Hello everyone, welcome back to

SemiAnalysis

weekly, episode number 13, lucky number

13. We're here with Joey, Jeremy, and

Kristen. We're going to talk about an

article that we put out recently

uh called Anthropic growth and bedrock

mix drive AWS margins higher while peers

lag. That means we're going to talk

about everything Anthropic, including

the uh recent announcement of their

series H, the release of Opus 4.8, but

with the spin on focusing on

infrastructure, like how exactly they

serve these tokens, especially with

their partnership with AWS. Guys, uh

welcome to the show. Excited to talk

through this.

>> Thanks, Jordan.

>> Man.

>> All right, so let's uh let's dig in with

the article itself and and talk a little

bit about the backstory. Like, I think

uh a lot of people understand the

concept of tokens, understand what GPUs

are, um but and and certainly like cloud

providers and neo clouds, but not

everybody is getting their tokens from

the same place. So, can one of you guys

give me a backstory on bedrock? Like,

what is AWS? Uh what are they, you know,

in terms of what they're doing for

Anthropic and how they're serving tokens

with bedrock? Joey, we'll start with

you.

>> Yeah, yeah, yeah. Yeah, I'll go through

that. And I'll do it kind of for all the

clouds as well. Um so so if we really

look across all the hyperscalers, and

especially the big three, um when we

look at Amazon, Microsoft, and Google,

there's kind of two big breakouts, uh

maybe even three. There's a bit on the

software side. So, if you look at

Microsoft like GitHub, um Copilot, you

know, that's an area that's like an AI,

you know, software-as-a-service type

product. They have AI, you know,

infrastructure-as-a-service where

they're just renting out um these

accelerator chips. Um and then they kind

of have this token-as-a-service

business, you know, which is where

they'll, you know, essentially expose

these outside models or their own own

models um

to consumers to interact with.

Uh so it's a little bit of a different

business uh because instead of just

renting the underlying chip, you're

you're renting the underlying model, you

know, buying it through your cloud, you

know, your CSP account and kind of that

enterprise, you know, spending

agreement, you know, you have all the

same benefits around security,

availability zones. Um and able to buy,

you know, third-party models, you know,

and also some of these first-party

models, you know, through your cloud

provider of choice.

Um so at a high level,

um those are kind of the three big, you

know, buckets right now at the large

hyperscalers that we see and we're

seeing some big, you know, changes or

differences between them and how they've

gone about it strategically. Um and then

the two main things was Anthropic's

growth um really led, you know,

Anthropic's growth and then uh Amazon's,

you know, strategy around Bedrock has

really driven their margins higher

recently. Um

was really the takeaway from the

article. And this token as a as a

service is obviously much better

business for the hyperscalers than

infrastructure as a service, you know,

in the AI side.

>> Yeah, and the if you look at the big

three, they're kind of going in opposite

directions right now. Um just in terms

of their operating margin, right? That

was the biggest chart from the article.

I'll put it up on screen right now. But

Crystal, can you maybe explain like when

we look at AWS versus Google and versus

Microsoft and you try and break out the

cloud business, like is this all to

blame on Anthropic from our perspective

or like what's driving AWS to improve

their operating margins while

Microsoft's is declining and Google's

kind of stays flat here?

>> So I feel like a lot of it is that cloud

usage, right? Like a lot of people are

using cloud more and it's mostly routing

through AWS. And then like what Joey was

saying, the token as a service um

business model, they just have a better

margin on that as opposed to like the

infrastructure as a service model.

>> Makes sense, yeah.

>> Yeah, so I I I I think like if you if

you take a step back, like the the world

of clouds has really changed a lot in

2023 as you as you started seeing these

neo clouds, these GPU as a service

businesses because uh in the old days,

which is basically 2022 and before,

um

cloud cloud service providers, you

basically had three. Right, Amazon,

Google Cloud, Microsoft Azure, they had

amazing margins, easy returns on on

capital. Some new entrants like Oracle

were trying to get in, but right the

market was dominated by three players

that had an amazing business.

And now you get to GPU as a service and

what we first started to realize is that

the barriers to entry are much lower.

So, Jordan, you're probably, you know,

the best person to talk about this. You

probably know the the CEO of 150 Neo

Clouds or 200 Neo Clouds.

How many [laughter] now? 300? And anyway

>> over 200. Yeah, yeah.

>> [laughter]

>> It is pretty insane.

Um, now obviously some are better than

than others, but the point is the the

market has much lower barriers to entry.

Uh, the moat that uh, cloud used to be

doesn't really exist in the AI era

because it's really more about

infrastructure and especially the end

users, the whole point of cloud

computing was you make the IT much

easier. Um, folks don't need to have

such a big IT department internally.

They can just rent through the cloud.

Uh, it's super easy. Everything is well.

