Former Tesla exec Drew Baglino: AI load growth won’t be 'as dramatic as people think'

Anxiety about the energy demands of AI persist — and Tesla’s latest strategic shift to autonomy probably won’t help matters.

The EV juggernaut’s renewed focus on autonomous robotaxis may eventually add to the pressure data centers place on the power grid. Meanwhile, the company has thrown its weight behind a particular vision of a decarbonized global economy, known as the Master Plan Part 3. 

The paper, which makes an economic case for extensive electrification and a reliance on wind and solar, was published in March of 2023 — after the release of ChatGPT, but before the magnitude of AI-related energy demand was clear. But as former Tesla executive Drew Baglino outlined on Catalyst this week, its authors think it holds up.

• The top line: Baglino, who recently resigned from Tesla amid a major shakeup at the company that saw the lay-off of its entire charging team, countered concerns about the energy demands of AI. He also expressed optimism about overcoming other hurdles to the company’s vision for decarbonization, including transmission bottlenecks and mineral requirements. 
• The current take: “There's a lot of noise out there about whether a sustainable energy economy is actually feasible, not only technically feasible, but commercially feasible.” Baglino said. “For a company like Tesla, where the mission is to accelerate the transition to sustainable energy, the broader feasibility needs to be settled,” he said. “It shouldn't be considered a question.” Baglino argued that the Master Plan Part 3 demonstrates this feasibility, down to specific investments and materials. 

In Baglino’s 18 years at Tesla, he worked on batteries, cars, and even its lithium refinery. In the aftermath of his resignation, though, he was most eager to talk about the Master Plan Part 3 — and to counter concerns about it.

He said load growth from AI “won’t be as dramatic as people think,” and predicted that several factors will temper demand, including technological innovation and load-shifting. He also anticipated an eventual slowdown in investor enthusiasm if AI fails to generate significant revenue. 

In addition, he expressed optimism for the buildout of transmission capacity in the United States, despite the slow pace of buildout.

AI-driven load growth

Tesla’s plan, released a year ago, did not predict rising power demand driven by data centers and manufacturing. AI in particular has become a top concern among utilities, industry observers, and elected officials. Microsoft, for example, has seen its emissions rise since providing data center services for OpenAI’s generative AI bot ChatGPT.

Baglino, however, predicted several factors would temper AI-driven load growth.

He noted ongoing advancements in chip efficiency, but also suggested opportunities to improve efficiency in the other circuitry, such as memory interfaces that “have not seen as much innovation.” Latency would also be a self-limiting factor, Baglino argued; as [processor] arrays get bigger and the energy penalty to keep them connected gets bigger, he expects that to drive innovation. 

“Any improvement in latency is going to be bringing things closer together, which is going to use less power,” he said.

He also predicted that data centers would find ways to soak up excess renewable power by placing generation behind the meter and using more computational resources when clean power is available, such as through time-of-day pricing.

“If you co-locate with wind or solar, especially solar that's going to be curtailed or seeing lower market prices during the middle of the day like in California, all of a sudden you have a higher value end use to send that power,” Baglino said.

He suggested that AI providers offer cheaper time-of-day rates for computation, such as between 10 a.m. and 2 p.m. when there’s more renewable energy being generated.

“And that's when people are working anyway,” he added, “so maybe it's actually complementary.”

Some data center operators shape computational load, for example, by moving less urgent tasks to times when the grid has more clean energy, but time-of-day pricing is not currently common practice in the industry.

Catalyst host and investor Shayle Kann agreed that shaping AI load could represent a potential “paradigm shift” in the industry, saying that hyperscale data center operators are currently offering “low latency and extraordinarily high reliability with flat pricing,” but AI-load shaping “may just get forced into existence” by the increasing scarcity of power in some regions, especially low-carbon power.

In addition, Baglino anticipated that investor enthusiasm might cool off. 

“At some point, the investment people flooding all of the investment into AI [are] going to be like, ‘well, where’s the return?’” he said. “And so that until there’s applications where there’s real money flowing in the other direction, we may see a little bit of a correction as well.”

Transmission and critical mineral bottlenecks

The extensive electrification and renewables buildout outlined in Tesla’s Master Plan Part 3 will also rely on a dramatic expansion of U.S. grid capacity, and of its access to critical minerals 

The Princeton Net Zero America modeling study estimated that transmission capacity would need to grow four times by 2050. 

Baglino, however, was also optimistic that the U.S. would build enough transmission, despite the slow pace of buildout, citing the viability of intra-state transmission, undergrounding lines, and using existing rights-of-way, like railroads and highways. 

He also pointed to a recent major rule change from the Federal Energy Regulatory Commission that is likely to make it easier to plan transmission projects.

In addition, he argued that incremental improvements in the transmission network will grow in the way that other networks have, like fossil fuel pipelines and the fiber-optic cables that make up the internet. Those interconnected systems, he said, get "built up over time and they just become more and more efficient and effective with every link that gets interconnected.” 

Lastly, he pointed to the “huge economic incentive” from high price differentials across the grid. He’s confident that actors involved will ultimately “close that nodal price gap” by building transmission capacity.

Meanwhile, Baglino is similarly optimistic that the major mineral requirements of a high-electrification, high-renewables scenario will work themselves out.

The International Energy Agency estimates that the global demand for critical minerals, like copper, lithium, and nickel, will grow six times by 2040 — and many of those resources are rendered inaccessible by the U.S. due to geopolitical and permitting challenges. 

But, Baglino said, manufacturers are able to substitute scarce materials for more abundant ones, as pricing signals evolve.

He was also optimistic about the prospect of recycling batteries, magnets, and solar panels, among other technologies. However, he called for the domestic development of capital-project expertise, something he said the U.S. lacked, but China has in spades.

The U.S. CHIPS Act and infrastructure law, he said, is designed to support the development of such domestic expertise, but it’s not well-developed yet: “The ecosystem just isn't really there of talented consultants [and] contractors.”

Lastly, he argued that permitting laws need to change to allow regulatory certainty and permitting by rule. He pointed to the way that Texas allows large capital projects, like a lithium refinery, to be built by default as long as they meet legal requirements, rather than by ad-hoc review by local elected officials. He argued that critical-mineral suppliers, like mines, refineries, and other processing facilities, need this kind of regulatory certainty to develop. 

“If the rules just stay the same, then innovation will happen and companies can go out and solve these problems, ” he said.

Listen to the full episode of Catalyst: