Intelligence Gets a Meter
AI felt like a subscription. Agents turn it into a utility bill. Chapter 4 follows the shift from unlimited experimentation…
AI felt like a subscription. Agents turn it into a utility bill.
The first phase of generative AI was easy to understand. A user paid a fixed monthly fee and gained access to a system that wrote text, generated images or explained code. The real cost remained hidden behind a familiar interface.
Agentic applications change that model. An agent reads documents, plans several steps, calls tools, checks results, corrects errors and starts again when necessary. The user sees a completed task. Behind it, millions of tokens, several models and many API calls pass through a variable cost system.
A new economic phase begins. AI moves from innovation budget to controlled operating expense. The key question is no longer only what a model can do, but how much a reliably completed task costs.
The Invisible Invoice
A single chat appears inexpensive because the human limits the scope. The user asks a question, reads the answer and decides whether another step is needed.
An agent makes that decision itself. It decomposes a task, reads large contexts, searches for information, calls software and repeats failed steps. Every pass consumes input, output and often additional reasoning tokens.
A 2026 study of agentic coding tasks found that complex agent runs could use up to one thousand times more tokens than simple code chats. Repeated runs on the same task sometimes differed by a factor of 30, while higher usage did not reliably improve outcomes.
AI consumption is therefore less predictable than traditional software. A license costs the same every month. An agent may consume almost nothing on one day and create a large bill through retry loops on the next.
The user sees the task. The CFO sees the loops.
From Innovation Budget to Cost Center
During experimentation, developers often choose the strongest model available. Speed and quality matter more than the cost of each call. When only a few teams are testing, the bill remains manageable.
Once agents enter support, coding, research, procurement or back-office production, volume multiplies. Thousands of small decisions become a permanent stream of consumption.
Ramp launched AI Token Spend Management in 2026. The company says the product was built with more than 1,300 businesses managing over 100 trillion tokens each month, with spending visible by provider, model, team and project.
That development reveals the direction of the market. AI expense will be monitored as granularly as cloud usage, travel or advertising. The CFO enters the chat.
At that point, the assumption that the best model is automatically the best choice disappears. The additional value must justify the additional cost.
A Token Price Tells Too Little
Providers publish prices per million input and output tokens. Those figures look comparable, but only partly are.
A model may offer cheaper input and more expensive output. Long context may be priced differently from short context. Cache hits, batch processing, tool calls, reasoning tokens and priority tiers all change the bill. Two models may also require very different token volumes to complete the same task.
The relevant comparison is therefore not: Which model costs less per million tokens?
It is: Which system completes the task at acceptable quality, reliability and speed for the lowest total cost?
That total includes:
- tokens and tool calls;
- errors, retries and human rework;
- latency and waiting time;
- storage, security and orchestration;
- the cost of a wrong or incomplete result.
A more expensive model may be cheaper if it solves the task correctly on the first attempt. A low-priced model can become expensive when it produces long loops or requires constant supervision.
Model Routing Becomes Operating Software
Enterprises will not send every task to the same model. They will build a selection system above the models.
complex, consequential task → frontier model
routine classification or extraction → small low-cost model
sensitive data → local or private model
large recurring volume → cache, batch or deterministic workflow
Routing becomes a new control layer. It determines which model receives which data, what level of quality is required and how much money a task may consume.
Power shifts from the individual model toward orchestration. A company that can distribute workloads flexibly can compare prices and switch providers. A company whose agents are deeply bound to one model loses bargaining power.
The most valuable AI system may not be the smartest model, but the router that knows when that model is unnecessary.
Caching, Batch and Smaller Models
Providers themselves reward efficiency. OpenAI, Anthropic and Google offer different forms of caching and batch processing. Repeated context can be reused more cheaply, non-urgent work can be grouped and routine tasks can move to smaller models.
Caching matters because agentic systems repeatedly read the same system instructions, documents or codebases. Recomputing the entire context on every call raises cost without adding knowledge.
Batch processing rewards patience. Not every analysis needs an answer within seconds. Grouping work and processing it outside peak demand can reduce cost and improve utilization.