Uh, now the big end users, they want to

have much more control. Uh, you shift

from uh, platform as a service in the

old days to now bare metal. Uh, folks

want just just the metal. Uh, Open AI

wants to think that the way they like

it, Microsoft, Meta. Um, and so this is

basically the token as a service

business is basically the first case uh,

at scale where you see a an AI cloud

provider having a a business with a

different profile. Uh, right, because

uh, obviously as you have less of a moat

in the GPU as a service era, margins go

down. Uh, Oracle, best example, they

have an RPO of half a trillion dollars,

uh, which is a few quarters ago that was

as of as of Q1 is still bigger than

Amazon's.

>> [laughter]

>> Their backlog is bigger than Amazon's,

but no one gives them the credit because

people know it's much more risky, the

returns are not the same. Um,

and we've seen empirically that every

every single GPU as a service cloud has

faced struggles when they started to

ramp up their business. But our view as

a firm is we actually think that GPU as

a service business model is sound

business. Companies like Coreweave have

a sound business, but there's this lag

effect where as you ramp up and bring

more capacity online, there's these lags

that make your margins go down

temporarily because your asset base

depreciates because you have to pay data

center leases and so on and so forth.

And so it's really interesting to see

that Amazon is basically the first cloud

provider that in a time of unprecedented

capacity expansion,

over a gigawatt per quarter now,

they expand margins.

And that kind of tells you that if at a

period of accelerated capacity delivery

they can expand margins, you start to

think of okay, what's the stabilized

margin of this business?

And that's one of the points that we

make is the stabilized margins of token

as a service for Amazon are actually

extremely rich.

The return on capital is fundamentally

much more sound.

>> Yeah, let me throw this chart

up on screen actually from the from the

article which is the percent of revenue

that is going to Bedrock. Bedrock being

the token as a service at Amazon, right?

You can obviously see it ramp up quite a

bit at roughly the the time that we're

in right now, right? Q4 of last year and

first quarter of this year. And then if

we compare that to the chart I had on

previously where we're seeing their

operating margins improve in the first

quarter of this year, they're saying

that's due to it's Can you explain maybe

in more detail why it's unprecedented to

say that Amazon can bring on a gigawatt

per quarter and still improve operating

margins?

>> So

to to to understand this, you basically

have to go back to like

why are the pure bare metal providers

seeing their margins go down? So you

look at Coreweave, which is a pure play,

so they're the the cleanest example.

Oracle actually is kind of kind of the

same. When you look at this chart, it

essentially what this tells you is that

their stabilized business

does something like you know 25%

operating margins 30 to 40% gross

margins but the the whole issue is

stabilized stabilized means that your

GPU cluster is fully functional

you're getting the the monthly rent or

whatever rent from your customer

take or pay so it's a flat fee you know

exactly how much revenue you're going to

make often times it's a 5-year take or

pay contract again fixed rate

so you know your revenue you know your

cost everything is stable but before

getting there

the obviously have to set up the data

center which is a huge capital expense

up front

and then you have this whole process or

if you have the data center built but

you need to fill it with equipment that

takes a few months with some new stock

types of equipment like GB200 which is

super complicated we've seen that lag

get longer and longer which means that

this period of time where you depreciate

your assets you pay data center rent you

pay some labor you pay a whole bunch of

stuff that period of time goes up and

you don't make any revenue because your

cluster is not yet rolled out right and

so that's the whole dilemma that these

guys are facing is they know their

business model is stabilized structure

but they're they've been facing

challenges to wrap it up some a bit

contractual again GB200 some just more

structural because there is this lag

and Amazon is the same thing right like

like everyone else they have the same

profile they're bringing on a whole lot

of data centers

a whole lot of XPUs these XPUs in theory

they should take time to bring online

yet despite this you're seeing their

margins go up which you know pretty good

sign for them

>> Yeah but it and it's clearly different

than the others

in this space where the percentage of

their total AI revenue that they're

reporting

as being from token as a service as

opposed to other products is much higher

than Google and Azure like they are not

just bringing on capacity they're

successfully selling it into

the labs that are using it to serve

tokens, right? For these models.

Um

>> Yeah, and I guess the the chart you you

showed earlier, like what I forgot to

mention is obviously the margin buffer

that you have uh when you when you sell

tokens with your infrastructure as

opposed to having a 5-year take or pay

fixed contract with a capped upside.

Right? So, that's kind of the key one of

the key points of of of the article.

>> Yeah, yeah.

Guys, can you talk a little bit about

the workload mix as well? Like obviously

um any provider could conceptually do

this, but not everybody is doing it suc-

it's not like Azure doesn't have a token

as a service business or Google. In

fact, even Crusoe, CoreWeave, Nabia are

all trying to get into this business,

too, but they really they need a

customer and they need a customer

serving the right type of workload for

it to really result in a bunch of

growth, I would say.

>> Yeah, I I I would say, you know, for

Amazon specifically, you know, they

benefit from the biggest customer base.

Yeah, they've been doing this for

20-plus years now. At AWS,

um people are very comfortable buying

that, you know, through them. You know,

people even buy infrastructure software,

you know, through them, a provider like

MongoDB, like Snowflake. Um pretty large

marketplace business. So, customers are

really comfortable with the security,

uh you know, at this point um buying

through them, having a single bill for

all this. Um and I and I think this

comes back to you some of the Anthropic

news today, you also on the ARR number

of 47 billion. You know, we look at Q1

and even in the Q2 here,

you know, a lot of the Anthropic's

um you know, the business mix is much

much different. So, Amazon is benefiting

from picking, you know, being 80 to 90%

cloud in Bedrock, you know, versus

Microsoft is is heavily OpenAI,

obviously.

