I invested in Nvidia in in the fall of

2016. There's been 20 years of lack of

investment in in the hardware capex

cycle. There's not an actual GPU

shortage anymore.

>> Does our portfolio positioning make

sense? Why are we so long semiconductors

if the true shortage is in powered land,

not chips other than memory? Got a

special conversation today. I'm joined

by Ben Pulandian of BEP Research. Ben is

a specialist investor and analyst in the

semiconductor world and the

semiconductor supply chain. Of course,

that is what powers AI. Ben, welcome to

Monetary Matters.

>> Thanks for having me, Jack. Excited to

be here. Love your podcast and would

love to dive in on some interesting

topics and some things that your

listeners care about.

>> I'm I'm really glad you're here, too. I

really like your work. You're very in

the weeds on the semiconductor uh world.

I want to start by asking kind of the

the I propose the bare argument. You

know, Ben, perhaps many people watching

this, many institutional investors are

who are a little skeptical about AI,

skeptical about semiconductors, they may

think, you know, that this is the repeat

of the dot bubble. And so my first

question to you is Cisco was a very

profitable company. The earnings and the

growth experienced in the do-com bubble

was extreme just like it is now in AI.

Why isn't Nvidia Cisco? Why isn't this a

replay of the dotcom bubble?

>> Thanks. Uh, a lot of people would bring

up Cisco overlay the Nvidia chart. The

the challenge with Cisco and comparing

it to fiber in my opinion, things like

that are sort of like a commodity. Uh,

so when you're talking about dataccom or

telecommunications, it's the idea of

transmitting the data, right? So it just

goes from point A to point B. With

compute and Nvidia, you have the chance

of actually creating intelligence,

something from nothing. Everybody wants

intelligence. It's not a commodity that

anyone else can make. And I think that

is a discerning difference. You needed a

company like Cisco to basically

overbuild or global crossing to die on

that hill to bring that capacity and

bandwidth to have the compute that we

have today. Little different.

>> So here's I'll start where I I disagree.

I think intelligence is kind of a

commodity like open AI creates geniuses

and anthropic creates digital geniuses.

What sometimes some geniuses are better

than others. Now the other one company

pulled pulled ahead and there some are

better at math, better at writing. Okay.

But that is kind of like it is kind of a

commodity. But here here's where I agree

with you that semiconductors are not a

commodity. The challenge is is people

are just still using AI to do routine

work like hey help me plan my trip you

know obviously write software or how do

I make this recipe what's this bug that

I took a picture of right the next

inflection of AI which we'll see coming

and what I'll be writing about soon is

the intersection of artificial

intelligence and the bigger sciences

mostly material science biotech and

other things once you start discovering

new materials, new medicines. That

impact of doing something 10 years in

the lab trying to figure out if you can

model it with AI and the actual uh GPUs

and computers and find drug targets that

is a big unlock. You compress 10 years

of work into one year and people need to

realize the speed up that you get with

intelligence. So you're saying that your

bullcase is not only 15% of the

population is using AI every day to plan

their trips and do routine tasks and

that 15% is gradually going to head to

70 or 80%. That is not your bullcase.

Your your bull case is that mainly the

intelligence is going to be advanced at

the frontier that can do tasks that

previously required millions, tens of

millions, billions of dollars in human

intelligence and now that can be done

much more cheaply and at scale. That's

what you're saying.

>> Yeah. I mean, going back to what Jensen

has been saying at GTC for a while from

Nvidia, the whole point of all this is

can you do your life's work in your

lifetime. So uh when you look at R&D for

medicine and people working in the labs

there's a lot of trial and error if you

can scale that with and compress the

time to get to better patient outcomes

with better options that is the unlock

you couldn't do that with classical

computing you can only do that with high

performance computing larger clusters

more data the more data points you throw

into it genomics like organs NHS data

from different uh places of the world.

Once you have multivariable data points,

you can all of a sudden triangulate into

things that you weren't able to do

before and find that needle in the

haststack to help that one patient.

>> Yes. And that bullcase for uh

semiconductors and AI is something that

I'm you know seriously entertain and I

encourage my my you know people watching

to to to entertain. I also think Ben

that a short-term 12 to 18month bullcase

is that a tremendous amount of capital

way more than has already been spent is

going to be spent on that goal with the

anthropic open AI basically trying to

build a digital god or you know a

digital IQ person that never sleeps and

has you know it's a billion years of a

160 IQ person in five minutes you can

you can do and whether or not that is

going to be achieved within a short

period of time so much capital is going

to be spent And that money is going to

go to Nvidia, Lamb Research, ASML, the

entire semiconductor supply chain. So I

think that that's something I want to

see is that the in over the next 12

months, the bullcase doesn't need, you

know, building the 100 IQ person in a

data center to become true. It just

needs the accelerated capex. That's what

I would say. What's your reaction to

that?

>> Yes. And the parallel to that also is

obviously the best, smartest, fastest AI

from a national defense perspective is

the best military. So if we step back

and say, "Oh, this is all a waste and

we're not going to do it," then you have

other countries or bad actor nations

that could basically have the best AI

and infiltrate all of our digital

systems and basically our

economy and virtually take over our

country, right? So people need to

realize it's just not again chat bots.

The the race of the best AI is the best

military and the best military in the

world basically sets policy around the

world. And that's where you know I've

written a lot about it in the the my

thesis of the token dollar how

everything's denominated in tokens how

the world has enough oil. You can't

close a straight of compute, right? It's

only power constrained and the country

that leads in the highest compute per

watt at scale wins and it's a race

against us against China. So, it's sort

of like the space race reversed. Uh

instead of Russia, it's China this time.

explain very briefly what a token is and

then why is there such a shortage of

tokens that you forecast for to for

there to be many many years and why

won't there be a an excess of tokens

such as there was an excess of oil you

after the the the increase in the price

of oil in 2022 there was a there was a

surplus that lasted for many years until

the Iran war why isn't there going to be

a shortage in tokens that ultimately

leads to a glo glut. You know, in most

commodities and most things, shortages

leads to gluts because there's overp

production. Are you saying that you

don't think that's ever going to happen

or it's going to be, you know, 5 10

years, it's unlikely? Explain what gives

you that that confidence in your view

here.

>> To draw a parallel to oil, all tokens

aren't created equal. So, when you start

with a barrel of oil, it's either West

Texas crude or Brent Brent. And then

from there you refine kerosene and or

gasoline or diesel. And the the way

compute is heading based on the scaling

laws if you see from anthropic and open

AI the more money and more compute they

throw out these bigger models they're

getting better more intelligent results.

And the idea is if you're doing compute

what model are you getting your

intelligence from? If you're getting it

from a free I don't know 30 billion

parameter llama model for meta from a

few years ago like that's like it's like

water at this point. It's like

worthless. Like it's it barely does

anything. But if you want frontier

intelligence with the best model at the

highest speed

answers and tokens, that's almost like

jet fuel, right? The most expensive type

of fuel. So you're going to see this

segmentation in the market on different

types of tokens and the speed and the

intelligence that that you get it at.

And will there be a supply glut of

intelligence? I don't know. What I know

right now is that me, just one guy

trying to use Fable 5 for the past

weekend, like I've hit my rate limits

over and over again and I had to buy

more credits and compute and it's it's a

scary feeling when like you're mid

through a project and it says sorry,

come back in four hours before I can

finish it. Like it it it's it's not

good. I mean, if they're doing that to

me, who else are they doing it to? And

it it just another data point that was

just very compute constrained.

>> Explain the difference Ben between the

frontier model like Fable from Enthropic

or OpenAI news model that is is going to

come out in early July. What is the

difference between what the frontier

models can do and what the decent models

like you know a model from 12 months ago

that was considered a cutting edge 12

months ago can do?

>> I think it just comes down to the data

that it's trained on. Obviously like

there's cut offs in intelligence. So

some stuff is cut off because they

closed the books like oh this I'm only

recent up until January of 2026. So the

more history and time you put into it

obviously the bigger the model is and

then there's also things called world

models that you can input audio and

video. So the more multivariable it is

and the more flexible it is obviously it

provides more intelligence. And then

some of these models are tuned for

specific tasks like agentic coding

software development where basically you

can turn English into computer language

to build things in the virtual world or

software. And the better the frontier

model is at these things the better

results you get and your code or your

software is stronger and made with

possibly less lines of code and less

bloat. So, so what percentage right now

of the value and maybe of the

consumption and and spending on the

frontier tokens is specifically in

software engineering

>> from an enterprise standpoint from what

you gather from companies. Almost all

software developers are using agentic

coding at this point because if you're

running an enterprise, it's not uh it's

not all you can eat. They're all running

APIs. So it's consumptionbased and

that's where you see the huge inflection

in anthropics ARR is they made a bet on

the enterprise and I remember I was

using last summer my friend showed me

Claude sonnet he's like oh I'm using

cloud I'm building all these things I'm

multiple windows open and I remember I

was trying to build a few things and

Claude would tell me he would do it and

then I'm like okay where is it and he's

like oh I'm sorry I lied and I was like

dude like come on like I spent hours

coding and I didn't get anywhere and

then I let it go and then in November

he's like hey they came out with this

thing called opus it's much better it

doesn't hallucinate as much and that

really was the inflection point for

anthropic once people realize that the

coding model improved and it

hallucinated less and it lied less and

obviously it has those agentic workflows

where it's controlling parts of your

computer from Xcode to other opening

windows and looking at things you saw

the future in front of you and it was

pretty amazing.

>> So you feel like you kind of have def we

have definitive proof that in the world

of coding the the advancements here are

are revolutionary and uh certainly are

going to generate a lot of revenue. What

do you think the odds are though that

it's limited to coding?

>> The idea of coding is just one

application. So the idea is to take this

and how do you make like a cloud code

for law? How do you make a cloud code

for finance? How do you make a cloud

code for science? How do you make a

cloud code for education? How do you

turn natural language into a tool to

build things and learning? I think

that's the idea. And then run loops

around that. There's a repetitive task

in your organization. How can somebody

who doesn't have a computing background

build a script with cloud code or some

coding assistant to automate that task

so they don't have to waste their time

on something mundane to work on

something more productive to grow the

business to grow their minds. I think

that's the unlock for a lot of the AI

stuff is like, you know, stop trying to

build spreadsheets that you that you've

already built a template once, right?

