The Meter Is Running: Forecasting and Governing AI Spend in the Token Economy
AI is the fastest-growing line in most technology budgets and the least predictable. Only 11% of organizations can forecast AI costs to within 10% of actuals. Why the cost of an AI outcome is decided at design time, and the governance changes that make AI spend predictable.
Executive summary
AI has become the fastest-growing line in most technology budgets, and the least predictable one. Only 11% of organizations can forecast their AI costs to within 10% of actuals, down from 15% a year earlier, and a quarter have delayed or cancelled AI initiatives because of unexpected costs. The cause is structural. AI is billed by the token, consumption grows non-linearly with every agent step, retrieval call and reasoning chain, and vendors are moving enterprise contracts from flat seats to metered, pre-committed consumption. The unit price of a token keeps falling while the enterprise bill keeps rising.
This paper explains why traditional forecasting fails for AI workloads, sets out five actions executives can take to reduce surprise and waste, and describes the governance changes needed to run AI as an economic system. Our view is that the cost of an AI outcome is decided mostly at design time: in model selection, context engineering and workflow architecture. Finance can see the bill, but only design changes it. Organizations that put FinOps discipline and engineering discipline in the same room before a workload ship will scale AI with confidence. Those that wait for the invoice will keep reacting to it.
| 11% of organizations forecast AI costs within ±10% of actuals | 25% delayed or cancelled AI initiatives due to unexpected costs | 52% have no clear owner for AI costs | 13% report even basic visibility into AI spend |
|---|
01 A new cost curve
Worldwide AI spending is forecast to reach $2.59 trillion in 2026, a 47% increase on 2025, with spending on AI models alone expected to more than double this year. Enterprises spent an estimated $37 billion on generative AI in 2025, 3.2 times the $11.5 billion of 2024. Nearly every FinOps practitioner (98%) now manages AI spend, up from 31% two years earlier, and many organizations are being told to fund AI from savings elsewhere in the technology estate. More than a third of IT leaders believe they are already overspending on AI.
Value has not kept pace. Although 88% of organizations use AI in at least one business function, only 39% attribute any EBIT impact to it, and about 6% qualify as high performers that attribute 5% or more of EBIT to AI. Roughly a quarter of AI spend is estimated to be wasted, and only about a quarter of organizations have a reliable way to measure the business value of that spend.
That is the underlying paradox: per-token prices are falling while total enterprise AI spending rises, because the number of users, the complexity of models and the intensity of workloads all grow faster than prices drop. Google alone has reported processing about 1.3 quadrillion tokens a month, roughly 130 times its volume of a little over a year earlier. A single token at a fixed capability tier may keep getting cheaper. The tokens an enterprise actually consumes, increasingly weighted toward reasoning and agentic workloads, do not.
02 Why AI costs resist forecasting
Traditional IT budgeting works from seats, servers and last year's run rate. AI breaks all three assumptions. Four forces explain most of the forecast misses we see in the research.
Consumption is non-linear
The tokens consumed by a single request depend on five variables: standing system instructions, retrieved context and conversation history, model choice, output length, and the overhead of retries and orchestration. These compound. A query routed through a retrieval pipeline, a reasoning model and several tool calls can consume one to two orders of magnitude more tokens than a direct prompt to a smaller model. Multi-turn conversations resend the full history on every turn, so a ten-turn exchange can cost ten times as much per turn as a single question. Agents typically make three to ten times more model calls than a simple chat application.
Output tokens usually cost three to five times more than input tokens, and their volume is set by the model at runtime, outside the organization's direct control. A user-facing metric such as queries per day therefore tells finance very little about cost.
Vendors are shifting consumption risk to buyers
In April 2026 Anthropic moved Claude Enterprise customers away from seat plans with bundled usage toward lower seat fees plus token billing at API rates, with committed monthly spend that customers pay whether or not they reach it. Existing customers transition at renewal. OpenAI moved Codex to token metering the same month, and on June 1, 2026, GitHub replaced request limits on every Copilot plan with token-based AI Credits covering input, output and cached tokens. The procurement question has changed from how many seats to buy into how much compute to forecast and pre-pay. Most organizations do not yet have the consumption forecast that question requires.
