No Gotterdamerung?

The Lies AI Gods Tell Themselves

The former Cloud Czars, now the AI Gods, got through the latest earnings season unscathed. (Illustrations from Google Gemini.)

Amazon, Microsoft, and Google all reported numbers Wall Street found acceptable. They all claimed that they’re re-selling the capacity they’re building for AI applications, at a profit, and thus they can continue to build new capacity at an accelerating rate.

This news helped most of the pack behind them stay in the game. Even Oracle, down 50% from its June peak, stayed alive on the news, as Michael Burry closed his short position.

But just how did Microsoft get $19.6 billion in free cash flow? People are asking questions.

For instance, Microsoft’s build out is based on a claim it has a $627 billion backlog of “remaining performance obligations” (RPO). This is computing capacity Microsoft says customers have contracted for, but which Microsoft hasn’t yet delivered.

The problem is that 45% of this comes from one company, OpenAI. If OpenAI crashes and burns, Microsoft’s RPO goes with it. I believe this has already happened. I’m not the only one. It’s an open secret the relationship between the two companies is crumbling. Which means almost half of Microsoft’s growth claim is bogus.

Question. How do you make money on what you give away? The Gemini art that accompanies this story was free, in that I didn’t pay for it. But it cost something, right? Who paid?

Bloatware Returns

A bigger point I have yet to see referenced comes from inside Microsoft software. AI demand is being forced into existing programs like Word, which is writing this piece. Word can now read this story back to me (why?), and it constantly tries to edit my stuff, in ways I find extremely annoying.

I’ve been using Microsoft software for 40 years and this forcing of new features has a name. It’s called bloatware. Microsoft is doing things with AI it could easily do in other ways, and for a lot less time, money, and compute.

How do I know they’re hurting from this? Just look at the box that now appears in the upper-right corner of an editing window. “Upgrade Your Plan.” Families are being asked for $13 more per month to support features they’re not using.

Microsoft isn’t alone. Look at the corner of your screen the next time you open Gmail. The word there is “upgrade.” Google is also creating bloatware, having AI do things it could easily do with other techniques, and just like Microsoft it wants you to pay for it.

Microsoft and Google have hundreds of millions of customers. What they’re asking for is hundreds of billions of dollars. I don’t think they’re going to get them.

Amazon has also begun to nickel-and-dime customers. Walmart’s online pricing is now a lot more affordable, for example. You may have also noticed a new streaming service, MGM+. That’s part of Amazon Prime, now repackaged at $7/month. (Believe it or not, it costs $8/month if you’re already a Prime customer.) The price for avoiding ads on Prime itself is also up, from $3/month to $5/month. Streaming bills are now cable bills.

Balance Sheet Shenanigans

The biggest lie sits deep in the balance sheet.

This is the depreciation schedule, the rate at which the Czars are writing off their investments. Microsoft has extended the “useful life” of its buildings from 15 years to 25, but more important it now claims its servers have a useful life of 6 years.

Physically, the chips should easily last 6 years. But practically? Uh, no. The year 2020 was when Nvidia first launched its Ampere cards. These claimed to be 20 times faster at AI inference than the Turing units they replaced. Since then, Nvidia has produced Hopper, Lovelace, Blackwell, and now Vera Rubin equipment with similar claimed improvements. The Rubin chipset alone is 10 times faster than Blackwell at AI Inference.

I’ve written about Moore’s Law for decades, but Huang’s Law puts it in the shade. As I have written, it does this by ignoring Moore’s limits on heat and electricity use. Fans have been replaced by water baths, which move the heat the chips produce more efficiently than air can. But that also means a lot of the electricity being put into Nvidia chips is being wasted.

Huang’s Law is inflationary, but it also makes the AI Gods’ depreciation schedules ridiculous. The nature of Moore’s Law still holds. Chips start depreciating the moment they hit the loading dock. The extraordinary demand for AI is currently keeping chip prices high, especially for memory now replacing hard drives. But it’s unsustainable.

Losing the Great Game

Microsoft, Google, and Amazon are doing all this in the name of what I call the Great Game of AI. If you don’t play, you lose. Think of it as a bike race, where each acceleration of the leading riders causes slower riders to drop back, the peloton becoming smaller-and-smaller until there is only one winner.

That’s the way everyone assumes this ends, the way previous generations of technology ended, with one winner. But there were five Cloud Czars. Why can there be just one winner in the AI game?

Or maybe, they will all wind up losers. I know investors will end up that way. I’m not playing.