For most of 2025 and into early 2026, the instruction coming down from enterprise boardrooms was simple and, in hindsight, dangerously unqualified: use AI more. Executives wanted usage charts climbing and quarterly reviews to show adoption. Some organisations went as far as gamifying it, rewarding the employees and teams who consumed the most tokens as evidence of productivity. The industry now has a name for the resulting behavior: tokenmaxxing.
The reckoning arrived faster than almost anyone predicted. Uber burned through its entire 2026 AI budget in four months, and has since capped spending at $1,500 per tool per engineer after internal leaderboards tied to Claude Code usage turned into an accelerant for runaway costs, according to reporting on the company’s internal governance changes. One enterprise reportedly spent $500 million in a single month on Claude after failing to set usage limits. A healthcare organisation burned through 1 trillion tokens in six months, translating into more than $6 million in unplanned costs before its finance team understood what was driving the number.
These are not isolated horror stories. Deloitte’s research found that half of enterprise leaders are now spending between 21% and 50% of their entire digital transformation budget on AI — a figure that, twenty-four months ago, would have read as a typo. The FinOps Foundation’s 2026 State of FinOps report, drawn from more than 1,190 practitioners overseeing upward of $83 billion in annual cloud spend, found that 73% of organisations reported AI costs exceeded their original projections. Forrester now predicts that a quarter of planned 2026 AI spending will be deferred into 2027 as CFOs push back on initiatives that cannot demonstrate a measurable return.
Meanwhile the average enterprise AI budget has grown from roughly $1.2 million a year in 2024 to around $7 million in 2026, with some Fortune 500 companies reporting monthly inference bills in the tens of millions of dollars. The paradox, industry analysts note, is that the underlying cost of intelligence is falling sharply — the blended price of a token dropped from $18.40 to $6.07 per million between Q1 2025 and Q1 2026 — while the cost of deploying that intelligence keeps climbing. Enterprises are not paying more for AI because models got more expensive. They are paying more because nobody is watching how much of it gets used, on what, and to what end.
A Cloud-Cost Crisis, Except It Compounds Faster
Narasimhadevara draws a direct line between today’s AI spending chaos and the early days of cloud migration, when teams spun up servers and consumed storage with little oversight until the invoice arrived. “Spiralling AI bills are very much similar to how cloud costs were 10 years ago,” he said. “Businesses didn’t realise the repercussions of unchecked usage until the invoice was generated.”
But the analogy has a limit, and it’s an important one. Cloud spend was tied to provisioned infrastructure — a server running idle was at least visible in a dashboard. AI spend is metered per token, per call, per model, across an unpredictable and often invisible chain of prompts, retries, and background agents. “AI consumption is API based and extensively tied to usage, particularly the number of tokens being consumed,” Narasimhadevara said. “That difference makes it a lot more challenging to predict spend in real time.”
The numbers back him up. Tokens do not scale linearly the way seats did. In agentic workflows especially, a single task can trigger dozens of nested model calls, each consuming context that compounds geometrically rather than additively. Only 43% of organisations have a formal AI governance policy in place at all, and just 21% report having mature governance specifically for agentic systems, according to industry survey data from 2026 — a gap that leaves most enterprises applying essentially none of the cost discipline to AI that they already apply to cloud compute.
Where Control Breaks Down: The Leap From Testing to Production
If there is a single inflection point where AI budgets go from manageable to unmanageable, Narasimhadevara places it precisely: the handoff between testing and production. “During the testing phase, AI usage happens in a controlled environment, but once it reaches production, it is a whole different ballgame,” he said. “Usage becomes unpredictable and depends on a lot of factors, such as traffic, that remain outside teams’ control.”
He points to two early warning signs that technology leaders routinely miss. The first is what happens to system reliability as AI usage scales. The second, more subtle signal, is where developers are actually spending their time after the code ships.
“I have seen tech leaders marvelling at and celebrating the speed at which AI writes code… In this scenario, a serious question often gets pushed into the background: how long does it take to review and validate that code? Because that human review can become the bottleneck,” says Ganesh Narasimhadevara.