Um you know, Google has a lot of Gemini,

which doesn't benefit as much from a lot

of these agentic coding tasks.

Um

and when we look at kind of the OpenAI

or you know, and the coding percentage

that's really driving Anthropic, um you

know, you're talking probably like 10

billion a month net new ARR in the last,

you know, 3 months, you know, of March,

April, and May

um, together. And, you know, a lot of

that, probably 80% of that uh of

Anthropic's net new ARR is in this API

business. You know, if we go to, you

know, OpenAI, you know, 60% of that

business in Q1 was was consumer,

consumer subscriptions, really. Um, so

um, there's kind of a mix of factors,

but but like Amazon was in the right

place at the right time

uh with Anthropic. And then two, you

know, they were able to give with

Trainium 2, I think, pretty you know,

interesting deal structure for both

parties. And I think that was the big

thing, too. We we kind of mentioned in

the article.

Um, obviously, like their mix of

Trainium, you know, they could there's

an infrastructure as a as a service like

fee component that Anthropic pays for

this infrastructure, like Bedrock

infrastructure. Um, but then there's

some interesting like hurdles around

revenue share, you know, and then and

really margin share that happened. And

because, you know, I think Anthropic was

probably at, you know, 25 million of ARR

per

um, you know, mid-one off the top of my

head in Q1 versus like probably six back

in Q2 four, like, you know, the numbers

really made sense for both parties. And

they really, you know, both benefited.

>> So, yeah. Eric, uh Crystal, can you talk

a little bit about those forecasts you

guys were making like going into this

end of Q1 and forecast for Q2? You don't

necessarily get disclosures the same way

from Anthropic, but we've at this point

kind of been bang on with the

disclosures with the disclosed revenue

figures and the margin figures, right?

>> Mhm.

>> [clears throat]

>> I think for

us, it's a little easier to forecast

Anthropic than it was OpenAI just

because so much of Anthropic um, ARR

comes from like the API side. And

because we also use Anthropic and

there's a lot of data out there about

like how people are using all of the

different uh Claude models, like, it's

so much easier to predict the workflow

and um, kind of see what the token

consumption is going to look like and

they've kept token pricing relatively

stable these past two releases more or

less, right? Whereas like OpenAI

all a lot of their revenue comes from

like these subscriptions and you never

know like if consumers are going to

switch over to another one, so it's a

lot harder to quantify the number of

users that are using the subscription

when they can just cancel anytime versus

for like API.

>> Yeah. Can you explain a little bit about

the

release on 4.8 and 4.7 cuz

pricing's remained the same but fast

mode has changed and maybe there's some

other there's been some other changes in

terms of how they're doing pricing on

the API.

>> Yeah, they said pricing is the same,

right? For regular but then for fast

mode it's different. Another cool thing

that they said apparently was that it

doesn't hallucinate as much and they had

like a pretty cool bar chart showing

that it's like it's rate of like

hallucination is a lot lower for 4.8 and

4.7 supposedly. I haven't tested it out

yet and they said that it's super close

to like mid those previews, so hopefully

it'll stop just, you know, pulling

numbers out of thin air from when we're

using it for our analyses.

>> That'd be good. That'd be good if

numbers weren't pulled out of thin air.

Yeah.

>> Yeah.

>> Um

in terms of the the fast mode and

consumer subscriptions, you have

comments there on like

I don't know what we've learned over the

past few weeks or months from digging

into how Anthropic is running their

business on AWS and like what's uh

maybe what's the biggest percentage of

their

revenue or their margin contribution

across those different mixes like those

different types of workflows that people

could be consuming tokens on the API

for?

>> Yeah, I mean we don't like we have that

in the tokenomics model like and we're

doing a big study right now on on what

percentage geo coding market you know is

of of current you know token spend

especially on the API side.

Um you know, we've done a lot of work

too on just the consumer and to some on

the B2B data of where Anthropic has has

just been taking a ton of share of net

new customers year to date both in

consumer subscriptions and enterprise.

Um but I mean it's really difficult and

a big question among a lot of our

clients like how big is the the coding

market currently. Um, I think Anthropic

recently said at their financial

services day that financial services was

the second biggest vertical.

Um,

and we kind of know there's a pretty big

gap between coding and and financial

services in terms of especially in API

spend.

Just of how people, you know, use this,

you know, anecdotally. Um,

so it's, you know, and you know, too,

there's been a lot of like recent

comments too on token maxing it and

especially the Fortune 500s like how do

you budget for that? How do you blow

through that spend over time? How do you

measure ROI?

Um, so things will have to change.

You know, I think and then people will

put some processes in place, but I mean

I know guys, you know, like Jeremy get

insane ROI at the data center model and

his team and and him

give with that and and you know, maybe

there's less policing out there at some

other you organizations that just let

people go crazy.

