Skip to content

AI Cost Optimization

Uber burned through its annual AI budget in four months. Here is what business owners should know about AI cost controls.

In April 2026, Uber's CTO disclosed that the company had burned through its entire annual AI budget in just four months. Uber then implemented spending tiers capped at $1,500 per month per employee, with higher access available only on request. The same week, Flo Crivello, CEO of AI startup Lindy, told CNBC that he had switched his 25-person company off Anthropic's Claude models entirely, moving all of the company's AI traffic to a cheaper alternative. The reason was simple. AI costs had exceeded Lindy's payroll. "It's a matter of survival for the business," Crivello said. "That's all it is." CNBC reported on June 26, 2026 that this pattern is spreading. Enterprise customers across the country are pulling back on AI spending, demanding clear ROI before renewing contracts, and routing work to cheaper models. The practice of maximizing token usage without cost constraints, which CNBC described as tokenmaxxing, is ending. Most medium-size businesses are not running the AI operations Uber or Lindy run. But the underlying pattern is the same: AI usage spreads faster than cost controls are put in place. And the time to set limits is before the bill becomes a problem, not after.

By Fabio Rabelo · Founder, ATLACIS ·

What CNBC reported

CNBC's June 26 story traced a pattern that has been building for months and is now accelerating across US businesses. The core dynamic: AI costs in the coding and agentic AI space grew far faster than companies expected when they told employees to use AI as much as possible. Developers using AI coding tools do not just send a few thousand tokens in a chat session. They run agents that spawn sub-agents and process entire codebases, burning millions of tokens daily without typing a word. That is how Uber exhausted a year of AI budget in four months. The response has been spending caps and tiers. Uber capped access at $1,500 per month per employee. Higher access requires an escalation request. Other companies are implementing similar controls. On the vendor side, companies like Lindy are switching from frontier models to cheaper open-weight alternatives for tasks where the output quality difference does not justify the cost gap. Lindy's CEO said the switch was not about preference. It was about whether the company could stay in business. D.A. Davidson analyst Gil Luria told CNBC: "Some of their largest enterprise customers may start limiting their out-of-control token spend." He noted that this creates urgency for AI companies approaching IPOs: getting to market while growth numbers still look strong. The broader picture is that the first generation of AI adoption, which was largely about exploration and tool experimentation, is being replaced by a second generation that asks what the AI spend is actually producing.

Why most business owners should read this differently than a tech CEO

Uber has thousands of engineers running AI coding agents at scale. Lindy is a software startup that built its entire product around AI model calls. Most medium-size businesses are not in that position. But the underlying pattern runs at any size. It usually looks like this. A business subscribes to one or two AI tools. Usage spreads from the person who set them up to a team. The team starts using the tools in ways that were not anticipated. A developer uses an AI coding assistant heavily. A marketing team runs a content workflow through an AI API. Someone connects an AI agent to a business process without notifying anyone. The bill grows. Nobody is tracking it closely because the amounts are small at first. Then the amounts are not small. The Uber story is notable because of the scale. But the structure, usage spreading faster than tracking or limits, is not unique to large technology companies. For a business with 20 to 150 employees, the same pattern often plays out over 6 to 18 months rather than 4. The result is a bill that is much harder to explain than the value being produced.

What business owners often misunderstand about AI cost

The first misunderstanding is that AI cost is a subscription problem. Many businesses think of AI spending as a fixed monthly subscription: one amount for ChatGPT, one amount for Copilot, one for whatever else the team uses. That is no longer accurate for most meaningful AI deployments. Models billed by token, API-connected agents, and AI coding tools can produce bills that vary by an order of magnitude month over month depending on how heavily they are used. A developer running an AI coding agent intensively for a week can generate more token spend than a full team using a chat interface for a month. That variability requires active monitoring, not passive subscription management. The second misunderstanding is that frontier models are always the right default. The Lindy case illustrates a cost-performance tradeoff that applies to most businesses: frontier models, the most capable and most expensive AI models, are justified for tasks where quality genuinely matters and cheaper models fall short. For simpler, higher-volume tasks, that gap often does not exist in practice. Using a frontier model for tasks a cheaper model handles equally well is a cost decision, not a capability decision. The third misunderstanding is that ROI appears on its own. Several companies interviewed by CNBC are pausing AI spend until they can prove a return. That pause rarely happens before costs have already grown. Measuring return on AI spend requires defining, before deployment, what success looks like and how it will be measured. Most deployments skip that step. The fourth misunderstanding is that spending controls slow down adoption. At Uber, spending caps applied per employee with escalation paths for higher access. That design is not a restriction on AI use. It is a forcing function that makes each request for higher spend justify itself. Controls applied thoughtfully do not slow down the workflows that are generating real value. They slow down the workflows that were never generating value in the first place.