That bottleneck is now showing up in hard numbers. New Relic’s 2026 State of AI Coding report — a survey of 200 U.S. technology decision-makers conducted with Hanover Research — found that 94% of leaders rate AI-generated code as higher quality than human-written code at the point of review. But once that code ships, 78% report a measurable spike in production incidents, 86% report an increase in senior-engineer firefighting, and 74% say at least a quarter of all AI-generated code needs significant post-deployment rework. Eighty-two percent had at least one major production failure tied to AI code in the prior six months.
Independent telemetry tells a strikingly similar story from a different angle. Faros AI’s 2026 Engineering Report, drawn from two years of systems data across 22,000 developers and 4,000 teams, found that the ratio of incidents to pull requests more than tripled under high AI adoption, bugs per developer rose 54%, and the ratio of code deleted to code added climbed 861% — engineers spending their time undoing work that was approved too quickly the first time. New Relic separately found that 62% of teams often or always ship AI-generated code without line-by-line manual verification.
“Even when AI dramatically speeds up coding, if engineers spend their valuable time reviewing that code, this review time can defeat the very speed advantage AI brings to the table,” Narasimhadevara said.
“In this case, AI could be making development faster, but it is also creating a lot of expensive busy work.” New Relic’s own researchers have taken to calling the resulting backlog of unvetted logic “agent debt” — a deficit that, like technical debt before it, accrues quietly and comes due at the worst possible moment.
The Models Aren’t the Problem – The Prompts Are
It would be easy to blame the frontier labs for enterprise sticker shock. Narasimhadevara doesn’t. “The AI models themselves are not a problem. In fact, they are actually getting cheaper and faster,” he said. “The real issue is how organisations use them.”
He offers a blunt illustration: teams building AI agents capable of ingesting documents of any length to answer a question — even when the question only requires a yes-or-no answer. “Do you really need to feed a 500-page PDF to the model every time?” he asked. “You can, but that is hardly the most efficient way to use AI. It spikes token consumption without necessarily driving any additional business value.”
Market data shows how expensive that habit can be. The spread between the cheapest and most expensive frontier models now exceeds 4,500 times on a per-token basis, and most enterprises default to whichever model is most visible or most capable — routinely routing simple FAQ-style tasks to models that cost fifteen to thirty times more than necessary. One enterprise cut its monthly API bill from $40,000 to $24,000 without a single product change, simply by routing simpler subtasks to cheaper models. Organisations running a tiered model architecture report a median blended cost near $2.31 per million tokens, versus $18.40 for those routing every workload to a frontier model by default — an eightfold difference for often-identical outcomes.
Where the Tokens Actually Go — and Why CIOs Are Surprised
Ask most CIOs which function drives their AI spend and they’ll point to customer-facing chatbots or coding copilots. Narasimhadevara says the real answer is less visible and, engineering teams included, less flattering. “As of now, software development, especially coding agents, contributes the most to AI token consumption,” he said. “Surprisingly, what I learned from our customers is that more than front-end, tokens used by background processes add more to the cost.”
An internal bot routing IT tickets. A summarizer chewing through hundreds of emails a day. A customer-service agent re-fetching the same answer for the hundredth time. None of these show up in a quarterly product review, and all of them run largely unsupervised. “These back-end processes often run with little human oversight, and teams can often miss it when tracking AI consumption,” Narasimhadevara said.
“When these processes run autonomously at a large scale, they can silently drive up consumption behind the scenes without leaving obvious traces. By the time they identify which processes are driving the consumption, it’s already too late.”
Observability as the Missing Bridge
New Relic’s central prescription — unsurprisingly, given its business — is observability purpose-built for AI, not bolted on after the fact. But Narasimhadevara’s case for it rests less on the vendor pitch and more on a simple gap he sees across nearly every enterprise he talks to: businesses can see the bill, but not what generated it.
“Observability is the bridge between what is being generated and what is actually running in production… Visibility at the individual code and trace level is what businesses really need. With the right level of visibility, teams will be able to correlate the exact prompt given by the end user, the number of tokens consumed, how long the model took to respond and the feedback from the end user, all in one place,” says Ganesh Narasimhadevara, Director of Solutions Consulting, New Relic India.
He offers a scenario that should unsettle any CIO comfortable with top-line dashboards: a single generative AI feature consuming 40% of the company’s entire AI budget while carrying an average two-star user rating. Without granular, trace-level telemetry, that mismatch stays invisible until someone in finance finally asks why the bill doubled. “If AI costs suddenly increase by 30% overnight, businesses should be able to immediately identify whether it is good news, such as a new product that has gained customer traction, or bad news, such as a single developer accidentally triggering a loop that runs the same prompt 50,000 times,” he said.