Um,

but that's, you know, too, another

contributing factor I think what was

just, you know, the success of coding

and then too, Anthropic like, you know,

starting to win a lot of net new share

on just the subscription side in both

B2B and and consumer, which we saw was

really interesting in Q1.

>> Yeah, I guess three things have happened

since the last time we talked about this

topic on the podcast, um, which is first

of all they signed that massive deal

with SpaceX AI Cursor, which added a ton

of uh, Jeremy Wyatt laugh at over there.

>> This SpaceX AI Cursor, beautiful,

beautiful.

>> Yeah, not Cursor not part of it yet

because they're trying not to change

their S-1, I think, but anyway, the uh,

the SpaceX S-1 revealed the many

billions that they're spending

Anthropic is spending with SpaceX with

the

uh, clause to let them back out. And uh,

meaning

let XAI has the ability to reclaim these

GPUs if they want, but that should be

some contribution to Anthropic's total

revenue. In other words, if they were

constrained by compute

uh for their ability to grow the the

business on the consumer subscription

side and enforce rate limits or things

like that, those those should go away

pretty quickly. The second thing that

happened is obviously that they raised

their series H $65 billion in funding at

a $965 billion valuation. So, raising 65

billion at 900 billion and resulting in

965 post-money. I don't know why they

didn't round that up to a nice even

trillion, but uh

we'll see.

And then

>> Isn't that like almost double from

February number, their valuation, right?

>> Yeah, what was the February number? They

were at 400 something?

>> 380 or 400, somewhere around there.

>> I guess they need to have their

valuation track with their ARR growth.

>> I mean, it's not a crazy multiple, 20x

like versus, you know, the the software

bubble back in '21, like

some of those were like 80x like

Snowflake, Cloudflare.

It's not like super expensive, and then

too, like, you know, from a Wall Street

Journal article, you know, recently my

financials, you know, that we also like

had in our model,

was, you know, they're they're

profitable now. You know, when you when

you exclude stock-based compensation,

it's, you know, Anthropic's a a pretty

good business model and like they're

seeing a ton of operating leverage.

Um and then we also know from the

Information article we read today like

you know, OpenAI's not seeing that um as

well and

um it's very clear like what the better

business is right now and and you know,

what the better business model is.

Um there's obviously a lot of like

operating, you know, deleverage if if

you know, people start cutting how much

they're spending on on coding tokens,

but

>> Yeah.

>> Um

>> But, we're not seeing that right now.

We're not

>> We're not.

>> I mean, yeah, they're still growing at

10 billion net new ARR a month the last,

you know, 3 months, which is, you know,

pretty crazy.

Um they're probably going to do another

10 in June.

And

I mean, who knows, you know, who knows

where they end up at the end of the year

like, you know, you know, Mythos getting

released as well, like that probably you

know, that does help.

Um

I mean

there's not like there's no train that's

slowing right now. And they're in all

the right places.

Um yeah, they're like seeing the

benefits of that like just throughout

the entire business model.

>> Big question here.

X XAI deal or SpaceX deal with

Anthropic?

Is it bullish or bearish for the markets

overall?

>> Which market? Neo cloud market or

compute demand? Overall, you know,

everything's related, right? Stocks,

compute demand,

um

Nvidia, all of it.

>> I think it's super bullish. Neo clouds,

man. Everybody wants to be a neo cloud.

Neo cloud is the terminal business. Even

AI labs want to be neo clouds selling

their compute to whoever they choose to

right now.

>> Well, but

I

I agree with this because this this deal

is basically

um

one player that was supposed to be a

source of demand becomes a source of

supply. So, now suddenly there's more

competition in the supply market, which

is GPUs as a service, and the

off-takers, there's one less.

I don't know, man.

It's uh

>> Yeah, I I disagree with this. I think

Cursor's training plenty of models on

Colossus right now. I think they

wouldn't have that provision in the

contract to take back their GPUs if they

really were going to be no source of

demand in the future.

Um I think that there's lots of demand

to go around and the reason that deal's

happening is because Anthropic's demand

is so overwhelming right now that they

need to do these crazy things like buy

compute from their competitors in order

to be able to serve that demand. If they

had a different way to be able to serve

that demand, they would be doing it, I

assume.

>> I mean, on the other hand, if you're a

you know, if you're XAI

uh or Meta or whoever uh other ones of

these labs that are uh lagging, you're

seeing Anthropic, you're like, "Whoa,

bro, this is what I could do if I was at

a frontier."

All right. Like if if Grok suddenly was

at the frontier, they could be a hundred

billion dollar AR business. So, to some

extent, you would argue this should make

them more bullish.

>> I think they said one has about 24

trillion of enterprise AI applications

carved out as future market TAM, right?

So, no don't even need to be a hundred

billion AR. You can just take that

forward a few more quarters and take it

to

>> [laughter]

>> 24 trillion, no problem.

>> saying So, their TAM is 24 trillion, but

uh that that's for what? GenAI

applications?

