The Information
Power scarcity is all anyone talks about in the AI world. Funny how we actually have plenty. GridCare, which emerged last year from a Stanford University sustainability accelerator, has surprised AI developers and even utilities by finding power that they can tap even in congested urban markets where the grid appeared sold out. Power grids are designed to handle the hottest and coldest days and other periods of peak demand. Utilities have historically built two to three times more capacity than they really needed because calculating the right amount of excess is hard. It turns out that the grid was maxing out less than they thought, GridCare says. The startup uses artificial intelligence to run quadrillions of simulations to find extra power that utilities could produce and deliver using existing assets. To get the data it needs, GridCare does the grunt work of combining siloed data sources inside utilities including operational data from power substations, weather patterns and grid incidents. It raised $13.5 million last year led by Xora, a venture arm of Singapore’s Temasek, and is announcing a new funding round in May. If GridCare succeeds, it would be yet another case of Silicon Valley finding invisible, idle capacity and putting it to use, akin to what Uber and Turo did for cars and Airbnb for spare rooms. The Information got exclusive peeks into GridCare’s work to get power to what CEO Amit Narayan calls stuck projects in impossible places. Location has become increasingly important in AI as more businesses adopt the technology and need rapid response to their queries. For example, enterprise AI services provider Gruve sought power in the super-saturated data center markets of Santa Clara, Calif., Portland, Ore., Seattle and Los Angeles. GridCare located 150 megawatts of potential power within data centers, other commercial buildings or raw land where, with the addition of some software and backup tools, Gruve could fit. Stanford researchers working with GridCare have found that U.S. power networks run at just 30% utilization on average. Narayan believes U.S. grid utilization can safely rise to 60% just by employing “code instead of copper.” That would unlock roughly 300 more gigawatts of power that is already paid for—the equivalent of 150 Hoover Dams—on a 1.3-terawatt U.S. grid, he said. If he’s right, then U.S. utilities might be wasting some of the $1.4 trillion they intend to spend on new capital projects by 2030, largely to power AI. Companies developing AI including Meta and Microsoft are also planning to shell out epic sums to build their own power and keep electricity rates for consumers from rising further. Narayan said he sees some of the added spending as the equivalent of an airline “buying a bunch of planes because it is Thanksgiving” while running well below capacity for the rest of the year. GridCare argues that adding new power demand will generate revenue for utilities. The company says that adding a one gigawatt data center to an existing system could generate $350 million in additional revenue and $142 million in earnings, which could fund a 5% rate decrease for customers. There’s a long history of innovators failing to cajole utilities into becoming smarter and more efficient, however. One reason is that utilities are risk-averse, because the only time they get in real trouble is when the power goes out. Another is that industry executives have become masterful at convincing regulators to approve new power plants, which generate a guaranteed rate of return for utilities. So the industry has proved stubbornly resistant to efficiency moves to make more of the resources they’ve already built. Still, if there’s ever been a moment when companies like GridCare could have an impact, it’s now. GridCare can find spare power in six to 12 months, versus multiple years to build new generating capacity or upgrade the grid. “This is trying to solve for the present,” said Ram Shriram, a founding Google board member and longtime Silicon Valley investor who recently joined GridCare’s board. He added that better utilizing the grid “also improves affordability for the payers already in the system.” Utilities might actually listen to Narayan, who became GridCare’s CEO after he sold another grid software firm he spun out of Stanford, AutoGrid, to Schneider Electric. AutoGrid helps utilities manage electricity flows from energy sources such as solar panels and batteries that are spread across a company’s territory. Much of GridCare’s work centers on helping utilities be certain they can meet peak demand even after they’ve added more customers in the form of data centers to their networks. GridCare began working last year with Oregon utility Portland General Electric to activate 400 megawatts of available power for customers willing to use tools like batteries to retreat from the grid when it was under stress. Last month, National Grid signed on as a partner to find hidden power in its New York territory. To make it work, a customer who squeezes onto a crowded grid may have to make concessions. Typically data centers demand “five nines” (99.999%) of uptime, so they may have to settle for three nines (99.9%) or maybe even two nines (99%) instead. Still, 99.9% can mean 8.76 hours a year of outages, enough to hurt a data center’s business. Those concessions though can be addressed in other ways. Battery backups have reached the point where they can prevent much of the downtime. Data centers can also take advantage of the scale of the U.S. power grid. When one region is facing peak demand, others may have surplus power. One startup we’ve cited before, Emerald AI , can transfer workloads in milliseconds between AI data centers in faraway locations. It just completed a pilot where it sent computing from Virginia to an affiliated facility in Illinois with virtually no computing time lost. Being able to do this can win a spot on a crowded grid. In an age of backbreaking AI capex, GridCare could help Silicon Valley once again go capital light. In other news: In a closely watched case, Wisconsin’s utility commissioners announced substantially strengthened safeguards Friday to make large data center customers pick up the entire tab for new power built to serve them. Microsoft, Oracle and Meta have large projects underway there. Public Service Commissioner Kristy Nieto said, “Wisconsin customers should not pay a single cent to subsidize the service of data centers or very large customers – not now and not decades from now.” A much-awaited IPO filing from enhanced geothermal darling Fervo Energy revealed that Google (a key long-term backer of the company) has negotiated veto power over Fervo’s partnerships and rights to its technology, potentially edging out deals other hyperscalers might want to make for its power. Benzinga called it “a peek into the high-stakes ‘Game of Thrones’ currently being played by Big Tech over the future of the power grid.” New From Our Reporters Exclusive Behind Cursor’s Deal With SpaceX, Anthropic and Compute Costs Loomed Large By Cory Weinberg, Julia Hornstein, Erin Woo and Katie Roof Crypto Hack Puts Big Lender in Turmoil By Yueqi Yang Exclusive Berkshire Hathaway, Chubb Win Approval to Drop AI Insurance Coverage By Laura Bratton
Go to News Site