And invest in banking analysts,

sensitivity tables. Like that was like

the hardest thing. Just give the inputs

and have Claude make it. Go focus on

something else. Focus on something

something more strategic. I think that's

that's where I see the the unlock and

the higher productivity for for

enterprises and and people. Now tell us

the role of semiconductors. The

semiconductor stocks have absolutely

been on fire. Just what is it about AI

that demands so much semiconductors in a

way that semiconductors have so much

pricing power such as as they have? I

mean Nvidia's uh revenues and profits

have just been unbelievably unbelievably

high. you know, you know, from in 2023

to 27 billion to last 12 months, $250

billion and in operating profit, you

know, over the next 12 months, I think

it's going to make $200 billion. Um,

that that would be a conservative

estimate. Just looking at it made more

than that in the last quarter. And by

the way, none of that included any

markups. I'm not talking about net

income. I'm just talking about operating

profit. So that is pre-tax. And by the

way, Samsung reported and they had a

quarterly operating profit that was

higher than Nvidia's because the squeeze

in memories is so high. So I I think

that the semiconductor stocks have gone

up, but the multiples have gone flat to

down. This has not been a at all a

valuation increase. This is this has not

been a a bull market driven by

valuation, the multiple going up. It's

been the earnings going up and going up

so much. So I just want to set the stage

for you there. I think you have to

really go back over 25 years. I

graduated college in ' 04 right after

the dotcom bubble of electrical

engineering. I got a job in investment

banking in San Francisco at Jeffrey's

Broadview. Uh the guy that hired me, the

associate, Alvin Lynn, the first day on

the job is like, "Hey, actually, you're

a great candidate, but I'm quitting. I'm

going to go work for this company called

Nvidia." He's still there. I think he

reported stuff to Jensen. Great guy. And

then at the same time we actually had a

mandate to sell this accelerated uh

digital media software company for

mobile processors to Nvidia and I think

they eventually absorbed it was called

Paceoft Silicon and Nvidia at that time

was just a graphics computing company.

There was a lot of hardware start

companies in in Silicon Valley. I mean

Intel was a leader and from ' 04

then on it was sort of the birth of the

new internet. That's when Facebook took

off with AWS and Microsoft and

Salesforce. So you've had this regime

give or take of 20 years of

classicalbased computing which was just

software as a service. So if you would

go up and down Sand Hill Road or to any

investor in New York, they're like, "Oh,

you want to raise $20 million to make a

chip? Oh, that's too expensive. you got

to like engineer it and then you got to

tape it out with TSMC and you know like

it's just it's just I can put less money

in this this like B2B SAS and I know I'm

going to get you know 10 times ARR and

it's going to be amazing and that's why

if maybe you or your readers or your

listeners can actually chart maybe the

Bessemer CLOU index or IGV over the last

20 years kind of see the multiples and

how it's gone up a lot and obviously

there's a lot of stockbased compensation

and rule of 40 and like all this stuff

that they would hype up into SAS and

ARR, right? And the hardware guys, all

the semi guys for the last 20 years were

just like no one loved them. Like I

remember in college I had an internship

at this company called Simer. Bob Akens

>> was the the founder CEO. He went to UC

San Diego. him and his partner Rick

Sandstorm developed the DUV Xrimer

laser. Right. So our customers then at

Syber were obviously ASML, Nikon and the

end users were TSMC, Intel and other

fabs around the world. And he's like at

that time the market cap was only a

billion dollars and he was he was like

stuck in small cap hell and he was like

struggling and eventually they sold the

company to ASML. So without that laser

and ASML making the fabrication

equipment, we wouldn't be where we are

today. So, you know, going full circle

to what I'm saying is there's been 20

years of lack of investment in in the

hardware capex cycle. And all we've been

doing the last 20 years is just building

regular data centers with Intel CPUs or

AMD CPUs or whatever Amazon's doing with

their with their CPU chip and just

hosting email and you know some browsing

and web pages and maybe the only company

that was actually doing AI at the moment

was Google with their TPU program

optimizing stuff for Alph Go and search

and recommendations. And then all of a

sudden in uh I mean I invested in Nvidia

in in the fall of 2016 when DGX came out

and all that other stuff. But

>> not bad. Not not bad. By the way,

>> I'm like it at that time it was called

par they had the the advertisement on

Facebook. It was like power your machine

learning or your AI journey. It was like

a $50,000 box. I have the photo. It was

like a Facebook ad. And I was like yeah

machine learning that makes sense. And

then but in in fall of 2022 when the

chat GPT app came out and everyone

downloaded was like holy like it's

actually pretty smart. It answered my

question. It's not stupid. And then that

was the inflection iPhone moment for for

AI. And I remember you know my 2007

iPhone moment. I remember when Steve

Jobs showed it and I went to the AT&T

store and I was like I'm gonna get a

Blackberry World because I'm a

businessman. businessmen use

Blackberries. But then I saw the iPhone

and I was like all of glass and it was

like beautiful. I was like, "Holy

this is like amazing." And like that was

the iPhone moment. No one looked back.

