Everyone claims to be 'repatriating AI from the cloud' in 2026, but the honest answer is: it depends on utilisation and data sensitivity, not ideology. Cast AI finds average enterprise GPU utilisation sitting at just 5% in real Kubernetes clusters — meaning most cloud GPU spend pays for idle silicon. Meanwhile a UK-listed NVIDIA H100 PCIe unit now lists at £32,050, VAT and import duties included. This piece works through the actual GBP and USD numbers so you can decide, workload by workload, whether to own or rent.
View the data behind this chart
| H100 PCIe | H100 SXM | H200 SXM | B200 SXM | |
|---|---|---|---|---|
| Low estimate | $k25 | $k35 | $k38 | $k30 |
| High estimate | $k33 | $k40 | $k45 | $k50 |
The great repatriation, reality-checked
Gartner puts new AI infrastructure spending at $401 billion for 2026 (VentureBeat, May 2026), and a good chunk of the 'repatriation' narrative is genuinely happening — but not because cloud got worse. It's because enterprises finally measured what they were paying for. Cast AI's April 2026 data shows average GPU utilisation across real enterprise Kubernetes clusters sitting at just 5%, which means the typical organisation is renting five, ten or twenty times the compute it actually consumes.
The flip side matters just as much: Chamber Blog's February 2026 analysis shows that with proper scheduling and workload management, achievable GPU utilisation for AI training can reach 85-95%. That gap — 5% observed versus 85-95% achievable — is the entire repatriation debate in two numbers. Below a certain utilisation floor, renting wins on unit economics every time. Above it, ownership starts to win, and the crossover point is where this decision actually gets made.
The 2026 price sheet: what GPUs cost to own vs rent
Start with the hardware. As of March 2026, an NVIDIA H100 80GB PCIe card starts around $25,000 and climbs to $30,000-$33,000 depending on supply (ElectronicsHub by Alibaba.com). The higher-bandwidth SXM5 variant runs $35,000-$40,000+ per GPU on the same date. Step up a generation and an H200 SXM 141GB unit costs $38,000-$45,000 (Fluence, November 2025), while Blackwell-generation B200 SXM 192GB GPUs carry an estimated $30,000-$50,000 per GPU when bought in clusters of eight or more (Thunder Compute, July 2026).
UK buyers need a different number for capex planning. A UK-listed NVIDIA H100 PCIe unit was priced at £32,050 (roughly $40,700 at the exchange used) as of June 2026 — reflecting VAT, import duties and localised certification on top of the raw US list price (ElectronicsHub, June 2026). That's the figure to put in a UK budget, not the US headline price most vendor decks quote for GPU accelerators.
Rental rates vary just as widely by provider tier. Hyperscalers price an on-demand H100 SXM5 at up to ~$6.88/GPU/hr, or ~$3.78/GPU/hr with a one-year Savings Plan (GPUaaS.com, May 2026). Specialised neo-cloud providers quote $2.20-$3.14/GPU/hr for the same class of hardware (GPUPerHour, July 2026) — hyperscalers can charge three to five times more for equivalent silicon, per the same market data.
Hidden costs neither side puts in the headline rate
Cloud bills rarely stop at the hourly GPU rate. Egress fees run $0.08-$0.15/GB and storage $0.08-$0.23/GB/month on major hyperscalers (GPUaaS.com, May 2026) — costs that compound fast for training pipelines moving large datasets, and that sit on top of the overprovisioning waste already implied by 5% average utilisation.
On-prem has its own hidden line item: power. UK industrial electricity prices are forecast to grow at an 11.1% compound annual rate through 2026-27 — a multi-year, compounding cost that a one-off GPU purchase price simply doesn't capture. The government's April 2026 'British industrial competitiveness scheme' aims to cut electricity bills by up to 25% for over 10,000 businesses, but the benefit lands mainly as a one-off additional payment in 2027 — likely too late for firms sizing power and cooling budgets against 2026 GPU purchases. Cooling design and specialist MLOps staffing add further operational load that a hardware quote never shows; see our AI server cooling considerations before finalising a rack plan.
