Markets ·
Where AI capacity is actually available.
The market nearest you is rarely the right answer. Power availability and time-to-energisation vary so sharply by region that geography is usually the largest lever a buyer holds.
Most capacity searches begin with a default: the market the organisation already uses, or the one with the deepest ecosystem. For latency-bound inference that instinct is often correct. For training workloads, which are the bulk of new AI demand and largely latency-tolerant, it is frequently the most expensive assumption in the plan.
The constraint is power. Land is available, buildings can be built and hardware can be ordered — energised megawatts on a date you can plan against are the scarce good. So the useful question is not where capacity exists but where power arrives soonest at a price you can live with.
The three questions that decide geography
- Is the workload latency-bound? Interactive inference needs to sit near users. Training and batch inference usually do not. This single answer eliminates or opens most of the map.
- What is the sovereignty and residency requirement? Regulatory, contractual or customer-driven. Often stated more absolutely than it actually is, and worth testing before it constrains the search.
- What is the binding date? If it is inside eighteen months, you are shopping existing capacity and the search is narrow. Beyond that, development and secondary markets open up considerably.
Market by market
North America
Northern Virginia remains the deepest market in the world and among the most constrained, with power the limiting factor rather than land or fibre. Dallas–Fort Worth and the wider ERCOT footprint offer materially different queue dynamics and a market structure that rewards flexible load. Phoenix, Atlanta, Chicago and Columbus carry more available power on shorter timelines than the headline markets. Secondary and tertiary markets frequently beat all of them on time-to-energisation, which for a training cluster is often the only metric that matters.
The Nordics
Iceland, Norway and Sweden present the strongest cost-per-megawatt case in Europe for latency-tolerant work: low-cost renewable generation, free cooling for much of the year, and a favourable power-price basis. The tradeoff is distance from European population centres, which is a real constraint for inference and close to irrelevant for training.
UK and Ireland
London and Slough remain the right answer for latency-bound inference serving the UK. Ireland is available where grid position allows it, and grid position — not land, planning or appetite — is increasingly the binding constraint.
Continental Europe
Frankfurt, Amsterdam and Paris form the FLAP-D core and carry the ecosystem depth that comes with it. Where FLAP-D power is unavailable at the density required, the credible alternatives are Nordic for cost and the Iberian peninsula for a combination of renewable generation and improving connectivity.
APAC
Singapore's constrained supply pushed a great deal of capacity across the strait, and Johor and wider Malaysia now carry much of the regional growth at a substantially better cost basis. Japan is significant for domestic latency requirements and sovereignty-driven demand.
Latin America
Brazil, Chile and Mexico offer renewable-heavy grids, improving subsea connectivity and a cost basis that rewards workloads with no regional latency requirement. Less mature ecosystems, and for the right workload the arbitrage is meaningful.
How to actually choose
Quantify the tradeoff rather than debating it. For a given cluster, model total cost of ownership across three or four candidate markets — power price, capacity rate, ramp, staffing and network cost — against time-to-energisation. In our experience the answer is rarely close once it is on paper, and it is rarely the default market.
For a latency-tolerant training workload, geography is usually worth more than every clause in the contract combined.
The second lever, once geography is settled, is load flexibility. A workload willing to curtail during system peaks is a materially easier load for a utility to accommodate, and AI training is more curtailable than most industrial demand. Offering that deliberately, in exchange for an earlier date or a better tariff, is the most underused move in this market.
Related: power procurement, capacity sourcing and powered shell and build-to-suit.