What's Blackberry? I mean, what's

Motorola? Like everyone forgot about

that stuff. So then you basically the

iPhone moment for for 2007 till now was

a huge supply chain shock in itself for

bassband processors, memory, cos image

sensors for the camera, capacitors like

they went through all that uh at Certino

with Tim Cook and they basically the

supply chain has matured. So fast

forward, you know, to 2000 15 years

later with this this Nvidia, was it H100

at the time that actually made all this

stuff and everyone was like, "Oh my god,

like this stuff actually works. You can

parallelize intelligence and the more

GPUs you put, the better answers you

get." And then like the AI boom took

off. So we're four years into that this

November. Okay, four years. And it's not

as easy as going to Taiwan and China and

build a supply chain and come out with a

new phone every year and have consumers

buy it. You have basically like the

largest companies in the world basically

as your buyer and then you have to go

get a piece of land, permit it, make

sure there's power, get a bunch of

tradesmen from all across the United

States that are very limited, build it,

energize it, make sure it works, and and

bring it online. So the two are not the

same. And that's why you see this lag

and this huge investment in

semiconductor capex regular

semiconductor stocks and obviously

Nvidia because there's been this 20-year

lag in investment and care for whatever

it's been doing because all the love and

attention has been going to capital like

businesses like software. I mean

>> and when you sorry Ben when you say the

two are not the same what are you

comparing to? I think I think I have an

idea but I just want to be clear. the I

mean the I mean I'm just drawing the

parallel between Apple and but they're

kind of like the same but they're not

but it's the same sort of inflection

when you're using as an inflection point

for iPhone moment. I think the idea is

the idea of a smartphone was always

there with like a web browser and like

all this other stuff but like it was too

clunky to use. I remember when I was at

UC San Diego, I was part of the Qualcomm

Kiosera uh like beta testing program and

they gave me this this like brick phone

that had a web browser and I was like

look I can check my emails but it was

like

>> it was this clunky crappy thing. It was

like it had no user experience, no user

feeling, no soul. But when the iPhone

came out it made sense like anybody can

use it. saving a chat GPT anybody can be

a data scientist anybody can ask a

question that's the idea is like tech

the technologies are all out there but

what does he unlock to open it up to

everybody and that is that moment

>> and I want to compare CPUs versus GPUs

CPUs are sequential

and GPUs that parallel processing so

they're running on on multiple nodes at

the same time and is that why there is

such a supply squeeze in the

semiconductor supply chain is that GPUs

are just so much harder to make or are

there other things in other words why is

there such a um why is there such a

shortage of chips right now in a way

that in most the time let's say okay

memory prices or chip prices go up in

1994 they went down in 1995 most of the

time they resolve quickly

What is it about about the specific

semiconductor supply chain about why it

is so hard? Because honestly, Ben, the

reality is and I, you know, I love

commodities, but oil like there's a

secret a secret about oil is that in

Saudi Arabia, like you stick a stick in

the ground and oil comes out. There's a

lot of oil in the world and of course

there are shortages and the straight of

our moves but you know if the price of

oil goes to $200 extremely intelligent

uh petroleum engineers around the world

are going to work extremely hard to try

and get oil out of very difficult to get

places and they're going to succeed but

and and the same is true over a

long-term time horizon about about

semiconductors but specifically what

about the semi supply chain is is so

hard over the over the past three years

looking back and then over the next

three years looking forward

>> Nvidia itself is sort of like a very

special company because it's not

strictly a semiconductor company. I

think more than half the company is in

the software space and there's a lot of

researchers and developers. They have

their own open- source frontierish type

of model. Neatron and the stuff that

I've written about is like they can

squeeze more performance per watt out of

the existing chips just based on

software optimizations. Okay, this is

like the same thing as getting an

overtheair iOS update on your phone over

the overtheair update on your Tesla,

right?

>> Yeah. By the way, there were times I saw

a like news that Tesla had to do a

product recall. I'm like, "Oh my god,

this is so bearish for this company."

But then I realized it literally is just

a software upgrade. Like they're not

going to have to send the trucks back,

>> right? So any hardware company that

obviously doesn't have like a software

layer that provides the customer benefit

over time commoditizes and obviously you

can make that argument for a lot of

parts in the the semiconductor supply

chain. Memory has always been like a

commodity. That's the argument.

Companies like Micron trade at a

multiple of book value. That was always

a metric when I would talk to funds and

people. But apparently this time it is

different because high bandwidth memory

is special. It's not standard. And you

need all this memory because the more

context you give your AI or your brain,

the more data points it has and the

better answer you'll get. That's all it

comes down to. It's just having a bigger

brain and remembering more things,

right? I mean, that's why you see this

huge inflection in in memory all of a

sudden. This huge demand for years. It's

been stable and going back to CPU. Yeah.

Because we were just running classical

computing every year. Like every 3 years

you had to replace your computer, right?

That was the the corporate upgrade cycle

for for desktops. Okay, it's been three

years. We're going to have a corporate

upgrade refresh cycle. Like all that

stuff's like out the window now.

Everything is changing. The dynamics are

changing. And I think you know Nvidia

it's like because they have the software

hardware co-design and then they work

obviously with anthropic and open AAI

and others like they're tuning the

software and hardware exactly to how

these models work to the spec and once

you do that you can squeeze out more

performance and I think that's where a

lot of people are missing uh those

things and it there's not an actual GPU

shortage anymore. there was at some

point, but I think I read the number. I

think Nvidia spent $110 billion on their

supply chain basically buying everything

up for I don't know how many years out

to make sure that say for example one

screw or one cable wire or whatever it

is. If that's missing and the rack is

like a million parts, that means your

whole data center deployment is delayed

and every day

>> you think there's not a GPU shortage

anymore. Then how come Nvidia has such

pricing power and as does to a lesser

degree the other GPU companies AMD comes

to mind if there's not a GPU shortage?

>> I think there's a shortage of powered

land and tradesmen to build data centers

as fast. I think that's where you the

disconnect is. People don't again you're

not building millions of phones and

selling them to consumers at AT&T stores

and Apple stores like this is

infrastructure. Infrastructure and

buildings take time. You have neighbors

uh petitioning. You have sequa. You have

uh water issues like like infrastructure

takes time and there's a lag, right? And

sequencing.

