AI Is Now a Utility Bill. Here Is What That Means for Engineering Teams.

In July 2026, J.P. Morgan's investment strategy team published a note with a title that should catch the attention of every engineering leader and CFO: AI Use Is Exploding. So Are the Bills. It was not a technical paper. It was an investor briefing, which is precisely what makes it significant. When Wall Street is writing about AI token costs, the conversation has moved out of engineering and into the boardroom.

The note identified a structural shift that has been building since late 2025: the way companies pay for AI is changing. Per-seat subscriptions are giving way to consumption-based billing, where costs scale directly with activity. The implication is simple and uncomfortable. AI now works like electricity. You pay for what you use, and the bill only arrives after you have used it.

The pricing shift that changed the problem

For most of the early enterprise AI wave, the budget conversation was relatively manageable. A $30 per user per month subscription for a coding assistant is a line item with a predictable ceiling. Finance teams understood it. Procurement knew how to handle it.

That model is breaking down. As AI agents have taken on more complex, multi-step tasks, the token consumption per session has climbed sharply. Providers have responded by shifting to usage-based pricing, where the bill reflects actual token throughput rather than seats. For teams running agentic workflows, the difference can be an order of magnitude.

GitHub Copilot made this concrete on 1 June 2026 when it moved to usage-based billing. The plan structure is now $39 per user per month, which includes $39 in AI credits (one credit equals $0.01). Code completions remain included. Everything beyond that, including agent mode, longer context sessions, and premium model access, draws down credits at token rates. The familiar flat fee is gone. What replaces it depends on how your team works.

This is not an anomaly. It is the direction of travel for the industry, and engineering teams that have been thinking about AI costs as a fixed line are now discovering they are managing a variable one.

The 600x problem

The most striking data point in the J.P. Morgan note came from Wolfe Research: the top 1% of AI users are consuming more than 600 times as many tokens as the median user. That gap is not a rounding error. It is a signal that token spend within most organisations is highly concentrated, highly uneven, and almost entirely invisible to the people responsible for the budget.

The intensity skews toward complex coding workflows. A short email draft is cheap. A multi-step agent that reads documents, compares options, writes code, iterates across several passes, and produces structured output burns tokens at every step. Teams running agentic coding workflows at scale are generating bills that bear no relationship to the number of developers on the platform.

J.P. Morgan estimated that a single software engineer at a firm running an enterprise AI coding subscription could generate up to $730 per month in token costs. For a Fortune 500 firm with 5,000 engineers, that trajectory would exceed $3.5 million per month before any optimisation. The teams discovering this are not doing so because they planned for it. They are discovering it on the invoice.

Why aggregate numbers give you the wrong picture

The standard response to a large AI bill is to look at the total and ask which team is responsible. But that question, while reasonable, tends to produce answers that are too coarse to act on. Knowing that your engineering team spent $85,000 last month on AI API calls does not tell you which workflows are driving it, whether those workflows are efficient, or which ones would benefit from routing to a cheaper model.

The J.P. Morgan note made an observation that maps directly to the core problem: companies are now adapting by moving away from scaling AI usage as quickly as possible and toward optimising where it delivers the best return on investment. But you cannot optimise what you cannot see at the task level.

Token spend is not like cloud infrastructure spend, where a consistent workload produces a consistent cost. AI API costs vary by model, by task type, by context window size, by whether prompt caching is active, and by whether an agent loop is running efficiently or redundantly. The same workflow run against two different models can produce costs that differ by a factor of ten or more. None of that variation is visible from a monthly invoice total.

Model routing as the response, and its limits

The technically literate answer to rising AI token costs is model routing: assigning simpler tasks to cheaper models and reserving frontier model access for work that genuinely requires it. The logic is sound. A classification task that costs $0.002 per thousand tokens on a lightweight model should not be routed through a frontier model at $15 per million tokens if the quality difference is negligible for that task.

This is where Chinese open-source models have added a meaningful dimension. J.P. Morgan noted that GLM-5.2, China's leading LLM as of July 2026, performs at the level of the global frontier model from March 2026, at over 80% lower cost. For teams making cost-based routing decisions, that cost floor matters.

But model routing done on instinct produces mixed results. Teams that have approached it by intuition, assigning model tiers based on a general sense of task complexity, often find that their routing decisions are either too conservative (expensive models doing cheap work) or miscalibrated (cheap models producing outputs that need rework). The teams with effective routing strategies built them from actual spend data, not from first principles.

To route intelligently, you need to know your current spend by task type, your cost per output unit by model, and which workflows have the highest variance in token consumption. Without that picture, routing decisions are educated guesses. The guess might be right. But it will take months of invoice cycles to find out.

What you need to see before you can act

The minimum viable visibility for an engineering team managing AI token spend consists of four things.

1. Spend by model. Which models are being called, and what is the cost distribution across them? A team using five models may find that 85% of spend is concentrated in two.
2. Spend by task type. What categories of work are generating the most tokens? Classification, summarisation, code generation, and agent orchestration have very different cost profiles, and the split matters for routing decisions.
3. Cost per output unit. What does each task actually cost to complete, at the model being used? This is the number that makes routing trade-offs concrete. Without it, you are comparing abstractions.
4. Efficiency signals. Are there patterns suggesting wasted spend? System prompt bloat, agent loop redundancy, prompt cache misses, and unnecessary model calls for simple tasks all appear as patterns in request-level data before they appear as line items on an invoice.

Most organisations have none of these in a usable form. They have invoice totals and, if they are using a platform like OpenRouter, a CSV export that contains the raw data but no analysis. The gap between having the data and understanding what it means is where the diagnostic value lies.

The moment the conversation changed

It is worth noting what it means that J.P. Morgan is now publishing investor notes on AI token costs. Two years ago, the question was whether enterprises would use AI at all. A year ago, the question was which models to use. Today, the question that is landing on CFO desks is whether the results justify the costs, and what to do when they do not.

The scale of that spending is not hypothetical. The FinOps Foundation reported that enterprise token consumption drove approximately $21 billion in annualised new revenue for Anthropic between October 2025 and April 2026. That figure covers one provider, over six months. It is the visible surface of a market where enterprise teams are committing significant budget to AI without the diagnostic tools to understand what they are actually paying for.

J.P. Morgan drew an explicit parallel to the cloud computing transition of the 2010s. What began as a scarce and powerful capability became, over time, a metered utility to be optimised. Pricing power shifted toward the parts of the stack that were harder to swap out. AI is following the same path. The teams building cost intelligence now are building a capability that will be table stakes within 18 months.

The diagnostic question is not whether to take AI costs seriously. It is whether you have the visibility to act on them before the bill arrives.

Data cited in this article: J.P. Morgan Wealth Management, "AI Use Is Exploding. So Are the Bills," 17 July 2026 (Wolfe Research token consumption data; Bloomberg LLM Token Expenditure Index; GitHub Copilot enterprise cost estimates). GitHub Copilot usage-based billing transition confirmed 1 June 2026. Artificial Analysis model cost benchmarks cited for GLM-5.2 comparison. FinOps Foundation, "Token Economics: The Atomic Unit of AI Value," FinOps X 2026 keynote report (Anthropic enterprise revenue consumption data, October 2025 to April 2026).