>> I'm going to bring up that chart now

from the S-1. Yeah, it's it's like

you know, there's like space and Telco

and then there's a really big section

for enterprise AI applications.

>> Okay, enterprise AI applications. So,

that would basically be the TAM for

Grok, right? And so, they're saying

actually we give up on the 24 trillion

TAM, you know, I'd rather give my

compute to Anthropic. That's a better

use case than fighting for a 24 trillion

dollars TAM.

>> Well, I I think in some ways it's just a

matter of when you can spend that money.

Like or when you can spend that compute.

In other words, there's potentially

some serial nature to the development of

AI progress, where you have to wait for

you or all of your competitors to run a

bunch of experiments to figure out what

the optimal model architecture and data

set mix or just creation of data,

synthetic data, whatever it is, is done.

Because if we were to just take a lot of

these models that people are running

today and try and run them on hardware

three years ago,

like actually the the model

architectures would run really well. Um

all the innovations in like sparse city

and attention and like these are um

these are huge improvements over like

dense models from three years ago,

right?

>> So, they they could have used that

compute uh if you're saying, "Hey, may

maybe they had a bottleneck because they

weren't able to figure out like new

architectures that would have enabled

them to like use that compute

efficiently." Then they could have used

that compute to

to do research on these specific topics,

right? Like less training, more

research. Uh or they could have used it

to like give a whole lot of tokens to

their employees and then make them much

more productive at, you know, doing a

bunch of research tasks.

Um you know, now we know that uh we know

that AI can do pretty complex science

problems, right? The OpenAI math stuff.

Uh which I know a few things about,

right?

>> [laughter]

>> Uh cuz

cuz my dad does math for a for a living.

Um

so uh yeah, I mean, I don't know. That

does tell you that uh it's a odd

decision uh when you're kind of in the

fight and giving up when

uh you're seeing the strongest signals

we've ever seen that this time is real

and accelerating and uh actually, you

know, $10 billion of ARR per month,

right? So,

>> [laughter]

>> it's sad and it's pretty odd. Uh

Cuz like I I I I I feel like this is

opposite to to to Meta. Like my sense is

like to some extent AI is giving up on

the frontier race, whereas Meta is if

anything getting more pulled up because

of what they're seeing from Anthropic.

They're like, "Hell yeah, this is what I

bet on. I'm going to [ __ ] double down

after having double down, you know, so

many times already." So, uh Meta is the

one that is seeing Anthropic and being

like, "Hell yeah."

Right? Uh I want that.

>> Well, I Okay, I think there's two

dynamics that you're overlooking here a

little bit. Potentially, one is that

Meta has a huge cash-generating business

that they can use to fund all of this

and SpaceX really just doesn't have this

business that generates hundreds of

billions of dollars of free cash flow

that they can pour into compute for the

research bets. So, they have to do it

based on venture capital, which is

unfortunately finite when you're talking

about the scale of tens or hundreds of

billions of dollars and needs returns on

some timeline, whereas Meta can do it on

a longer timeline. The second thing is I

think that the

optionality of having access to compute

that you can then take back and pour

into something is actually quite

powerful, which is to say, if they're a

cash-generating neo cloud business that

can do some research on the side and

then in the future have some

breakthrough or have some distribution

mode with X or you know, something in

the

you know, Starlink or something in Tesla

relationship or something that just

means they can take advantage of it.

They should be in theory be able to then

pour that compute into that that thing

that's just not ready yet. And

I guess what I'm saying is that well, I

I really actually love Joey to to cover

a little bit about like the earnings

before training concept which is say

like other labs are spending a whole

bunch of money training models right now

that they need some return on and right

now they are getting them namely

Anthropic and and Open AI is getting a

return on these models. But Meta is

getting no return on their models

outside of the Rexis stuff. There's no

return on new spark for example.

XAI has very limited. They had like a

million subscribers to Supergrok or

something. It's quite quite different

than

approaching a billion, you know, miles

for for some of these consumer

applications, right? And so I think that

the like the the play to say well,

cursor is doing pretty well training on

Kimmy. Why don't we just let the open

source guys build us a model for the

next year and then we'll take our

compute back and go run a bunch with it

instead of spending a bunch right now

just to keep being in fourth or fifth

place is is

uh well, I I think it plays into the

>> I have huge disagree.

>> earnings before training argument,

right?

>> Massive disagree, but we've been

monopolizing the the speech for a bit.

So I'll let Joey take in, but uh

I have huge disagree here.

>> Well, you got to you got to explain now

after he says something.

>> [laughter]

>> Yeah, sure. Well, yeah, I mean like

look, it's it's pretty simple, man. Like

uh

what we're seeing right now is that

there are tremendous return returns to

training compute.

I think it's pretty clear.

And you basically want to make sure you

have more than others if you want to

stay in the race.

Open source versus frontier, I think

it's pretty clear that the gap is

expanding not closing which everyone was

saying last year open source is going to

catch up or the gap is going to close.

The gap is closing. China is getting

closer. No, that's not happening. The

frontier is like beating to a massive

extent the open source models as

demonstrated by Anthropic they are our

trend.

