The Investment Case for GPU Infrastructure

The affirmative case for GPU infrastructure as an investment asset rests on three pillars. First, demand visibility: AI adoption across enterprises, governments, and research institutions is early-stage. Global AI spending is projected to reach $632 billion by 2028 (IDC, 2024).

GPU compute underlies all of it;training, fine-tuning, and inference all require GPU capacity. Demand is structural, not cyclical. Second, supply constraints: building GPU-capable infrastructure requires power (constrained by grid capacity), data centre space (constrained by cooling requirements and land in prime locations), and GPU hardware (constrained by TSMC and HBM supply chains).

These constraints are real and long-duration;new power connections take 3-5 years, greenfield data centres take 2-3 years to build. Supply cannot expand as fast as demand, creating persistent pricing power for existing capacity. Third, long-term lease structures: serious GPU deployments;sovereign AI programmes, enterprise training capacity, hyperscaler overflow contracts;are underpinned by 3-5 year contracts with credit-quality counterparties. This creates infrastructure-like cash flow characteristics: predictable, long-duration revenue from creditworthy tenants.

The risks are real and deserve honest treatment. Technology obsolescence is the primary risk: GPU generations turn every 18-24 months, and customers with flexibility migrate to newer hardware. An investment in H100 infrastructure today competes against B200 alternatives within 2 years and whatever follows Blackwell within 4 years.

The asset's revenue-generating life is shorter than its physical life. This risk is manageable but requires explicit modelling;operators who assume indefinite revenue streams from specific hardware generations will be disappointed. Concentration risk is significant: NVIDIA controls 80-90% of AI training accelerator supply.

An NVIDIA pricing decision, supply disruption, or architectural change propagates immediately through the entire GPU infrastructure market. Customer concentration risk also exists in neocloud models: CoreWeave's early revenues were heavily concentrated in a small number of large AI lab customers. Finally, market normalisation risk: the 2022-2024 GPU supply shortage inflated pricing and margins. Operators who built business plans on 2023 peak pricing have found their models challenged.

The investment case for GPU infrastructure is strongest when: the investment is underpinned by committed contracts from creditworthy counterparties, the infrastructure is GPU-generation-agnostic (designed for refresh rather than a single hardware generation), depreciation and obsolescence assumptions are conservative (3-4 year primary period, 15-20% residual value), and the operator has demonstrated track record at scale. Infrastructure funds entering the GPU market in 2024-2025 at $19B+ CoreWeave valuations were buying into fully priced assets.

The better entry points have historically been either earlier (before the AI boom inflated valuations) or at specific assets with locked-in long-term contracts trading at predictable yields. Buying an asset at 6x revenue expecting infrastructure-like stability requires believing the revenue is genuinely stable, which requires scrutinising the underlying contracts carefully.

Sovereign AI infrastructure;GPU clusters funded and operated by national governments;represents the most defensible segment of the market. Sovereign programmes are not subject to commercial competition in the same way neocloud capacity is.

They have committed public funding, long time horizons, and political support that insulates them from short-term demand fluctuations. France's €109M national AI compute commitment, Saudi Arabia's $40B AI infrastructure investment, and the UK's expanding national compute capacity are backed by government budget commitments;not venture capital with a 5-year exit horizon. For operators able to win sovereign contracts;typically requiring local presence, compliance credibility, and technical capability at scale;the investment characteristics are superior: longer contracts, lower churn risk, creditworthy obligors, and insulation from neocloud pricing competition.

GPU infrastructure warrants a distinct framework from both conventional data centre investment and early-stage AI company investment. The correct comparators are: infrastructure assets with shorter technology cycles (satellite capacity, fibre networks), real assets with supply constraints (merchant power plants, LNG terminals), and technology-dependent recurring revenue businesses.

A disciplined framework models three scenarios: base case (current demand trajectory, hardware generation on historical 18-24 month cycle), downside case (demand growth slows 30%, hardware generation accelerates to 12 months, pricing compresses 20%), and upside case (sovereign AI drives demand 50% above base, supply constraints persist 12 months longer than expected). IRR sensitivity across these scenarios determines whether the investment is compelling at the proposed entry price. For detailed financial modelling, deal comparables, and investment case review, speak to our advisory team at disintermediate.global/services.

Bull case rests on structural AI demand growth, multi-year supply constraints (power, land, hardware), and long-term lease structures with creditworthy tenants

Key risks: hardware obsolescence (18-24 month generation cycles), NVIDIA concentration, customer concentration in neocloud models

Conservative model parameters: 3-4 year primary period, 15-20% residual value, 5-15% annual revenue decline in years 3-4 as competition intensifies

Sovereign AI programmes offer the most defensible investment characteristics;long contracts, government-backed obligors, insulated from neocloud competition

Scenario analysis (base/downside/upside) across demand, pricing, and hardware cycle assumptions is essential;single-case models will mislead