AI Infrastructure Financing 2026: GPUaaS vs Capex vs Lease

The AI Compute Conundrum: Why Financing Now Decides the Build

NVIDIA's H100 sits in the Hopper family, shipping in PCIe and SXM variants with up to 80GB of HBM3 memory on the PCIe part — the memory that most AI workloads lean on hardest. That memory carries its own price tag: one 2026 financing guide put H100 purchase prices at $25,000-$40,000 per GPU against $30,000-$40,000 for the higher-memory H200, a floor that doesn't drop even at the budget end of the range.

Layer fast depreciation on top and the capex maths changes shape entirely. Financial-planning models built for AI servers assume a 3-4 year useful life against 5-7 years for traditional servers, and reckon 30-40% of economic depreciation can land in year one alone. That's a materially different profile from the laptop and switch refresh cycles most UK finance teams already budget for.

None of this happens in a demand vacuum. Goldman Sachs noted in March 2024 that data-centre build-out assumptions depend heavily on the pace of AI infrastructure expansion — which is exactly why financing terms, not just unit prices, decide whether a build ages into a bargain or a write-off.

Beyond the GPU Hour: What UK Buyers Actually Pay For

Hourly rates make for tidy headlines, but they hide the real bill. GPUnex's 2026 guidance put full three-year total cost of ownership for a single 8×H100 node at $570,000-$1.2 million — a range wide enough on its own to swallow most of the argument about which hourly rate looks cheapest, once networking, storage, power and support are layered on top of raw compute.

For UK buyers specifically, three extra layers change the ownership break-even versus cloud or lease pricing: ex-VAT sterling quotes rather than headline USD figures, import duty and shipping on any direct hardware purchase, and local power costs that vary by site. Regulated or security-sensitive deployments add a fourth: data-residency and supplier-risk review can tip the decision toward on-prem or UK-hosted GPU capacity even where overseas public cloud looks cheaper on paper.

There's also an accounting layer that's easy to miss. Many equipment leases create right-of-use assets and liabilities under IFRS 16 even when the supplier retains ownership of the hardware — so "moving to opex" doesn't automatically mean moving off the balance sheet, and finance teams need to model the leverage and EBITDA optics before they sign.

GPUaaS in 2026: Hyperscaler vs Specialist Pricing

The spread across GPUaaS providers is wide enough to change a business case on its own. AWS's on-demand H100 and H200 pricing fell 44% in June 2025, from $7/hour to $3.90/hour. A separate 2025 financing guide quoted general on-demand H100 pricing at $2-4/hour, while Hyperbolic was listed in the same period's coverage at $1.49/hour for H100 and $2.15/hour for H200 — specialist-budget pricing sitting well below hyperscaler on-demand rates.

A separate 2026 provider table added two more data points: specialist cloud H100 pricing at $2.00-$3.00/hour and GPU marketplace pricing at $1.50-$2.50/hour, both undercutting typical hyperscaler on-demand rates. Commitment structures move the number further still — reserved capacity on 1-3 year terms can discount on-demand pricing by 30-40%, while spot pricing runs 50-70% below on-demand but carries interruption risk that suits batch training rather than latency-sensitive inference.

One caution: every figure above is a USD-quoted, largely US-market rate. None of it is UK-specific pricing, which is precisely why the sterling-quote step in the previous section matters — always convert to an ex-VAT GBP number before comparing a cloud rate against a lease or direct-purchase line item.

• •AWS on-demand: $7/hour cut to $3.90/hour in June 2025 — a 44% reduction
• •General on-demand H100: $2-4/hour (2025 financing guide)
• •Hyperbolic (budget specialist): H100 $1.49/hour, H200 $2.15/hour
• •Specialist cloud (2026 table): $2.00-$3.00/hour; GPU marketplace: $1.50-$2.50/hour
• •Reserved 1-3yr: 30-40% off on-demand; Spot: 50-70% off on-demand (interruption risk)

Financing Models: Capex, Leasing, GPUaaS and What Sits Between Them

Direct purchase means full cash outlay upfront at $25,000-$40,000 per H100 or $30,000-$40,000 per H200, and it sits at the sharp end of depreciation risk given the 3-4 year useful-life assumption. It's the right call only where utilisation is high and sustained enough to outrun that depreciation curve.

