We imagine AI as something immaterial, software in the cloud. The reality is brutally physical. In July 2026, Anthropic signed a $19 billion data center lease, while Meta is training its next model on the equivalent of around one million chips. The real battle of AI is fought in concrete, silicon, and electricity. Let's break down this invisible but decisive race.
When you chat with an AI, everything seems to happen in an abstract cloud, without substance. That's an illusion. Behind every response lie vast warehouses filled with blazing machines, consuming amounts of electricity comparable to those of entire cities. The most strategic resource of AI in 2026 is neither talent nor even data: it's compute. Let's understand why.
Compute, the new key battleground
Training a large AI model requires colossal computing power, provided by tens of thousands of specialised chips (the famous GPUs, originally designed for video games, now repurposed for AI) running in parallel for weeks or months. This power can't be improvised: it's built, bought, leased, and it's scarce.
The figures from July 2026 are dizzying. Anthropic signed a $19 billion data center lease with TeraWulf, adding to more than a dozen other data center commitments representing over one gigawatt of capacity, the equivalent of a large nuclear reactor. Meta, for its part, is reportedly training its next model on the equivalent of around one million chips. These sums are no longer those of a software company, but those of a heavy industry.
Why lease rather than buy?
One strategic detail deserves attention. Anthropic isn't building its own centres; it's leasing them via long-term agreements. This approach, also adopted for its deals with Amazon's and Google's infrastructure and the Colossus supercomputer, has a clear logic: locking in guaranteed access to compute without tying up gigantic sums in hardware that ages quickly.
This predictability is all the more valuable as Anthropic prepares its stock market listing, scheduled for autumn 2026. Long-term locked-in compute capacity reduces operational risk and gives investors the visibility they seek. In this economy, securing your access to compute is as important as securing your revenue.
This compute race is redrawing the map of global power. US export restrictions on advanced chips to China, which we've mentioned on several occasions, are precisely aimed at depriving rivals of this raw material. That's what makes it all the more remarkable that a model like LongCat-2.0 was reportedly trained entirely on Chinese chips, without US hardware. Control of silicon has become an instrument of national power, just as oil was in the last century.
The hidden cost: energy
All this power has an unavoidable physical counterpart: electricity, and lots of it. AI data centers are energy hogs, raising serious environmental questions that we explored in our article on the environmental cost of AI. Notably, Anthropic chose a partner, TeraWulf, specialised in centres powered by nuclear and hydroelectric energy, i.e. low-carbon power.
This choice isn't just ethical; it's also strategic. Access to abundant, stable, and ideally clean energy is becoming a limiting factor for AI growth. Some players are even investing in their own energy sources. In the years ahead, the question may no longer be "who has the best models?" but "who has enough electricity to run them?".
What to take away
Behind the apparent software magic of AI lies a massive industrial reality: a global race for compute, silicon, and energy, swallowing tens of billions of dollars and redrawing the balance of power between companies and between nations. AI isn't immaterial; it's one of the most physically intensive industries of our era.
This reality has concrete consequences. It explains why only a few heavily capitalised players can compete at the top, why control of chips has become a major geopolitical issue, and why energy is becoming a decisive factor. Next time you ask an AI a question, remember: somewhere, in a gigantic warehouse, thousands of chips just heated up to answer you. The magic has a very real weight, consumption, and price.