RBC Capital Markets just published a chart that would make any AI vendor's marketing team weep with joy: 100% of surveyed CIOs are allocating budget to AI and large language model initiatives. Not 95%. Not 99%. Every single one.
That chart is real. But it is also dangerously incomplete.
In the same month that RBC's survey of 100+ chief information officers declared "broad-based enterprise spending momentum into H2 2026," Uber exhausted its entire annual AI coding-tools budget in four months. Microsoft began canceling Claude Code licenses across its Windows and Microsoft 365 divisions. Coinbase set token caps. Walmart imposed limits on duplicative AI requests. Amazon shut down its internal token leaderboard.
The enterprise AI spending story of mid-2026 is not a contradiction. It is two things happening simultaneously: organizations are committing more money to AI than ever before, and they are losing control of how that money gets spent faster than they can build governance around it. The CIOs answering "yes, we're increasing AI budgets" and the CFOs discovering six-figure monthly token bills are in the same companies, looking at the same numbers from different floors.
This is the AI FinOps crisis. And the data from the past two weeks — RBC's survey, Ramp's spending benchmarks, the FinOps Foundation's State of FinOps 2026 report, and half a dozen real-world budget blowouts — tells a story that every enterprise technology leader needs to understand before the H2 planning cycle closes.
What RBC's Survey Actually Found
Rishi Jaluria, the RBC tech analyst who has been one of the more cautious voices on enterprise AI adoption, came away from this survey encouraged. The findings are striking:
- 100% of respondents are allocating budget to AI and LLM initiatives
- 91% said these are entirely new budgets, not reallocated from existing software spend
- 90% expect to increase AI spending in 2026
- Nearly 9 in 10 said token budgets are "manageable"
- More than half already have AI in production; another 35% expect to reach production within six months
- Not a single respondent expects to spend less on software overall
On the vendor landscape, OpenAI dominates: 57% named ChatGPT as the AI model-based service they use most, compared with 12% for Anthropic's Claude. OpenAI also leads on perceived performance, with 44% naming it the highest-performing provider versus 24% for Anthropic.
The long-predicted "SaaSpocalypse" — AI killing traditional software spend — has not materialized. The vast majority expect to spend more on software, and not one respondent expects to spend less.
If you stopped here, the story is simple: enterprise AI is working, spending is accelerating, and the market is healthy. That's the analyst read. Here's the operator read.
What's Actually Happening on the Ground
While 90% of CIOs tell RBC that token budgets are manageable, the FinOps Foundation reports that 98% of practitioners are now actively managing AI costs — up from 31% two years ago. That is the fastest adoption of a cost discipline the industry has ever seen. You don't build the fastest-growing cost governance practice in enterprise IT history because spending is "manageable."
The real-world case studies from H1 2026 tell the ground truth:
Uber rolled out Claude Code to approximately 5,000 engineers. By March 2026, 84% were classified as agentic coding users. The company burned through its entire 2026 AI coding-tools budget in four months. The response: a hard cap of $1,500 per employee per month per tool — Claude Code and Cursor each capped separately. Power users had been running $500 to $2,000 monthly; the average sat at $150 to $250.
Microsoft began canceling Claude Code licenses across its Experiences and Devices division — the teams behind Windows, Microsoft 365, Outlook, Teams, and Surface — redirecting engineers to GitHub Copilot CLI by June 30. The cost structure mismatch was clear: Copilot Enterprise bills a flat per-seat rate; Claude Code charges a base seat plus variable token usage. Finance chose cost certainty over raw capability.
Pylon, a Y Combinator-backed AI software company, discovered it was approaching 150 employees on its Anthropic plan, a point where the bill would triple to $1.4 million. CEO Marty Kausas declared "the era of unlimited spending is over" and began setting token ceilings for non-technical employees.
The word "tokens" appeared in 129 earnings calls in Q2 2026, up from 57 calls the prior quarter — a 126% increase in a single quarter. When CFOs start talking about tokens on earnings calls, the bill has moved from IT budget to board-level concern.
OpenAI CEO Sam Altman himself acknowledged the shift: at the beginning of the year, "people were totally happy with the amount they were spending." Now, these costs are "a huge issue."
The Disconnect Explained
How can CIOs call spending "manageable" while their companies are blowing budgets and setting emergency caps? The answer lies in three structural gaps:
1. The Budget vs. Consumption Gap
CIOs approved budgets. Engineers determined consumption. The gap between planned spend and actual spend is where the crisis lives. Ramp's data shows that 73% of enterprises exceeded their original AI spending projections in H1 2026. The RBC survey captures budget intent. The invoices capture budget reality.
