AI Is Actively Destroying the Revenue Streams That Pay For Its Existence

Someone has released pre-alpha clones of Microsoft Office and Adobe Creative Suite, built using Claude and written in Rust under claimed clean-room conditions. They're buggy and incomplete, but they work well enough to make a point. The author describes them as "clean-room reimplementations." Whether that framing holds up legally is another matter, but here's the problem for Microsoft and Adobe: companies that got rich training AI on the largest unauthorised IP grab in history are going to struggle making the case that AI shouldn't be turned on their own products. The chutzpah required for that argument is almost admirable.

You can't copyright functionality. Clean-room implementations have a long legal history. And the code wasn't copied, it was generated. There may be angles for the lawyers to chase, but the ground is not solid.

The broader implication is more interesting than any individual lawsuit. If AI can clone mature commercial software to a usable degree in pre-alpha, then effectively everything becomes open source whether the original authors like it or not. Lock-in, which has been a cornerstone of big software revenue for decades, stops meaning much. The obvious counter-move is pushing everything to subscription cloud services where no one can get their hands on a binary. The problem with that plan is that AI can drive a browser. It doesn't need a binary. It just needs a screen.

This is the pattern now: AI disrupting the very systems that make AI viable. Microsoft and OpenAI's own internal communications acknowledged a version of this with advertising. AI answers kill search clicks, search clicks fund publishers, publishers fund the creation of the content that trains the models. Strangle the ad economy and you eventually strangle the data pipeline. It's the kind of slow-motion self-destruction that's hard to arrest once it's started.

Then there's the money problem, which is large and not going away.

Gartner has been saying for a while that enterprise AI ROI doesn't justify the infrastructure spend. SemiAnalysis, which is far more bullish on AI's trajectory and considerably better informed than most, has put total global AI capex at something approaching $11 trillion by 2030 and is asking the obvious question: where does that actually come from?

The current investment structure is strange. Neocloud startups have enormous appetites for compute but no investment-grade credit rating. So Nvidia effectively backstops them, guaranteeing to absorb unsold capacity at fixed rates long enough to make the startups look creditworthy to other investors. Large cash-rich tech firms do variations of the same thing. SemiAnalysis calls this "balance sheet as a service," which is a tidy phrase for something that concentrates risk in a small number of large companies. That works until it doesn't, and it works less well if AI itself is eroding those companies' core revenues.

Older datacenter hardware is also behaving oddly. Normally, chips depreciate over five years or so down to scrap value. Right now, vintage silicon is generating comparable revenue to what it earned when new because demand is outrunning supply of newer kit. Export controls on high-end chips have had the unintended consequence of making older hardware more economically durable. This sounds like good news. It isn't, really. It's another variable that doesn't fit standard financial models, in a market already full of variables that don't fit standard financial models.

Institutional investors, the large conservative pools of capital that any mature asset class eventually needs to attract, historically dislike things that don't fit their models. The people who remember what happened when clever financial modelling met poorly understood risk in 2007 and 2008 are not yet all retired. Getting that class of capital comfortable with AI infrastructure investment is a non-trivial problem.

The honest answer to "where does the money come from" is that nobody has a convincing one. The spend is happening regardless. What the returns look like, and whether they arrive before the credit runs out, is genuinely unclear.

Meanwhile, someone just shipped a working pre-alpha clone of Photoshop. Built by one person. Using AI. In their spare time.

We're somewhere in the middle of this, not at the beginning, and not yet anywhere near the end.