UK GPU Rental Price Index 2026: H100 to B200 Hourly

The UK's GPU Rental Landscape in 2026

Mid-2026 finds UK AI teams caught between two forces: a genuine global GPU shortage and a wall of new domestic investment designed to ease it. High Bandwidth Memory (HBM3e) and CoWoS packaging constraints continue to choke supply of H200 and Blackwell-generation silicon, and industry trackers expect GPU pricing pressure to persist through at least mid-2026.

At the same time, the five largest hyperscale cloud providers — Amazon, Google, Meta, Microsoft and Oracle — have committed a combined $600–630 billion in capital expenditure for 2026, with roughly 75% of that aimed directly at AI infrastructure. That scale of spending doesn't automatically translate into cheaper rental rates for UK buyers; if anything, the data below shows the opposite is often true.

NVIDIA H100, H200 and B200 Rental Prices, Mid-2026

Every figure here is a verified per-GPU-hour rate; scopes differ by billing model and node structure, so we've kept each one attributed to its source and date rather than blending them into a single average.

On the H100 side, AWS on-demand pricing sat at roughly $6.88 per GPU-hour as of March 2026, while Azure's ND H100 v5 instances were priced around $12.29 per GPU-hour in the same month. CoreWeave's European on-demand rate for a full HGX H100 8-GPU node was $49.24/hour in July 2026 — equivalent to $6.16 per GPU-hour, though it comes with no single-GPU option. Neocloud specialist Spheron listed H100 on-demand at $2.01/hour in May 2026. Across the wider market tracked by GetDeploying, on-demand H100 pricing has ranged as high as $14.90 per GPU-hour, with spot rates as low as $0.36.

H200 pricing followed a similarly wide band. Octagon AI's market scan put on-demand H200 rates between $2.50 and $10.60 per GPU-hour as of May 2026, while Jarvislabs offered single-GPU H200 instances — unlike most hyperscalers, which only sell H200 in 8-GPU HGX nodes — at $3.80/hour in January 2026.

B200 is the newest and scarcest tier. GPUaaS.com's tracked range shows Blackwell B200 rates spanning $2.12 per GPU-hour for spot instances to $14.24 per GPU-hour for on-demand instances. Most B200 hardware orders were backordered through mid-2026, with an estimated 3.6 million units in the global queue as of April 2026 — making cloud rental, not procurement, the realistic route to Blackwell capacity for almost every UK buyer this year.

• •Hyperscale providers typically charge 3–6 times more than neocloud alternatives for equivalent H100 hardware, per Spheron's May 2026 analysis.
• •The widest verifiable spread — comparing Azure's $12.29/hr on-demand rate against the market's lowest H100 spot rate of $0.36/hr — runs to roughly 34x, though this compares two different billing models, not like-for-like on-demand pricing.
• •Comparing on-demand rates only (Azure at $12.29/hr vs Spheron at $2.01/hr) still leaves a 6.1x gap for identical H100 hardware.

Beyond the Hourly Rate: Hidden Costs for UK Buyers

The advertised per-GPU-hour figure is rarely the number that lands on a UK finance director's desk. Several factors compress or expand that headline rate before a project's true cost is known, and none of them show up in a provider's pricing page.

Currency exposure is the first: nearly every rate above is quoted in USD, so a UK buyer's effective cost moves with sterling regardless of what the provider charges. VAT applies to cloud services consumed by UK businesses, and import duty considerations arise only if hardware is purchased outright rather than rented — a distinction worth checking against current guidance before budgeting a project. Data egress fees, storage charges and support tiers are typically billed separately from the GPU-hour rate itself, and hyperscalers' node-bundling model (paying for an entire 8-GPU HGX node even if a workload only needs one GPU) can multiply the effective cost of smaller jobs several times over. Latency and data residency also carry a cost that doesn't appear on an invoice: routing training data through a non-UK or non-EU region raises GDPR compliance questions that many regulated UK sectors — finance, health, public sector — simply cannot absorb, regardless of price.

Worked Example: 200 Hours of Fine-Tuning, UK vs Hyperscaler

Take a common scenario: a UK software team needs 200 GPU-hours to fine-tune a mid-sized language model on a single H100. Using Azure's verified March 2026 on-demand rate of $12.29 per GPU-hour, 200 hours of compute alone comes to roughly $2,458. Using Spheron's verified May 2026 on-demand rate of $2.01 per GPU-hour for the same hardware class, the identical 200-hour job comes to roughly $402 — before either provider's storage, egress or support charges are added.

That gap is before currency conversion to GBP, before VAT treatment, and before any GDPR-driven decision to keep data inside a specific jurisdiction rather than route it through a US-based hyperscaler region. For a UK business running iterative fine-tuning cycles rather than a single job, that multiple compounds fast — which is exactly why our self-hosting vs cloud GPU cost comparison is worth running alongside any multi-month compute plan.

