Almost every 2027 AI plan being written this quarter contains an assumption nobody states out loud: that compute will be available, affordable and instant.
For the last few years that held. It is now the shakiest line in the plan.
Gartner expects 40% of existing AI data centres to be operationally constrained by power by 2027. Not by chips — by electricity. AI data centres are heading for roughly 500 TWh a year, up from 260 TWh in 2024, and 83% of operators believe AI workloads will exceed their current capacity within two years. Half are responding by retrofitting existing sites for efficiency rather than building new, for the obvious reason: you can add racks in weeks, but a grid connection takes years.
The bottleneck moved from GPUs to power, and power has lead times measured in years.
Why this matters even if you'll never build a data centre
You don't need to own infrastructure to feel this. Constrained supply shows up in the things a cloud buyer does see:
- Regional availability — the model or instance type you designed around isn't available in the region you need it in.
- Price movement — the assumption that per-token costs only ever fall stops being reliable for premium, always-on, low-latency capacity.
- Queueing and throttling — "instant" quietly becomes "usually instant", which matters enormously if you've built a synchronous customer-facing experience on top of it.
None of that stops an AI programme. All of it breaks a plan that assumed otherwise.
Four moves for a capacity-aware 2027 plan
- Treat right-sizing as a capacity strategy. We've argued for a while that you should use the smallest model that does the job because it's cheaper and faster. The 2027 argument is stronger: a small model is the one you can still run when capacity tightens. Right-sizing per task is now resilience, not just thrift.
- Budget for the curve, not the quote. Unit costs may keep falling while your total bill rises, because usage grows faster than price drops — the classic trap of usage-based economics. That's the same discipline behind the FinOps of AI: track spend against value, per use case, continuously.
- Architect for portability. If your provider's region is constrained, can you move the workload? Abstracting the model behind your own interface, avoiding provider-specific features in the hot path, and keeping an evaluated fallback model are cheap insurance today and expensive to retrofit later. Lock-in has become a capacity risk, not only a commercial one.
- Batch what doesn't need to be real-time. A surprising share of enterprise inference — document processing, summarisation, enrichment, classification — has no genuine latency requirement. Moving it off premium synchronous capacity creates headroom for the work that genuinely needs to be instant.
The thing not to conclude
This isn't an argument to slow down on AI, and it isn't a reason to over-provision defensively. Constraint rewards efficiency, and efficiency is mostly engineering discipline you should want anyway: the right model per task, measured cost per outcome, portable architecture, and honest separation of real-time from batch.
It is an argument to stop treating compute as infinite in your planning assumptions — and to put a capacity line in the 2027 budget next to the licence line.
The bottom line
The constraint on enterprise AI in 2027 is shifting from model quality, past cost, to physics. Plans that assume unlimited cheap instant compute will meet reality somewhere in the year. Plans that right-size per task, budget for the usage curve, stay portable and batch the batchable will absorb it without drama — and spend less along the way. That engineering pragmatism is what our machine learning delivery is built on.
Want a capacity-honest view of your 2027 AI architecture? Let's look at it.