Hayden Field

said to reporters, I've just never seen him so stressed when he was talking about this, about how the compute constraints were stopping them from scaling appropriately and how they couldn't really deliver what clients wanted unless they just could somehow get their hands on more compute.

I've never seen him more stressed out.

And so, yeah, that's playing out now months later.

Sora took up a bunch of compute and there wasn't really a big return there.

And so they abruptly decided to cancel it, apparently 30 minutes after working with Disney on a related project, and then just suddenly pulled the plug with no notice.

So, you know, things over there, it seemed like are in kind of a tailspin.

It's like, if you're pulling the plug on a project with a huge company like Disney, 30 minutes after talking to them about like how it was going great,

There's some real issues there.

So in the numbers that leaked from investors in OpenAI and Anthropic this week, you know, there are a ton of crazy charts showing just how much of their money and profit is going or lack of profit is going towards inference, training.

The way they break out these categories is just really interesting.

It's not always apples to apples, but it's like Anthropic is going to spend one third to one quarter on model training and

compared to OpenAI.

But OpenAI's revenue is expected to hit like $275 billion in 2030.

But Anthropic says its revenue will hit $150 billion in 2029.

So it's kind of like Anthropic seems to be spending a lot less on the same categories as OpenAI and growing slower.

But OpenAI is spending a ton and then just hoping that's going to lead to a return on investment, which kind of tracks with the way that companies have been operating for the past couple of years.

Yeah, but Anthropic isn't worried enough that it'll stop because they just took out their frontier safety pledge and said, oh, actually, we're going to stay competitive even if we think it's a little dangerous.

Sorry.

So when I talked to an economist about this this morning, he said that agents have just changed everything.

And I talked to another couple of tech leaders about this, and they said agents are consuming hundreds of thousands more tokens than basic chat models have been.