Open Model Trade-Offs and Geothermal Power's Role in AI Infrastructure
The AI industry is simultaneously navigating two structural tensions that will shape how systems are built and operated over the next several years. The first concerns the practical costs and risks of deploying open-weight models — systems that offer flexibility but shift significant operational burden onto the organizations using them. The second involves energy: specifically, whether geothermal power can offer a more reliable and scalable alternative to the energy sources currently straining AI infrastructure.
These two pressures are not unrelated. As organizations adopt open models and run inference workloads internally rather than through managed APIs, their compute and energy footprints grow. The infrastructure question and the model deployment question are, in practice, the same question asked from different angles.
The case for open-weight models has centered on cost, customization, and data privacy. Organizations that fine-tune and self-host models avoid per-token API costs, retain control over sensitive data, and can adapt systems to domain-specific tasks without dependency on a third-party provider's roadmap. These advantages are real, but they come with trade-offs that are often underestimated at the adoption stage.
Open models require internal expertise to evaluate, fine-tune, and monitor. Safety and alignment properties that are handled upstream by frontier model developers must be addressed independently. Organizations deploying open models absorb the operational surface area of the model itself — including misuse risk, output quality control, and version management. For enterprises without dedicated ML infrastructure teams, these costs can quickly offset the savings from avoiding commercial APIs.
The energy dimension compounds this challenge. Running large open models internally, particularly at inference scale, demands sustained compute that draws heavily on local or regional power grids. Data centers have already encountered capacity limits in markets like Northern Virginia, Ireland, and Singapore. Power availability — not chip supply — is increasingly the binding constraint on AI expansion.
Geothermal energy has entered this conversation as a candidate solution. Unlike solar or wind, geothermal generation is continuous and not subject to intermittency. It draws from thermal gradients in the earth's crust and, where geologically viable, can provide baseload power with a small physical footprint relative to output. Iceland and Kenya have operated significant geothermal capacity for decades. The question for AI infrastructure is whether geothermal can be deployed at sufficient scale in regions where data center demand is concentrated — and at what cost.
Current geothermal development is expanding through enhanced geothermal systems, which use drilling and hydraulic techniques to access heat in regions without natural hydrothermal activity. Several projects in the western United States are now in development under this model. The technology is capital-intensive upfront and carries geological risk, but operational costs are low and the power output is stable — attributes that align well with the requirements of large-scale AI compute facilities.
The convergence of these two trends points toward a more distributed and operationally complex AI landscape. As open models proliferate and organizations internalize more of the compute stack, the energy and infrastructure requirements shift from centralized hyperscaler facilities to a broader set of regional operators. That transition demands more diverse energy sourcing, not less.
From an operational standpoint, enterprises evaluating open model adoption should treat energy and compute infrastructure as first-class considerations, not afterthoughts. The flexibility gains from self-hosting are real, but they are realized only when the underlying infrastructure — power, cooling, network, and hardware — is stable and cost-effective. Organizations that move to open deployment without solving the infrastructure layer will find the economics less favorable than projected.
Geothermal's promise is that it could provide one piece of that infrastructure puzzle at regional scale. Whether it develops fast enough to meet the rate of AI expansion is a separate and harder question — one that depends on permitting timelines, capital availability, and drilling technology as much as on demand.
Sources: — MIT Technology Review (https://www.technologyreview.com/2026/10/13/1146382/the-download-open-models-trade-offs-geothermal-power/)