AI × Climate: Adaptation, the Grid, and What Comes Next
Takeaways from CLAI Ventures' SF Climate Week 2026 event at DLA Piper, San Francisco
For the second year running, CLAI Ventures hosted a community event during SF Climate Week, this time with a sharper focus: how is AI being deployed at the frontier of climate adaptation, and what will it take to power AI sustainably? We brought together practitioners from utilities, national labs, grid infrastructure, industry, and climate tech ventures for two panels that went deep on questions the industry is still figuring out. Here is what we heard.
Panel 1: Climate Adaptation × AI
We were joined by Ravi Jain, whose career spans Google X (CTPO of Tapestry), Amazon (VP of Search AI and Alexa), Meta, and Vulcan Inc., and who is now working on AI for climate; Sean Gilleran, Chief Machine Learning Scientist at PG&E, where he leads AI efforts to stop catastrophic wildfires; and Mayur Mudigonda, Founder and CEO of Vayuh, a CLAI portfolio company, and former climate AI researcher at Lawrence Berkeley National Laboratory.
We are at the tipping point, and AI is the moonshot tool
Ravi opened by situating climate change as not merely an environmental challenge but the single most pervasive problem facing humanity today and we are at literal tipping points. What gives him optimism is that the three key elements of radical solutions, namely policy, finance, and technology, are finally converging. AI sits at the center of the technology piece: for forecasting, for grid optimization, for financial decision support systems capable of handling the high variability and deeply interrelated unknowns of climate risk. He called this category “AI for the Hard Stuff,” meaning problems at the intersection of the physical world and complex systems. The bottleneck, he argued, is not technical. It is what he calls the Imagination Bottleneck: getting engineers, business leaders, and policymakers to direct their ambitions toward the gnarly, hard problems that really matter, rather than the glamorous advances at the frontier of core AI. The potential is enormous, as these are UN SDG scale problems with the capacity to unlock trillions in economic value.
Why AI native weather models beat physics models
Mayur walked through the fundamental limitation of traditional forecasting built on physical simulations: you can neither write all the physics of weather and climate, nor simulate it fast enough to generate the many scenarios that forecasting actually requires. AI models overcome both constraints. They can integrate large volumes of disparate data (satellite, ground observations, radar, reanalysis, and gridded datasets), and their nonlinear learning functions improve accuracy by roughly 20% over conventional models. Crucially, fast and cheap inference makes it possible to generate thousands of scenarios, which is the core value of probabilistic forecasting. AI models can also generalize to new geographies or low data environments through fine tuning. Vayuh applies this specifically to extreme weather events and climate risk for the insurance sector, where precision at the asset level matters most.
What AI driven wildfire risk actually looks like inside a utility
Sean grounded the conversation in operational reality. PG&E has a simple but absolute commitment: catastrophic wildfires will stop. Achieving that requires multiple layers of protection, including physical investments in grid hardening and undergrounding, operational mitigations like more sensitive protection settings, and measures of last resort such as Public Safety Power Shutoffs (PSPS). AI models inform every one of those layers: weather models, outage and failure prediction, ignition probability, and fire spread modeling. The hardest unsolved problems, he said, are the inherent uncertainty and chaos of the atmosphere and the sheer scale of physical infrastructure that AI recommendations must map onto in real time. From Mayur’s perspective, what utilities still are not getting from AI weather intelligence is hyperlocal precision at the asset level, meaning forecasts specific enough to drive decisions about individual spans of line, not just regional weather patterns.
Beyond insurance: investing in good outcomes, not just protecting against bad ones
The panel also turned to the question of risk pricing powered by AI. The insurance industry is beginning to price climate risk at the asset level, a development that Vayuh’s models feed directly into, and that companies like our portfolio company FutureProof are building entire business models around. Ravi pushed the conversation further: insurance is ultimately a mechanism for protecting ourselves against bad things that could happen, but it does not address the underlying causes. A more durable approach is to invest in what he called “good things that must happen,” meaning economically viable solutions designed to last. He cited mangrove forests, salt marshes, and coral reefs as natural buffers against storm surges, with wetland and reef restoration in the Gulf of Mexico yielding benefit to cost ratios greater than seven to one. AI can accelerate these solutions not just through climate forecasting but through intelligence, planning, and operational support for complex, unconventional projects rooted in nature.