Two worked examples: inference vs training
Take an illustrative 8-GPU H100 PCIe cluster, sized independently of any vendor TCO deck. On US list pricing that's $200,000-$264,000 in capex; on UK landed pricing it's roughly £256,400 before power, cooling or rack costs. Rent the same eight GPUs instead and, run near-continuously for a sustained inference workload, hyperscaler on-demand pricing costs roughly $482,000 in year one; a one-year Savings Plan roughly $265,000; a neo-cloud roughly $154,000-$220,000. On these independently sourced unit prices, an on-prem PCIe cluster's capex is recovered within a single year against hyperscaler on-demand rates, and lands close to one year of committed hyperscaler pricing — provided the workload runs at high, sustained utilisation. This is the scenario where self-hosting LLMs versus cloud GPU costs genuinely favours ownership.
Vendor-commissioned studies push this further: a Dell-commissioned analysis cited via Lyzr claims on-prem inference can run up to 62% cheaper than public cloud over four years; a Lenovo 2026 TCO analysis cited via Tilkal claims self-hosted inference can be up to 18 times cheaper per million tokens over three years; Lenovo Press separately claims a breakeven under four months for workloads above 20% utilisation. Treat all three as directionally consistent with the independent maths above, but as vendor-commissioned upper bounds — not numbers to build a board paper around without your own unit-price check.
Now flip the workload. If the same cluster only runs jobs a fraction of the time — consistent with the 5% utilisation Cast AI observes across real enterprise clusters — the capex sits idle for most of the year, and a neo-cloud paying only $2.20-$3.14/hr for hours actually used will typically beat ownership outright. The breakeven only tips toward buying once scheduling discipline pushes utilisation toward the 85-95% range Chamber Blog reports as achievable — which is an MLOps capability question as much as a hardware one.
The UK regulatory and sovereignty layer
There is no dedicated UK AI Act in 2026; the government continues to signal a light-touch, innovation-first regulatory model (Scaffold Digital, June 2026). But techUK's January 2026 analysis notes that digital sovereignty is rapidly becoming a central pillar of UK national security, driven by how much data modern AI systems process.
The concrete legal shift began on 5 February 2026, when several significant provisions of the Data (Use and Access) Act 2025 came into force, with further measures following later in 2026. The provisions already in force expand lawful automated decision-making while tightening safeguards around it — meaning organisations doing more automated decisioning on personal data need demonstrable control over where and how their models run. For regulated sectors handling sensitive personal or financial data, that control argument can justify repatriation even for workloads that wouldn't clear a pure cost-utilisation threshold on their own.
View the data behind this chart
| $/GPU/hour | Hyperscaler | Hyperscaler 1yr | Neo-cloud | Neo-cloud 1yr |
|---|---|---|---|---|
| H100 SXM5 rental rate | 6.88 | 3.78 | 3.14 | 2.2 |
Building the decision framework
Put the pieces together and the decision stops being a binary 'cloud vs on-prem' call and becomes a scoring exercise across three axes: sustained utilisation level, data sensitivity, and operational maturity to run GPUs in-house. Run every workload through this before committing capex — our Cloud vs. On-Premise TCO Calculator is built for exactly this comparison using your own rates rather than vendor assumptions.
The hybrid playbook and future-proofing
The practical default for most UK enterprises in 2026 is hybrid, not a single winner-take-all migration. Rent bursty, experimental and R&D workloads from neo-cloud providers at $2.20-$3.14/hr rather than locking them into capex. Repatriate sustained, high-utilisation inference and any workload touching regulated personal data, where the Data (Use and Access) Act 2025 raises the bar on demonstrable control.
Avoid lock-in on both sides: keep repatriation candidates on portable, open tooling rather than hyperscaler-proprietary APIs, and keep on-prem procurement flexible given how fast GPU generations are turning over — H100, H200 and B200 pricing already overlaps in the market, so buy for the workload in front of you, not a speculative future one.
Sources
Every figure in this article traces to the sources below.
- •ElectronicsHub by Alibaba.com — NVIDIA H100 PCIe/SXM5 and UK landed GPU prices
- •Fluence — NVIDIA H200 SXM 141GB purchase pricing
- •Thunder Compute — NVIDIA B200 SXM 192GB cluster pricing
- •GPUaaS.com — hyperscaler cloud rental rates and hidden cloud costs
- •GPUPerHour — neo-cloud GPU rental rates
- •Cast AI — average enterprise GPU utilisation
- •Chamber Blog — achievable GPU utilisation with scheduling
- •Lyzr — Dell-commissioned on-prem inference cost analysis
- •Tilkal — Lenovo 2026 self-hosted inference TCO analysis
- •Lenovo Press — on-prem TCO breakeven point claim