Right like I think anyone who does

production or clothes sees the

difference between Claude and No Kimi or

Deep Seek V4. Um

if you were to take a guess, would you

imagine that the gap is going to expand

or is going to narrow? I would assume

that it's going to expand because one

has much more compute than the other

which goes back to the fundamental point

which is that you know training compute

has tremendous returns and so not having

compute means that you're disadvantaged

relative to competitors.

>> I think we're agreeing about one thing

which is that training compute has

massive returns if you're in the first

place, but not necessarily if you're in

fifth place. Like

>> Not necessarily. And I I think the point

is okay, let's imagine this. Let's say

Meta goes like completely crazy next

year. They secure um

an amount of compute so they have let's

say they have five x more training

compute than Anthropic. I guess do you

think they catch up? And they're they're

they're really far behind, but if they

have five x more compute than Anthropic,

do you think they catch up?

>> No, probably.

>> No?

>> No.

>> Obviously they need talent as well, but

>> Yeah, please go ahead.

>> I think it I think they need the talent.

And so I do think there's a potential

for I think you need both compute and

the talent basically.

Um and so I I think that like the maybe

the these go hand in hand where when XAI

gives up all of their talent or close to

it with all the co-founders leaving and

then they give up all their compute, it

kind of goes hand in hand there.

>> Yeah, I know 100% agreed, but I think

the point is these two things tell you

that they're basically out of the race.

Then it's going to be incredibly tough

for them to

to like

you know come back and uh extract value

out of the $24 trillion and write AI

applications market.

>> Okay, that came up again. So, I'm going

to pull that up on the S1 and then Troy,

maybe you can get us away from this

argument and talk about the uh

Yeah, this stuff here. Here's their

Here's their TAM. It wasn't 24, it was

22.7 trillion dedicated enterprise

application.

Little down here is Starlink.

>> [laughter]

>> That is insane.

>> Anyway.

>> Um

>> Does the race even matter though? Cuz I

feel like at a certain point if Meta

gets so much more compute and their

model gets so much better, even if

they're like number three, number four,

number five, or whatever, it's good

enough for most people to use, right?

And it's probably good enough to be

replacing a lot of jobs already. That

you don't have to be number one or

number two to be winning.

>> No. No. No. No.

>> Uh No, I Okay. I I think this kind of

comes down to your perspective on how

they actually use the models. So, like I

don't know. We're We're saying number

three, four, five here just to define it

from my perspective, which others may be

disagree with, is like Anthropic's uh in

first place right now because I believe

coding is the only thing that matters. I

think they've been proven correct.

Coding is not coding, it's computer use

and everything a human can do with a

computer and AI can do with a computer.

And so, therefore, this is interface to

the computer, not coding. So,

Anthropic's in first, OpenAI's in

second. I put Cursor in third. My

experience using Composer is

significantly better than using Gemini

or using Muse Spark because it's like

you can't use it

um and uh from using the Grok models

from xAI. And so, I I In some ways, I I

think that they have both the

distribution and the model to be in

third place right now. And the question

is just how much compute does Cursor

need to stay in the race? And I think

that having the optionality to feed them

more compute in the future is super

compelling. Whereas, they can't use it

right now. And so, why have it on your

balance sheet if you can't actually use

it? Why not turn it into a

revenue-generating asset and use it

later once you have more distribution or

once you build out the training stock to

improve it to the point where you can

actually do these hero runs. But we'll

see because there's there's it's really

interesting that there's four labs, five

labs in the US testing this theory from

different angles. Some are stacking

compute and have a bunch of revenue,

some are stacking compute and have no

revenue and some actually have quite a

bit of revenue if you look at cursor and

don't have that much compute right now

on a relative basis.

And I'd like to see all three pursue it

that way because

I'm not sure what the right playbook is

or

who the winner will be.

Um but we'll see. I mean it's going to

be interesting for Anthropic to attempt

to defend their number one position

because that's not a position they've

been in before. They've only had to play

catch up.

Um and I think that's actually quite

hard. I think it's quite hard to retain

talent. I think it's quite hard to

um keep pressing a compute advantage. I

think it's it's quite hard to motivate

users and and consumers to keep

consuming more instead of getting, you

know, grass is always greener with some

new feature from some competitor. And uh

I think it's up to them to to maintain a

trillion-dollar market cap.

We'll see.

>> This is Yeah, Jordan, I think it

>> Jeremy, you want to go?

>> I said we've been talking a lot, man. I

want other people to talk, but I had a

response for Crystal, but you go first

and then

>> it goes into like the the earnings

before, you know, training interest and

taxes. But I think it's really

interesting like, you know, we think of

ETA like before training interest taxes

is like the, you know, the cash

operating profits that they you generate

for running inference. Um and if you

want to think of training and research

as CapEx, you go back to this

conversation like I think this is a big

investor question, like even corporate

strategy question is that return on

invested capital. Um and like right now

we know like Anthropic is obviously

having massive massive returns on the

invested capital they not only put in

these models, but also like just for

coding applications, these computer

applications specifically. And then, you

know, we're seeing more and more news of

other

um you know hyperscalers uh or not

hyperscalers at their labs um trying to

get into this coding you know market. I

think there was some Microsoft news this

morning on that. Uh you know mixed mixed

opinions I think here of of how

successful that might be. Um but

training more and more models for this

um because obviously people do want to

use frontier models to Jeremy's point.

