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.