The loudest AI conversation right now is about agents, and it is a distraction from a harder one. Frontier models are becoming expensive to run at scale, and very few leaders have worked out where their own line sits between renting that capability and owning it.
The question nobody is costing
Ask most leaders how they will keep AI from quietly draining the company as usage grows, and there is no answer, because the conversation has been about capability, not cost. As demand moves from occasional to constant, running everything on a third-party frontier model can become the single largest line nobody planned for. This is the waste I pointed at in Gartner’s trillion-dollar IT spend forecast, most of it wasted, and the reflex to simply cut the budget, which I argued against in AI Budgets Are Being Cut. This Is the Moment to Move, is the wrong response to a real problem.
There is a crossover point, and most companies have not found theirs
Let me be precise, because this is where the argument is usually made badly. It is not true that every company should stop using the frontier labs and build a data centre. That advice gets dismantled by the first competent CTO, and rightly. Renting frontier models is cheap and correct at low or spiky volume. Owning the platform wins at heavy, steady, around-the-clock inference, running many models in parallel where the company profits from every second of use.
So the real question is a crossover calculation: at what volume does owning become cheaper than renting for us? Most companies have never done that maths, which is why the decision is driven by fashion in both directions, either overspending at the frontier or refusing to look at ownership at all.
The honest cost of owning it
If you do cross that line, be honest about the full bill, not the flattering version. It is not only hardware, electricity and cooling. The GPUs cost real money and depreciate fast. The talent to run the platform is scarce and expensive. There is security to manage, and there are model updates to keep pace with. And self-hosted open models may still fall short of the frontier on the hardest tasks, though they are now good enough for a large and growing share of what a business actually runs day to day. This is the open-source pressure I wrote about in AI Bubble: Navigating the Future of AI.
I am not arguing this from a slide. As my own demand has grown to a genuine around-the-clock load, I have started pricing hardware to run open models locally, and even against that full cost, at my volume it still comes out ahead. That is the test each leader has to run for themselves, not assume.
Operational maturity is the real dividing line
Here is the part that ties everything together. The companies that will handle the cost wall well are the same ones that will be calm about regulation and steady under pressure. It is not the cleverness of the AI that separates them. It is operational maturity: the discipline to do the boring maths, to plan the infrastructure, to know their own numbers. That single trait decides who thrives and who is quietly overrun, more than any amount of spending without direction, which is the failure I described in You Didn’t Lead Your AI Project. You Just Bought Shovels.
Agents are the conversation everyone is having. The cost wall is the one that will decide who can still afford to run the company in a few years.
If you are leading your organisation through this, I work with a limited number of senior leaders each quarter. Get in touch at Anglero.com.
Thomas Anglero is a Strategic AI Advisor, keynote speaker and author of Intro to Artificial Intelligence. He has delivered over 450 keynotes across 30 countries for organisations including IBM, the WHO, the World Government Summit and the European Commission. He founded the IBM Watson AI Lab for Cancer at the Oslo Cancer Cluster and closed over $500 million in enterprise transformation deals as CTO and Chief Innovation Officer at Cognizant.
Frequently asked questions
Should companies stop using frontier AI models and self-host instead?
Not as a rule. Renting frontier models is cheap and correct at low or spiky volume. Owning the platform only wins at heavy, steady, around-the-clock use. The right answer depends on where your crossover point sits.
What is the crossover point for AI infrastructure?
The volume at which owning and running your own models becomes cheaper than renting them from a frontier lab. Most companies have never calculated theirs, so the decision gets driven by fashion in both directions.
What are the hidden costs of self-hosting AI?
Beyond hardware, electricity and cooling, there is fast GPU depreciation, scarce and expensive talent, security, and constant model updates. Self-hosted open models may also fall short of the frontier on the hardest tasks.
What actually separates companies that handle AI cost well?
Operational maturity. The discipline to do the boring maths, plan the infrastructure and know their own numbers is what decides who thrives, and it is the same trait that keeps a company calm about regulation.
Just over half of employees admit to using AI tools their company has not approved, rising to about two-thirds in the United States. At the same time, most executives say they are confident they can see how AI is used across the organisation. Both of those cannot be true, and the gap between them is a leadership signal, not an IT problem.
Why people reach for the tools you did not give them
Employees are not being difficult. They have been handed a handicapped tool, often an older, limited version of Copilot with little of the capability of what is freely available, and they want the best tools to do better work faster. That is the whole story. The company bought a tool so leadership could say it provided AI, and nobody wants to use it.
