By Lars Mikelsen, Avanto Capital: Since the invention of the microprocessor, the demand for computational power has increased exponentially. Over the past decade, much of this data processing has been moved from local machines to the “cloud,” resulting in the rise of large data centers. While it’s long been known that these data centers consume large amounts of electricity, energy consumption remained relatively stable for years despite a surge in computing power. According to a Goldman Sachs report, this was largely due to advancements in hardware efficiency and the transition of smaller data centers into more efficient “hyper-scale” ones. However, once the low-hanging fruit was harvested when most of the smaller centers were replaced by support data centers, power consumption began rising again, driven by the ever-growing demand for computing power.
Artificial Intelligence: A Game Changer for Energy Consumption
The use of artificial intelligence is, as is widely known, even more power-intensive than traditional computer use. For instance, a search conducted via ChatGPT consumes approximately ten times more electricity than a Google search. Importantly, even before the major breakthroughs in AI, data centers were experiencing almost insatiable growth in power consumption. Thus, the rise of artificial intelligence adds to an already substantial and rapidly increasing demand for electricity.
The rise of artificial intelligence adds to an already substantial and rapidly increasing demand for electricity.
The potential increase in power consumption due to artificial intelligence is fraught with enormous uncertainty, quickly leading to unfathomable scales. According to specifications, an NVIDIA H100 chip has a power output of 700W, which corresponds to approximately 6 MWh (6000 kWh) annually per chip at full utilization. In total, NVIDIA sold 1.5 million H100 chips in 2023 and is estimated to sell 2 million H100 chips in the current and coming year. The sales of this model alone could consume up to 21 billion kWh, or 21 terawatt-hours (TWh), annually. For reference, Norway’s total electricity consumption is around 106 TWh per year.
NVIDIA’s latest model, the Blackwell, has been promoted as significantly more energy efficient, offering substantial energy savings compared to the H100. However, this efficiency is measured per unit of compute. The Blackwell model claims a power output of 1000W, reflecting a 40 percent increase.
A critical point of contention will be whether those involved in the AI race, particularly the management of technology giants, will maintain the current level of processing power or purchase the same number of processors to double their processing capabilities. Historical trends indicate that computing power tends to increase in line with the latest technological advancements. Furthermore, if one believes that current technology can evolve into general artificial intelligence (AGI) or even narrow AI developer intelligence, it is not only rational but also existentially crucial to invest heavily in the development of the most advanced models.
If one achieves general artificial intelligence, as a natural logical conclusion, it logically follows that superintelligence (ASI) – a form of omniscient, god-like power – will almost immediately follow, marking “Game Over” for all competitors.
If the tech aristocracy is not paralyzed by political bans, they will first want to secure as many of the most advanced processors as possible, then secure access to electricity and other necessary and scarce input factors. The race to reach AGI/ASI means that virtually any estimate of electricity consumption could be significantly understated.
The risk is definitely on the upside!
It is therefore not surprising that the same tech elite are actively seeking to secure access to electricity. Both Bill Gates and Sam Altman of OpenAI are leading the charge in their respective companies to deliver the next generation of nuclear power plants (Terrapower and Oklo). Coincidentally, Altman has been lobbying in Washington for the establishment of more 5 GW data centers, and news surfaced that Microsoft has agreed to reopen the decommissioned Three Mile Island nuclear reactor. The signaling effect of both developments is nothing short of astonishing.
The very notion of a demand for data centers with capacities in the Gigawatt range – let alone 5 GW – is incredible in itself. For context, a traditional nuclear reactor typically has a capacity of about 1 GW, suggesting that each data center could effectively require its own nuclear power plant. The fact that Three Mile Island is being restarted at Microsoft’s request carries immense symbolic weight. It was precisely a meltdown of another reactor at this power plant in 1979 that turned American public opinion against nuclear power. These examples are not isolated; Oracle is planning a 1 GW data center powered by three small modular reactors (SMRs), while Google and Amazon both announced large investments in nuclear power to fuel future datacenters. Other tech giants are likely engaged in similar initiatives behind the scenes.
Nuclear Power: The Tech Elite’s Preferred Solution
The techno-nobility’s preference for nuclear power is no coincidence. These companies require stable, carbon-free energy production that can operate around the clock 24/7, as data centers cannot rely on solar, wind, or other forms of intermittent energy generation. However, significant challenges accompany these plans. Constructing nuclear power plants is a time-consuming process – even in China, it takes around five years – and none of the small modular reactors (SMRs) proposed have been built yet. Additionally, the nuclear fuel supply chain is already overloaded, particularly if Russian conversion and enrichment services are to be avoided. This creates a scenario where, if uranium prices were previously influenced by a limited and price-inelastic supply, the introduction of a new player with substantial price-inelastic demand only exacerbates the situation.
These companies require stable, carbon-free energy production that can operate around the clock 24/7, as data centers cannot rely on solar, wind, or other forms of intermittent energy generation.
However, the realization of all the planned data centers powered by nuclear power plants is unlikely to occur within this decade. Specifically, the construction of the power plants themselves is not expected to materialize soon. First, small modular reactors (SMRs) are still in the design phase, and even Bill Gates’ Terrapower does not anticipate becoming operational before 2030. Furthermore, significant bottlenecks need to be addressed regarding both infrastructure and access to uranium enriched to 19.6 percent (HALEU), which is nearly impossible to acquire outside of Russia and all Russian nuclear material will be illegal to import into the USA starting in 2027. Therefore, it seems most likely that while the data centers will be built, they will initially rely on alternative power sources, as Big Tech cannot afford to wait five to ten years.
Until SMR power plants can be rolled off the assembly line, data centers will need to rely on fossil fuel-based electricity production. Natural gas will likely be the preferred choice for powering these facilities, as coal-fired power plants are a “no-go” PR-wise. In any case, the increased demand for electricity suggests that coal power will not be phased out according to the current plans. Consequently, it seems that all available sources of electricity will have to be utilized until the benefits of renewed investment in nuclear power become evident.
Implications for Energy Markets and Commodities
All of this bodes well for energy commodities. In addition to the obvious uranium trade, U.S. natural gas looks attractive. The U.S. has the highest density of data centers globally, and its natural gas is the most affordable in the world. Additionally, the expansion of export capacity through new liquefied natural gas (LNG) terminals is likely to tighten the market switching the market balance from oversupplied to undersupplied. This combination of factors positions U.S. gas to benefit significantly from the growing demand driven by artificial intelligence.
All of this bodes well for energy commodities. In addition to the obvious uranium trade, U.S. natural gas looks attractive.
The various compliance carbon markets are also poised to emerge as potential winners. In recent years, EU carbon prices have experienced relatively weak performance, influenced by Germany’s economic challenges, significant advancements in wind and solar energy, and the softening of regulations amid the energy crisis stemming from the Ukraine situation a couple of years ago. However, Germany hosts numerous data centers, and a renewed demand for electricity, combined with the expected reversal of the aforementioned regulatory softening from 2027, could significantly improve the situation.
The electricity market is expected to remain tight for many years, leading to sustained high electricity prices. This situation often results in subsidies in most countries, making investments related to electricity generation even more attractive than they would otherwise be. The “green shift” may ultimately not be a true shift but rather additional “green” energy layered on top of the existing energy infrastructure. Historically, humanity has proven to be remarkably inventive; each time new energy sources have emerged, innovative uses for them have quickly followed. This trend seems to persist even today, despite the belief that all possible inventions have already been made – an idea famously attributed to Charles Duell in 1889. It’s likely that this perception will continue until the advent of artificial superintelligence is achieved.