This is structurally different from traditional software. SaaS costs are predictable — $X per seat per month. AI costs are consumption-based, model-dependent, and driven by behavior rather than headcount. An engineer who runs an agentic coding session with Opus 4.8 can generate $50 in token costs in a single afternoon. Multiply that by 5,000 engineers, and you have Uber's problem.
2. The Model-Tier Ignorance Gap
The single biggest driver of unexpected AI costs is model-tier migration — teams upgrading from lightweight models to frontier models for quality reasons, often without finance visibility. The price spread between tiers is enormous:
- GPT-5.4-nano: $0.20 per million input tokens
- Claude Haiku 4.5: $1.00
- Claude Sonnet 4.6: $3.00
- Claude Opus 4.8 / GPT-5.5: $5.00
- GPT-5.5-pro: $30.00
That's a 150x spread from cheapest to most expensive. Routing a simple classification task to a flagship model isn't a small mistake — it's paying two orders of magnitude more for work the cheapest tier handles cleanly. Ramp's data confirms that premium models represented 45.8% of tokens consumed but 55.9% of total cost in April 2026, with premium cost share rising from 5.7% in June 2025.
3. The Agentic Amplification Gap
Traditional AI usage is request-response: one prompt, one answer, one charge. Agentic AI usage is recursive: the agent decides how many steps to take, each step generates charges, and the user has no visibility into the cost until the task completes. You approved the task; the agent determined the bill.
This is why Uber's budget burned fastest: 84% agentic usage means 84% of consumption was running on autopilot. The agent doesn't ask permission before making its fifteenth API call to solve a coding problem. It just keeps going.
Framework #1: AI Spend Health Assessment
Before you can optimize AI spending, you need to know where you stand. Use this assessment to benchmark your organization against the data from RBC, Ramp, FinOps Foundation, and the H1 2026 case studies.
Section A: Spending Visibility (Score 0-5)
| Question | Yes = Score | No = Score |
|---|---|---|
| Can you attribute AI costs to specific teams or projects? | 5 | 0 |
| Do you have real-time visibility into token consumption by model tier? | 5 | 0 |
| Can you identify your top 10 AI cost drivers by workflow? | 5 | 0 |
| Is AI spend reported to finance monthly with variance analysis? | 5 | 0 |
| Do you track cost per AI-resolved task across use cases? | 5 | 0 |
Section B: Cost Governance (Score 0-5)
| Question | Yes = Score | No = Score |
|---|---|---|
| Do you have per-employee or per-team token budgets? | 5 | 0 |
| Is model-tier selection a finance-visible decision for workflows >$500/month? | 5 | 0 |
| Do agentic workflows have per-run cost ceilings? | 5 | 0 |
| Is there an escalation process when spend exceeds projections by >20%? | 5 | 0 |
| Have you tested top use cases on cheaper model tiers? | 5 | 0 |
Section C: Value Attribution (Score 0-5)
| Question | Yes = Score | No = Score |
|---|---|---|
| Can you calculate ROI for your top 3 AI use cases? | 5 | 0 |
| Do you measure productivity gain per $1,000 of AI spend? | 5 | 0 |
| Is there a formal approval process for new AI tool adoption? | 5 | 0 |
| Have you deprecated any AI tool/model that failed to deliver ROI? | 5 | 0 |
| Do you benchmark AI PEPM against industry data? | 5 | 0 |
Scoring:
| Score | Rating | What It Means |
|---|---|---|
| 60-75 | Optimized | You're ahead of 95% of enterprises. Focus on continuous improvement. |
| 40-59 | Maturing | Solid foundation but gaps in governance or attribution. Priority: close the model-tier visibility gap. |
| 20-39 | Reactive | You're managing by invoice, not by strategy. Priority: implement per-team attribution and model-tier routing. |
| 0-19 | Blind | You're in Uber territory — the budget blowout is coming. Priority: immediate spend audit and emergency caps. |
Industry benchmarks (Ramp, April 2026):
| Metric | Median | Average | Top 10% |
|---|---|---|---|
| Monthly AI spend | $2,246 | $140,842 | $73,030+ |
| Per-employee-per-month (PEPM) | $46 | Varies | $442+ (26+ models) |
| Models in use | 9 | 16.5 | 26+ |
If your spend is near the median, AI is a productivity tool. If you're approaching the average, AI is a COGS line item that warrants a dedicated FinOps function.