The arithmetic nobody has done
Here is a question worth sitting with, and I am putting it as a question, not a claim.
Companies report that employees burned through an entire year’s AI budget in a few months. Separately, a majority of the real work is running on employees’ own personal tools, not the company’s. So how much of the official AI spend actually produced anything?
Nobody would pay out of their own pocket for a private tool to do their job if the company tool did the job. People use shadow AI because it works and the sanctioned tool does not. Put the two facts side by side and the uncomfortable read is this: the internal effectiveness of the company’s own AI programme may be far worse than the public numbers suggest. The official spend bought the appearance of adoption. The actual value showed up on tools the company is not paying for and cannot see. It is the value-gap I set out in Why Your AI Adoption Isn’t Driving Business Value, and the mirror image of releasing a budget with no direction, which I described in You Didn’t Lead Your AI Project. You Just Bought Shovels.
Why this is a trust problem, not a control problem
You can lock the tools down, and you will simply push the best work further into the shadows, or push your best people out of the door. The instinct to control is the wrong instinct. The signal to read is that your people do not trust the tools you chose, and by extension the judgment behind them. Those are often the same quiet high performers I described in What Happens When Your Best People Start Using AI.
A visibility figure of near-total confidence, sitting on top of half the workforce hiding its most-used tool, is not visibility. It is a comfort blanket, and governing on it means governing a fiction. If your reflex right now is to cut the AI budget rather than fix the tooling, I would read AI Budgets Are Being Cut. This Is the Moment to Move first.
The honest version of the visibility question is not can we see AI use. It is why is half our company hiding its best tool from us.
If you are leading your organisation through this, I work with a limited number of senior leaders each quarter. Get in touch at Anglero.com.
Thomas Anglero is a Strategic AI Advisor, keynote speaker and author of Intro to Artificial Intelligence. He has delivered over 450 keynotes across 30 countries for organisations including IBM, the WHO, the World Government Summit and the European Commission. He founded the IBM Watson AI Lab for Cancer at the Oslo Cancer Cluster and closed over $500 million in enterprise transformation deals as CTO and Chief Innovation Officer at Cognizant.
Frequently asked questions
What is shadow AI?
The use of AI tools that a company has not approved. Just over half of employees report doing it, rising to about two-thirds in the United States, usually because the sanctioned tool is weaker than what is freely available.
Why do employees use unapproved AI tools?
Because the approved tool is often a limited version with little of the capability of the best tools, and people want to do better work faster. Nobody pays for or risks a private tool if the company one does the job.
Is shadow AI a security problem or a trust problem?
Primarily a trust problem. Locking tools down pushes the best work further out of sight, or pushes your best people out. The real signal is that people do not trust the tools leadership chose.
What does shadow AI reveal about official AI spend?
Possibly that it achieved less than reported. If a year’s AI budget is consumed in months and most real work runs on personal tools, the company’s own programme may be far less effective than the public numbers suggest.
For the first time, executive surveys show CEOs ranking AI as their single biggest business risk, ahead of war, cyber attack and recession. The real fear underneath that ranking is not the technology. It is being exposed for spending on AI without ever having a strategy for it.
The fear they will not name in the boardroom
A CEO cannot control a war or the economy. They can control their budget, and their AI budget has exposed them. Money went out so the company could be seen doing AI, so the leader could stand up and say this too is an AI-first company, and now the bills have arrived with nothing built.
It is the Ferrari in the driveway. Bought to impress, and for that first moment it works, every friend sees it. Then come the payments, the insurance, the mechanic every time it breaks, the maintenance that never stops. AI behaves exactly the same for a leader who bought it so as not to be accused of not doing AI. You have to get something back. You have to build from it. What these leaders are afraid to tell their board is one sentence: I have no AI strategy, I did it so I would not be accused of not doing AI.
The two confessions
Depending on the leader, the confession takes one of two forms.
The first is the reckless one. I had no plan, I moved because of the hype and the pressure, and the money is gone.
The second sounds far more responsible, and it is the more dangerous of the two. I saved us a fortune by not overspending like everyone else, and I have never put this company at greater risk. No counterforce. No ability to react in any reasonable timeframe. It is the second CEO, the one who is quietly proud of holding back, who should worry you most, because prudence is the disguise.