>> There's a shortage of infrastructure and

powered land, but there's not a shortage

of GPUs. Why is that a bullish scenario

for Nvidia? Nvidia, by the way, actually

on a forward multiple trades at a

discount to its past five or 10 years.

if we can, you know, put up a chart of

that. So, the market is kind of pricing

in that that lack of pricing power or

that their their pricing power is going

to go from extremely high to very high.

It's still going to be very high, but

just less so. And you know, their gross

margin is going to go down, their

operating margin is going to go down.

Why do you think that is or is not going

to happen? I think the the bearish

argument with Nvidia is that obviously

every large hyperscaler from Google to

Amazon, Meta, was it even Bite Dance?

Everyone's going to make their own

custom accelerator to try to not pay the

Nvidia tax. Obviously, you're paying a

higher price because you're getting a

dependable product that ships on time

with the software stack on top of it.

But maybe if you can use it as a way to

leverage on pricing or maybe you'll get

better performance. The the jury is

still out if you're actually saving

money on doing it yourself. I think the

only company that really has full use of

their internal silicon is Google and

their TPU, but they also buy billions of

dollars a year of Nvidia GPUs and racks

and serve it on the Google cloud. So,

and they make money on that, too,

because there's so much demand. So,

either way, there's more demand for

compute than supply. And it's just not

the chips. I It's just energized, ready

to go compute because like I said,

getting the data centers up is the lag.

It's not making these in factories or

like everything at Taiwan Semiconductor

is like fully automated to the max.

Like,

>> yeah,

>> you don't have to worry about that. Like

they're good. All these chip companies

are good. The lag is deploying it in the

real world.

>> Okay, Ben. So, I'm going to venture and

assume that that you own a lot more

semiconductor stocks, you know,

notionally in terms of value than you do

the infrastructure companies like iron

or cororeweave. I mean, prologus is, you

know, something like that or or the

power suppliers eaten, verdive. Okay.

And by the way, that is true about

myself. But if what you said is true

that the true shortage is in powered

land and not in chips, does does our

portfolio positioning make sense? Why

why are why are we so long

semiconductors if the true shortage is

in powered land not chips other than

memory?

>> Because I think the backlog is measured

in years and not quarters. So like I

said if if a macro investor doesn't want

to is like semiconductor agnostic is

just scared of semiconductors because of

the whole cyclicality then definitely

the way to really play this is power and

infrastructure power is everything. If

you can generate power profitably or

have the infrastructure for it, then

there's value today. And I think, you

know, I've been writing about one of the

most interesting companies that I think

since December is Bloom Energy. They've

been around for 20 years, gone nowhere,

and the guy wanted to like powerhouses

with like this small fuel cell box that

takes natural gas and converts it to

electricity. And now I think in less

than like 80 days they can deploy

modular stuff at a data center site,

hook up to the the natural gas and

produce electricity and at 800 volt DC

without all the the transformers and all

this the crap in in the middle of it and

it doesn't make any noise, right? So I

think a a lot of the challenge between

with General Electric and Verova with

their turbines and like that stuff is

noisy and weird. That's not the data

center of the future. That's to me

that's like a temporary solution because

you the grid doesn't have enough

capacity. We need to find long-term

better sustainable power solutions

because power is everything

>> and and so so so many data centers that

the the grid doesn't support because it

has to be approved by everything. So,

they want to go behind the meter. So,

okay, we're going to turn natural gas

into electricity. We're going to buy the

the turbines from the three companies

that do that. You know, Mitsubishi, uh,

G Venova, and okay, Genova, their

backlog is out until 2028, 2029. Okay,

what else we're going to do? We're going

to do turn this company that, you know,

Ben Padian has been talking about called

Bloom Energy. That stock has done uh,

tremendously, tremendously well. a

little bit of pullback and I actually

don't know that that much about it. But

why it seems to me like semiconductors

are so much more difficult to do than

this this energy thing because is it you

know is it really that hard to turn

natural gas into

um into energy?

>> I mean there's even the there's this guy

Boom Aviation who's making a supersonic

airplane. instead of work focusing on

the airplane, he's converting his whole

manufacturing plant to making turbines

because he thinks he can make turbines.

Like everyone's running into this type

of uh place. Uh but I do think that

sustainable efficient energy generation

is a net benefit for society and for

these data centers and for the future.

Obviously, you have guys like Elon

saying if we covered was it a part of

Nevada with solar panels like it would

be enough to power all the United

States. It's possible. I think there's a

combination of natural gas, uh, solar

and batteries that I think will

definitely be powering the future with

some nuclear, not SMRs, but just

probably just traditional nuclear

reactors, which is a clean energy source

because our electricity demand is just

only going to go up. It's not going to

go down. If we're going to add more

electric cars, more electricity, more

electronics, more electricity, more data

centers, more GPUs, more electricity.

So, like, we're not going backwards

anymore. We're only going forwards.

>> So, Ben, the question that I asked of if

the shortage is in the powered land, not

the chips, why invest in the chips?

Shouldn't you invest in the in the

powered land? You know, that's what uh

famous now hedge fun manager Leopold

Ashen Brener has has done with

situational awareness. I don't think

there are, you know, there are not as

many chip stocks nearly as there are

powered land stocks like iron or Bloom

Energy. Do I I asked you for a push back

to that. I don't it sounds like you're

not giving that much of a push back. So

maybe is is that sounds like you think

that that actually is wise.