Visibility is fragmented
AI spend lands in API invoices, cloud bills, SaaS renewals, developer tool subscriptions and corporate cards. Provider invoices show totals by key or project, with no native concept of business unit, application or workload. AI embedded in SaaS products hides token consumption entirely. The infrastructure around each model call (vector databases, embeddings, orchestration, caches, data egress and observability) often accounts for 40-60% of a feature's total cost and never appears on the model provider's invoice.
This is why token cost is rarely the biggest surprise. When enterprises are asked where unexpected AI cost comes from, data platforms rank first and model token costs rank fifth. And while 73% of organizations have AI cost policies, only 13% have basic visibility into AI spend.
Incentives reward consumption
When AI is free at the point of use, people use more of it. More than half of engineers say their organization actively encourages "tokenmaxxing," maximizing AI usage regardless of value. Uber's experience shows what this looks like at scale.
Case in point: Uber's 2026 AI budget. Uber rolled out Claude Code to its engineers in December 2025. Usage rose from roughly a third of engineers in February to 84% in March, helped by internal dashboards that ranked people by consumption. In April, the company's CTO disclosed that it had used its entire 2026 budget for AI coding tools. Typical per-engineer bills ran $150-250 a month, with the heaviest users generating $500-2,000. Uber's COO later said the company could not draw a clear line from rising token spend to more consumer features. Uber then capped spending at $1,500 per employee per month for each agentic coding tool, with exceptions by approval. The tools were not the failure. The budget assumed seat-like costs for a metered product, and the incentives pointed the wrong way.
The aggregate effect shows up in analyst forecasts: more than 40% of agentic AI projects are expected to be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. For some engineering teams, compute already costs more than the people who use it.
03 Five actions to reduce surprise and waste
1. Build a complete inventory first
Most organizations underestimate how many doors AI spend comes through. Start with every model provider account, API key, payment method, AI developer tool and SaaS contract with AI features. A common finding is engineering teams paying provider invoices on team cards that finance has never seen. Bring all of it into one view, normalized with the FOCUS billing specification so token costs sit beside cloud and SaaS costs. Where provider-native tagging is not enough, route calls through an LLM gateway that attaches team, application, feature and environment metadata to every request.
2. Forecast from drivers, not from run rate
Run-rate forecasts fail because AI demand changes shape when a product adds retrieval, switches to a reasoning model or deploys an agent. Build the forecast from its drivers instead, and assign each driver an owner. Model at least three adoption scenarios, because adoption is the variable that broke Uber's budget. A practical rule is to instrument workloads for 30 to 60 days, set budgets at 110-120% of the measured baseline, and alert at 80% and 100%. Track blended cost per token alongside total volume. A stable invoice can hide fast-growing consumption that surfaces once volume outruns price declines.
Figure 1. The drivers of an AI cost forecast
| Driver | What moves it | Natural owner |
|---|---|---|
| Adoption | Rollouts, incentives, new features, users and tasks per user | Business owner |
| Tokens per task | Context size, retrieval design, agent steps, retries | Engineering |
| Model mix | Routing rules and default model choices | Platform team |
| Price per tier | Contracts, commitments, batch versus real-time | Procurement and FinOps |
| Supporting infrastructure | Vector stores, orchestration, egress, logging | Cloud and FinOps |
3. Engineer the cost down at the workload level
The largest savings sit in architecture, and most of them require no change to the user experience. Figure 2 shows typical ranges for the main levers. Routing can go further still: published research on model cascades and routers has reported cost reductions of more than 85%, and up to 98%, against a single frontier-model baseline. Healthcare offers a concrete example of context engineering. One health data company cut model context by 92% by translating FHIR queries into SQL before calling the model, instead of passing raw clinical records into the prompt.
