Conference Presentation, Panel
The Trillion Dollar Question: Financing AI's Infrastructure | Modular, IREN & More | RAISE 2026
RAISE SummitTim Davis, Kent Draper, James B. Krellenstein, Wayne Nelms, Erwan Menard, Jeremie Eliahou Ontiveros, Jeremy, Tim Buckley, James Kallenstein
- The industry is projected to become increasingly capital intensive across all layers, including power acquisition, with financing requirements rising to support gigawatt-scale expansions over the 12 months leading up to May, necessitating a mix of GPU-specific debt, lease-style financing, and convertible debt.
- Capital markets are developing a willingness to attribute value to the post-contracted period of GPU assets and fund sub-investment grade entities, potentially unlocking significant debt capital for vertically integrated providers and smaller AI labs, though supply chain pressure points currently limit the ability to serve total demand.
- H100 on-demand pricing is forecast to rise from approximately $1.70 to closer to $2.50 per hour in the fall of 2025, while memory prices are expected to continue fluctuating with a noted 12x increase in the last year as providers optimize cost per token.
- Software is anticipated to be the primary driver for radical changes in industry financing and compute fungibility, with requirements for unified platforms enabling easier movement between different silicon to de-risk investments for hyperscalers adopting alternate compute.
- Market strategies are shifting toward multi-tenant managed exchanges and serving smaller frontier labs, as future growth vectors include disaggregation between pre-fill and decode processes on the inference side within the next 24 months.
- Nuclear technology deployment aims to solve continuous supply issues via uprates on existing reactors across 41 U.S. sites on a three-to-five-year timeline, contrasting with the approximately one-decade timeline required for new nuclear builds, which are expected to pave the way for future infrastructure.
- High utilization rates are identified as a primary requirement for alternate compute adoption, addressing current enterprise utilization of 50 to 60% caused by sinusoidal traffic patterns, while long-term token sales for 2032-2033 are noted as difficult to secure given commodity market dynamics.
- Some GPUs are predicted to become less valuable over time once software enables better compute fungibility, and hyperscalers seeking to sell alternate silicon must rely on unified software to overcome the risks associated with current infrastructure heterogeneity.