UK enterprises are getting bill shock as AI vendors quietly swap flat-fee subscriptions for usage-based token pricing. New survey data shows nearly a third of senior executives can't predict what scaling AI will actually cost — and infrastructure pricing pressure is only building.
View the data behind this chart
| Struggled with cost… | Re-phasing AI deployments | |
|---|---|---|
| % of senior execs | %29 | %50 |
Why the AI invoice suddenly looks different
A KPMG survey of more than 2,000 senior executives across 20 countries, reported by The Register, found that 29 percent struggled to understand operating costs as their enterprise AI deployments scaled, while nearly half were re-phasing rollouts once costs outweighed expected value.
The mechanism behind this is simple but has caught many buyers off guard: Anthropic, OpenAI and GitHub have each moved away from flat-fee, all-you-can-eat subscriptions towards usage-based, per-token billing. For UK finance and procurement teams used to predictable SaaS-style invoicing, that shift converts AI from a fixed opex line into a variable cost that scales with adoption success — precisely the outcome IT leaders were told to chase.
Inference vs training ROI: the maths is getting harder
The cost pressure isn't confined to chatbots. Gartner research cited in the Register discussion found a lack of transparency from vendors over coding-agent costs and an absence of the cost-optimisation tooling enterprises expect from cloud platforms. The projection: costs per developer for AI coding agents will exceed the average global developer salary by 2028 — and in lower-salary markets such as India, agent costs are already overtaking developer pay, because token pricing is global while wages are not.
UK buyers sit in the middle of that spectrum: salaries here are higher than in emerging markets but well below Silicon Valley, meaning the crossover point for coding-agent economics will vary sharply by team and location. Anyone running an inference-heavy coding assistant at scale should be modelling this now, not waiting for the invoice. It's worth comparing that maths against your own infrastructure choices — you can calculate your AI GPU requirements to see where self-managed inference capacity might undercut per-token vendor pricing.
On-premise vs cloud economics: the calculation is shifting
Behind the per-token squeeze sits a much bigger number: one major investment house puts industry-wide AI datacentre capex at roughly $1.5 trillion over five years to 2030, according to the Register's discussion. That capital has to be recouped somewhere, and rising GPU rental rates and metered API pricing are the most visible mechanisms so far.
For UK buyers, this reframes the on-premise versus cloud debate. Cloud AI capacity looked cheap when subscription pricing masked the true compute cost; usage-based billing now exposes it directly. Before committing further budget to hyperscaler AI services, it's worth running the numbers to compare cloud versus on-premise TCO over a realistic three-to-five-year horizon, factoring in the same capex pressures now driving hyperscaler prices upward.
Right-sizing AI infrastructure spend
The Register's reporting notes that nearly half of surveyed executives are already responding by mixing lower-cost and high-fidelity models instead of running everything on the most expensive frontier model available. That's a sensible starting discipline for UK buyers too: not every workflow needs the top-tier model, and workload-matched model selection is one of the few cost levers enterprises fully control.
Practically, that means auditing which use cases genuinely need frontier-grade inference versus those that could run on smaller, cheaper models — or on owned hardware where volumes justify it. It's also worth reviewing how sizing decisions are made in the first place, since undersized or oversized clusters both erode the return on AI investment.
- •Segment workloads by model tier needed, not by what's easiest to deploy
- •Track per-token spend against business value delivered, not just usage volume
- •Reassess deployments quarterly given how fast vendor pricing is moving
Negotiating opex models with vendors
Given that AI labs themselves are still working out whether usage-based billing can be profitable without triggering customer flight to open-source or lower-cost alternatives, UK buyers have some leverage right now. Locking in capped-spend agreements, tiered discounts, or hybrid subscription-plus-usage contracts before pricing hardens further is a reasonable defensive move.
This is also the moment to bring procurement and finance into AI decisions that were previously left to engineering teams. If you haven't already built AI cost governance into your vendor contracts, it's worth taking the time to optimise your IT procurement approach specifically for usage-based AI billing before renewal cycles lock in unfavourable terms.
View the data behind this chart
| Subscription… | Usage-based… | Mixed/hybrid | |
|---|---|---|---|
| Billing basis | Flat monthly fee | Per-token metering | Blended by workload |
| Cost predictability | High | Low | Moderate |
| Vendor examples | Early-stage vendors | OpenAI/Anthropic/GH | Emerging option |
| UK budget risk | Low overrun risk | High overrun risk | Needs FinOps |
| Best suited to | Fixed-scope pilots | Mature monitored use | Scaling deployments |
What this means for UK IT budgets in 2026
None of this means AI adoption should stop — but it does mean treating AI infrastructure spend with the same rigour applied to any other major capex decision. The KPMG data suggests plenty of peers are already recalibrating, and the capex figures underpinning industry-wide pricing aren't going away soon.
Buyers who model true unit economics now — rather than after the next invoice cycle — will be better placed to negotiate, right-size, and avoid the re-phasing scramble nearly half of surveyed executives are already going through. For ongoing coverage of how these cost dynamics are reshaping procurement decisions, it's worth continuing to explore more AI infrastructure insights.