Where is AI creating real economic value?

00:04.004 --> 00:08.008

I think before last year, the value of a chatbot to

00:08.008 --> 00:10.543

the consumer was reasonably high.

00:10.543 --> 00:14.114

It was kind of, it replaced search in a sense for some people, for some tasks,

00:14.114 --> 00:18.118

but it was more information gathering, as opposed

00:18.118 --> 00:22.689

to information finding necessarily, it allowed you to go deeper.

00:22.689 --> 00:26.626

That is not a particularly monetizable use case for AI over

00:26.626 --> 00:30.630

time, especially in the prize, which as we've

00:30.630 --> 00:33.266

seen can ramp a lot quicker than the consumer.

00:34.467 --> 00:37.103

For the use cases that AI is best at.

00:37.103 --> 00:41.107

But then when coding was unlocked, I think that that

00:41.107 --> 00:45.645

can drive a lot of really efficient spend

00:45.645 --> 00:49.582

and worthwhile spend, high margin revenue for these model

00:49.616 --> 00:53.620

companies. One of the ways I think about it is, I think the numbers are there's

00:53.620 --> 00:57.791

roughly a trillion of software and services wage

00:57.791 --> 01:00.026

spend globally.

01:00.026 --> 01:03.563

If you think about that, really the gating item on that spend.

01:03.596 --> 01:05.665

Was supply, not demand.

01:05.665 --> 01:08.902

There weren't enough software developers, there weren't of service providers to

01:08.902 --> 01:10.503

really increase that spend.

01:11.905 --> 01:15.909

And so, and you know, one of the ways you saw that was, I think wages in

01:15.909 --> 01:20.413

India over three or four years in the early 2020s

01:20.413 --> 01:24.584

from call it 2021 to 2023 or four,

01:24.584 --> 01:26.786

I think they like doubled or triple, right?

01:26.786 --> 01:30.290

Because there wasn't enough supply, they just had to go find anyone that could

01:30.290 --> 01:32.859

code and bring them into the ecosystem.

01:32.859 --> 01:37.497

With the advent of coding agents and

01:37.497 --> 01:41.901

AI's proficiency at coding, now you're

01:41.901 --> 01:43.636

no longer supply constraint.

01:43.636 --> 01:47.574

Let's say in reality they spent a trillion,

01:47.574 --> 01:51.344

they would have rather spent two trillion if there was a supply of software

01:51.344 --> 01:53.012

developers available.

01:53.012 --> 01:54.681

That's a really big tam.

01:54.681 --> 01:57.784

Then the other piece is these software developers, maybe they're not coding for

01:57.784 --> 02:01.855

six hours a day effectively, Maybe it's 24 because their agents can run 24

02:01.855 --> 02:05.058

7 or they have multiple agents and so they're more productive.

02:05.058 --> 02:08.495

And so you can kind of make the argument that one trillion should have been two

02:08.495 --> 02:12.599

trillion with productivity that could be six or $8 trillion.

02:12.599 --> 02:16.636

So the TAM is there and it is a really kind of monetizable use

02:16.636 --> 02:22.575

case and that's part of the reason you've seen the labs ARR.

02:22.575 --> 02:26.646

So if you think about, they call it annualised recurring revenue.

02:26.646 --> 02:30.984

I don't think it's recurring necessarily but let's say for this debate it is.

02:30.984 --> 02:32.785

It's about a hundred billion dollars.

02:32.785 --> 02:36.823

And to go from, you know, at the end of last year, 30 trillion to a hundred

02:36.823 --> 02:39.826

trillion at the, there's a hundred million at the middle, in the middle of this

02:39.826 --> 02:43.396

year, it tells you there's real product market fit and there's really demand.

02:43.396 --> 02:46.432

And the margin should be there, especially because the margin is really betting

02:46.432 --> 02:50.436

on Moore's law that effectively, semiconductors will produce

02:50.436 --> 02:54.841

tokens cheaper over time and that the lab's algorithms will get more efficient.

02:54.841 --> 02:58.578

So if I just step back, that's the bull case is, you have this use case that is

02:58.578 --> 03:02.782

potentially, trillions and trillions of dollars of

03:02.815 --> 03:05.285

value to companies and they're willing to spend.

03:06.920 --> 03:10.056

That I think is what the hyperscalers are pursuing.

03:10.056 --> 03:14.127

The bare case on this as a bubble is that might be

03:14.127 --> 03:18.364

true, but what if the economics don't really accrue to anyone other than

03:18.364 --> 03:19.365

the enterprise.

03:21.568 --> 03:25.572

If you have to believe the labs are the marginal buyer of all the data centre

03:25.572 --> 03:29.542

capacity today, the frontier labs, that could be

03:29.542 --> 03:33.813

problematic if open source comes along and negatively impacts

03:33.813 --> 03:37.350

their margins because open source is good enough or the Chinese models are good

03:37.350 --> 03:41.354

enough or the hyperscalers can develop their own models or models switching

03:41.354 --> 03:44.991

away. That really takes economics away from frontier labs for really financing

03:44.991 --> 03:48.695

this build out. Maybe that's a much more challenging place to be and you can

03:48.695 --> 03:50.430

make a bubble argument that.

03:50.430 --> 03:54.033

Even if eventually you get to a good spot, the build-out's happening too fast

03:54.033 --> 03:57.036

relative to the lab economics, which are compressing.

03:57.036 --> 04:01.074

So I think that's kind of the push-pull is this is a really valuable, really

04:01.074 --> 04:05.245

big use case, but there needs to be economics to support the

04:05.245 --> 04:08.748

buildout and we're in the early stages of seeing them.