If competition in artificial intelligence is most of the time seen through the prism of models: size of LLMs, quality of agents, performance of benchmarks or sophistication of multimodal capabilities. The agreement concluded yesterday between Anthropic and SpaceX around the Colossus 1 data center shows that the battle is no longer just about algorithms, but is now about energy, GPUs, physical infrastructure and the ability to build computing systems on an unprecedented industrial scale.
Anthropic has announced that it wants to use the full capacity of Colossus 1, the gigantic computing complex developed by xAI near Memphis, Tennessee. Called “Colossus”, this infrastructure is now among the most powerful ever built for artificial intelligence. Initially designed to train xAI’s Grok models, it would bring together more than 220,000 interconnected NVIDIA GPUs within an architecture optimized for intensive AI loads, with an estimated power consumption of around 300 megawatts. This volume illustrates the change in scale of the sector. At this level of power, an artificial intelligence laboratory begins to look less like a software publisher and more like an industrial operator consuming volumes of electricity comparable to those of national critical infrastructures. Built in just a few months, Colossus embodies the emergence of a new generation of infrastructure where calculation, energy and industrial capacity become inseparable.
Another particularly revealing element of the announcement: Anthropic has expressed its interest in a collaboration with SpaceX around future space data centers. A perspective that would have been science fiction only a few years ago is now beginning to appear as a credible response to the physical constraints facing the artificial intelligence industry. Explosion in energy needs, progressive saturation of electrical networks, tensions over land and limits of terrestrial cooling are pushing players in the sector to consider radically new computing infrastructures. In this logic, space no longer appears only as a field of scientific exploration, but as a potential extension of the global computing infrastructure.
Because the economy of frontier models now relies on four resources that have become structurally limited: GPUs, electricity, cooling and land. Data centers capable of accommodating several hundred thousand graphics processors require gigantic electrical capacities, complex cooling infrastructures and suitable land near energy networks. In the United States, connection delays are already becoming a slowing factor for several hyperscale projects.
It is in this context that Elon Musk is pushing a much more radical vision than that of other market players. Where Google, Meta or OpenAI mainly seek to secure their own terrestrial capabilities, Elon Musk seems to envisage AI infrastructure as a global system combining energy, space, connectivity and calculation production.
This logic gradually brings xAI closer to an infrastructure operator rather than a simple AI laboratory. The economic model evokes that of the new “neoclouds” specialized in GPUs, like CoreWeave. But Musk is clearly seeking to integrate many more elements of the industrial chain: accelerated construction of data centers, energy management, potential development of proprietary chips, and ultimately extension of the infrastructure outside of Earth.
The idea of orbital data centers addresses several physical constraints that have become critical for the AI industry. In orbit, solar energy can theoretically be captured almost continuously, without day-night alternation or weather constraints. Above all, moving part of the computing infrastructures outside of terrestrial networks would make it possible to escape increasingly visible limits: electrical saturation, land pressure, tensions over cooling and conflicts over energy uses. The vacuum of space also opens up new possibilities in terms of heat dissipation by radiation, even if this issue remains today one of the main technical challenges of future orbital data centers.
The credibility of this vision obviously rests on SpaceX’s launch capabilities. No other player today has a comparable integrated space infrastructure: reusable rockets, Starlink constellation, orbital station projects, logistical control and verticalized industrial ambitions.
The situation is reminiscent in certain aspects of the large industrial infrastructures of the 20th century: railway networks, energy pipelines, power plants or submarine cables, where the strategic value no longer resides only in the applications visible to the end consumer, but in the invisible layers which power the entire system.
From this perspective, Colossus probably constitutes not only a new hyperscale data center, but the prototype of an integrated computational infrastructure, conceived from the outset as a global industrial system. The challenge no longer consists only of hosting AI models, but of controlling all of the physical layers which condition their existence: energy production, access to advanced semiconductors, cooling capacities, transmission networks, industrial logistics and, ultimately, orbital infrastructures. Artificial intelligence is gradually ceasing to be a software industry and becoming a power industry.
This transformation could profoundly redefine the hierarchy of the global technology sector. For two decades, value was primarily concentrated in software platforms and user interfaces. The AI era shifts some of this value to heavy infrastructure capable of supporting exponential computational demand. The truly dominant players of the next decade will perhaps not be those who develop the most efficient models, but those capable of orchestrating on a very large scale the physical resources essential to their operation: energy, GPUs, networks, industrial capital and space capabilities. As computing needs increase, technological competition begins to come closer to the historical logics of strategic industries, steel, oil, nuclear or telecommunications, where control of infrastructure lastingly conditions the balance of economic and geopolitical power.