Every optimization needs a quality check beside the cost check. A change that cuts spend by 80% and halves task success has only made failure cheaper.
Figure 2. Typical savings from common token cost levers
| Lever | Typical savings | Effort | Best fit |
|---|---|---|---|
| Model right-sizing and routing | 60-90% | Medium | Most workloads |
| Prompt caching | 50-90% of cached tokens | Low | Stable, lengthy system prompts |
| Batch processing | 50% | Low to medium | Work that tolerates delay |
| Context window management | 20-60% | Medium to high | Conversational and agentic apps |
| Output length control | 10-40% | Low to medium | All workloads |
| Volume and commitment discounts | 10-30% | Low | High, steady volume |
4. Buy AI like capacity
Committed-spend contracts are capacity commitments and deserve the same analysis as reserved cloud capacity. Gather 60 to 90 days of instrumented usage before negotiating. Calculate the break-even utilization for any provisioned throughput, commonly between 50% and 80% depending on the model and provider. For some frontier and reasoning models, provisioned capacity costs more per token than pay-as-you-go even at full utilization, which makes it a performance purchase rather than a discount. Where providers support spillover to pay-as-you-go, size reservations for average load instead of peak.
For sustained high volume, model self-hosting early. Once token production passes a threshold, an on-premise AI factory can deliver more than 50% savings over three years compared with API and neocloud options, with roughly half of the cost sitting outside the GPUs in networking, power, facilities and software. That decision is hard to unwind, so the crossover point should be calculated before spend forces the question.
5. Measure cost per outcome
Cost per token tells you what you paid. Cost per resolved ticket, per completed prior authorization or per drafted clinical note tells you whether it was worth paying. Pair consumption metrics (cost per inference, cost per workflow run, and token yield, the share of tokens that contributed to a usable result) with a business measure such as cycle time, cost to serve or error rate. A model that consumes ten times the tokens but produces an outcome worth a hundred times more is the better buy. A cheap model that needs three retries and a human rewrite is not a saving.
04 The governance changes AI spend requires
Most organizations now have AI policies. Few have cost governance that works at the speed of consumption. More than half have no clear owner for AI costs, with responsibility split across engineering, FinOps, finance and IT. Five changes close that gap.
Name an accountable owner. Token spend sits across finance, which owns budgets; engineering, which makes architectural choices; security, which governs access; and procurement, which manages vendors. Three operating models work in practice: extend the existing FinOps practice, give a platform engineering team ownership of the gateway and cost controls with FinOps reporting on top, or stand up an AI center of excellence once scale justifies it. The model matters less than having one named person with the authority to enforce decisions.
Gate new workloads on economics. Forecast token demand, enforce ROI thresholds and approve only projects that meet defined economic criteria. In practice every production AI workload should carry a token budget, a cost-per-outcome target and an architecture review before launch. Agentic workflows warrant spend modeling of their own, because they carry disproportionate cost risk.
Set guardrails engineers can work within. Publish an approved model list by use case and environment, token limits by request type, data classification rules for what may go to external providers, and a spend threshold above which purchases leave the developer's card and go through procurement. Put cost in the tools engineers already use, including cost estimates in pull request reviews for AI-heavy changes.
Enforce limits at runtime. A runaway agent loop can consume millions of tokens in minutes, so a policy reviewed monthly cannot stop it. Use the gateway to apply per-workload budgets, per-user caps, iteration limits on agent loops and anomaly alerts measured in tokens per minute.
Align incentives with accountability. Move from showback to chargeback as attribution matures, so the teams that drive consumption see it in their own budgets. Retire dashboards that rank people by raw usage. Report AI unit economics to the executive team on the same cadence as cloud spend.
