Gartner Says Enterprises Now Spend More Running AI Than Building It, and a Quarter of Projects Are Getting Killed by Surprise Bills

Enterprise AI spending just crossed a line, according to Gartner: organizations are now spending more money running AI systems than they spent building them in the first place. That sounds like good news. It means the pilots are graduating into real production tools. It also means the invoice most finance teams are staring at doesn't look anything like the one they budgeted for.

Gartner's research, reported by CIO Dive, frames this as a maturity signal. Tools that companies piloted in 2024 and 2025 are getting pushed into full production in 2026, and production costs are running well past pilot costs. What is catching teams off guard is how the pricing model itself works.

The seat-price instinct doesn't survive contact with a token bill

For thirty years, enterprise software pricing meant a number per seat, budgeted once a year, revisited once a year. That's the model IT leaders trained on, according to a CIO.com account from a technology builder who watched a marketing-support AI system move from pilot to daily use. Add ten users, the license line moves in ten predictable increments.

AI doesn't work that way. Every query, every retrieval, every step an autonomous agent takes to finish a task burns tokens, and tokens get metered like electricity, not sold like a subscription, according to that same account. Two employees on the identical seat license can generate wildly different bills depending on what they ask the system to do. Someone summarizing a short email costs almost nothing. Someone running a multi-step research task across a stack of documents can rack up costs many times higher on the same license.

Token prices have actually collapsed. Stanford's 2025 AI Index Report found the inference cost for a system performing at GPT-3.5's level fell more than 280-fold between November 2022 and October 2024. AI looks dramatically cheaper on the price sheet, but adoption itself is the cost driver. More people using the tool for deeper tasks means more tokens burned, regardless of whether the price per token keeps falling.

A quarter of AI projects are getting derailed by cost surprises nobody saw coming

This isn't an abstract budgeting inconvenience. Research from Mavvrik, cited by CIO Dive, found that poor visibility into AI spending is causing roughly one in four businesses to delay or cancel AI projects outright. The FinOps Foundation's State of FinOps 2026 report found that 98% of FinOps teams now manage AI spend as part of their job, up from just 31% two years earlier, according to CIO.com. An entire discipline is scrambling to catch up to a cost model finance departments were never built to track.

The root problem, according to Mavvrik's research, isn't sticker shock from one bad vendor contract. It's accumulated opacity: usage-based pricing stacked on overlapping tool subscriptions stacked on shadow AI adoption happening outside any procurement process finance can see. Nobody signed a purchase order for the employee who discovered a free-tier AI tool and started running it against company data. Nobody budgeted for the agent that decided a task needed forty API calls instead of four.

The infrastructure bill is also climbing, and enterprises will eventually pay for it

This isn't happening in a vacuum. Google raised its capital expenditure guidance to $205 billion, with CFO Anat Ashkenazi citing demand growth and capacity constraints as the drivers, according to CIO Dive. That's one hyperscaler's bet on how much compute enterprise AI adoption is going to consume. It also sets a rough floor for what enterprise buyers will eventually pay once that infrastructure gets monetized through usage fees. Whatever savings show up in falling per-token prices, rising infrastructure demand is pulling in the other direction.

The finance model most companies inherited from thirty years of seat-based software licensing is the wrong tool for pricing something that behaves more like a utility bill. Gartner's framing treats the operational-cost shift as evidence of maturity. Mavvrik's numbers suggest a lot of companies are hitting that maturity milestone with their eyes closed, and roughly a quarter of them are pulling the plug on projects rather than figure out where the money went.

The open question is whether FinOps teams, now managing AI spend at a 98% adoption rate according to the FinOps Foundation, can build the visibility tools fast enough to keep pace with agents that can generate a month's worth of unplanned token consumption in a single bad afternoon.