Why the cautious CEO is in more danger than the reckless one
The reckless spender at least learned something. Many are now tweaking the model and putting the right people in the right place to produce an outcome. The cautious non-adopter is betting that the future will look like today. When a competitor implements AI properly, the loss of market share and revenue can be permanent, and closing the gap takes years the non-adopter does not have. My related argument on why cutting the budget now is the wrong reflex sits in AI Budgets Are Being Cut. This Is the Moment to Move.
Then there is the entrant you never see coming. Google did not win as a better email company. It walked into markets no one expected it to enter, funded a free product with a different model, and took the whole market while the incumbents watched. AI lets new entrants do precisely that, at speed, from directions your competitive map does not show.
What this is really about: the board
Boards are risk-averse by generation and by habit, and a cautious CEO can talk a cautious board into waiting. The mistake is at board level. Boards are relying on a CEO who is too busy to think in this new way, and they trusted existing partners who sell consulting hours rather than build understanding. I wrote about that failure directly in Your Traditional Partners Are Failing You in the Age of AI, and about why the board is so often the furthest behind in Why Boards Are the Furthest Behind on AI.
What a mature board does now is build a relationship with a trusted advisor: someone knowledgeable, who has the board’s trust, acting as the bridge from board to CEO. That is the differentiator, not more hours of slideware. It is also why the pressure is real for the individual at the top, a point I made in Why a CEO Will Be Fired Over a Failed AI Implementation.
The question is not whether you bought the Ferrari. It is whether you can afford to keep it, and whether anyone on your board is asking.
If you are leading your organisation through this, I work with a limited number of senior leaders each quarter. Get in touch at Anglero.com.
Thomas Anglero is a Strategic AI Advisor, keynote speaker and author of Intro to Artificial Intelligence. He has delivered over 450 keynotes across 30 countries for organisations including IBM, the WHO, the World Government Summit and the European Commission. He founded the IBM Watson AI Lab for Cancer at the Oslo Cancer Cluster and closed over $500 million in enterprise transformation deals as CTO and Chief Innovation Officer at Cognizant.
Frequently asked questions
Why do CEOs now rank AI as their biggest business risk?
Because it is the one major risk they created and control. War, cyber attack and the economy are external. The AI budget is theirs, and in many companies it was spent without a strategy and returned little, which exposes the leadership rather than the technology.
What is a for-show AI investment?
Money spent so a leader can say the company is doing AI, with no plan to get anything back. It buys the appearance of an AI-first company and leaves the bills, the maintenance and the exposure.
Is a cautious CEO who avoided AI spending actually safer?
Often not. The reckless spender at least learned something. The cautious non-adopter is betting the future will resemble today, and when a competitor implements AI well, the loss of market share can be permanent and slow to reverse.
What should a board do about AI in mid-2026?
Stop treating an announced strategy as the finish line, and build a relationship with a trusted, knowledgeable advisor who can act as the bridge between the board and the CEO, rather than buying more consulting hours.
AI budgets are being cut. This is the moment to move.
The wave of companies cutting their AI budgets this summer is not proof that AI failed to deliver. It is proof that most of them deployed it badly. And it is the clearest signal yet that now, while almost everyone else retreats, is the moment to move.
If you are a senior leader reading the headlines about abandoned AI projects and feeling quietly relieved that you held back, that relief is the most expensive feeling in your business right now. The companies pulling back are not closing the door on AI. They are handing you a map of exactly what not to do, and most leaders are too busy feeling vindicated to read it.
What actually failed was leadership, not the technology
When a company announces it is cutting AI spend because the return was not there, the failure being described is almost never the tool. It is leadership, in the same way bad AI output is an accountability problem and not an AI problem. It usually takes two forms.
The first is that people were handed AI with no real training. Not a three-hour online course. Real training is a person nearby who answers the questions as they come and holds people’s hands through the discomfort. Think back to the first time a personal computer landed on your desk and the typewriter was taken away. Nobody knew how to set a margin or delete a line. We were not slow. We were untrained. AI is a far stranger arrival than that, because this tool answers back, and a tool that writes better than you do does not inspire confidence in an untrained person. It triggers insecurity. Hand it out with an instruction to “just use it” and you have not started a transformation. You have started a quiet panic.
The second failure is that nobody named the advantage. Let me give you the picture I trust most. As a boy playing baseball, I watched a weighted ring sit on the dugout floor for half a season. We stepped around it every game. None of us knew what it was for. Then one day a player from the other team slid it onto his bat, took a few heavy swings, pulled it off, and walked to the plate swinging a bat that now felt like nothing. He could hit anything. The advantage had been lying in front of us all season, useless, because no one had named it. That is what most companies did with AI. They put the most powerful tool of our working lives in front of their people and never showed them what it was for.