We even saw like the meta token maxing

article like

you know all that spends external. Uh

token spends external because of

you know because to Jordan's like

computer and coding point.

Uh that's where like you know there's a

ton of obviously product market fit and

they're seeing their own ROI when they

use the product.

Um

but yeah, that's heavy. I I'm guessing

Jordan's still on the call. So Jeremy

I'll I'll send it back over

>> Yeah, yeah.

>> to you.

>> J- just wanted to respond to like

Crystal's point because

yeah, I mean it obviously depends like

how you think of the the market evolves,

but

I really like the micro framework that

Malcolm has which is that okay, I think

like you know 2030 2035 like uh what

type of uh tasks are going to drive the

bulk of the total addressable market,

you know the dollars that people are

actually spend on AI. And there's tasks

that are like uh have a certain amount

where they're good enough that's kind of

finite like translation maybe at some

point there's only time it doesn't make

any sense uh to spend more on

translation, but then there's like these

very open-ended task where the spending

is pretty much infinite. And then on

legal for example, you could assume that

if AI is really good at at legal, if you

want to make sure you beat your

competitor, you want to probably want to

spend more on AI than he does, have more

intelligence than he does uh because you

want to gather more evidence, you want

to think through many different ways of

uh you know coming up with a defense and

whatnot. Um scientific research research

is typically the very open-ended use

case. Um you know health care um for us

as analysts, we try to get insights from

gathering a lot of data, very

open-ended. Um I would assume these use

cases are going to drive much more

spending than this the finite sort of

it's good enough, right? And and and

really when you think of these

open-ended use cases, what what matters

to be able to do like what you want to

do uh

uh in the cheapest way. And if you're

only doing the cheapest way, you're

going to have to use frontier models.

It's the whole point we made we made

like Mythos is actually 2 first cheaper

than Opus.

The model in terms of token pricing is 6

times, I think, 5 times more expensive

than Opus.

But you know, if it's like 10 times

smarter, if it requires 10x less tokens

to answer a given task, then it's

actually way cheaper to complete that

task with the the model, right? And so I

think like at least that's that's kind

of the way I view it. And so

I I I think the bulk of the market is

going to concentrate to

the frontier models. I think if you're

three or four or five, you're not going

to get any dollars. And I think like if

you take step back and think, "What are

the signals that we've seen in 2026? Did

the AI market has it beaten or has it

missed versus expectations we had in

2025?" You know, massive beat. But then

what I think is super interesting is

what is the composition of this beat?

Did everyone beat or is it just a few

companies, right? And I I think that is

actually kind of interesting. On the

frontier side, it's basically one

company, it's just Anthropic, massive

beat.

Google is probably tracking behind to

some extent when you look at just the

Gemini adoption and how much people were

spending on Gemini. OpenAI is tracking

behind. Obviously xAI, Meta are nowhere

to be seen. You could have hoped that

they would have had something. They

don't really. To be fair, on the

open-source side, I think that it's also

been a beat. I think there's been some

good adoption, but the dollars uh spent

are still pretty small. So I think it's

an interesting composition of massive

beat where it's basically all driven by

one company, which kind of gives gives

you a sign that it's pretty much winner

takes all. And if you're state of the

art, you get the bulk of the volume, and

if you're not,

yeah, people don't spend on you. I know

Joey stopped playing. What do you think?

>> Can you repeat the last part of that?

>> Holy [ __ ] you didn't listen to my

beautiful prose. I'll jump off.

>> I heard that I heard the most of it.

>> I got to jump off and respond to this.

So, Jeremy, I I think you're this is

totally this totally makes sense, but

we're also seeing massive beats or

massive reported ARR numbers from

startups serving open-source models

right now. It's not just Anthropic

growing.

There's there's huge growth for

Fireworks that's tied to Cursor.

>> that big, man.

>> Really big.

>> They're not that big, man. That is the

thing. They're they're not that big.

Some encouraging signals. I think Cursor

is at what? 2 2 billion now?

Uh and they were at what? Maybe

1.something at the end of 2025. There's

still like really good growth.

Definitely you could say Cursor is a

beat. I don't know WinZO, I guess,

they're in Google, but they're worth to

be seen.

>> Yeah.

>> No, but I I I I I agree with you. Some

of the open-source guys have had a beat,

but in terms of like dollar amount,

still like not very meaningful.

>> Yeah, I mean the claim from Fireworks is

$315 million of ARR, right? Like that's

>> What is 300 million between between

friends?

>> [laughter]

>> Okay. Like okay.