Framework #2: AI Model Routing Decision Matrix
The highest-impact cost optimization in enterprise AI is not spending less — it's spending smarter. Production teams report 60-80% cost reduction from model-tier routing without visible quality loss. The principle is "default cheap, escalate on evidence."
The Routing Matrix
| Task Category | Recommended Tier | Example Models | Input Cost ($/M tokens) | Escalate When |
|---|---|---|---|---|
| Classification / extraction / lookup | Cheapest | GPT-5.4-nano, Haiku 4.5 | $0.20-$1.00 | Accuracy below eval threshold |
| Structured summarization | Cheapest → Mid | Haiku 4.5, GPT-5.4-mini | $0.75-$1.00 | Source is long or nuanced |
| Multi-step reasoning / planning | Mid | Sonnet 4.6, GPT-5.4-mini | $0.75-$3.00 | Chains break or hallucinate |
| Long-context synthesis | Mid → Flagship | Sonnet 4.6, GPT-5.5 | $3.00-$5.00 | Cross-document accuracy required |
| Agentic orchestration | Flagship | Opus 4.8, GPT-5.5 | $5.00 | Default high; downgrade subtasks |
| Creative / high-nuance output | Flagship | Opus 4.8, GPT-5.5-pro | $5.00-$30.00 | Brand voice or judgment is the product |
Implementation Playbook
Gate 1: Default Cheap. Start every task category at the lowest tier that could plausibly clear your quality bar. Most enterprise AI workloads — email classification, document extraction, FAQ resolution, data formatting — run cleanly on nano/Haiku tier models at 5-25x lower cost.
Gate 2: Escalate on Evidence. Promote only the specific calls that fail an evaluation, not the whole pipeline. If 90% of your summarization tasks succeed on Haiku, route only the 10% that fail to Sonnet. This is the single change that moves the bill most.
Gate 3: Cache Aggressively. Workflows that reuse context — system prompts, reference documents, static instructions — achieve 80%+ cache hit rates. Ramp's data shows Claude Sonnet 4.6 enterprises paid $0.62/M tokens in April 2026 versus the $3.00/M list price. That gap is caching.
Gate 4: Batch the Non-Urgent. Not every AI task needs real-time response. Batch processing during off-peak hours, queuing non-critical analysis, and consolidating prompts reduce both token count and per-token cost. If a workflow can tolerate 30-second latency, it can tolerate a cheaper model.
Cost Impact Example
A mid-size enterprise running 10 million tokens per day across mixed workloads:
| Scenario | Model Mix | Daily Cost | Monthly Cost | Annual Cost |
|---|---|---|---|---|
| Frontier-by-default | 100% Opus 4.8 ($5/M input) | $50 | $1,500 | $18,000 |
| Naive routing | 50% Opus, 50% Haiku | $30 | $900 | $10,800 |
| Optimized routing | 20% Opus, 30% Sonnet, 50% Haiku | $19 | $570 | $6,840 |
| Full playbook (routing + caching) | Same mix, 60% cache hit | $8 | $240 | $2,880 |
Annual savings from frontier-by-default to full playbook: $15,120 per 10M daily tokens. At enterprise scale (100M+ tokens/day), savings exceed $150,000 annually — from routing alone, without reducing capability.
The Vendor Power Shift
The RBC survey reveals a vendor landscape that should concern every enterprise making long-term AI commitments.
OpenAI's dominance is clear but fragile. At 57% most-used and 44% highest-performing, it has the kind of market position that attracts both customers and pricing power. But the same survey shows hybrid pricing models — combining seat licenses with usage-based pricing — have quickly become the preferred way enterprises want to buy AI. That preference favors vendor diversification, not concentration.
Anthropic at 12% usage and 24% performance perception has an interesting position: CIOs who use it think it's nearly as good as OpenAI, but far fewer have adopted it. That gap represents either a sales/distribution weakness or an opportunity — depending on which side of the negotiating table you sit on.
For enterprise buyers, the strategic implication is clear: multi-model is now a cost strategy, not just a resilience strategy. The 25x price spread across model tiers means that locking into a single vendor's flagship model is the most expensive architectural decision you can make. The companies that will spend the least per unit of AI value delivered in H2 2026 are the ones routing dynamically across providers and tiers.
What the "No SaaSpocalypse" Finding Really Means
The RBC finding that traditional software spending isn't declining is significant — but not for the reason most coverage suggests.
The narrative has been that AI would cannibalize existing SaaS budgets: why pay for Zendesk when an AI agent resolves tickets? Why pay for Salesforce when an AI handles customer outreach? RBC's data says that's not happening yet.
