2GW of AI compute requires roughly 1.5x the energy used by every household in London combined. Anthropic and AMD just shook hands to deploy that. Let’s talk about the new AMD and Anthropic partnership announced yesterday.
It runs something like this:
- Anthropic commits to deploy 2GW of AMD’s flagship Instinct MI450 GPUs (based on public data, I estimate the hardware investment alone to be around $30B - $35B)
- The first GW is set to go live in the first half of 2027
- AMD will use Anthropic’s models to optimise workloads for their GPUs and speed up ROCm development by adopting Claude across engineering and product teams.
- Based on certain sales milestones, AMD committed to make an equity investment of up to $5B in Anthropic. As an unrelated fun fact, AMD are currently trading at a trailing P/E ratio of 181 (~6x NVIDIA), and a forward looking P/E ratio of 75 (~3.5x NVIDIA). What I want to zoom in on is what this deal implies - and remember, this is rather small compared to similar commitments (NVIDIA and OpenAI have committed to 10GW of infrastructure, and there’s plenty more.).
Deploying 2GW of compute requires roughly 17.52TWh of power a year. Let’s put that into perspective. Based on the latest published figures, the power usage of London, a city of more than 9.5 million people, is about 39.6TWh per year, which is about 13% of total UK power demand. Only 11.8TWh of that is for domestic use.
This means that, through this partnership, Anthropic and AMD want to deploy GPUs that draw about 45% of the total energy used in London, or 148% of the energy used domestically by its inhabitants. Of course, they will claim that most of this will be powered by renewable energy.
The problem is that most of the renewable energy is produced somewhere else, so the actual energy used by those will be… from dinosaur juice and plankton gas presumably? The next issue is how quickly current gen GPUs become obsolete… but all that in another post. Until then, do you think the tens of GW in compute commitments are actually justified by projected AI demand given what we currently see in the market?