This is also where the runaway cost comes from, the cost the headlines blame. A trained person narrows quickly. They ask a broad question, then a sharper one based on the answer, then sharper again, closing in on what they need. Costs fall as the questioning gets better. An untrained person stays at the wide top of the funnel, asking general question after general question, all day, all month. The bill climbs and nothing lands. The expensive AI bill was not a technology problem. It was the sound of people who were never taught how to think with the tool. The numbers behave the same way: the P&L only moves after the leader does the work, never after the tool is simply switched on.
The retreat is the opening
Here is the part almost nobody is saying. The pullback is the best news a serious leader has had all year.
A year ago there was no map. Today there is. The market is now full of public, specific failure: which companies overspent, how large they were, where the money went, how the usage ran away. That is not bad news to a leader who is paying attention. It is a free education paid for by your competitors. You can sit down with that data and build a strategy that is sharper, more specific, and far cheaper than anything that was possible before the failures existed, the kind of project that actually moves through its three phases instead of stalling. You can decide which team uses which tool, where an open model fits, where a negotiated rate on usage belongs, and where to put the rails that the early movers forgot. Someone else’s failure is your map.
The timing is the whole point. The time to move is not when the market is charging. It is when the market is retreating. Right now the market is retreating and frightened. That is the opening, and openings close.
You cannot un-give the tool
There is one more reason the “AI was overhyped” conclusion is wrong, and it is the one leaders most often miss. Budgets are being cut. Usage is not.
The people in your organisation who learned to work with AI are not going back, in the same way we never went back to the typewriter once we had the PC. Cut the corporate licence and they will spend twenty dollars a month of their own money. Block it on the work computer and they will use their phone. Once a person has had the experience of doing in four minutes what used to take a month, you cannot take that back. The adoption you are looking at is already permanent. The only open question is whether it happens inside your strategy or around it, and that comes down to who you actually put in charge of it.
What an honest leader does on Monday morning
If any of this lands close to home, the first move is not to panic and cut. It is to find out whether it is you, and to use the tool itself to find out. That is the honest reckoning that has to come before any AI strategy, and it is rarely comfortable. By the end of a single morning you can know which team ran away with the spend, whether there were any rails at all, and where the real failure sat. More than half the time it sits in the same place: no rules, no training, no leadership.
And before you cut a single budget line, ask the question almost no one asks. What did the overspend actually buy? Sometimes a team blows the budget because they are about to ship in three months what should have taken three years. Going over budget is not automatically failure. Sometimes it is the first sign that something is working. Find out before you switch it off.
If you are leading your organisation through this, this is the kind of work I do with a limited number of senior leaders each quarter. Work with Thomas
Questions this article answers
Why are companies cutting their AI investment in 2026? Most companies cutting AI spend are not seeing poor returns because the technology failed. They deployed it without training their people and without rules on how it was used, so costs climbed while results did not. The cut is a response to a leadership failure, not a technology failure.
Is the failure of AI projects a sign that AI was overhyped? No. AI budgets are being cut while AI usage keeps growing. Employees who learned to work with the tool do not stop when a corporate licence is removed; they pay for it themselves or use their phones. Adoption is already permanent, which is the opposite of a technology that failed.
Why do AI costs run out of control in large organisations? Untrained users ask broad, general questions repeatedly, which consumes far more usage and produces weaker results. Trained users narrow their questions quickly and reach the answer with less waste, so their costs fall over time. The runaway bill is a symptom of missing training, not of an expensive tool.
Is now a good time to invest in AI, or should we wait? Now is the stronger moment to move, because the market’s recent failures have created detailed public data on what does not work. A leader can use that data to build a more specific and cheaper strategy than was possible before, while competitors are retreating. The time to move is when the market is hesitating, not when it is charging.
What should a leader do first when an AI initiative overspends? Before cutting the budget, find out where the failure actually sat, usually in the absence of rules and training, and use AI itself to get that answer within a morning. Then ask what the overspend bought, because a team that exceeded its budget may be about to deliver years of value in months. Cutting before understanding can switch off the one thing that was working.
Thomas Anglero is a Strategic AI Advisor (MerkabaPhi AS, Oslo), with 450+ keynotes across 30+ countries. Enquiries: anglero.com
Gartner now expects worldwide IT spending to reach $6.31 trillion in 2026, a rise of 13.5% on 2025, raised from $6.15 trillion earlier in the year. Almost all of the new money is going into AI infrastructure.
