In December 2025, Uber gave its engineers access to AI coding tools. Adoption was fast. By March, 84% of engineers were classified as agentic coding users, around 95% were using AI tools every month, and roughly 70% of committed code was coming from those systems. By most measures, the rollout worked.
Then the bill arrived. Uber spent its entire 2026 AI coding budget in the first four months of the year. Average spend ran between $150 and $250 per engineer per month, with heavy users between $500 and $2,000. The company is now capping employee use.
Two details make this more than a procurement story. The first is how adoption was driven: through an internal leader board ranking teams by total AI tool usage. Measure adoption by volume and you get volume. The second is what the company still cannot say. Uber’s COO has said it is “very hard to draw a line” between rising AI usage and more useful features for customers. Adoption is easy to count. Value is not.
For anyone running a fund, or sitting on the board of a portfolio company being told to adopt AI, this is the question underneath the enthusiasm. The time savings are visible. Whether they are ROI-positive is a separate matter, and the way AI is priced is about to make that question harder, not easier.ways that help users realise its full potential.
The shift: from seats to meters
Whether token consumption is even your cost depends on how the AI is billed. There are two regimes, and they behave completely differently.
Per-seat or subscription tools, such as copilots and chat assistants. You pay a fixed fee per user. Feeding a 100-page PDF into the chat does not cost you more; the provider absorbs the compute and protects itself with usage caps. Cost is fixed and token consumption becomes someone else’s problem.
Usage-based or metered models, including API access, agentic coding tools, and AI embedded into your own products and workflows. You pay per token. Every document read, every reasoning step, and every agent loop is on your bill.
The trend that matters is where the heavy use cases sit. The highest-consumption work falls overwhelmingly into the metered camp. Uber’s bill came from a usage-based coding tool, not seat licences. Microsoft cancelled most of its direct Claude Code licences after heavy usage pushed per-engineer cost to between $500 and $2,000 a month.
And the flat-rate refuge is shrinking. Today’s subscription pricing is widely understood to be subsidised, with providers absorbing heavy-user compute to win adoption. Economics like that do not hold, and repricing has already begun. Even where you are on a fixed seat today, you carry repricing risk tomorrow.
But that only works if the underlying environment can support it.
What drives the meter
When you are paying by usage, the cost is not just in the question you ask. It sits in everything the model reads and generates around it.
Long inputs and long outputs. Large documents, data-room PDFs, and transcripts cost more in a single pass than hundreds of short questions, and generated output is usually priced higher than input.
Reasoning depth. “Thinking” models produce large volumes of intermediate reasoning that you pay for but never see.
Agentic workflows, the biggest driver. An agent plans, calls tools, reads results, and retries, re-sending its accumulating context at each step. One task becomes dozens of calls, each larger than the last.
Coding. Agents read across whole codebases and run iterative write, test, and fix loops, which is why engineering is usually where bills appear first.
Always-on use. Embedding AI into every email, ticket, or query turns occasional cost into continuous cost.
Importantly, not all usage needs the same level of compute. One of the fastest ways to control cost is through efficient token use and model selection. Routine, repeatable tasks should be handled by smaller, cheaper models, while frontier models should be reserved for complex reasoning. Organisations that fail to differentiate quickly see costs rise without a corresponding increase in value.
There is also a behavioural pattern emerging. For simple information retrieval, such as searching people, companies, or basic facts, AI is often used in place of traditional tools. Over time, this replaces low-cost or free search behaviours with metered AI consumption. At scale, even these “small” queries compound into a material cost line.
The number that matters
That single shift explains much of the cost story. It sits alongside a second reality that is easy to misread. The unit price of tokens is forecast to fall sharply over time. Yet enterprise AI spend is rising, not falling, because total consumption is increasing far faster.
The maths is simple:
- Cost per unit goes down.
- Units consumed go up significantly.
- Total bill goes up.
This is the cheap-token trap. Lower prices do not reduce spend. They enable more expensive behaviour: deeper reasoning chains, more iterations, and more embedded use cases. They also mask inefficiency. Cheap tokens are not the same as cheap capability.