Four questions to answer before AI costs become uncontrollable

The time to answer these questions is before usage spreads, not after the bill arrives. First: what is your total current AI spend and where is it coming from? Include subscription fees, API usage charges, AI features bundled into software you already pay for, and any usage-based billing on tools like GitHub Copilot or Microsoft Copilot. If you cannot answer this question accurately, that is the first problem to solve. Second: which workflows are producing clear, measurable value and which are experimental or habitual? The workflows that are producing value deserve continued investment. The workflows that are experimental or habitual are worth examining before they scale. Third: do you have a monthly spending limit in place, and does your team know what it is? This does not need to be a rigid cap. It needs to be a number that, if exceeded, triggers a review. Without a limit, there is no signal that the spend has moved beyond what was intended. Fourth: which tasks in your AI workflows require the most capable models, and which could run on a cheaper alternative? This is a routing question, not a capability question. Frontier models are the right choice for tasks where quality genuinely matters and cheaper models fall short. For high-volume, simpler tasks, the difference in output quality often does not justify the cost difference.

What a serious business should do next

Start with an audit of what you are actually spending on AI today. Pull the numbers from every tool that bills by usage, every API connection, and every subscription that includes an AI component. This usually takes a few hours, not weeks. Then categorize the spend by workflow. Which AI use cases are clearly tied to a business outcome? Which are speculative? Which are the result of individual experimentation that has not been reviewed? Set a monthly limit that reflects what you are willing to spend before a review is triggered. This is not a budget ceiling. It is a review trigger. When the number is hit, someone asks whether the spend was worth it. Apply model routing logic where appropriate. If you have high-volume AI tasks that use a frontier model, test whether a cheaper model produces acceptable output for that task. The cost difference can be significant. This is especially relevant in coding workflows, content generation at volume, and structured data processing. Review AI spend monthly the same way you review any other variable cost. AI bills that grow unchecked for six months are harder to rationalize and harder to cut than AI bills that are reviewed monthly. For businesses that have not yet mapped their AI spend at all, or that are unsure whether current AI deployments are producing a return, the right first step is a short review of the current picture before the next tool purchase.

The Atlacis view

The Uber and Lindy stories are worth reading carefully, not because they are representative of what most medium-size businesses spend on AI, but because they show what happens when spending is not tracked until it has already become a problem. Most business owners are not facing Uber-scale AI bills. But the pattern, adoption spreading faster than the cost model is understood, is exactly the same at smaller scale. The businesses that manage AI cost well are not the ones who spend the least. They are the ones who know what they are spending, what each workflow is producing, and what their limit is before they need to find out the hard way. That clarity does not require complex infrastructure. It requires answering the four questions above before the next expansion of AI use inside the business. At Atlacis, we help business owners understand what AI is costing them, which workflows are generating real value, and how to build a cost model before spending becomes difficult to explain. If your AI bill is growing faster than your visibility into it, that is the right time to map it, not after the first oversized invoice.

The short version

  • Uber burned through its annual AI budget in four months, then capped employee AI spending at $1,500 per month with escalation paths for higher access. The pattern: AI usage spread faster than spending controls were in place.
  • Lindy CEO Flo Crivello switched his 25-person company off Anthropic's Claude entirely because AI costs had exceeded the company's payroll. He described it as a matter of survival, not a technology preference.
  • CNBC reported on June 26, 2026 that enterprise customers across the country are pulling back on AI spending and demanding clear ROI. The era of maximizing AI usage without cost constraints is ending at the top of the market.
  • AI costs are not subscription costs. Token-billed models and AI agents can produce bills that vary by an order of magnitude depending on usage intensity. That variability requires active monitoring.
  • Frontier models are the right choice when quality matters and cheaper models fall short. Using them for high-volume simple tasks where output quality is comparable is a cost choice, not a capability choice.
  • The right time to put spending controls in place is before usage spreads, not after the bill becomes a problem. A monthly review trigger, not necessarily a hard cap, is usually enough to keep the cost model visible.
Tags:AI costAI spendingtoken costsAI budgetAI cost controlsbusiness AIAI decision-makingAI cost optimizationAI vendorsAI ROI
FAQ

Common questions

Do I need to set hard caps on AI spending, or is a review process enough?
For most medium-size businesses, a monthly review trigger is more useful than a hard cap. A hard cap without an escalation path can block workflows that genuinely justify higher spend. A review trigger, a number that prompts someone to ask whether the spend was worth it, keeps costs visible without restricting legitimate high-value work. Hard caps make more sense in environments where developers have access to high-volume API connections or AI coding agents that can generate large bills quickly without oversight.
Does switching to a cheaper AI model mean accepting lower quality?
Not necessarily, and not for all tasks. The cost difference between frontier models and less expensive alternatives is significant. For tasks where that gap is justified by output quality, frontier models are the right choice. For high-volume tasks where a cheaper model produces acceptable output, the quality difference often does not justify the cost difference. The practical step is to run a structured comparison on a sample of your most common high-volume tasks before deciding. The result is usually task-specific, not a blanket answer.
Is this the right time to cut AI spending, or is it a bad time to slow down?
The right frame is not whether to slow down but whether your current spending is mapped. Businesses that can identify which AI workflows are producing measurable value and which are not are in a position to increase spending in the right places while reducing it in others. Cutting AI spending as a blanket response risks cutting the workflows that are working. Auditing spend to understand where the value is coming from is the step that makes the next increase defensible.

Make better AI decisions, starting with one call.

Book a free AI Fit Call. We will tell you what to use, what to avoid, and where to start. No jargon, no pressure.