Microsoft, one of the most well financed and wealthy companies to ever exist on Earth, started charging for Copilot Code tokens at their actual monetary value yesterday instead of heavily subsidizing them on June 1, 2026. Today, many Copilot Code users have discovered that their day-one use of tokens was at times more than half the tokens allowed by their subscription tier.
Microsoft can't make AI work from a financial standpoint. Cost of Inference does not go down. Newer models require more hardware that needs more power and more water to operate, and even the datacenters theoretically under construction now for Blackwell architectures will not be adequate for next-generation Rubin hardware, nor is it expected that support could be retrofitted. This was the case for Blackwell as well. Patterns in markets also mean that there's not much incentive to fire up older hardware. It takes up too much space and costs too much to operate compared to the newer systems, so the second the next-gen thing is released, all spending will move to that.
Oracle is on the hook for
$50 billion annually for datacenter construction in the name of AI. That's the GDP of Libya, and it's doing that by taking on debt. Oracle isn't exactly a growing concern. Most of its income comes from rapacious license agreements from the same customers.
Current estimates say that 50% of purported datacenter construction, at least in the US, has either never started or has been paused shortly after starting. There are almost no planned datacenter construction sites in the US with no ongoing opposition from the communities where they're planned.
Ride share company Uber spent its annual AI token line-item budget in under four months.
Does this sound like a healthy industry to anyone? Anyone?
