The approval gate most teams add to an AI budget in its second year is aimed at the wrong part of the bill, because the bill moved into the margin. In October, 82.0% of measured spend at the median company ran production workloads, up from 78.8% in September [1]. Production spend is cost of revenue — the line that moves with work delivered to a paying customer — so a shipped feature cannot be paused without someone outside the company noticing. What remains is development, testing and experiments, and that is the only part an approval gate can reach. CloudZero asked 252 finance leaders, and the teams that lead with gates killed or paused 20% of AI initiatives [1].
At a glance
- The shift: AI spend now runs mostly in production, which is cost of revenue [1].
- The number: 82.0% of measured spend ran production in October [1].
- The risk: A gate cuts the fifth that can pause, not the four fifths that cannot [1].
Where the bill changed lines
The two documents arrived two days apart. CloudZero published its October panel on 10 November, and on 12 November Gartner forecast global AI spending at $4,050bn for 2027, 47.5% above the $2,746bn it projects for this year [2]. Budgets for the year the forecast describes lock in over the next few weeks. Read together, the market total and the company total show one movement from opposite ends: the money goes into hardware, and inside the company the spending moved from a project line into the margin.
The market money is not in doubt. Hardware and data centres take 59% of the 2027 total, up 53% to $2.4 trillion. Gartner’s John-David Lovelock ties that growth to “faster than anticipated expansion of user requirements into scaling up inference workloads” — the compute that answers each request [2]. That is a forecast about demand, not about seats, and it explains why a company’s own bill keeps growing while its headcount does not.
The volume now comes from software. On OpenRouter, agentic workloads ran about 131 trillion tokens a week by late September, against 18 trillion from human users [3]. Agents resend the same long context at every step, so most of the unit cost sits in the cache hit rate — the share of a request answered from memory instead of being recomputed — and in routing. Models launched this quarter serve 88% of agent traffic from cache, against 55% for models launched in spring 2025 [3]. I keep coming back to who set the default. A pilot is a call one person makes; a shipped feature is a call software makes thousands of times an hour.
By the numbers
- $957.2bn of AI servers: The largest hardware segment, and the forecast’s steepest jump [2].
- $4,050bn of AI spending: Gartner’s forecast total for 2027 [2].
- 9.8% of a cost panel: Share now spending at least a quarter of the cloud bill on AI [1].
Where the gate lands
The survey is the closest thing this month to a comparison group. Twenty-seven percent named new approval gates as their main change for the year. Teams that led with gates killed or paused 20% of AI initiatives, against 7% for teams that tied funding to a unit cost [1]. A gate is not the cautious choice. It aims at the one part of the bill that can be cut, which is the part where a programme learns something. The 23% who can both split AI spend by team and measure what it returned plan to grow at the same rate and lost far less: 3% failed to win an investment, against 19% of the rest [1].
| Control | What it lands on | What it protects | Killed or paused |
|---|---|---|---|
| Approval gate | The fifth of the bill that can be paused | The budget line | 20% of initiatives |
| Unit-cost funding | The four fifths already in production | The programme | 7% of initiatives |
The technical form of that meter is dull, and nobody owns it by default. Agents resend context on every step, so cost per finished unit of work depends on which model a request goes to. It also depends on how much of that context comes from cache, and neither choice shows up in a project plan. My read is that the decision with the largest effect on next year’s margin is a routing default set once and never looked at again, usually by whoever set up the gateway.
The catch
The panel is CloudZero’s own customers: a cost-management vendor measuring companies that bought cost management, so 430 companies is not the whole economy [1]. The split between gates and unit-cost funding is self-reported, and the page shows no method behind it. It describes what finance leaders say, not a controlled result. Gartner’s total is a forecast, and a forecast revised upward twice in a year is still a forecast [2].
None of these numbers says the AI work was worth doing. That question is older, and I have argued here before that the ROI percentage is a definition, not a measurement. What has changed is where the definition now sits: on the margin line, in front of a customer, rather than in a project review a committee can defer.
Related reading
- Your AI Bill Grew 1.65% — the same bill read from the finance side
- Your $200 Cap Still Buys 2.85x — a daily AI budget written in a falling unit
- Cloud Bill Break-Even — more on Mental Models & Strategy
Sources
- [1] CloudZero, “Your AI Economics Pulse for November 2026” (2026-11-10) — https://www.cloudzero.com/blog/ai-economics-pulse-november-2026
- [2] Gartner, “Gartner Forecasts AI Infrastructure Spending to Represent 59% of Total Global AI Spending in 2027” (2026-11-12) — https://www.gartner.com/en/newsroom/press-releases/2026-11-12-gartner-forecasts-artificial-intelligence-infrastructure-spending-to-represent-59-percent-of-total-global-artificial-intelligence-spending-in-2027
- [3] OnlyCFO’s Newsletter, Peter Walker (OpenRouter), “The Current State of AI Tokenomics” (2026-10-22) — https://www.onlycfo.io/p/ai-tokenomics