Leasing sits in between, and the terms move the number a lot. One 2026 guide puts monthly operating-lease payments at $900-$1,500 per H100 depending on term length and credit quality; a separate 2026 leasing guide puts H100 leases at $400-$600 per month over 48-72 month terms, for a total lease cost of $19,200-$43,200 per H100 across the full term. Both are legitimate quotes for different credit profiles and durations — see our Hire Purchase vs. Leasing IT Equipment comparison for how the structures differ beyond GPUs specifically.

Reserved capacity is the hybrid worth remembering: a 1-3 year cloud commitment for 30-40% off on-demand pricing, without taking on ownership or depreciation risk at all — useful where usage is predictable but the balance sheet appetite for capex isn't there yet.

Worked Examples: Three UK Scenarios

The utilisation forecast, not the hourly rate, is what actually decides the winner in each of these scenarios.

Model your own numbers against these before committing to a route — the AI GPU Calculator and Own vs. Rent AI Inference Break-Even Analysis both work from the same utilisation logic used below.

• •Startup pilot, single 8×H100 node, uncertain demand: owning would mean a $570,000-$1.2 million three-year commitment for a workload likely below the ~20% sustained utilisation point a 2026 decision guide flags as where renting wins — specialist cloud at $2.00-$3.00/hour or marketplace at $1.50-$2.50/hour preserves cash instead.
• •Scale-up, growing real-time inference, 20-40% utilisation band: an operating lease smooths cash flow versus a full purchase, whether at $900-$1,500/month per H100 on shorter or lower-credit terms, or $400-$600/month over 48-72 months on longer, stronger-credit terms.
• •Enterprise fleet, 1,000 GPUs at roughly 80% sustained utilisation: the three-year model puts direct purchase at $34 million against $43.2 million leased and $52.5 million cloud — effective costs of $0.48, $0.61 and $0.75 per GPU-hour respectively — comfortably inside the >40-60% sustained-utilisation band where owning wins.
• •For inference-heavy workloads specifically, one 2026 decision guide's heuristic is that above roughly 500 million output tokens per month, a managed API can cost more than running owned hardware — worth cross-checking against the utilisation math above rather than relying on either signal alone.

Risk and Future-Proofing: Hardware Moves Faster Than Most Finance Terms

With 30-40% of economic depreciation landing in year one and only a 3-4 year useful life assumed overall, a 48-72 month lease can span more than one hardware generation. Before signing, check whether the contract includes a refresh or upgrade option rather than locking the business into ageing capacity for the full term.

The demand side carries its own risk: Goldman Sachs's March 2024 note on how data-centre build-out assumptions depend on the pace of AI infrastructure expansion is a reminder that capex is a bet on your own utilisation forecast being right, not just a bet on GPU prices holding still.

The practical mitigation is a blended portfolio: reserved capacity for predictable base load, on-demand or spot for burst and batch work (accepting spot's interruption risk), and any lease reviewed for its IFRS 16 balance-sheet impact before it's signed rather than after.

The Verdict: A Financing Framework for UK AI Infrastructure

Treat utilisation as the anchor, not the hourly rate. One 2026 decision guide's heuristics put renting ahead below roughly 20% sustained utilisation and owning ahead above roughly 40-60% sustained utilisation, with leasing often considered the natural middle ground for the 20-40% band — but these are heuristics from cited guides, not universal break-even points, because workload mix, electricity costs, tax treatment and residual-value assumptions all shift the crossover in practice.

• •Model utilisation honestly before modelling cost — the financing decision follows from it, not the other way round
• •Get an ex-VAT sterling quote for every route before comparing it to a USD-denominated cloud or lease rate
• •Check the IFRS 16 treatment of any lease before assuming it's off-balance-sheet
• •Compare three-year totals ($34m/$43.2m/$52.5m-style figures), not headline hourly rates alone
• •Build refresh flexibility into any lease or reserved-capacity contract given the 3-4 year useful-life assumption
• •Separate predictable base load (reserve or buy) from bursty demand (rent or spot) rather than financing both the same way

Sources

Every figure in this article traces to the sources below.

• •NVIDIA — H100 Hopper family specifications
• •Introl — AI infrastructure financing, capex vs opex, GPU investment guide 2025
• •Introl — GPU procurement strategies: leasing, buying, reserved capacity 2025
• •GPUnex — Rent vs buy GPU server, 2026 pricing and TCO guidance
• •Digital Applied — GPU buy vs rent vs cloud, 2026 AI inference decision guide
• •Goldman Sachs — Assumptions shaping the scale of the AI build-out