>> You have to kind of really look at the

assets of each powered land player and

at the end of the day powered land is a

quintessential real estate play. You

need the powered land needs a quality

tenant and that quality tenant needs to

be paying a good price. So if you can

line all those things up together and

you think all those things will work out

then it'll it will be a great

investment. But at the end of the day

it's a simple landlord tenant real

estate play which could be uh lower on

on the risk curve, right? But you need

but the question is there's a

bifurcation in in how

powered land players are playing. I

think I went to the Nebius event.

They're buying different software

companies. They're trying to build a

full stack different solutions. They're

trying to be like the hyperscaler versus

other guys who were like doing Bitcoin

before. The Bitcoin market is dead.

Who's doing Bitcoin?

>> Yeah. The IR name Ben. I just struggle

when you use a term like less risky. I

mean, some of these companies are

borrowing billions of dollars to build

out physical infrastructure as opposed

to Nvidia, which is actually going to be

borrowing like $20 billion for for

liquidity purposes, but is just printing

so much cash or even companies like ASML

or Lamb Research that optically are a

very high valuation, but are just so

critical and their their uh their market

position is so dominant, just

igopolistic, or in the case of the EUV

for ASML, literally a a monopoly that I

I think that those sound a little bit

less risky than borrowing $20 billion to

build data centers in in Texas.

>> There is a lot of risk in that. And I

think the Do you remember that scene

from the movie Batman where Bane is in

the plane and then he looks at the guy

he's like one of us has to be die in the

wreckage brother.

>> So figure there's a bunch of neoclouds

on that airplane. A few of them are

going to die in the wreckage or get

absorbed. I don't know who they are, but

the whole idea of of a NeoCloud and

obviously a high leverage. I think what

Fermy is one of them. There's a crazy

IPO and the CEO did all these weird

things and they have power land. It's a

ground lease. Like I got pitched it a

few times. Like it was just like way

over my head. Like I think from an

investing standpoint, KISS, keep it

simple stupid sometimes is easier. But

if you really really want to like you

have an itch for for powered land or

data centers maybe buy utility maybe

maybe buy Blackstone I don't know

>> black okay no one is really talking

about that Blackstone's position in the

data center market is very large and

that is there okay what do you think

about the

um the the powered land names like Texas

Pacific land or landbridgeidge which

owns land out in West Texas where data

centers may or may not be developed and

in the case of TPL uh a former executive

of of Google is trying quite hard to to

build a data center there. Do what do

you think of that?

>> I mean those things just like like I

said you're it's a quintessential real

estate development play high risk high

reward real estate in itself as an asset

is filled of debt. I mean it's a

leveraged asset. So, if you want to put

leverage on leverage and hopefully get

that awesome tenant and win and make it

a home run, then go in that direction.

But for me, sometimes it's a little bit

too uh hard to to quantify because I

don't have I don't know what's going on

in that market specifically who's there.

I'm not looking at every county report.

There's a lot of guys analyzing what

things are filed. I mean, I even had

this guy put together a bunch of land in

Northern California. He's like, "This

this this land will be amazing for a

data center." But the funny thing is,

I'm like, "How are you powering it?"

He's like, "Oh, we're powering up a

bunch of Bloom Energy." So, I was like,

"Oh, that's an interesting data point."

So, I think uh that to me was like,

"Okay, another data point for Bloom

Energy." So, like people are using it in

these these different places like I

rather just invest in that a little

easier for me.

>> That makes it. So, you've been an

investor in Bloom Energy. Yeah, it's it

to me it's obviously misunderstood. 20

years, no product market fit and all of

a sudden it became like the killer app

because it's quiet energy for the data

center and it ships fast, right? Time to

power is everything. So if you're a data

center developer and you're borrowing at

7 8% you want to wait for this power

thing to come or do you want to wait for

the utility to finally hook you up when

you have the GPUs and everything ready

to go or you going to be like you know

what Oracle is doing like give me the

bloom let's be online let's energize

this project let's start charging the

customer and let's move on. So that's

that's the way I look at it. Yes, for

for Texas Pacific land and landbridge,

they probably are not going to do the

development. People would be building it

on their land. So less debt, less capex

that they would be doing. And for full

disclosure, I have been an investor in

TPL. I actually sold my entire position

as I record this, but that may or may

change. I may reenter for for for

various reasons. Just want to have no

position. So

>> yes. Yes. Ben Ben, have you heard of I

don't own this series Energy or S series

C. They're trying to compete with Bloom

Energy. Uh, no I have not. I think the

one that's kind of run up recently is

Fuel Cell or FEL. Yes, that's like an

older company which basically it's like

the same type of technology. Fuel cells

I I guess are getting business. I

haven't like looked.

>> So, so in the chip world, Ben, how are

you allocating your capital to invest in

semiconductors and share your philosophy

behind that?

>> All chips aren't created equal. I think

people need to really understand that.

There's obviously memory chips, there's

FPGAAS, there's AS6, there's even like

diodes. Remember, I started a commercial

LED lighting company from 2005 to 2019.

An LED diode, a light is like the

dumbest type of LED you could use. It's

just off or on. It's light, right? And

you can go all the way up to the the

harder stuff which is obviously like a

GPU which is processing massive

computational loads and is creating

intelligence. So you got to pick where

you want to go but I think obviously you

have players like Nvidia then you have

their MI series. I think they have an AI

event in a few weeks in San Francisco.

