AI Inference Energy Cost 2026: Pounds per Million Tokens

The True Cost of AI Inference: Why Pounds-per-Million-Tokens Matters Now

Every AI vendor prices tokens to several decimal places on a public rate card. Almost none of them publish the electricity bill sitting behind those tokens, and fewer still convert that bill into pence for a UK-based deployment. That absence matters because inference — not the one-off cost of training a model — is the line item that recurs on every query, every day, for as long as the service runs.

This article builds that missing conversion. It starts from verified GPU power specifications, a UK colocation electricity benchmark, and a structural formula for turning watts into cost, then compares that against what UK businesses are actually charged per million tokens through managed APIs in 2026. The aim is a transparent framework a procurement team can run with its own numbers, not a single 'correct' price that applies to every deployment.

From GPU Watts to Kilowatt-Hours: The Physics Behind Every Token

The starting point for any inference energy calculation is the accelerator itself, and here the spread is wide. NVIDIA specifies the H100 PCIe card at 350 W total board power, while the H100 SXM version — the variant used in most dense inference servers — is rated at 700 W. That is a doubling of power draw for what is nominally the same GPU generation, purely because of the form factor and interconnect chosen.

Moving to newer silicon does not automatically lower that draw. NVIDIA specifies the H200 in the same 700 W-class power envelope as the H100 SXM, so upgrading Hopper-generation hardware for inference does not, by itself, cut electricity consumption per accelerator. The Blackwell-generation B200 goes further in the other direction, specified at 1,000 W — moving flagship inference GPUs into the kilowatt-per-card class.

None of this happens in isolation. A formal energy-footprint model published on arXiv expresses the energy cost of a single query as inference time multiplied by GPU power plus non-GPU power, all multiplied again by the data centre's Power Usage Effectiveness (PUE). That third term is why the same GPU can cost noticeably different amounts to run depending on the facility it sits in, quite apart from the electricity tariff itself.

Building the UK Formula: Converting GPU Power into Pounds

Turning that physics into a pounds-and-pence figure starts with electricity price. There is no single regulated UK tariff for AI colocation, but a benchmark cited in UK-focused commentary from January 2026 puts commercial and colocation electricity around 30p per kWh for expensive UK regions — an indicative industry benchmark, not a national utility rate, and one to swap for your own contracted price wherever possible.

Applying that benchmark to the GPU wattages above gives a useful ceiling figure: the cost of running the accelerator alone, at full power, for one hour, before any facility overhead is added. An H100 PCIe card (350 W) works out at roughly £0.105 an hour in electricity for the GPU alone; an H100 SXM or H200 (700 W-class) at roughly £0.21 an hour; and a B200 (1,000 W) at roughly £0.30 an hour. Scale that to an eight-GPU inference node and the H100 SXM/H200 configuration reaches roughly £1.68 an hour in GPU electricity alone, while an eight-GPU B200 node reaches roughly £2.40 an hour — before the arXiv formula's non-GPU power and PUE terms are added on top.

That is deliberately not yet a cost per million tokens. John Snow Labs' guidance on measuring LLM inference energy is explicit that the missing variable is throughput: energy per token should be calculated as joules consumed divided by tokens produced, measured at GPU or server level, and that figure needs benchmarking per workload because tokens generated per joule varies materially by model, quantisation and batching configuration. In practice that means logging the token count for every request alongside GPU or server-level energy draw, computing energy per token as joules divided by tokens produced, and repeating that measurement separately for each model, quantisation setting and batching configuration you intend to run — rather than assuming one benchmark run transfers across configurations, since tokens-per-joule varies materially across exactly those three variables. Two UK deployments running identical GPUs at the identical 30p/kWh tariff can land on very different £-per-million-token figures purely because one serves far more tokens per second than the other.

The practical calculation for a UK buyer is: take your GPU's rated power, multiply by hours run and your contracted electricity rate to get a baseline electricity cost; add non-GPU server power and your facility's PUE per the arXiv structure; then divide the resulting £ figure by your own measured tokens-per-second throughput, aggregated as John Snow Labs recommends, to arrive at a genuine £-per-million-tokens number — rather than relying on a single industry-wide constant that doesn't yet exist.

What UK Businesses Actually Pay: API Pricing Across the Tiers

For businesses buying inference through a managed API rather than running their own hardware, the vendor's rate card already bundles electricity, infrastructure and margin into one number — and that number varies enormously by tier. A February 2026 inference-economics guide describes three broad bands: budget-tier APIs sitting around $0.06 to $0.30 per million tokens, mid-tier models around $0.55 to $15 per million tokens, and frontier models around $15 to $75 per million tokens — more than a thousand-fold spread from cheapest to most expensive.

Named-model pricing from a January 2026 UK-pound comparison shows where that spread sits in practice. Claude 4 Opus was listed at £12 per million input tokens and £60 per million output tokens, firmly at the expensive end of the frontier band. Gemini 2.5 Pro, by contrast, was listed at £1 per million input tokens and £8 per million output tokens — a fraction of the Opus figure for a model still marketed as frontier-capable. GPT-4 Turbo's historical pricing of $10 per million input tokens and $30 per million output tokens is a useful anchor for how far the market has moved: a March 2026 unit-economics analysis found that GPT-4-equivalent inference, which cost roughly $20 per million tokens in late 2022, had fallen to around $0.40 per million tokens by early 2026.

The commercial implication is straightforward: model choice inside the API market moves cost far more than any electricity-side optimisation can. Selecting a mid-tier model over a frontier one, or routing simple queries away from a premium model, changes the bill by a much larger factor than anything achievable through PUE or tariff negotiation alone.

API vs Self-Hosted: Where the Hidden Costs Really Sit

A managed API call folds vendor infrastructure, electricity and margin into a single published rate, which is convenient but opaque — you cannot see how much of that $0.06-to-$75-per-million-token spread is GPU electricity versus provider margin. Self-hosting or colocating your own inference hardware exposes the real cost stack instead: GPU choice (350 W to 1,000 W per accelerator, per the specifications above), your facility's PUE, your contracted UK electricity rate, and your own measured throughput all become variables you control directly rather than accept as given.

That control comes with trade-offs the raw electricity arithmetic doesn't capture. Data residency is one UK-specific consideration procurement teams should weigh alongside cost: some mainstream hosted models and enterprise suites are offered with UK data residency and some are not, which affects both compliance posture and which vendors are realistically available for regulated workloads. Before committing capital to either route, it's worth working through the numbers properly — you can compare own vs. rent AI inference economics directly, and the fundamentals of running inference in-house are covered in our guide to on-premise AI inference.

Cutting the Bill: Optimisation Levers That Actually Move the Number

The single most reliable lever available today is measurement itself. John Snow Labs' guidance on quantifying inference energy recommends logging energy at GPU or server level alongside token counts per request, then aggregating across the specific combination of model, quantisation setting and batching configuration in use — because tokens generated per joule varies materially across exactly those three variables. A UK business that has never measured this baseline has no way of knowing whether its current deployment is efficient or badly under-batched.

Hardware-generation choices are a second lever, but one that needs careful scoping. NVIDIA's GB200 NVL72 rack-scale system is described as delivering 30x faster real-time trillion-parameter inference compared with the prior-generation Hopper platform — a throughput claim, not a stated reduction in watts per token. Whether that gain lowers your £-per-million-tokens figure depends on how the accompanying power draw compares, which is exactly the kind of number that needs measuring on your own workload rather than assumed from a headline multiplier. Before scaling hardware for an inference workload, it's worth checking real requirements against the AI GPU calculator and reviewing where power, not compute, becomes the binding constraint — a pattern covered in our analysis of the AI power bottleneck.

Model-tier discipline is the third and often largest lever available without touching infrastructure at all: routing routine queries to a budget or mid-tier model instead of a frontier one, or reserving a premium model like Claude 4 Opus for genuinely complex work rather than an entire workload, exploits the same thousand-fold pricing spread described above.

Environmental Reality Check: What We Can — and Can't — Yet Measure

Converting AI inference into a genuine environmental footprint figure for UK sustainability reporting is harder than the token-pricing comparisons above suggest. Epoch AI's analysis of ChatGPT energy use makes a point that applies across the industry: reliable measurement linking model usage to actual energy consumption is still an emerging discipline, not a settled standard with agreed public figures for every model and workload.

That doesn't mean UK businesses should skip the exercise — it means the credible starting point is the same structural formula used earlier in this piece: inference time multiplied by GPU and non-GPU power, multiplied by facility PUE, as set out in the arXiv energy-footprint model. Applying your own measured watts, your own PUE and your own UK electricity contract to that formula produces a defensible internal figure, even where an industry-wide public benchmark for carbon, water or hardware lifecycle impact doesn't yet exist in the way it does for GPU wattage or API pricing.

Methodology

This data study draws on GPU manufacturer specification pages, UK-focused colocation and electricity-cost commentary, published API pricing comparisons and unit-economics guides, and a peer-reviewed energy-footprint modelling paper, compiled through mid-2026. Source dates range from NVIDIA's original H100 specification (March 2023) through to a John Snow Labs measurement guide published in April 2026, and every figure in this article retains the exact scope, date and source stated by the original publisher.

Figures are deliberately not merged where they measure different things: a vendor's published input-token price, an output-token price, a GPU's rated board power, and a UK colocation electricity benchmark are four distinct categories of number, and this article keeps them separate rather than blending them into a single implied cost. Where the UK-specific electricity or carbon data needed for a fully worked, region-by-region cost model was not available in the source material, that gap is stated explicitly rather than filled with an assumed figure, and the formula is presented instead so buyers can substitute their own contracted electricity rate, PUE and measured throughput.

Sources

Every figure in this article traces to the sources below.

• •NVIDIA — H100 GPU specification (350W PCIe, 700W SXM)
• •NVIDIA — H200 GPU power envelope
• •NVIDIA — B200 GPU specification (1,000W)
• •NVIDIA — GB200 NVL72 inference throughput claim
• •360Strategy — UK colocation electricity benchmark and named-model token pricing
• •Mirantis — budget/mid/frontier API pricing bands
• •Introl — GPT-4-equivalent price collapse, 2022–2026
• •Epoch AI — measurement challenges in ChatGPT energy use
• •John Snow Labs — measuring energy and tokens-per-joule for LLM inference
• •arXiv — AI energy footprint modelling formula