If there is an AI line in your budget this year, you are part of that $6.31 trillion. Here is the uncomfortable part. Most of it buys infrastructure that changes nothing about how your business actually works, and the leaders who confuse the two will spend a fortune and have nothing to show their people for it.
A forecast that keeps climbing
When I wrote about Gartner’s outlook in late 2024, the open question was whether enterprises would pull back. Through 2025 many did, in what Gartner described as an uncertainty pause, a deliberate hold on net-new initiatives. Spending rose anyway, and the 2026 number has since been revised upward twice. The growth is concentrated where the AI build-out is loudest: data centre systems, where Gartner expects spending to grow by more than 55% as hyperscalers race to stand up AI-optimised compute, and software and services, where generative AI is quietly lifting the cost of tools companies already run. On Gartner’s figures, IT services alone will pass $1.87 trillion and software will approach $1.44 trillion.
Spending is up while expectations have come down
The detail most leaders skip past is the split Gartner keeps naming. Expectations for generative AI have slid into what it calls the trough of disillusionment, even as spending on it rises. The hard numbers explain the mood: MIT’s 2025 report, The GenAI Divide: State of AI in Business, found that about 95% of enterprise generative AI pilots had delivered no measurable return. Read that again. Organisations are putting more money into AI in the same year they have quietly lowered what they expect it to do. That is not a contradiction. It is exactly what a real technology shift looks like from the inside: after the launch excitement, before the compounding returns.
Buying infrastructure is not the same as getting value
The overwhelming share of this $6.31 trillion is infrastructure: servers, data centres, memory, the plumbing of AI. I have led enterprise transformation from inside this market, at IBM and as Cognizant’s Nordic CTO and Innovation Officer, and the pattern never changes. The organisations that see a return are not the ones that bought the most technology. They are the ones that did the human work alongside it. AI is a culture project before it is a procurement line. The hardware arrives in weeks; the change in how people decide, trust and work takes leadership, and that is the line item that never shows up in Gartner’s table.
What this means for you
If you sit on a board or lead a team, three things follow from this forecast. First, separate the infrastructure decision from the value decision. Knowing the market will spend more than $6 trillion tells you nothing about what your people should do on Monday morning. Second, treat lowered AI expectations as the buying signal, not the warning. The noise is leaving and the real work is starting, and this is the window in which moving first compounds. Third, budget for the human side as deliberately as the technical side, because that is where the return on all this spending is won or lost.
The number will keep rising, and the next wave, the move to embodied AI and robotics, will raise the stakes again. The leaders who benefit will be the ones who invest in capability and culture, not only in compute. If the human side is the part you are wrestling with, that is the conversation I have with leadership teams, and it is worth starting before the budget is spent, not after.
Frequently asked questions
How much will global IT spending reach in 2026?
Gartner expects worldwide IT spending to reach $6.31 trillion in 2026, up 13.5% on 2025, and the figure has been revised upward twice. Almost all of the new money is going into AI infrastructure, with data centre systems alone forecast to grow by more than 55%.
If spending is rising, why does Gartner say expectations have fallen?
Because the two move independently. Expectations for generative AI have slipped into what Gartner calls the trough of disillusionment even as spending climbs. MIT’s 2025 report found about 95% of enterprise generative AI pilots delivered no measurable return, which explains the mood, and it is exactly what a real technology shift looks like after the launch excitement and before the compounding returns.
Why will most of the $6.31 trillion be wasted?
Because the overwhelming share buys infrastructure, servers, data centres, memory, that changes nothing about how a business actually works. The organisations that see a return are not the ones that bought the most technology, but the ones that did the human work alongside it. AI is a culture project before it is a procurement line.
What should a leader actually do in response to this forecast?
Three things: separate the infrastructure decision from the value decision, treat lowered AI expectations as the buying signal rather than the warning, and budget for the human side as deliberately as the technical side. That human line item never appears in Gartner’s table, but it is where the return is won or lost.
Thomas Anglero is a Strategic AI Advisor, keynote speaker and author of Intro to Artificial Intelligence. He has delivered over 450 keynotes across 30 countries for organisations including IBM, the WHO, the World Government Summit and the European Commission. He founded the IBM Watson AI Lab for Cancer at the Oslo Cancer Cluster and closed over $500 million in enterprise transformation deals as CTO and Chief Innovation Officer at Cognizant.