They want a piece of the market. Intel

is trying again with investments in

Sambanova and other companies, but you

have a whole host of startups like

Etched or Posatron

or Unconventional AI. There's like

everyone's like everyone sees like a $5

trillion market cap and it's like every

entrepreneur's dream or investor's dream

is like let's get a crazy team together.

We're all smart. we're ex Google or ex

Nvidia, whatever it is, and let's just

build a better chip than they can. And

if we can just get 1% market share,

we'll be successful. And then one of

these other companies will just buy us

because we'll be a nuisance. But can

they really scale the supply chain? Do

they have a hundred billion dollars to

buy all this memory and fly to Taiwan

and move with all these suppliers and

cables and racks and like like like have

President Trump on speed dial and like

it's just very multifaceted. It's a it's

a rich man's game and it's hard to pick

smaller companies to think that they can

really grow in but there's different

parts of the stack that win. I think the

one of the most interesting parts for me

is obviously optics has been interesting

with communication and lasers and moving

the data faster between GPUs. And then

there's been this big push in uh it's

called CXL which is memory pooling. The

idea is that because memory prices are

so high instead of buying more memory

you kind of pull your memory in one

place and you share it and you process

it at a central location. So companies

like Astero Labs was like Left for Dead

and then all of a sudden shot up 400%

and the same thing with Credo and Marll

which focus on that space as well. So

but these are all like inflections and

moments in time. uh you know holding

Nvidia for 10 years coming up uh in

November you know I've been through a

lot of draw downs the big crypto draw

down in 2018 and then the 2020 boom in

the 2022 I remember every day was going

down these things aren't easy to hold

everyone thinks it's like glorious but

you know you just have to pay attention

to the company the long-term road map

management what are they saying are they

being honest started doing what they're

saying and just look at how things are

executing. Obviously, the next when I

was at GTC in March, all this stuff is

being built for this huge token

explosion and I think the next big

explosion in tokens is physical AI and

edge AI which is basically robotics

doing agentic work, right? So like you

can think of robotics or robots as

like when you're running claude code or

cloud the CPU is telling these agents to

do certain things orchestrating. So what

if the CPU is orchestrating physical

agents which will be robots? That's the

way that's the way it's headed. And

you're telling a robot a sequence of

things to do on the factory floor or a

sequence to do in the lab makes these

chemicals, right? It's sequencing in

parallel with tokens and a loop. I mean

that's basically what what it is. So

Ben, if I you your number one choice for

a stock or company in the AI supply

chain is what?

>> It's Nvidia. I mean

>> I knew you'd say that. Okay. What is

what is number two?

>> I mean I like Apple.

>> Really interesting. Tell me about that.

>> Apple is always going to be the company

that's going to bring useful AI to the

world, right? In a way where it's

hopefully humanistic and relevant to

your everyday life, right? I think AI is

not going to be some sort of product.

It's just going to be an upgrade to

whatever we have. And you can already

see uh I mean the iPhone is already

using AI. You can already sort your

album based on faces that you trained it

with that it recognizes and you can

create videos on I was with my friends

in this country. That's that's like

simple on the device AI. You just have

to

>> Ben I um actually I'll give a shout out

to the compound. I was listen you know

Josh Brown and Michael Batnik. was

listening to their podcast I think

yesterday about just how much money

Apple is going to make in services from

okay I'm a user and I'm using Claude on

my phone and I'm paying Claude and

paying anthropic Apple is going to get

20% of that 15 20% of that in the same

way they do that through you know Apple

has a hundred billion dollar it's called

services but really it's just taking a

cut of a lot of it is taking a cut of

you know money that of the apps that

people buy on the the app store and I

think that is going to be a big market

so even if Apple isn't an innovator AI,

which it very well could be. I think

that is going to be a driver for the

stock. That being said, Ben, I think

Apple Intelligence is has been very

underwhelming to me in my personal

experience as a consumer.

>> Yeah, it sucks.

I thought they had they have all the

data and they and they want to keep it

private but for some reason something

hasn't clicked which is weird but

hopefully with Google and whatever doing

with Gemini or even anthropic or people

can use their own models we you can get

some ondevice AI that's like relevant to

what you're doing in your life and your

social context and it's private to you.

I think that's that's the most

important. I don't people need to

realize like even with meta my people I

was doing the photos yesterday with the

the Instagram app like the the thing

like we're basically voluntarily

training their models of of our own

data. Does that matter to you or not? I

mean from a privacy perspective some

people really care. I mean but you have

to understand Apple's ethos from the

beginning is your data belongs to you.

It won't train on your data. And I think

that over time gets important when you

have photos and text messages and things

like that that should all belong to you.

It shouldn't belong to any frontier

model lab,

>> right? So you like Apple as a is with

this AI exposure. Technically though it

is not you know in the semiconductor

supply chain even though it it does

design it its own chips and stuff. So if

strictly sticking to like stocks that

would be in the you know VNX

semiconductor ETF or the the ICE semi

index or Philly whatever index like pure

semi stocks what is your number two?

