Artificial intelligence continues to develop at a rapid pace, with every benchmark seemingly saturated not long after it is created. Recently, even the “LLM-proof” ARC challenge—the “Abstraction and Reasoning Corpus”—suffered a near-fatal blow by OpenAI’s new o3 model, which is a variant of its o1 model available on ChatGPT. ARC was introduced by François Chollet, a French software engineer and AI researcher, as a test of generalization and abstract reasoning. It reportedly cost over $3,000 to solve just one of the challenge problems.

Industry leaders such as OpenAI’s Sam Altman believe massive investments in the computational power of AI systems are worthwhile, with Altman suggesting “Infrastructure is Destiny,” and a “New Deal”-sized opportunity awaits if hundreds of billions are spent on energy infrastructure. Their evaluations are driven by the empirical observation that increasing the amount of “compute” available to these AI systems leads to regular improvements of their performance on a logarithmic scale.

This phenomenon, often called AI scaling “laws,” has been observed repeatedly, not only with overall computation but also the number of model parameters and data volumes. A recent discovery made by OpenAI in the o-series models further shows that scaling AI compute can be done per problem, and scaling this per-problem compute or “inference time” compute also obeys a scaling law that increases performance as a log of the computation used.

These AI scaling laws are likely to continue to hold for the foreseeable future and represent one of the most important engines driving technological progress today. This is certainly something the AI industry is betting will be the case, with massive investments made into acquiring larger and larger data centers for both training runs and serving ever more computation-hungry inference. Elon Musk’s xAI recently developed the Colossus supercomputer with an astonishing 100,000 NVIDIA H100 GPUs and an estimated cost of up to $4 billion dollars. Microsoft is projecting to spend $80 billion dollars on AI data centers just this year. Just this January, Amazon Web Services announced plans to invest at least $11 billion in Georgia for its AI infrastructure.

These clusters come with considerable energy demands—Colossus in Memphis, Tennessee alone requires 150 megawatts (MW), which is comparable to roughly 7% of San Francisco’s power needs. Today AI workloads globally consume around 20 terawatt-hours (TWh) per year and account for roughly 10% of total data center power consumption. These workloads are growing much faster than non-AI workloads and are on track to double by 2028. If scaling laws do hold, this means we can expect a massive increase in energy demand, outstripping current infrastructure.

This projected energy demand has at least, among technologists, seemingly broken a decades-long taboo on building more nuclear power plants. Big tech companies are no longer averse to being seen investing in one of the few carbon-neutral sources of new baseload grid power rather than the intermittent “green” energy sources fashionable among environmentalists.

Microsoft made a significant statement by entering a twenty-year agreement with Constellation Energy, which plans to reopen the Three Mile Island nuclear plant, the site of the 1979 partial nuclear meltdown. Constellation Energy plans to invest $1.6 billion to refurbish and restart the reactor by 2028 with an estimated 835 MW of capacity. Microsoft entered the agreement to provide the energy demands for its AI data centers. Amazon also recently purchased a nuclear-powered data center in Pennsylvania for $650 million with 2.5 gigawatts (GW), i.e., 2500 MW of capacity. This trend has already caught the eyes of various economic players, with Goldman Sachs estimating that by 2030, data centers will see a 160% increase in power consumption. Some industry experts even project as much as 4.5% of global energy to be used by AI workloads.

There will be continuous pressure to throw more compute at important problems. OpenAI recently outlined five levels of autonomy for AI systems starting with chatbots at the first level, agents that can do tasks like building a feature for a software product at level three, and ending with AIs that can do the work of an organization at level five. As we climb the stages of autonomy, AI systems will require increasing “nines” of accuracy in order to sequence together many operations towards achieving a complex objective with many steps over a long horizon. To see why this matters, a system with only one nine, or 99% accuracy, if tasked with achieving a task which requires one hundred serial simple operations will only achieve the objective 36% of the time.

Climbing to the last level of autonomy would mean we could potentially add trillions of dollars to the economy by spinning up multi-billion dollar companies from just silicon. The upside of having a “country of geniuses in a data center,” especially in biology, is given some treatment by Dario Amodei, the cofounder of Anthropic, in “Machines of Loving Grace,” where he speculates powerful AI systems could, for example, solve open medical problems in cancer research, longevity, and fertility. Currently, in the U.S. alone healthcare spending neared $5 trillion dollars.

If the scaling laws hold and AI fulfills the economic potential all major economic decision-makers attribute to the industry, then the explosive growth in energy demand will double or quadruple decade after decade. This is a notable speed-up from the historic trend of energy consumption doubling every thirty to fifty years! At this faster decade-over-decade pace, we would, by the end of the 21st century, see an energy economy completely dominated by the insatiable energy demands of artificial intelligence. More accelerationist estimates posit that some individual clusters will require 100 GW of electricity, as much as some individual U.S. states, by the end of this decade. If this comes to pass, it will be fair to say our ongoing intelligence revolution was as impactful as the Industrial Revolution itself.

