From: redpointai
The intersection of artificial intelligence (AI) and energy is a critical and fascinating area of discussion, especially concerning the massive energy demands of AI compute buildouts [00:00:44]. Mike Schroepfer, former CTO at Facebook and founder of Gigascale, a venture capital firm investing in climate change solutions, has shared insights on the energy requirements for broadly democratizing access to AI globally [00:00:47].
AI’s Growing Energy Demand
The current demand for AI compute is significantly accelerating the need for energy infrastructure [00:00:07]. This is seen as a positive development, as the demand from technically savvy and well-funded customers is pulling forward the deployment of new energy solutions [00:02:09].
The Need for Grid Expansion
Even without AI, the United States needs to expand its energy grid by approximately five times by 2050 to meet goals such as converting all gas-powered vehicles to EVs and supporting domestic manufacturing for industries like steel, cement, and concrete, all of which require immense energy [00:02:18].
Democratizing Access and Comfort
Technology is the key to enabling eight billion people to live in comfort and safety [00:02:27]. Many people globally lack basic necessities like air conditioning or clean water, which are fundamentally power problems [0003:57]. If the cost of energy could be reduced tenfold and deployed universally, it would unlock the possibility of bringing many more people to a desirable standard of living [00:04:11]. The aspiration for an AI agent dedicated to an individual 24/7 highlights the immense power requirements this would entail globally [00:04:26].
Current and Future Energy Solutions
The focus for meeting these escalating energy demands is on generating significantly more energy in a sustainable and cost-effective manner [00:04:56].
Solar Power
Solar energy is currently the most cost-effective way to add new electrons to the grid, accounting for 80% of new energy on the US grid in 2024 [00:05:09]. However, solar power only works about 25% of the time, making it insufficient for data centers that need to run 24/7 [00:05:22].
Fusion
Fusion is considered a promising future energy source due to its incredible power density [00:05:54]. A single supertanker could potentially fuel the entire United States power grid for a year with fusion, or a pickup truck could fuel a major power plant for a year, demonstrating the vast energy output from minimal matter input [00:06:12]. Fusion also has the advantage of not producing waste concerns [00:06:25].
Innovative Offshore Compute Platforms
Companies like Panacea are developing offshore compute platforms that float in the ocean, generating energy by harnessing wave power [00:06:44]. These platforms combine energy generation with natural immersive cooling from the ocean, potentially creating the cheapest AI inference platform [00:06:59].
Intersection of AI and Energy Sectors
Hyperscalers (large cloud providers) are already making significant investments in energy, announcing purchase agreements for new and existing nuclear power plants [00:07:51]. Meta, for instance, has issued RFPs for next-generation fusion plants [00:08:01]. This deep involvement by AI companies underscores that energy is now a major conversation point, alongside data, computation, and algorithms [00:08:20].
Shift from Training to Inference Compute
The compute demand for AI is increasingly moving from training to inference time, especially with reasoning models [00:08:35]. The cheaper these inference requests become, the higher the demand, indicating a near-infinite demand for AI inference compute at the right price point [00:08:48]. Companies that can provide abundant, low-cost energy will be crucial enablers for the growth of hyperscalers [00:09:01].
Historical Parallels and Future Outlook
The historical trend of decreasing compute costs, from assembly language to high-level programming languages, has enabled massive progress [00:09:31]. Similarly, if energy costs decrease by 10% year-over-year for the next 20-30 years, it could lead to transformative advancements, from universally accessible AI compute to widespread air-conditioned comfort and new manufacturing capabilities [00:10:25].
Climate Change and Short-Term Decisions
While some in the AI community believe that AGI will eventually solve climate change, leading to short-term acceptance of natural gas for data centers, this view is seen as “punting the problem down the hill” [00:11:03]. The next five years are expected to be challenging, with some new gas assets coming online for AI to meet immediate power needs [00:11:24]. However, long-term planning must involve simultaneous investment in sustainable solutions like solar, battery backup, next-generation geothermal, and fusion [00:12:15]. Hyperscalers are actively making bets in these diverse energy areas [00:13:07].
Lessons from Capacity Planning
Predicting future compute needs is inherently difficult, as it involves forecasting product adoption years in advance [00:24:48]. However, underpredicting capacity can be more detrimental than overpredicting, as a lack of capacity can hinder product and technical progress [00:25:32]. While repurposing excess compute capacity (like GPUs) is possible, the physical world’s long lead times for infrastructure (e.g., data centers) pose a significant “impedance mismatch” for decision-makers [00:24:18].
AI Applications in Climate Change
AI can address critical problems in climate change:
- Exploration: Using AI to identify optimal drilling locations for geothermal energy or to find deposits of critical materials like copper and hydrogen, reducing the need for extensive physical exploration [00:30:30].
- Prediction: Improving weather prediction and risk assessment for insurance purposes [00:31:14].
- Materials Discovery: Inventing or optimizing new materials for carbon capture, catalysts, and other reactions [00:31:27]. This involves searching vast multi-dimensional spaces to find materials with specific properties, accelerating the discovery and testing process [00:31:37]. This approach also applies to pharmaceutical target discovery [00:31:59].
Challenges in Scaling AI-Driven Climate Solutions
While AI can accelerate discovery, it often addresses only a small portion (e.g., sub-10%) of the overall time and cost involved in bringing new materials or solutions to market [00:33:11]. The primary challenge often remains scaling manufacturing and securing customers [00:33:23]. Solutions that provide direct, end-to-end impact are more compelling, such as AI for geothermal exploration, which directly impacts the core activity that people care about [00:33:34].
An example of a direct consumer application is using AI with thermal cameras on phones to analyze home efficiency, providing actionable recommendations for insulation or appliance replacement with clear payback periods [00:34:04].