For years, we told ourselves a comforting story. Compute was elastic. Storage was infinite. Bandwidth was “someone else’s problem”. If you needed more, you clicked, scaled, paid, and moved on. That story is over.
In 2026, compute feels less like a utility and more like a supply chain. Not the sexy kind. The real kind, with bottlenecks, queues, physical constraints, and unpleasant surprises that don’t care about your roadmap. GPU scarcity is the obvious headline, but it’s only the first domino. What’s really happening is that AI has dragged IT back into the world of atoms. Power. Cooling. Real estate. Grid capacity. Water. Lead times. Latency. The things we spent twenty years abstracting away are suddenly back on the CIO’s desk.
And there’s a strange irony in that. We digitized everything, and the limiting factor became… electricity.
Data centres already account for a meaningful slice of global electricity consumption, around 1.5% in 2024 (about 415 TWh), and projections point to a sharp increase by 2030, driven in large part by accelerated computing for AI.
That’s the macro view. The micro view is even more concrete: in Europe, new data centre capacity is not slowing because demand disappeared, but because power availability and grid constraints are delaying what can actually be built and connected. So “compute” stops being a line in a cloud invoice. It becomes a planning problem, an energy problem, and sometimes a political problem.
Cooling is the next reality check. High-density AI racks are pushing beyond what traditional air cooling can handle, which is why liquid cooling and other thermal strategies are moving from “interesting” to “necessary” in modern designs. At this point, a CIO can’t pretend the infrastructure layer is abstract. If your workloads depend on accelerated compute, then your delivery capacity depends on power density and thermal engineering. Not just architecture diagrams.
Then comes latency, the quiet killer. The more GenAI becomes embedded into real workflows, the less tolerance people have for “it’ll answer in a few seconds”. For some use cases, latency is not a user experience detail, it’s the difference between adoption and abandonment, and sometimes the difference between a safe decision and a rushed one.
Vendor lock-in also changes flavour. It’s no longer only about software licenses or cloud contracts. It’s about who can actually deliver capacity, where, under what constraints, with what lead times, and with what dependencies in the stack. When compute is scarce, procurement starts looking like risk management.
Finally, added to this challenge is an element that is becoming increasingly important and sensitive in the current geopolitical context: energy independence. If compute becomes the lifeblood of business, companies will inevitably have to ask themselves how dependent they are on energy ecosystems over which they have little or no control. They must ask themselves whether this risk is acceptable.
What does this mean for the CIO role. It means we now own a chunk of physical reality again. Not because we want to, but because the business is betting on capabilities that are constrained by physics. It also means the right conversations shift. Instead of “should we do AI”, the question becomes “where do we place scarce compute, and what do we stop doing so our most valuable workloads can run”. Instead of “cloud strategy”, it becomes “capacity strategy”. Instead of “FinOps”, it becomes “compute economics plus energy economics”.
I’ve seen organisations treat this like a temporary inconvenience. It isn’t. Even the energy community is explicitly framing data centres as a material driver of electricity demand growth in Europe toward 2030.
Which leads to a pragmatic conclusion.
If compute is the new supply chain, then we need supply chain discipline. Visibility on demand. Prioritisation mechanisms. Multi-sourcing where it’s realistic. Standardisation where it reduces fragility. And a ruthless focus on what actually creates value, because scarcity has a way of making strategy honest. The good news is that this is also an opportunity. When IT returns to the world of constraints, it also returns to the world of trade-offs that executives understand instinctively. Scarce resource. Competing demand. Risk exposure. Resilience. Cost of delay. In other words: a language the board already speaks.
Maybe that’s the real shift of 2026. Compute is no longer a background utility. It’s becoming a strategic asset with a supply chain, and the CIO is once again one of the people who must make the hard calls about how that asset is allocated.
Not in PowerPoint. In the machine room.
What does “compute is the new supply chain” actually mean?
It means compute capacity is no longer an infinitely elastic utility. For AI and high-performance workloads, it behaves like a constrained resource with bottlenecks, lead times, dependencies, and allocation trade-offs. You don’t “just scale.” You plan, source, prioritize, and sometimes compromise.
Why are GPUs scarce, and why does it matter for CIO strategy?
GPU scarcity is a mix of explosive demand (AI training and inference), limited high-end manufacturing capacity, and long procurement cycles. For CIOs, it changes the game from “cloud consumption” to “capacity strategy”: what gets priority, what gets delayed, and how you secure access without locking the business into one vendor or one geography.
How do power constraints become an IT problem?
Because accelerated compute is power-hungry, and data centres are often limited by grid availability, permits, and connection timelines. When your AI roadmap depends on infrastructure that depends on electricity and grid capacity, IT is inevitably part of energy planning and corporate risk management.
Why is cooling suddenly such a big topic in AI infrastructure?
High-density AI racks generate heat levels that make traditional cooling approaches inadequate at scale. Cooling becomes a design constraint, a cost driver, and sometimes a deployment blocker. If you can’t cool it, you can’t run it, no matter what your architecture slides say.
What is the link between AI, latency, and adoption?
As GenAI moves into day-to-day workflows, latency stops being a “nice to have” and becomes a productivity and trust factor. If responses are slow or inconsistent, users abandon the tool, revert to old habits, or create shadow solutions that increase risk.
How does vendor lock-in change when compute is scarce?
Lock-in is no longer just about licenses and contracts. It becomes about physical and operational dependency: who can provide capacity, where they can provide it, with what lead times, and with what constraints in networking, data residency, and integrated tooling. Scarcity makes dependencies more expensive and harder to unwind.
What is “capacity strategy” for a CIO?
It’s the discipline of forecasting compute demand, securing supply, and allocating capacity to the highest-value use cases. It includes portfolio prioritization, workload placement (cloud, on-prem, edge), sourcing models, resilience planning, and governance that prevents “GPU sprawl”.
How do CIOs avoid “GPU sprawl” and uncontrolled AI costs?
By treating compute like a scarce asset with rules: clear intake, prioritization, quotas, observability, and chargeback/showback models that reflect real unit economics. If everyone can spin up accelerated compute freely, you will fund experiments indefinitely and starve production.
What’s the role of FinOps in an AI-heavy world?
FinOps becomes necessary but insufficient. You need “compute economics” and “energy economics” together: cost per inference, cost per transaction, utilization rates, idle capacity, and the hidden costs of networking, cooling, storage, monitoring, and governance.
Does this push more workloads back on-prem, or deeper into cloud?
Both can be true. Some organizations will go hybrid to secure predictable capacity and control sensitive workloads. Others will leverage cloud flexibility for burst and global reach. The strategic question is not ideology. It’s risk, time-to-capacity, economics, and constraints in each region.
Why does compute scarcity force better AI portfolio decisions?
Because scarcity makes trade-offs unavoidable. When compute is constrained, you can’t fund everything. You must pick the use cases with clear business value, measurable outcomes, and operational readiness. Scarcity has a way of turning “AI ambition” into actual strategy.
What does “good” look like for CIOs managing compute as a supply chain?
Good looks like visibility, prioritization, and resilience: clear demand signals, workload classification, sourcing options, standardized platforms, utilization discipline, and governance that scales internationally. It’s less “cloud dream” and more “industrial planning,” which is exactly why it works.