That granularity matters more as enterprises deliberately diversify their model portfolios. “A multi-model strategy is no longer a choice today, rather a necessity,” Narasimhadevara said, noting that most large organisations already route workloads across providers such as OpenAI, Anthropic, and Google, alongside open-source models, to avoid single-vendor dependence. The governance cost of that diversification is real: “Every provider has its own methods, devices and ways of tracking token consumption. Managing all of this through spreadsheets is not practical… the amount of data generated across multiple AI models makes manual tracking feel like finding a needle in a haystack.”
It’s a challenge the industry is now trying to standardise around. The Linux Foundation launched the Tokenomics Foundation in mid-2026, explicitly modeled on how the FinOps Foundation professionalized cloud cost discipline a decade earlier — and the FinOps Foundation itself now reports that 78% of FinOps practices sit inside the CTO or CIO organization, up eighteen points since 2023, a sign that AI cost governance is being pulled squarely into the technology leader’s remit rather than left to procurement.
Why AI Observability Will Outgrow APM
Application performance monitoring has anchored IT operations for more than a decade by watching deterministic systems: code that behaves the same way twice. Narasimhadevara argues AI breaks that assumption entirely, and with it, the adequacy of legacy monitoring. “AI applications are different as they are inherently non-deterministic,” he said. “They can give a completely different answer to the exact same prompt at different times. Agents can hallucinate, get stuck in loops or behave in unexpected ways.”
The consequence, he says, is that AI observability has to do everything APM already does and then layer on two dimensions APM was never built to track: financial spend and output quality. “Together, they provide the bottom-line safeguard enterprises need as they embed AI into every business application,” he said.
Three Moves for a CIO Whose AI Spend Has Doubled
Pressed on what he would tell a CIO staring at AI expenditure that has doubled without a corresponding lift in business value, Narasimhadevara offers a sequence rather than a single fix: define, observe, optimise.
Define how success gets measured — and stop rewarding raw consumption. “Don’t celebrate token consumption. If one engineer is consuming a lot of tokens, don’t give them a Rockstar of the Quarter award,” he said. “Instead, define the cost per successful outcome and tie AI initiatives directly to the number of valuable features shipped and the business outcomes they deliver.”
Observe before acting. “Stop flying blind and instrument your AI applications as soon as possible. You need visibility into performance and token consumption and user feedback alike to understand which AI features are actually delivering value and which ones may need to be reworked or removed.”
Optimise what’s already built, rather than simply spending less. “Audit and refactor the AI architecture you have already built. Avoid ‘kitchen sink prompts,’ where you give the model everything as part of the prompt,” he said. “Match the model to the task and use the most efficient model for the specific job. Don’t use an expensive model or an unnecessarily large prompt when a simpler, more efficient approach can deliver the same outcome.”
The goal, he is careful to add, is not austerity for its own sake. “It is necessary to understand that the goal is not to reduce AI consumption for the sake of reducing consumption. It is to make sure every token consumed contributes to a meaningful business outcome.”
The Bottom Line
The tokenmaxxing era was, in retrospect, a predictable phase of a new technology’s adoption curve — the same over-provisioning, under-governing pattern that defined the early years of cloud, except compressed into months instead of years and metered at a granularity finance teams have never had to model before. The difference this time is speed: where cloud waste accumulated gradually enough for FinOps to mature alongside it, AI waste is compounding inside agentic workflows that multiply calls geometrically, inside background processes nobody is watching, and inside a review burden that quietly cancels out the productivity gains everyone was celebrating.
The data leaves little room for complacency. Three in four organisations have already blown past their AI cost projections. Nearly eight in ten are seeing more production incidents from the very code that looked flawless at review. And the enterprises still treating token consumption as a badge of productivity, rather than a cost to be justified against outcomes, are the ones most likely to be explaining a surprise nine-figure line item to their board next quarter.
The organisations pulling ahead are not the ones spending less on AI. They are the ones who can answer, for every dollar of AI spend, the only question that actually matters: what did the business get back for it. As Narasimhadevara puts it, the era of rewarding usage for its own sake is over. What replaces it is a discipline the cloud industry already had to learn the hard way — except this time, enterprises don’t have a decade to figure it out.