For most of the past two years, the argument for sovereign AI was made on risk, not on price. An institution kept inference inside its own boundary because the data could not safely leave - clinical records, process recipes, privileged correspondence - and it accepted that the arrangement cost more per request than routing the same work to a hyperscaler API. Data residency was a premium line item.
That framing is going out of date. Three developments, none of them individually decisive, have converged to compress the premium - and for some workload profiles, to invert it.
The hardware bar dropped
The first shift is architectural. Sparse mixture-of-experts models decouple what a model knows from what it costs to run. A recently released open-weight model carries 284 billion total parameters but activates only around 13 billion on any single inference pass, while still holding a one-million-token context window. The memory and compute it demands resemble a far smaller dense model; the reasoning quality does not.
The consequence for infrastructure planning is structural rather than incremental. An institution sizing hardware for sovereign deployment can now provision against the active parameter count rather than the total. Capability that would have required a substantial cluster last year fits comfortably inside a modest dedicated footprint. The procurement question stops being “can we afford enough accelerators to run a frontier-class model?” and becomes “which architecture matches our workload on the hardware we were already going to buy?”
Commodity accelerators are the ceiling, not the floor
The second shift is a signal from an unlikely direction. Providers whose entire business rests on selling inference are now designing their own silicon rather than buying general-purpose accelerators outright. The engineering rationale is plain: a general-purpose GPU carries a transistor budget shaped by training workloads, much of which sits idle during token generation. Purpose-built inference hardware strips that overhead away.
Custom silicon is not available to most enterprises, and that is not the point. The point is what the move reveals - that commodity accelerator pricing represents the ceiling of inference cost, not the floor, and that the floor is falling. An institution that commissions dedicated infrastructure today gets the current performance tier immediately, and sits in a market where competing vendors are actively driving cost-per-token down. That is a healthier position than depending on an API price set by a provider with no structural incentive to compress its own margin.
Cost control became a discipline
The third shift is operational rather than technical, and it is the one that closes the remaining gap. Inference cost governance has matured from ad-hoc monitoring into a structured practice: token budgets set per workload, routing policies that send routine queries to small models and reserve larger ones for genuine reasoning, quantisation applied where it does not measurably degrade output, and cache optimisation to relieve memory pressure on long-context work. Enterprises applying this discipline report inference cost reductions in the range of 30 to 60 per cent.
This matters disproportionately for sovereign deployment. Dedicated infrastructure serves smaller batches than a hyperscaler, which means a higher baseline cost per token. Quantisation and intelligent routing compensate directly for exactly that disadvantage.
What follows
Put together, the picture is a threshold being crossed rather than a trend continuing. For workloads with consistent volume - a legal team’s daily document processing, a manufacturer’s inspection pipeline, a compliance-monitoring system that never sleeps - the breakeven point against API pricing has moved close, and in some configurations has already passed.
The institution that builds inside its own boundary now captures two things at once: the residency guarantee it needed anyway, and a cost structure that improves with each hardware generation rather than one set by someone else’s pricing decision.