The next step goes further. Once a recurring agent workflow has succeeded, parts of it can be converted into deterministic software. The model plans once; the stable sequence then runs without repeated reasoning.
AI does not become less important. It moves toward the parts of the process where uncertainty and judgment are actually needed.
Local Inference Changes the Bill
Enterprises can increasingly run sensitive or repetitive workloads on their own hardware. A local model carries upfront equipment cost but avoids a variable API charge for every token.
The economics improve where utilization, privacy and predictable workloads coincide. Internal search, document classification, coding assistance and recurring agents can partly run on workstations, private clouds or edge systems.
Local inference will not replace the cloud. Frontier tasks, peak demand and very large contexts remain data-center workloads. But local systems create a price ceiling for routine work and strengthen the customer’s bargaining position.
The hardware chain gains a new source of demand: not only HBM inside cloud clusters, but also more DRAM, unified memory and powerful local systems.
The memory story becomes broader—and more differentiated. Not every AI task needs the most expensive infrastructure.
The Jevons Effect of Intelligence
Lower cost per task does not automatically reduce total consumption. Cheaper AI makes new applications economical.
When an analysis falls from ten dollars to one, companies may not simply spend one tenth as much. They may run one hundred times more analyses. Employees automate work that was previously never done. Products place AI inside every customer interaction.
This is the Jevons effect: efficiency lowers the price of a resource and can increase its total use.
That is bullish for data centers, chips and electricity. It is more ambiguous for individual model providers. Volume rises while price per task falls and customers route more aggressively.
The market can therefore experience two truths at once: AI usage explodes, and frontier-lab margins come under pressure.
The New Test of AI ROIC
For hyperscalers, the valuation test is changing. High capex was often read as proof of future demand. As cost control spreads, investors will ask more directly how much cash flow each new data-center block generates.
One chain remains powerful:
more agents → more token consumption → more cloud revenue → more data centers and chips
A counter-chain now develops beside it:
higher bill → CFO control → routing, caching and local inference → lower revenue per task
The balance decides the outcome. If the number of productive tasks grows faster than unit costs decline, the infrastructure story remains strong. If prices fall faster than volume grows, revenue and capex expectations will be revised.
This is where token economics connects to the HBM deleveraging loop in Chapter 3. Monetization doubts can delay hyperscaler projects, weaken memory expectations and pressure already leveraged semiconductor markets.
Who Benefits From Cost Control?
The rationalization of AI spending creates new winners:
- Model routers and orchestration platforms: They allocate tasks by price, quality, security and availability.
- Small and specialized models: They absorb large routine volumes.
- Cloud and FinOps tools: They measure usage, detect anomalies and enforce budgets.
- Local hardware and private AI: They provide predictable cost and data control.
- Caching, observability and evaluation vendors: They reduce repetition and make quality measurable.
- Applications with clear economic value: They connect AI cost to productive outcomes.
Business models that confuse token volume with value creation will come under pressure. An agent that computes ten times more is not automatically ten times more productive.
Frontier labs also need more than technological leadership. They need distribution, data integration, customer retention and a way to offset falling unit prices through volume or adjacent products.
The Gridizer Research Watchlist
For token economics and AI monetization, we will monitor:
- AI spend by provider, model, team and task;
- cost per successfully completed workflow;
- retry rates, agent loops and human rework;
- the share of cached, batched and locally executed workloads;
- migration from frontier models to smaller or open alternatives;
- gross margins at model providers;
- AI revenue and savings relative to capex and opex;
- hyperscaler capex, depreciation and free cash flow;
- HBM, DRAM and data-center demand as cost per task declines.
Intelligence Becomes an Operating Expense
The first AI wave won users with spectacular capabilities and simple subscriptions. The next wave will be decided by cost per task, reliability and control.
That changes the list of winners. Models remain important, but their economic value increasingly depends on efficient integration into real workflows. The CFO becomes one of the most powerful AI regulators.
AI can become cheaper, more useful and more pervasive—and thereby put pressure on the margins of the companies that invented it.