The gap between prediction and action is a human problem as much as a technical one
The panel closed on the gap between forecast and action. Inside PG&E, Sean described the path from an AI recommendation to an operator decision as requiring rigorous validation, backcasting, and a genuine transfer of ownership to the humans who will act on the model’s output. People have to trust the tool before they will stake safety decisions on it. Mayur took a broader view: climate change is anthropogenic, and solutions that do not address the human condition will remain incomplete. He pointed to movements like Save Soil and Rally for Rivers as examples of community level engagement involving multiple stakeholders, bringing together model vendors, utilities, and local governance in ways that technical tools alone cannot substitute for. Raising human consciousness, he said, is what keeps him up at night more than any technical bottleneck.
Panel 2: AI × Energy: Powering the Grid in the Age of AI
For the second panel we were joined by Chris Richardson, Partner at Realize 2050, which partners with utilities, industry leaders, and governments to accelerate the energy transition; Pratha Pawar, Applied Scientist at AWS working at the intersection of AI and energy, focused on building practical AI systems for data center growth, clean energy, and sustainability.; and Tom Nudell, CEO and cofounder of Piq Energy, which builds grid planning software and data systems to accelerate the interconnection of renewable energy and dynamic loads like data centers.
The grid’s interconnection crisis, and what flexibility can do about it
Chris opened with a broad view of where US electric utilities stand: aging infrastructure, regulatory complexity, misaligned incentives, and an interconnection queue that has become a structural bottleneck for the energy transition. Tom then made the clarifying point that the root cause of the interconnection crisis is that utilities have historically studied a single snapshot representing the worst possible conditions, effectively one hour of the year, and designed solutions like new lines and line upgrades that take years to permit and build. That one hour constrains 99.9% of the year. Piq Energy’s insight is that if you are willing to manage the grid dynamically for that 0.01% of the time, you can unlock enormous latent capacity far faster than traditional infrastructure buildout. Tom calls this “flexible interconnection,” drawing on point flexibility from shifting load directly, aggregated flexibility from distributed assets, and network flexibility from technology that enhances the grid, all of which can be deployed in under a year. This reframing, moving from static worst case planning to dynamic optimization, may be the most important conceptual shift in grid management of the decade.
The physics of the grid is changing, and AI has to keep up
Tom also raised a less discussed challenge. Inverter based resources (IBRs), the power electronics that connect solar, wind, and batteries to the grid, do not respond based on their inherent physics the way traditional generators do. They respond to programmed control logic. As data centers powered by AI, which are themselves predominantly IBR based, connect to the grid at scale, the modeling challenge becomes acute. NERC issued a rare Level 2 alert last fall telling utilities they need to do a better job modeling these assets. Piq’s software is designed to manage all types of grid models and simulations, making it easier for both human engineers and AI agents working alongside them to navigate this new complexity.
AI agents for the grid: what practical looks like in the energy industry today
Pratha brought the industry perspective. At AWS, he leads agentic AI initiatives across the sustainability organization, building tools that optimize Amazon’s renewable energy portfolio while supporting the sustainable growth of its global data center infrastructure. His NeurIPS work addressed a fundamental constraint, namely data visibility. You cannot optimize what you cannot see, and the grid’s data infrastructure was not built for the resolution or latency that optimization powered by AI actually requires. On practical deployment, he argued that the biggest near-term wins for AI agents in energy are not in running the grid — they are in helping experts move faster through the planning, analysis, and operational workflows that surround it. Start with the bounded, specialist work. Earn trust. Then go harder.
The inertia that technology alone cannot overcome
Chris closed with a diagnosis of the nontechnical barriers. Even when the electrons are available and the capacity exists on the grid, the inertia of large organizations, regulatory structures, and entrenched operating norms slows deployment dramatically. Utilities operate under strict reliability mandates and are deeply suspicious of the inherent unpredictability of AI recommendations. Hyperscalers face their own organizational complexity. The companies and advisors that can navigate both the technical and the institutional, disrupting industries without breaking them as Chris put it, are the ones that will move fastest. For 2026, all three panelists agreed on one theme: the question is no longer whether AI will reshape energy, but whether the organizational and policy environment can absorb change at the speed the climate requires.
A Note from CLAI Ventures
Both panels reinforced what we have believed since we launched: the intersection of AI and climate is where some of the most important and most tractable problems of our generation will be solved. The founders in our portfolio, including Brightband, Vayuh, Powerline, Rhizome, FutureProof, and others, are building exactly the tools that the panelists described as missing or nascent. We thank our speakers, DLA Piper for hosting, and the SF Climate Week community for a great event.
We invest in AI for a more sustainable world. If you are working on something at this intersection, we would love to hear from you.
This post was written by Ajay Gupta and the CLAI Ventures team, a Silicon Valley-based fund investing at the intersection of AI and climate. You can reach us at ajay@clai.vc and team@clai.vc.