’ll have procurement criteria that include governance, attribution, and ROI measurement alongside capability.

Build for that buyer. Not the one who said yes to everything in 2024 because the budget was small enough not to matter.

Pricing, policies, and exemption details from Electrek, Benzinga, MLQ.ai, and CryptoBriefing.