The PE lens: where this becomes material
1. Adoption is easy to measure. Value is not.
Across portfolios, adoption is accelerating. But the headline numbers tell two different stories.
- 64% of organisations say AI is enabling innovation.
- Yet only 39% report enterprise-level EBIT impact.
Source: McKinsey, The State of AI 2025.
Uber’s experience reflects a broader challenge. Despite widespread AI adoption, the company has publicly questioned how directly increased AI usage translates into customer value and business outcomes.
Source: Uber COO comments reported by MSN.
The gap is the issue.
Board packs tend to show activity: percentages of employees using AI, adoption rates, prompts executed, and hours saved. These are inputs, not outcomes.
- What did margin actually improve by?
- What revenue changed?
- What cost disappeared in practice, not just in theory?
Uber’s experience mirrors what operators are seeing more broadly: high usage, visible productivity gains, and an increasingly unclear line between AI activity and commercial impact.
2. The hiring substitution thesis
The implicit PE bet is straightforward: grow output without growing headcount. There are early signals this is happening, but expectations remain mixed.
The comparison underneath the bet has shifted.
- Old comparison: AI spend versus the software budget.
- New comparison: AI spend versus the fully loaded cost of the employee it is meant to replace.
This is where the economics tighten. Substitution only works when usage is efficient, costs are stable, and consumption is actively governed. Without that, AI becomes incremental cost rather than replacement cost.
3. A new diligence question
AI cost exposure is becoming an underwriting issue. Many companies are currently operating on subsidised AI economics, which mask the true cost structure.
The diligence lens needs to ask:
- What is the mix of seat-based versus metered AI?
- What happens to margin if token prices increase materially?
- What happens at significantly higher usage levels?
- Is the margin story dependent on today’s pricing holding?
The question is not just what a company spends now. It is how that cost behaves as the business scales and as pricing models mature.
This is not theoretical. It is the same transition cloud already went through. Early adoption looks cheap. Scaled adoption rewrites the cost base.
Governing the meter
The response is not to slow AI down. It is to run it with discipline.
Introduce token governance, effectively FinOps for AI. Track spend at task, team, and workflow level. Shift the question from “how much are we using?” to “what does each outcome cost?”
Route intelligently. Reserve frontier models for complex reasoning, and direct routine, high-frequency work to smaller, more efficient models. Misallocation here is one of the biggest hidden cost drivers.
Design for efficiency, not just capability. Redundant prompts, overly long agent chains, and unnecessary processing are the AI equivalent of cloud sprawl.
Build multi-vendor leverage. Pricing power sits with providers, and concentration in one model means your cost base moves when their pricing does.
Stress-test budgets. Model cost structures at higher pricing and higher usage levels. Fragility shows up early when you look for it.
Closing
AI adoption is succeeding. That is not in question. What remains unresolved is the economics of that success.
The shift from seats to meters introduces a discipline that sits between engineering, finance, and operations, and most organisations do not yet have a clear owner for it. When usage scales faster than understanding, cost becomes the constraint rather than capability.
A pattern is starting to emerge in the market. As teams push AI harder, a behaviour some are beginning to call “token maxxing” is appearing, where usage is optimised for output and speed with limited regard for cost efficiency. It drives rapid adoption, but also rapidly inflates the bill.
The firms that create value will not be the ones that deploy AI fastest. They will be the ones that understand how it behaves at scale, how to control it, and how the cost grows alongside it.
Where to go next
Managing AI cost is not a tooling problem. It is a data and architecture problem first.
The organisations getting this right have two things in place: clear visibility over how AI is used, and control over how it is deployed.
If this is becoming a live issue in your portfolio:
Understand your data foundation
Most cost inefficiency starts upstream. Fragmented data, duplicated workflows, and unclear ownership drive unnecessary consumption. Fixing these first reduces usage and waste.
Introduce structured AI governance
AI requires disciplined governance across model routing, cost monitoring, and scalable deployment, with clear guardrails, accountability, and oversight from day one onward.
