Excess Compute

First (this one started a bit earlier): Anthropic’s top models are allowed to be used again. Second: OpenAI’s IPO has been postponed indefinitely, and third: Meta ordered and received a shipload of compute — and has no idea what to do with it.

So, how are things in the computer world? For us normal computer users — meaning not AI companies — it’s been quite a remarkable year: hardware costs have exploded. RAM is hard to come by, GPUs are extremely expensive (especially the Pro variants), and SSDs and HDDs are likewise in enormous demand from data centre operators and priced accordingly.

All of this in anticipation of the earth-shattering, world-changing AI miracle that simply refuses to happen. Sure, there’s something about AI that promises a brilliant future. Only nobody knows what that is. And that’s now becoming a problem. An incredible amount of money has been poured into AI — and it’s not coming back. That’s obviously also a problem. The AI companies acted as though this fact simply didn’t exist, but when you want to go public to collect your trillions, you need to show numbers. And that’s now becoming an issue. As it turns out, OpenAI and Anthropic are buckling under debt, and only a fraction of revenue is being earned compared to expenditures. On top of that, these companies are running out of use cases. In other words: both seem to have no idea what to do with their technology. Yes, it has potential — everyone can see that. Only in the form it’s being presented, “it” simply isn’t there.

Meta was one of the main sponsors and has sunk billions into compute. Only Meta has no use case for it. Now this compute is being dumped on the open market with the cry: “We’re becoming the new AWS.” The new business model: we’ll rent it out. Great, except there’s already no demand — who’s going to rent it?

Another phenomenon: SoftBank wanted to take out a loan using its stake in OpenAI as collateral. Not a single bank could be found for it. What does that say about OpenAI’s (financial) state of affairs? The companies that went all-in on AI usage — tokenmaxxers — are now backpedalling very quickly as the actual token costs are being tallied up. It’s too expensive.

All in all, a picture is emerging where hype is being caught by reality — and apparently it hits hard.

And then of course the question arises: what happens next? If you don’t need the compute, are you still going to build all those already-redundant data centres with depreciation hanging over them?

Meanwhile, the case for open-source AI keeps getting stronger. The models have become enormously capable. What’s missing is: compute — but local.

Interesting times.