The Dark Data Centre

Home / Blog / The Dark Data Centre

NVIDIA has a production ceiling. There are only so many data-centre GPUs manufactured in a year, and the queue for the top tier is long. If a handful of the largest tech companies buy up most of the annual run, no one else can buy them until the next run.

Over the last two years the largest tech companies have done exactly that. Tens of billions of dollars of top-tier GPU have been purchased and warehoused in data centres that are not powered on. The chips are not running workloads. They are sitting in racks in buildings that draw no meaningful power and route no meaningful traffic. It is inventory, not infrastructure.

The purpose is not utilisation. The purpose is denial. A GPU in a competitor’s rack is compute that could be used to train a model that leapfrogs yours. A GPU in your rack, unpowered, is not doing any work — but it is not doing any work for your competitor either. Buying the year’s supply off the market is cheaper than betting nobody else finds a training breakthrough with it.

This is why the compute-rental deals started appearing. Data-centre GPU depreciates on a roughly five-year schedule to effectively zero — the workloads that pay for it will not run on it once it is two generations behind. The companies that bought the chips to keep them off the market are now trying to extract some cash flow before that clock finishes. Renting the racks out to each other is what that looks like from the outside: the same three or four names taking turns being landlord and tenant.

The implicit story sold with every purchase is that the company buying more GPU is the company whose AI is winning. If the GPU is powered off, the story is illusion. The purchase counts. The usage does not.

None of this is on the shipping manifest. GPUs shipped are shipped. Data centres built are built. The distinction between “built and running” and “built and dark” is not a line item anyone reports.