But the "yet" is doing heavy lifting. The survey also shows that 91% of AI budgets are net-new — not reallocated from existing software. That means enterprises are adding AI spend on top of existing software spend, creating a structural cost increase. The total technology bill is going up, not staying flat. The question isn't whether AI replaces SaaS — it's whether the combined spend delivers enough ROI to justify the increase.
Gartner's projection adds urgency: more than 40% of agentic AI projects will be canceled by end of 2027 due to cost overruns, unclear business value, and inadequate risk controls. The projects that survive won't be the ones with the biggest models. They'll be the ones that can draw a clean line from spend to delivered value.
What Enterprise Leaders Should Do Before H2 Planning Closes
The H2 2026 budget cycle is the first where AI spending will be evaluated as a mature line item rather than an experimental allocation. Here's how to prepare.
This Week
-
Run the Spend Health Assessment above. If you score below 40, you need an emergency AI spend audit before committing to H2 budgets. The data from Ramp, FinOps Foundation, and the case studies gives you the benchmarks; your invoices give you the reality.
-
Pull your actual token spend for Q1-Q2. Compare it to what was budgeted. If you're in the 73% that exceeded projections, understand why before projecting H2. The variance is almost always model-tier migration (teams upgrading to more expensive models) or agentic amplification (autonomous workflows running uncapped).
This Month
-
Implement the model routing matrix. Start with your three highest-spend use cases. Test each on the cheapest viable model tier. If 80% of tasks pass quality evaluation on a model that costs 10x less, you've found your H2 savings.
-
Set per-team token budgets with escalation paths. Uber's approach — $1,500/employee/month/tool — is aggressive but directionally correct. The goal is not to limit AI usage but to make consumption visible and create accountability. Engineers who know their budget will self-optimize to cheaper models for routine tasks.
-
Demand outcome-based pricing from your AI vendors. Both Salesforce and Zendesk are moving to pay-per-resolution models. If your vendor is still charging per-token or per-seat without outcome attribution, you're paying for activity, not results.
Before H2 Budget Lock
-
Build a FinOps-for-AI function. The FinOps Foundation reports that 98% of practitioners now manage AI spend — the fastest adoption of any cost discipline in enterprise IT. If you don't have dedicated AI cost governance, you're in the 2% that's flying blind. This doesn't require a new team; it requires extending your existing FinOps practice to cover token economics with the same rigor you apply to cloud spend.
-
Model the Jevons Paradox into your forecast. AI token prices are falling — Google, Anthropic, and OpenAI are all competing on price with smaller, more efficient models. But cheaper resources tend to get consumed more, not less. Budget for 2-3x the consumption growth you project, even as per-token costs decline. The companies that under-budget for volume while celebrating price drops will be H2's budget blowout stories.
The Bottom Line
RBC's 100% chart is not wrong. Every enterprise is investing in AI. The market momentum is real, the pilot-to-production transition is happening, and the long-predicted SaaSpocalypse hasn't arrived.
But the same survey that shows universal adoption also reveals a market where nearly half of enterprises have already exceeded their original spending plans — and say they plan to spend even more. That's not fiscal discipline. That's an industry telling itself the tab is fine while the bill keeps growing.
The enterprises that will win in H2 2026 are not the ones spending the most on AI. They're the ones who know — precisely, by team, by workflow, by model tier — what they're spending and what they're getting for it. The 100% are budgeting. The question is how many are governing.
Global AI spending is projected to reach $2.59 trillion in 2026, up 47% year-over-year. AI infrastructure spending alone will hit $487 billion, more than triple 2024's level. The FinOps Foundation now ranks AI cost management as the single most-needed skill for 2026 — displacing every other priority.
The money is flowing. The question is whether it's flowing to outcomes or just flowing.
Continue Reading
- Salesforce Paid $3.6B for the AI That Was Beating Its Own Product. — How Salesforce's acquisition reveals the outcome-based pricing shift
- 5 Acquisitions in 14 Days. The AI Agent Stack Is Being Bought. — The consolidation wave reshaping enterprise AI spend
- ServiceNow Just Launched AI Employees With Managers and KPIs. — Autonomous agents with built-in cost governance
- $60B Bought Cursor. Your Dev Team Is the Product Now. — The AI coding tool market where token costs first exploded
- FinOps for AI Spending: 98% of Teams Now Manage AI Costs — The governance framework for the token economy
Rajesh Beri is Head of AI Engineering at Zscaler and writes about enterprise AI strategy at beri.net.