In the past, a startup unicorn was

interesting when they were a

billion-dollar valuation. Now they start

to approach billion-dollars ARR, and you

go, "Ah,

ah, whatever. Fly on the wall. Like

>> Relative to the size of the market, you

know, which is like already above 100

billion dollars. It's like in the grand

scheme of things, not that big. Think of

the the positioning of the different

hyperscalers. I mean, we've talked about

we've talked about like, you know,

business models and whatnot,

but the beauty of the the economics 2.0

this

magic model is just so accurate,

is that it covers

>> Yeah,

I mean, you know, in 2 weeks we've

gotten some pretty good feedback so far

from some of that

you know, these you know, hyperscalers

customers, things like that. so um it's

been it's been solid, but I mean, I

think right now like just given like

Amazon continues to win.

Yeah, I think the things that benefit

them you know, around their customer

base, you know, benefit you know, Azure

you know, pretty similarly.

Especially as you know, you look at you

know, more and more you know, token as a

service type models that come on the

foundry.

And then obviously I think you know,

Google if they get a coding model you

know, right now you know, when we look

at what was formerly Vertex and is now

Gemini H enterprise platform.

You know, we we still think you know,

Gemini API is a pretty decent percentage

of that. So we're not benefiting a ton

from

you know, Claude and some of those

things, but like that's probably the

biggest thing right now. I mean, we

still see Amazon like Bedrock you token

as a service is a pretty you know, I

think by the end of the year it could be

the majority of the AI business

at Amazon and even though Amazon like

lags Google and or GCP and Azure like as

you know, their AI mix. Yeah,

at at AWS AI is a lot smaller percentage

of the business. But with Bedrock going

to the majority of the AI business and

then infrastructure as a service being

you know, 80 90% at Azure and GCP.

Um

you know, it's really really

advantageous. But then too to that

point, it's a very easy for Azure to

then come in you know, add Claude as a

model.

Um and kind of you know, this token as a

service business isn't that hard for you

know, these people with massive customer

bases to implement. That's much harder

as you go down to like Oracle.

Um

and then to the neo clouds like like to

implement this at the same level just

given they don't have you know, the

massive inertia customer bases.

Um

that you know, really benefit from like

the old school software distribution

enterprise software distribution like

you know, emotes.

Um

>> Yeah, let me throw

let me throw this chart on on screen

just to make the point Jeremy was making

and you're making right now which is

like bunch of rounding error for

anything but the top three.

Hyperscalers when it comes to the

inference play business, right?

Um yeah, it's uh

>> [laughter]

>> Yeah, very small portion of the market

that's going to ever to everybody else

bucket.

>> And it's also I mean that that kind of

goes to the to the same thing the same

point I was mentioning earlier with

regards to like the composition of the

market.

One of the important point he makes in

that article is that um

the key to being a successful token as a

service business

is that actually just to have

partnerships with the big labs and have

access to frontier models. It's pretty

simple, right? So, I guess it's uh the

big disadvantage that currently Nebius,

CoreWeave, Iron, all of these guys have

is they don't yet have the partnership

and they also don't have maybe the

capital to be able to deploy GPUs

without a 5-year contract. Um so, it's

kind of like as a as a function of the

way the market works that a company like

CoreWeave or, you know, Nebius have the

bulk of I mean maybe not Nebius

actually, but Iron have the bulk of

their business contracted over multiple

years. Um whereas Amazon, you know, is

free to be more spec

and better pick more demand that don't

have that 5-year off-take uh take or pay

locked in.

>> Yeah, well

>> And that I think too, Jordan, I think on

like who wins is is definitely the

custom silicon portion.

Um You know, around Tranium and you

know, TPUs at GCP that I think is a big

thing versus like in the tokenomics

model that and then when we look at like

brain accelerator in

um and some of the data center model

numbers

you know, it's really at Azure you know,

just as mostly in video and then you

look at that vertical integration at

um

AWS and GCP, like that's another big

advantage for them.

Uh that that I think, you know, probably

you know, especially on the margin side,

I should say. And then as we've seen a

lot of, you know, inference get more

efficient and you know, gross margins on

inference

you know, drastically improve at the

labs, you know, the two you know,

frontier labs, major frontier labs over

the last two years like that's

definitely been

you know, another like key consideration

when you think about who's going to win

you in this market.

>> Yeah. Yeah, I mean that it's from the

technical perspective, everything that

we've criticized all the chip startups

about and TPU and Trainium is always

about usability. But if the entire chip,

you know, a gigawatt of Trainium at

Rainier is all just serving tokens from

one to three models. Well, the customer,

the end user customer doesn't even

necessarily need to know that much about

what chip is running if they're only

buying tokens.

And certainly not the end users, like

the actual terminal user of the tokens,

you know, when I'm using Cloud Code, I

have no understanding of whether my

token is coming from a TPU or a Trainium

accelerator or a GPU. Doesn't make a

difference, right? Okay, guys, well,

really interesting stuff today. Anything

left unsaid on the the topic of

uh Bedrock, tokens as a service,

Anthropic's growth.

>> Hey, I think we got it all, Jordan.

Winners win, losers lose, and it was a

clear trend.

>> [laughter]

>> Winners win. Awesome. Okay. Well, thanks

guys for coming on the show. Good job

today.