The Trillion Dollar Question: Financing AI's Infrastructure | Modular, IREN & More | RAISE 2026

Industry Capital Intensity: The AI infrastructure sector is experiencing a structural shift toward extreme capital intensity, where securing power, data center construction, and GPU deployment are all becoming increasingly expensive.

• Data center construction costs for tier-1, liquid-cooled facilities now range from $40 million to $50 million per megawatt.
• Electric power load growth in the US is currently split evenly: 50% driven by data centers and 50% by diffuse building electrification and EVs.

Financing Structures & De-risking: Neoclouds and AI factories are employing varied financing models to match asset lifecycles, moving beyond traditional equity-only funding for non-NVIDIA silicon.

• Iron secured a five-year contract with Microsoft for GB300s, financed through a debt instrument rated A investment-grade (one of the highest in the market).
  • This structure funded 96% of the capital commitment via a blend of Microsoft prepayment and debt at a 3% blended cost.
• Crusoe utilizes a three-prong monetization strategy to align financing with revenue streams:
  • Selling megawatts via lease-guaranteed cash injections for hyperscalers (e.g., OpenAI, Microsoft).
  • Selling GPU hours via standard compute leasing.
  • Selling tokens via managed inference APIs with SLAs, which command the highest margin profiles.
• Lenders are increasingly willing to attribute value to "post-contracted" periods for data center assets, allowing vertically integrated providers to borrow significantly more against assets than current contract values alone would permit.

Asset Lifecycle & Financing Mismatch: A significant disconnect exists between the long life of physical assets and the shorter economic life of compute hardware, though market dynamics are evolving to address this.

• GPU Asset Life: The useful life of high-end GPUs (e.g., H100) is extending beyond initial 5-7 year predictions due to rising on-demand pricing (approx. $2.50/hour in late 2025, up from ~$1.70).
• Data Center Asset Life: Facilities typically have a 15-20 year useful life, creating a mismatch with shorter GPU contracts.
• Solutions: Participants advocate for futures markets and forward curves for compute pricing and GPU residual values to enable financing against end-of-life hardware (years 4-8).

Software-Driven Fungibility: The "moat" of legacy silicon providers is shifting from hardware lock-in to software ecosystems, with new platforms aiming to make compute fungible across different silicon architectures.

• Tim Buckley (Modula/Qualcomm): Argues that software is the critical factor enabling debt underwriting for non-NVIDIA silicon by allowing assets to be easily moved or resold between data centers.
  • Without software fungibility, alternative silicon must be financed via equity rather than debt due to lower liquidity.
• Vertical Integration Trend: Industry players (e.g., Google, OpenAI, Anthropic) are striving to be vertically integrated to capture the highest margin: selling tokens rather than raw compute.
• Hardware Agnosticism: Enterprises require unified software abstraction layers to run workloads across heterogeneous silicon (CPUs, GPUs, NPUs) without maintaining multiple proprietary stacks.

Energy Supply & Nuclear Uprates: Addressing the energy bottleneck, Alva Energy and others are pivoting to "behind-the-meter" power solutions, specifically nuclear uprates, to bypass the slow timelines of new builds.

• Alva Energy has developed technology to uprate existing nuclear reactors by 200-300 MW electrical per reactor.
  • Deployment timeline is 3-5 years, comparable to gas turbines, but without carbon emissions or fossil fuel price volatility.
  • Projects are structured as off-balance sheet, project-financed entities with fixed-price EPC contracts to mitigate cost-overrun risks.
• Hyperscaler Engagement: Major players are securing nuclear capacity at significant premiums (e.g., Microsoft's Three Mile Island deal at $116/MWh for 15+ years).
• Regulatory Path: Uprates utilize existing licenses and environmental impact statements, effectively bypassing the regulatory bottlenecks that delay new nuclear builds (which take 5-10 years).

Chipmaker Backstops & Market Dynamics: Chipmakers (NVIDIA, Broadcom, Google) are entering the financing space by backstopping capacity, though panelists view this as a response to demand rather than a circular economy.

• Sustainability: These backstops provide credit support for upfront capital, allowing owners to service shorter-term, higher-margin contracts; they are not necessarily "circular" if the underlying demand is genuine.
• Demand Context: Overwhelming demand is the primary driver, evidenced by 5-year reservation contracts for future chips (e.g., Blackwell/Vera Rubin) sold before deployment.
• Scarcity Translation: GPU and megawatt scarcity is directly translating into token scarcity, as seen in rising prices on platforms like OpenRouter.

Market Trends & Future Outlook: The industry is moving toward granular, fungible compute markets and expanding financing to sub-investment grade counterparties.

• Funding Sub-Investment Grade: There is a noted increase in lender willingness to fund smaller AI labs and enterprises previously deemed too risky, enabling higher margin token-based service models.
• Commodity Trading: Ornn (Wayne's company) positions compute as a commodity, aggregating excess capacity and seeking to create a futures index for GPU pricing to standardize financing.
• Utilization Challenges: Standard enterprise GPU utilization sits at only 50-60%, creating pressure to maximize asset usage through software optimization and heterogeneous workload orchestration.