Research Triangle Energy Consortium

Artificial Intelligence (AI), and its most recent avatar, generative AI, holds promise for industrial efficiency. Few will doubt that premise. How much, how soon, may well be debated. But not whether. In the midst of the euphoria, especially the exploding market cap of Nvidia, the computational lynchpin, lurks an uncomfortable truth. Well, maybe not truth, but certainly a firmly supportable view: this development will delay the decarbonization of industries, especially clean energy alternatives such as hydrogen, and the so-called hard to abate commodities, steel and cement.

The basic argument is simple. Generative AI (Gen AI) is a power hog. The same query made on a conventional search uses nearly 10 times the energy as when it is using Gen AI. This presumption is premised upon an estimate made on an early ChatGPT model, wherein the energy used was 2.9 watt hours for the same query which used 0.3 watt hours on a Google search. The usage gets worse when images and video are involved. However, these numbers will improve, for both categories. Evidence for that is that just over a decade ago, data centers were the concern in energy usage. Dire predictions were made regarding swamping of the electricity grid. The power consumption in data centers was 194 terrawatt hours (tWh) in 2010. In 2018 it was 205 tWh, a mere 6% increase, despite the compute instances increasing by 550% (Massanet et al, 2020¹).  The improvement was both in computing efficiency and power management.

More of that will certainly occur. Nvidia, the foremost chipmaker for these applications claims dramatic reductions in energy use in forthcoming products. The US Department of Energy is encouraging the use of low-grade heat recovered from cooling the data centers. A point of clarification on terminology: the cloud has similar functionality as a data center. The difference is that data centers are often physically linked to an enterprise, whereas the cloud is in a remote location serving all comers. We use the terms interchangeably here. Low-grade heat is loosely defined as heat at temperatures too low for conventional power generation. However they may be suitable for a process such as desalination with membrane distillation, and the regeneration of solvents used in direct air capture of carbon dioxide.

Impact on De-carbonizing Industry

The obvious positive impact will be on balancing the grid. The principal carbon-free sources of electricity are solar and wind. Each of these is highly variable in output, with capacity factors (the time spent by the capital in generating revenue) less than 30% and 40%, respectively. The gaps need filling, each gap filler with its vagaries. AI will undoubtedly be highly influential in optimizing the usage from all sources.

The obverse side of that coin is the increasing demand for electricity by the data centers supporting AI of all flavors. The Virginia based consulting company ICF predicts usage increasing by 9% annually from the present to 2028. Many data center owners have announced plans for all energy used being carbon-free by 2030. Carbon-free electricity capacity additions are primarily in solar and wind, and each of these requires temporal gap filling. Longer duration gaps (over 10 hours) are dominantly filled by natural gas generators. A major effort is needed in enabling the scaling of carbon-free gap fillers, the most viable of which are innovative storage systems (including hydrogen), advanced geothermal systems and small modular reactors (SMRs).

The big players in cloud computing have recognized this. Google is enabling scaling by purchasing power from the leading geothermal player Fervo Energy and is also doing the same with Hermes SMRs made by Kairos Power. An interesting twist on the latter is that the Hermes SMR is an advanced reactor of the class known as pebble bed reactors, using molten salt cooling (as opposed to water in conventional commercial reactors). It uses a unique fuel that is contained in spheres known as pebbles. The reaction products are retained in the pebble through a hard coating. This is not the place to discuss the pros and cons of the TRISO fuel used, except to note that it utilizes highly enriched uranium, much of which is currently imported from Russia. Google explicitly underlines part of the motivation being to encourage scaling SMRs. This is exactly what is needed², especially for SMRs, whose promise of lower cost electricity is largely premised upon economies of mass production replacing economies of scale of the plant.

Microsoft has taken the unusual (in my view) step of contracting to take the full production from a planned recommissioning of the only functional reactor at the Three Mile Island (conventional) nuclear reactor facility. In my view, conventional reactors are passe and the future is in SMRs. The most recent new conventional ones commissioned are two in Georgia. The original budget more than doubled, and the plants were delayed by 7 years. Par for the course for nuclear plants. Microsoft is certainly aware of the importance of SMRs, in part because founder Bill Gates is backing TerraPower, using an advanced design breeder reactor with liquid sodium as the coolant, and molten salt for storage. The “breeder” feature³ involves creation of fissile Pu²³⁹ by neutrons from an initial core of enriched U colliding with U²³⁸ in the surrounding depleted U containment. The reactions are self-sustaining, requiring no additional enriched U. The operation can be designed to operate over 50 years without intervention. Accordingly, it could be underground. The design has not yet been permitted by the NRC but holds exceptional promise because the fuel is essentially depleted uranium (ore from which fissile U has been extracted) and the issue of disposal of spent fuel is avoided.

Impact on Green Hydrogen, Steel and Cement

            Hydrogen as a storage medium, alternative fuel, and feedstock for green ammonia, has a lot of traction. One of the principal sources of green hydrogen is electrolysis of water. But it is green only if the electricity used is carbon-free.

            Similarly, steel and cement are seeking to go green because collectively they represent about 18% of CO2 emissions. Cement is produced by calcining limestone, and each tonne of cement produced causes about a tonne of CO2 emissions. Nearly half of that is from the fuel used. Electric heated kilns are proposed, using carbon-free electricity. Similarly, each tonne of steel produced causes emission of nearly 2 tonne CO2. A leading means for reduction of these emissions is the use of hydrogen to reduce the iron oxide to iron, instead of coke. Again, the hydrogen would need to be green and only high-grade iron ore is suitable, and it is in short supply worldwide. A recent innovation drawing considerable investor interest is electrolytic iron production.  This can use low-grade ore. But for the steel to be carbon-free, the electricity used must be as well.

            The world is increasingly electrifying. It runs the gamut from electrification of transportation to crypto currency to decarbonization of industrial processes. All these either require, or aspire to have, carbon-free sources of electricity. Now, AI and its Gen AI variant is adding a heavy and increasing demand. Many of these share a common trait: the need for electricity 24/7/365. In recognition of the temporal variability in solar and wind sources, the big players are opting for firm carbon-free sources such as geothermal and SMRs. That is the good news, because they will enable scaling of a nascent industry. The not so good news for all the rest is that these folks have deep pockets and are tying up supply with contracts. Remains to be seen how a startup in green ammonia or steel will compete for carbon-free electricity.

AI could well push innovation in industrial de-carbonization to non-electrolytic processes.

Vikram Rao

October 31, 2024

¹ Massanet et al. 2020 Recalibrating global data center energy-use estimates. Science, 367(6481), 984–986.

^2,3 pages 53 and 12 https://www.rti.org/rti-press-publication/carbon-free-power

Tagged: energy, nuclear, nuclear energy, nuclear-power, technology