AI budgets are burning out before year-end, and CFOs are rethinking every token

Annual AI budgets are running dry in one to two months at major U.S. companies. That is not a projection; it is what enterprise technology buyers are reporting directly to their vendors right now, according to CNBC. The dynamic is reshaping how operations and finance leaders think about AI spending and, increasingly, about headcount.

Arvind Jain, CEO of enterprise AI company Glean, told CNBC that AI cost is now the top concern across every enterprise conversation his team has. The culprit: frontier model pricing has not fallen the way buyers expected. Instead, each successive model release has come in at roughly twice the per-token cost of the previous one, putting what Jain described to CNBC as an unsustainable trajectory in front of CFOs who set budgets based on older, cheaper baselines.

Tech priced like people

The cost dynamic has opened a comparison that simply did not exist in previous technology cycles. Jain told CNBC this is the first time he can recall that the price of technology is comparable to the cost of a person, forcing a direct trade-off that historically never came up because technology was always a fraction of operating costs. Growing AI budgets are now coming, at least in part, in lieu of future headcount additions.

Matan Grinberg, CEO of Factory AI, framed it as a concrete resource allocation question playing out inside leadership teams: whether to optimize the number of employees or the AI spend per employee. Speaking to CNBC, he described three distinct phases companies have moved through in roughly a year. First, boards demanded action on AI. Then came what he called tokenmaxxing, deploying AI by any means necessary regardless of cost. Now, leadership teams are in a third phase, actively reassessing whether premium frontier models are necessary for every task.

95% on the expensive tier, and an obvious fix

One concrete number stands out in the CNBC reporting: roughly 95% of enterprise AI usage is still running on the most expensive frontier models, even for tasks that cheaper alternatives could handle without meaningful quality loss. Jain put the available savings from smarter model routing at roughly 10x for organizations willing to direct simpler workloads to lower-cost tiers.

Factory AI's product is built around exactly that premise, automatically matching each task to the most cost-appropriate model. Grinberg illustrated to CNBC how marginal the real-world performance gap often is between successive frontier releases, comparing the difference to that between a professor with 13 years of experience and one with 15. For most enterprise tasks, the distinction is imperceptible.

Adoption is not the same as value

The cost squeeze lands on top of a separate but related problem: widespread AI adoption has not reliably translated into measurable business impact. Commencis, in its 2026 State of AI: Hype, Reality, and What Comes Next report, argues that individual productivity gains, the typical first-wave benefit, do not automatically compound into enterprise-level value. The report's thesis is that realizing that next level of return requires organizations to redesign the structure of work itself, not just deploy AI tools within existing workflows.

Taken together, the two pressures point in the same direction. Buying more AI access at current prices, without changing how work is organized or how models are selected for each task, produces neither cost control nor scalable value. Operations leaders who treat AI as a software subscription rather than as a variable-cost infrastructure requiring active management are the ones most likely to hit the budget wall.

What this means for your team

• Audit model usage now: identify what share of your AI workloads are running on frontier-tier models and whether simpler tasks could be routed to cheaper alternatives without quality impact.
• Rebuild the budget model: per-token pricing has been rising, not falling; finance and IT teams should stress-test AI line items against continued cost escalation rather than assuming the relief that was previously expected.
• Move beyond tool deployment: evaluate whether current AI initiatives are changing how work flows at the process level, not just adding AI features to individual roles, as system-level redesign is where measurable value is more likely to emerge.
• Tie headcount planning to AI cost projections: with CFOs now explicitly weighing token spend against future hiring, operations and HR leaders need visibility into AI consumption data to participate meaningfully in that trade-off conversation.