Figure 3. How AI cost governance needs to change
| From | To |
|---|---|
| Annual budget built from seat counts | Driver-based forecast refreshed monthly |
| AI spend scattered across invoices and cards | One attributed view, normalized with FOCUS |
| Usage leaderboards | Cost-per-outcome targets for each workload |
| Policy documents reviewed after the fact | Gateway-enforced budgets, caps and circuit breakers |
| Cost review after launch | Economic gate before launch |
| A shared AI bill nobody owns | A named owner, then showback, then chargeback |
05 What is at stake
Prices will keep moving, sometimes against buyers. When they do, finance and engineering will report the same event in different ways. In a common scenario, the FinOps report shows the AI bill sharply over forecast and traces it to a handful of workloads, while the tokenomics report shows tokens per resolved outcome falling steadily as unit prices rose faster. Both reports are accurate. Acting on the first alone freezes a workload that is improving. Acting on the second alone leaves an efficiency program that nobody has funded. Executives need both views on one page.
Healthcare and life sciences feel this pressure sharply. Ambient scribes alone cost an estimated $300 to $500 per clinician per month, and cumulative software costs may exceed the savings those tools deliver through efficiency, reduced administrative burden and lower attrition. In one healthcare enterprise, token usage grew 8-10% a month, adding more than $6 million in unplanned annualized cost before finance could see the cause. Regulated settings also carry a heavier layer around the model: PHI controls, audit logging, evaluation and clinician review all add cost that a token-only view misses.
06 The Quantum Rise perspective
Quantum Rise builds AI, machine learning and advanced analytics systems, for mid-market companies and enterprise clients. From that vantage point, the most useful thing an executive can know about AI cost is when it gets decided. By the time an invoice arrives, the cost of an outcome has already been set: which model handles which step, how much context each call carries, whether a workflow calls a frontier model fifty times or a small model forty-five times and a frontier model five, and whether anyone measured quality before cutting cost.
That is why we treat unit economics as a design requirement, alongside accuracy, latency and compliance. Our approach rests on a few working habits.
- Define the outcome and its cost target first. Agree what one unit of value is (a coded encounter, a summarized dossier, a resolved query) and what it can afford to cost before anything is built.
- Instrument from the first prototype. Every model call carries attribution metadata from day one, so the pilot produces the baseline the forecast needs.
- Engineer context, not just prompts. In clinical and research data, most retrieved text is noise to the model. Structuring data before it reaches the model cuts tokens and often improves answers.
- Route by step and test every change. Reserve frontier models for the steps where they change the outcome, and evaluate each routing change against quality as well as cost.
- Hand finance a model it can run. Deliver each system with a driver-based cost model and the gateway controls that keep it inside budget.
FinOps asks the right questions about budget, ownership and variance. Engineering holds the answers about models, context and architecture. We work with both sides, and with our technology partners, to connect those questions to the decisions that answer them.
Conclusion
AI will keep getting cheaper per token and more expensive in total, because the value is real and demand follows value. The organizations that scale it well will be the ones that can predict their AI bill and explain it in terms of outcomes. Start with an inventory and a driver-based forecast this quarter. Put an accountable owner and runtime controls in place before your next renewal moves you onto committed consumption. Then make cost per outcome a design requirement for every new AI workload.
Start the conversation. If you want to pressure-test your AI cost forecast, or the economics of a workload before it scales, talk with Quantum Rise at quantumrise.com.
Sources
- Client-provided source material: Token_economics.docx (2026). A compiled research file on token economics and FinOps for AI, drawing on the FinOps Foundation, Deloitte, NVIDIA, SemiAnalysis, Menlo Ventures, McKinsey, Finout, Microsoft (EvalAgentic) and practitioner publications.
- FinOps Foundation, Token Economics working group. "Tokenomics: Managing AI Value in SaaS Model Token Costs." 2026. Included in client source material. https://www.finops.org/wg/token-economics-saas/
- Fuller, M. "Tokenomics and FinOps: How do they fit together?" FinOps Foundation, August 4, 2026. Included in client source material.
