Catalyst: 2025 Climate Tech Trends

Start with the split between solar and wind, because the same cause runs through most of what follows. Globally 2024 was a record year for both, 599 gigawatts of solar and more than 130 gigawatts of wind, which Bullard notes would have looked like a fantasy to an analyst in 2008 who assumed it would take a $200 per ton global carbon price to get there. In the United States the two have come apart. Wind investment is well off a peak he places around late 2019 to mid-2020, solar is climbing past $40 billion in trailing twelve-month investment, and storage now attracts more US investment than wind does. His explanation is not primarily political: solar is simply faster and simpler to develop absent transmission reform, and storage is complementary to a high-solar grid. Politics makes it worse rather than causing it, and Kann’s framing is that it was already a tale of two markets before the new administration, which is unfavorable to renewables generally and has a long-standing particular animus toward wind. The tell is ERCOT in Texas, the least constrained large market in the country, where the same shape appears without the policy overlay: cumulative solar likely passing wind in 2025, and batteries growing faster still.

What the grid does with all that solar is the second thread, and it is a market-design failure rather than a technical one. California curtails more than 750 gigawatt-hours of solar in a typical spring, when it is sunny, cool and demand is low, and as of 2024 the state’s minimum net load across the entire spring was negative in aggregate, meaning intermittent generation exceeded total demand. Europe’s version is priced rather than dumped: instances of negative hourly power prices rose from 500 to 550 in 2022 to more than 9,000 last year, summed across member states and grids, which is more hours than exist in a year. Bullard’s point is not that this is unfixable but that nothing is currently rewarded for fixing it. We cannot see the counterfactual, since we do not know what California would look like without the roughly 10 gigawatts of storage it has already built, and the four-hour duration standard that shapes what gets built there is an artifact of written regulation rather than of physics, so the market has no incentive even to find out what longer-duration storage or flexible load would do. Kann’s version is blunter: hundreds of gigawatt-hours of clean electricity wasted every year and growing, an inefficiency staring you in the face, fixable by policy, by a storage business model or by a business model for intermittent load. Transmission is the same diagnosis in harsher form. The US built 3,200 miles of high-voltage long-distance transmission in 2013 and 125 miles last year, and Bullard argues the shortfall is not for lack of funding or of demand to use it but because building is difficult or impossible, which makes it a planning and permitting question rather than a market one.

Load growth is where the episode does its most useful reframing, and it cuts against the panic rather than feeding it. Twenty years ago the official reliability forecasts assumed ten-year compound growth of 1.5% to 2%. That eroded steadily to roughly half a percent by 2020 and has now unwound back above 1%. Kann’s point is that if you round up to about 1.5% you get triple the 2020 expectation, which explains the whiplash, but still less than the roughly 2% the system was running at in the early 2000s. This is not historic load growth; it is load growth the country did historically, arriving at a sector that had spent a decade getting used to stagnation and, in Bullard’s phrase, doing a very good job of making itself hard to build into. Two qualifications travel with it. Bullard notes the chart is a chart of expectations, and that in the period when half a percent was expected the realized growth in total electricity demand was zero. And both allow that today’s forecast may be too low rather than too high; Kann floats 3%.

Then DeepSeek, recorded the Monday its release knocked the market down and released that Thursday. Bullard’s angle is training hours as a direct proxy for electricity, since a training hour is GPU time and GPU time is power. Meta’s Llama 3.1 used just under 31 million training hours at 405 billion parameters; DeepSeek reached comparable results with 671 billion parameters and fewer than 2.8 million training hours, built by a Chinese hedge fund’s $5.5 million rather than billions in venture capital and under constrained access to chips. Kann puts it at roughly ten times more efficient on training and about thirty times on inference, flagging the second as his own understanding, and lays out the two readings explicitly: either the gigawatts are not needed, which is what the stock market priced that day, or the Jevons paradox applies and cheap compute gets consumed in far greater volume, leaving energy demand flat or higher. Bullard observes that Microsoft’s chief executive invoked Jevons by name while reaffirming $80 billion of capital expenditure, and Kann notes that he has an interest in that reading. Bullard expects the truth in between and says the harder questions are who captures the value and whether lightweight models get hosted in third-party facilities or on a company’s own. The thread Kann pulls out is granularity: the entire reason for gigawatt-scale campuses is that bigger training runs wanted to sit on one site, and 20 megawatts is an interconnectable load while a gigawatt is not, so if this efficiency buys the same performance at 20 megawatts it changes where compute can be sited and how it can be powered. That the answer is undetermined is the honest conclusion, and it is reinforced by the cost structure: energy is a small line item in training, so the buyer is one that cares about whether power is available and when, not what it costs.