Yet energy demand isn’t just driven by operating the data centers themselves. At every step of the way it will increase the energy demands of our existing industrial civilization, driven by the incredible complexity and extreme energetic demands of manufacturing the necessary clusters of GPUs and other powerful computer chips. Modern GPU manufacturing relies on packing transistors at the nanometer scale to seemingly impossible densities using what is known as ultraviolet photolithography, which uses ultra-short wavelengths of light to etch patterns into silicon. It is so complex that there is only one company in the world—the Netherlands-based ASML—that can make what are known as “extreme ultraviolet” machines that can etch the fine patterns needed for the highest-end chips. That is just one component that goes into fabrication plants, also known as “fabs,” that manufacture these chips, and they, in turn, demand vast amounts of energy.

The world’s most valuable semiconductor company is Taiwan’s TSMC, which is the sole producer of Nvidia’s most advanced chip products, including the H100, H200, and the next-generation Blackwell series, which are seen as essential chips for AI and briefly led to Nvidia surpassing Apple as the world’s most valuable company as a capstone on an already meteoric trajectory of 2,000% growth over the past five years. TSMC currently already consumes around 8% of Taiwan’s total electricity consumption. This percentage is, of course, expected to rise. To accommodate this, Taiwan plans to add 12% to 13% more capacity during peak hours.

Once AI systems begin to automate AI research itself, an “intelligence explosion” could see an unprecedented mobilization of compute that dwarfs the historical expansion rate. In this scenario, timelines to scale up energy consumption by orders of magnitude, that normally would have taken at least a century to cause appreciable warming, would be accelerated to happen within a few decades.

Having made it possible to put a price on solving the most important problems facing humanity, be it in software, mathematics research, manufacturing, or potentially life-saving domains like cancer research, our energetic demands will grow to be enormous, even planetary scale, or beyond. There is one problem with this: we cannot continue to scale energy usage like this without making the Earth inhospitable to organic life.

This is not due to carbon emissions and their resultant greenhouse effect; rather, it is due to the laws of thermodynamics. It is impossible to produce free energy which does the useful “work” to power everything from our toasters to AI clusters without creating waste heat. The more energy we consume, the more heat our machines generate. Even today this waste heat is already warming the planet at a rate of approximately 2% of what can be attributed to the greenhouse effect. Eventually our machines will heat up the Earth more than even the Sun does. This will happen independently of the type of technology we use.

The only way for the Earth to radiate heat to outer space is through radiation governed by the Stefan-Boltzmann law, which radiates power to space as a function of temperature to the fourth power. The Earth can radiate more energy out as we increase the heat we create, but not without increasing the temperature of the surface of the Earth to allow for a commensurate amount of radiating power.

The thermal efficiency of our computers has improved over time, of course. We get more compute while producing less waste heat over time. The gains in the thermal efficiency of compute can be plotted on an empirical curve in what is known as Koomey’s law, named after Jonathan Koomey, a researcher known for his work on the energy efficiency of computing systems. The trend observes that the number of computations per joule of energy dissipated has historically doubled approximately every 1.57 years, indicating significant improvements in the energy efficiency of computing hardware over time.

However, that growth has slowed down and we know it is not a true law, as it will reach Landauer’s limit, which is the minimum amount of energy required to erase one bit of information, establishing a fundamental thermodynamic limit on the energy consumption of computational processes. Even with the pre-AI workload demands for computation, Landauer’s limit was estimated to end Koomey’s law in 2080.

The waste heat of our compute and, in fact, even the waste heat of all our manufacturing and transport is more or less benign now, but supposing an increase of two orders of magnitude of the Earth’s energy usage today, we will start to make the Earth downright uninhabitable. GPU energy usage as a percentage of the planet’s energy capacity is around 0.1% today, but even factoring in thermal efficiency gains, if we wanted ten million times more compute—so seven orders of magnitude more than we have today—we’d be in serious trouble. We would increase the temperature of Earth by at least 5 degrees Celsius, which is probably unsurvivable or undesirable at the very least.

So what to do? To space!

Thanks to companies like SpaceX, payload costs have fallen dramatically, using cost to low Earth orbit (LEO) per kilogram as a proxy of progress. Costs have fallen from approximately $50,000 per kilogram to approximately $3,000 per kilogram thanks to SpaceX’s Falcon 9. In fact, cost per kilogram represents something of a North Star for SpaceX and has motivated innovations such as reusable rockets. In the company’s quest to reach Mars, it has delivered the means for us to return to the Moon at a fraction of the cost of the Apollo missions.

