Microsoft is, by reports, canceling Claude Code licenses across its most prominent divisions. Is this to promote GitHub Co-Pilot or the signaling of a deeper problem?
Microsoft has decided to end the Claude Code licenses inside its experiences and devices group, a.k.a the teams behind Windows, Microsoft 365, Outlook, Teams, and Surface.
For Anthropic, this has to be nothing but bad news. Microsoft, which has been heavily investing in OpenAI and other AI tools, has always been optimistic about the tech. Trying this new thing out in their divisions could be part of a larger experiment to see what is working.
But we suspect there is something bigger here at play. The cost of running an AI system is not efficient, as these tools are developing into smarter versions of themselves- they are becoming more and more energy inefficient.
The tokens that companies use cost a lot of money because they require a lot of money. And the usage is drying up.
Claude’s usage windows have been reducing at a steady pace, sometimes getting over in one or two prompts or actions.
Uber, as every outlet has reported, has faced similar problems. The token budget Uber thought it needed was blown away in just 4 months. Of course, programming or doing any real work is complex. It requires experience, and thinking, and clearly thinking does not come cheap, even when organizations think it does.
Computing, thinking, and intelligence, and the ability to synthesize information, are a scarce resource. And AI might find it difficult to replicate it across multiple instances. It can do certain tasks very well, but it needs many tokens to do it.
That is difficult en masse.
However, this raises a question: where is AI heading, not in terms of if it will get better or more intelligent, but rather, how much energy would it need?
AI data centers cannot be called efficient. They are guzzling energy, and with everyone using it, that rate may be exponential. Every hard problem, every doc generated, every code written, every long-chain task it performs, the tech eats energy on a large scale.
The hard problem here isn’t managing costs, but scaling down energy costs. The question is how? We are currently using the LLMs and agents to solve problems by making them amazing at guessing what comes next, and that guesswork is draining every compute token dry.
It’s time we move beyond discussing alternatives and put research into finding them. Or this might end up as a cost that cannot be recovered.