- Chen, L., Zaharia, M., and Zou, J. "FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance." Transactions on Machine Learning Research, 2024. https://arxiv.org/abs/2305.05176 (cited in client source material).
- Ong, I., et al. "RouteLLM: An Open-Source Framework for Cost-Effective LLM Routing." LMSYS Org, July 2024. https://lmsys.org/blog/2024-07-01-routellm/ (cited in client source material).
- Flexpa. "How We Used SQL on FHIR to Shrink LLM Context by 92%." Flexpa Engineering Blog. https://www.flexpa.com/blog/sql-on-fhir-for-llm-context-reduction (cited in client source material).
- Gartner. "Gartner Forecasts Worldwide AI Spending to Grow 47% in 2026." Press release, May 19, 2026. See also "Gartner Says Worldwide AI Spending Will Total $2.5 Trillion in 2026," January 15, 2026. https://www.gartner.com/en/newsroom
- Gartner. "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027." Press release, June 2025.
- Menlo Ventures. "2025: The State of Generative AI in the Enterprise." December 9, 2025. https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/
- McKinsey & Company (QuantumBlack). "The State of AI in 2025: Agents, Innovation, and Transformation." November 2025. Findings as reported by CFOtech, November 12, 2025.
- FinOps Foundation. State of FinOps 2026. Findings as reported by Virtasant ("State of FinOps 2026 Signals Expansive Future for Practitioners," March 26, 2026) and Octopus.cloud (2026).
- Harness. 2026 State of AI in FinOps. As reported in "Harness finds AI costs outpace governance in firms," IT Brief UK and CFOtech UK, July 30, 2026. https://www.harness.io/state-of-ai-in-finops-2026
- Mavvrik and Benchmarkit. 2026 State of AI Cost Governance Report. Press release via EIN Presswire, July 29, 2026.
- Mavvrik and Benchmarkit. 2025 AI cost survey of 372 enterprises. As reported by Gross, G., "AI cost overruns are adding up," CIO.com, 2025, and CFO Dive, 2025.
- Merizzi, N., Smith, T., Mittal, N., Churiwala, G., and Kearns-Manolatos, D. "AI tokens: How to navigate AI's new spend dynamics." Deloitte Insights, January 19, 2026. https://www.deloitte.com/us/en/insights/topics/emerging-technologies/ai-tokens-how-to-navigate-spend-dynamics.html
- Deloitte. "The pivot to tokenomics: Navigating AI's new spend dynamics" and "Navigate the economics of AI," 2026. Healthcare example as cited in "Australia's AI budgets are bleeding tokens," iTWire guest research, 2026.
- Anthropic enterprise pricing changes: "Anthropic squeezes enterprises by ejecting bundled tokens from seat deal," The Register, April 16, 2026; "Anthropic shifts enterprise billing to token-based pricing," IT Brief UK, April 17, 2026.
- GitHub. "GitHub Copilot is moving to usage-based billing." GitHub Blog, April 27, 2026. https://github.blog/news-insights/company-news/github-copilot-is-moving-to-usage-based-billing/
- Axios. "AI can cost more than human workers now." April 26, 2026. https://www.axios.com/2026/04/26/ai-cost-human-workers
- Quartz. "Uber COO says AI spending is getting harder to justify." May 26, 2026, citing reporting by The Information, Business Insider and Forbes. https://qz.com/uber-coo-ai-spending-harder-justify-052626
- TechCrunch report on Uber's per-employee spending caps for agentic coding tools, June 2, 2026, as summarized in "Uber Caps AI Coding-Tool Spend After Exhausting 2026 Budget," Let's Data Science, 2026.
- JAMA Health Forum. Commentary on ambient scribes and health care spending. Published January 9, 2026.
- Peterson Health Technology Institute. Report on AI ambient scribes, March 2025. As reported by STAT News, March 27, 2025.
- Flexera. 2026 IT Priorities Report. As reported in "FinOps steps up as AI spend explodes," iStart, 2026.