Building GPUs and using GPUs in LEO may seem attractive, but we’d quickly run afoul of Kessler syndrome, whereby the density of space debris in low Earth orbit becomes so high that collisions between objects generate even more debris in a cascading effect that can make operating satellites and spacecraft, or in this case GPUs, increasingly difficult or impossible in these orbits.

So where else can we go? The Moon.

The Moon is rich in silicon from the regolith deposits on its surface, a necessary ingredient for the manufacture of GPUs. Advances in robotics mean that creating a robot that can assemble itself, or a smaller base of machinery that can reproduce itself, is within reach. Currently, the limiting factor in robotics is the lack of good-enough robots—especially humanoid robots, which are not reliable enough to do precise dexterous work. China is betting heavily on industrial automation, with many specialized robots in their factories today. They will continue to invest heavily in their development, and the U.S. will as well just to keep up.

Advances in humanoid robotics, coupled with extensive investment by many players including Tesla, mean that in twenty years, we will have robots that can perform any mechanical action, no matter how fine or subtle, that the human body can. Once there is mass adoption of humanoid dual-arm mobile manipulator robots—mobile robots that can move freely in any direction, or, at least, robots with two arms that do basic dexterous tasks a human can do—it will kick off a data flywheel: the more they are used, the more data we will have on their movements to train models on, leading to the robots becoming even more useful, and we will soon after get strong robotic foundation models akin to the GPT series, but for robotics.

Assuming progress in industrial automation, humanoid robotics, artificial intelligence, and space technology continues as currently envisioned by these industries, we will in just a few short decades be able to deliver payloads of a self-assembling farm of robots to mine the Moon, create chip fabs, build, and ultimately tile the Moon with GPUs. The Moon has a surface area of 14.6 million square miles, roughly the size of Asia. If we very conservatively tiled even half the Moon with GPUs and solar panels, the Moon could sustain a billion times the compute of the Colossus cluster and, with a few turns of Moore’s law driving chip technology forward, even a trillion times the compute.

The Moon thus transformed will come to resemble something out of science fiction concepts of planet-sized factories or Factorio—a popular video game among engineers working on these very technologies. This level of compute would effectively turn our Moon into a planet-scale supercomputer and represent a giant leap in Man’s capacity to control our destiny.

While it’s difficult to estimate when we’d absolutely need a Moon-sized computer, we know that we cannot increase our planet’s energy consumption by two orders of magnitude without inducing a catastrophic warming—which we were on track to do even before the AI revolution—in around 200 years. This is a leisurely timeline when AI scaling laws have accelerated to mere decades.

There are other reasons to build a Moon computer, including to avoid the regulatory hurdles to build the energy centers needed to power centralized AI clusters and as a sovereignty play in the increasingly fraught geopolitical game on the road to developing Artificial General Intelligence (AGI)—a system with the ability to solve virtually any cognitive task a human can. We of, course, don’t have to mine the Moon just to replicate a single human mind, but we might have to do so to power superintelligence, same in kind but vastly superior to our own.

In a world with sanctions on chips and espionage as a backdrop to developing systems that vie with collective human cognitive output, it is likely state actors may resort to increasingly bold attempts to sabotage each other’s AGI efforts. While some have advocated for the Moon to be a state, it perhaps may become the Earth’s or perhaps America’s special computation zone, relatively sheltered from competitors without a robust national space program. This is one of the few remaining areas of high technology where the United States—thanks only to SpaceX—remains far ahead of China.

What kind of problems can we solve with a Moon computer? That is up to us. Having solved intelligence, we can use that to “solve everything else,” as Demis Hassabis has quipped. Cure cancer? Live forever? Solve every problem in mathematics and the sciences that we can postulate? Likely.

In our prehistory, we started our relationship with the Moon as an illuminated object of wonder, often with religious significance. Later, as one understood by science as a conventional Moon around an unconventional planet. It was a special and permanent moment when we first stepped on it. And now, our relationship to the Moon can culminate in a triumph, of not merely having stepped on it, but of making it truly ours, while continuing to make our home planet habitable.

Perhaps we will turn back to the age-old questions of philosophy armed with our new technology. Even computational theology is not out of the question; perhaps ultimately, our quest for intelligence was simply a quest for revelation. But maybe some questions will forever remain incompressible and unanswered, even by our own intelligent Moon. In Douglas Adams’ Hitchhiker’s Guide to the Galaxy, a group of hyper-intelligent beings build a supercomputer named “Deep Thought” and ask it the question “What is the Answer to the Ultimate Question of Life, the Universe, and Everything” and, after applying inference scaling to the question and thinking for seven and a half million years, the computer answers 42.

Let’s find out if that’s right.