AI Additionality Is the Wrong Solution to a Real Problem
The United States needs to think seriously about the energy demands of AI. On this point, Brian Deese and Lisa Hansmann, who’ve penned an interesting proposal on regulating the energy use of AI, are exactly right. Electric demand growth is a paradigm shift for the industry and a strategic scale challenge for the nation. Whatever the outcome of November’s elections, the incoming administration and Congress should make addressing electricity demand growth a top priority.
Despite agreement on this basic premise, the approach articulated by Deese and Hansmann arguably takes the policy conversation in the wrong direction. Pursued as the lynchpin of a policy response, this approach risks introducing uncertainty and delay into the AI sector, does little to address the real challenges on the supply side, and misses the big opportunity to radically scale ambition in the electric power sector.
Additionality for AI?
Deese and Hansmann propose a “National AI Additionality Framework.” This would extend the concept of additionality—originally developed around the Inflation Reduction Act (IRA)’s tax credit for low-carbon hydrogen production—onto AI data centers.
To limit the emission impacts of new electricity loads, this policy places the standards of additionality, time matching, and deliverability on electricity purchasing. Additionality says that the clean power must be from new zero-carbon sources; a data center couldn’t buy power from, for example, existing nuclear plants. Furthermore, since that power must be time matched on an hourly level to actual consumption, the current practice of offsetting consumption with renewable energy credits on an annual basis would be insufficient. Deliverability requires that the power bought be within the same region: wind energy from North Dakota cannot be used to power a data center in Virginia.
If unable or unwilling to co-locate clean energy, procure green energy directly from the market, or negotiate additionality-compliant tariffs from utilities, data center owners would have to pay a fee to a Department of Energy (DOE) strategic investment fund. These funds could then be redirected into technology demonstration projects, infrastructure investments, and other DOE programs.
A Misguided Approach
At face value, this policy model ensures that the AI data center boom goes hand in hand with a boom in new clean energy investment. What’s the problem?
First, though green hydrogen and AI data centers both represent large electric loads, the two industries operate on wildly different commercial and political planes. With hydrogen, the three pillars of additionality were introduced to address the specific problem posed by the exceedingly rich 45V tax credits which, without strict emissions accounting, could lead to vast federal outlay, electricity consumption, and negative emission impact through economically useless hydrogen production. There is little point in using federal tax dollars to subsidize decarbonization technology that increases emissions. Hence additionality is a requirement to access a very attractive policy “carrot.” AI database investment is entirely driven by private commercial decisions independent of federal subsidies, and whose output has immense commercial, innovation, and economic security value. Additionality here is just a policy “stick,” which will restrict the realization of this value.
Second, caution is warranted regarding the “success” story of green hydrogen pointed to by Deese and Hansmann. Two years after the IRA was signed into law and unlike other sectors that have seen rapid results—new solar factories, booming battery storage deployment, and many new electric vehicle (EV) and battery manufacturing plants—the United States still produces nearly zero green hydrogen. A bruising policy battle over the implementation of the additionality framework has created enormous uncertainty, delayed investment, and stalled development of a nascent industry.
It is exactly the bruising politics of additionality and ensuing uncertainty that must be avoided for the AI sector. The proposed backstop fee only increases the prospects for politicization and, besides, it’s not clear why strategically vital sectors should be targeted with what amounts to a tax.
Policymakers should be very weary of setting the precedent of targeting specific industries with electricity procurement regulations. The nation’s power grid faces a wide range of new loads in the coming decades including gigawatt scale semiconductor fabs, electricity-intensive battery manufacturing plants, and the vast electricity requirements of EVs. Deese and Hansmann’s proposal counterproductively seeds new ground for politicization and policy uncertainty over the exact industries and technologies we are trying to scale.
Third, additionality is a theoretically suspect concept in today’s power grid. Thanks to the IRA’s production tax credits, there is no shortage of private capital seeking to build wind, solar, and battery storage projects in the United States; vast and growing capacity waiting in interconnection queues points to this. If public policy bottlenecks were resolved, capital would gladly flow into transmission and nuclear investments as well. To a very real extent, as much clean energy capacity that can be built is being built.
Additionality frameworks drive demand for new clean energy resources, but in a world with rate-limited clean energy expansion, demand for clean energy isn’t the problem. All additionality does is pour fuel on a raging fire.
An AI additionality mandate will only drive database owners to outbid all other buyers for whatever supply of new clean energy resources exists, and all other buyers of clean power would disappear (if they were to face a strict additionality mandate) or default to buying power from the incumbent resource mix of existing renewable, gas, coal, and nuclear resources. While an additionality-compliant data center may result, the emissions impacts have still been shifted elsewhere on the grid and some potential demand (and economic, social, and strategic value) has been destroyed along the way.
To some extent, Deese and Hansmann recognize this. They’ve designed their additionality framework to affect change on supply-side constraints by squeezing demand-side participants. This approach is inefficient and misses the point. The tech companies building AI data centers are some of the most sophisticated and innovative buyers in the electric power sector and have been for well over a decade now. They are already pushing hard on many of these issues, but the reality is that demand-side customers have a very limited ability to alter on their own the public policy problems, which truly constrain clean energy expansion.
Target the Core Problems
Federal policy should not target AI data centers, nor any demand-side customers with procurement regulations or fees. This imposes costs, uncertainty, and delays on the exact investments that federal policy should be seeking to encourage and accelerate. The United States cannot afford to usher in two-plus years of uncertainty over AI investment the way additionality has done for green hydrogen production.
Instead, federal policy response should directly address the supply-side constraints that today give states, utilities, and corporate buyers a poor and narrow set of choices. There is a lot to this challenge, but it boils down to expanding transmission capacity, building nuclear reactors, and radically altering the permitting landscape for energy infrastructure.
The recently proposed bipartisan Energy Permitting Reform Act of 2024 applaudably targets some of these issues and serves as an excellent starting point. But scope and ambition must grow if the United States is to meet the strategic scale of the challenge. The United States should be thinking about transmission investment tax credits for national interest projects (widely defined to include greenfield, reconductoring, and advanced grid technology projects), direct federal offtake or guarantees for new reactor projects, continued efforts to reform and accelerate Nuclear Regulatory Commission permitting, and ambitious interconnection queue reform that allows more resources to connect under the existing grid. Congress should require grid operators to deploy hourly renewable energy accounting systems, which lay the groundwork for increased voluntary 24/7 clean power purchases by buyers and improve price signals in the sector.
Real progress from U.S. policymakers on these issues will unleash new investment and dynamism in the sector and flip the electric demand growth story from one of challenge to one of opportunity. An AI additionality policy takes the United States in the wrong direction, what the country needs is a national electricity growth strategy.
Cy McGeady is a fellow with the Energy Security and Climate Change Program at the Center for Strategic and International Studies in Washington, D.C.