Choosing the Right GPU — and the Right Commitment Model

Not every UK workload needs the newest silicon. H100 remains the workhorse for most fine-tuning and mid-scale training jobs and is now the most widely available tier across both hyperscalers and neoclouds. H200 adds memory bandwidth that matters for larger context windows and retrieval-heavy inference, and is increasingly available across major cloud providers in H1 2026. B200 is the performance step-change — NVIDIA's Blackwell generation delivers up to 4x the inference throughput of H100 — but its 3.6 million-unit backlog as of April 2026 means rental, not ownership, is the only realistic access route for most UK teams this year. Our H100, H200, B200 comparison breaks down which tier suits which workload in more depth.

On commitment: neoclouds typically offer per-minute billing and single-GPU granularity, which suits short, bursty workloads and experimentation. Hyperscalers' node-bundled, longer-duration pricing structures can work in a UK business's favour only when utilisation is sustained and predictable enough to justify the node's full capacity — otherwise the effective per-GPU cost quietly climbs back towards hyperscaler on-demand levels.

The UK Investment Picture Reshaping Supply

UK policy moved decisively in mid-2026. At London Tech Week in June, government announced a £1.1 billion AI Hardware Plan, comprising £750 million toward a national AI supercomputer targeted for delivery by 2030, £400 million for advanced AI chip development, and £150 million earmarked for inference hardware procurement in summer 2026.

Private capital is moving in parallel. Nebius committed £1.7 billion to NVIDIA GPU clusters across three new UK sites, directly targeting the compute-access gap facing UK-based AI developers. NVIDIA's partners, including Nscale and CoreWeave, plan to deploy what would be Europe's largest GPU cluster in the UK — 120,000 Blackwell GPUs by the end of 2026 — alongside government-promoted 'AI growth zones' in Culham, Teesside, Newcastle and North and South Wales, each earmarked for up to 500MW of data centre capacity. None of this resolves the near-term shortage, but it is the clearest signal yet that UK-based capacity, rather than permanent reliance on US hyperscaler regions, is the direction of travel.

• •Nebius: £1.7 billion committed to NVIDIA GPU clusters across three new UK sites — a direct option for buyers prioritising UK-based capacity.
• •Nscale and CoreWeave: NVIDIA's deployment partners for a planned 120,000-Blackwell-GPU cluster in the UK, positioned as Europe's largest.
• •Government-backed 'AI growth zones' in Culham, Teesside, Newcastle, and North/South Wales, each earmarked for up to 500MW of data centre capacity, signalling where future UK-region compute availability is likely to concentrate.

Future Outlook: Late 2026 and Into 2027

The constraints shaping today's price index aren't easing quickly. HBM3e supply remains the primary bottleneck on H200 and Blackwell production, and GPU shortages are projected to persist through at least mid-2026 with upward pricing pressure continuing. Against that backdrop, the 120,000-GPU UK Blackwell deployment and the government's £150 million summer 2026 inference hardware procurement are the two developments most likely to move UK-relevant pricing over the next two to three quarters — though neither guarantees the neocloud-to-hyperscaler spread narrows before new capacity actually comes online.

For now, the practical takeaway for UK buyers is unchanged: check the verified rate, the billing granularity, and the data-residency terms before assuming a hyperscaler's brand name justifies its price. If ownership is under consideration instead of rental, our coverage of the refurbished H100 and H200 secondary market is a useful next read.

Methodology

This index compiles verified, dated per-GPU-hour figures published between January and July 2026 by hyperscale cloud providers (AWS, Azure), specialised GPU cloud platforms (CoreWeave, Spheron, Jarvislabs), and independent market trackers (GetDeploying, Octagon AI, GPUaaS.com), alongside UK government and investment announcements from GOV.UK, Computing UK and ITPro dated April to June 2026.

Each figure retains its original scope — GPU model, node structure, billing type and publication date — and no rate has been averaged, blended or extrapolated across sources. Where two figures could appear contradictory, such as the on-demand versus spot spread for H100, both are presented separately with the scope explicitly stated rather than merged into a single headline number.

Currency figures are presented in their originally published denomination (USD for cloud rates, GBP for UK government and investment figures) and have not been converted, given the volatility currency conversion would introduce to a rate-sensitive index.

Sources

Every figure in this article traces to the sources below.

• •Spheron Blog — hyperscaler vs neocloud pricing multiple and H100 on-demand rate
• •VESSL AI — AWS and Azure H100 on-demand pricing, March 2026
• •CoreWeave — European H100 HGX node on-demand pricing, July 2026
• •Octagon AI — H200 cloud pricing range, May 2026
• •Jarvis Labs Blog — single-GPU H200 on-demand pricing, January 2026
• •Inworld AI — B200 inference throughput and hardware backlog, April 2026
• •GPUaaS.com — B200 on-demand and spot pricing range
• •GetDeploying — H100 on-demand and spot rate extremes
• •Fusion Worldwide — GPU shortage persistence and pricing trend, March 2026
• •Barrack AI — HBM shortage constraints and hyperscaler capex commitments, 2026