>> Oh that's a good one. uh

>> why

>> I still think even though

we're in this crazy

memory is sort of constrained and the

only US company that really does memory

is Micron but we're I think highinix is

going to list in the United States. So I

I still see three companies dominating

and memory is obviously a very

interesting name but it just trades very

weirdly. So, but I think you need to

have a small allocation to memory. What

do you make of the semicap equipment

companies like ASML, KLA, Lamb Research,

Tokyo Electron that applied materials

that actually need are totally necessary

for memory to buy and to to expand and

if Samsung and SKH highix and Micron are

going to expand their capacity, which

they're you know trying to do, they need

to buy so many machines so many machines

and there's recurring revenue that Lamb

researched like 30% of their revenue is

recurring revenue. venue because they

need to sell this they need to replace

the spare parts which obviously you know

they're investors who in in like

AutoZone or something they they get that

but they should know that this is also

true of of lamb research and also to to

maintain the machine software revenue

and the like so you know if lamb

research revenue is going to scale

massively which I think it very likely

is perhaps higher than expectations then

like it's going to get revenue on that

continuously like in the many you know

many many years into the 2030s even if

this is a giant bubble that you know

that that that does burst you know, in

this year, this year or next. That being

said, they are they do trade at by far

the highest valuations in the semi-. So,

how do you how do you think of these the

semicap names? They're like literally

five.

>> I mean, the the the two cannot be the

same. I don't think Nvidia can be at

like such a low multiple and then

semicap which has always been the low

multiple guy be at such a high multiple.

So, there must be some sort of

mean reversion at some point. Uh things

have obviously run up a lot. Uh like I

said because there's been maybe two

decade two decades of a lack of

investment in any of these companies

like no one like no one cared if you

were about any of these companies for

the last 20 years. It was all about

Salesforce and Adobe and all this like

Twilio or whatever like that like that's

that was was sexy and like Splunk right?

>> Yeah. Yeah, the companies that their

names were in some instances were just

ridiculous. Like you should not be

naming a company that

>> didn't matter because they were capital

efficient, right? They were printing

money. But like now like applied

materials or ASML like like they need to

spend millions of dollars building out

more fab uh space and construction.

Yeah, applied materials is like super

busy. But obviously uh I feel like there

is some sort of

quasi cyclicality when it comes to these

types of names and people need to be

cognizant of that. There will be at some

point some sort of stasis and overbuild.

Trees don't grow to the sky and I think

maybe this generation of younger

investors have forgotten that. And

remember what you're investing in and

why and what your long-term thesis is. I

think that's important.

>> Yes. I I think we can safely say that

the 30 to 100% growth or over 100%

growth in some like micron like that

literally cannot continue for the next

50 years. Like I'd say the chance of

that happening is zero. Um of course

people have predicted that the the

semicycle would turn for the past two

years and and it hasn't happened yet.

What are you going to be looking for

that when you see A, you see B, you see

C, you say, "hm, Ben, this is kind of

looking like the cycle is turning and

maybe I am either going to sell my semi

stocks or definitely not buy more and

just just hold out and go into

protection mode like a little porcupine

because is this is this is the cycle is

turning. What are you going to be what

is ABCD those?"

>> Yeah, BAP research for my institutional

subscribers. If I'm tracking GPU

capacity, GPU rental rates, uh,

gigawatts of construction planned,

obviously looking at Taiwan, what are

what are the contract manufacturers

building, what are what is their growth?

Also looking at obviously the

Bellweather Nvidia, what is their road

map, how is that tracking? And then

looking at the uh, Frontier Labs,

Anthropic, OpenAI, what are they

building? How are they growing? How much

more computer they taking? Are they

growing headcount? And you kind of look

at all those different things and you

try your best to triangulate where you

are in this cycle. I don't know. But,

you know, based on my research and

industry conversations and speaking with

customers, feels like AI hasn't

penetrated as much as everyone thinks it

has. And I I just saw Anthropic just

leased 160,000 square feet on Hudson

Street in New York for their new New

York headquarters. You know, they're

going to put a thousand employees there.

Like, so San Francisco, New York, where

you are, are going to be the two big AI

hubs. They'll be there. So, you get to

say hi to them.

>> Yes, I do. And the AI has not penetrated

as much as people think. Is that bearish

or bullish? Just to be clear,

>> I think it's bullish. Uh, like I said,

Chad GPT was November 2022.

We're hopefully going four years in this

November. I mean, and this isn't like

the iPhone where you can scale

production and just ship a bunch of

stuff to to Apple stores and people just

pick them up. Uh, this is more

intricate. There's more parties

involved. There's fewer bigger pocket

buyers. Uh, and there's planning and

road maps involved. So I think the race

for comput is still there and like I

said classical computing died and what

Jensen is just saying it's like the era

of HPC and there's still time to really

grow and go in that direction until God

knows when you can do data centers in

space which is something for another

time. Yes, that that that is a a really

key point that the semiconductor supply

chain is so complicated and I'm not even

saying that this is the hardest thing,

but just to give people a sense, there's

there's a company like atomic layer

deposition, ASM, they used to own ASML,

I think, and they literally are putting

layers onto the chip that are one atom

thick or less than one atom thick. So

like I think that it's really really

really hard to do and there are certain

maybe false bare signals that people

could see of like oh shipments have

flattened that actually are a sign of a

constraint or a shortage rather than a

lack of demand.

>> Yeah. I mean there's there's a lot of

that going on. I I tried not to there's

a lot of noise. Just just look for

signal. Look for the productivity. Look

for the demand. Look how look how you're

using these these AI models how you're

getting or it's not allowing you to use

it. I think those are better signals

where we are until more capacity comes

online and you see more of this AI

actually diffuse into your everyday

life. I think that's where a lot of

people are missing. A lot of a lot of X

is concentrated in tech and obviously in

the Bay Area. So over there everybody is

doing AI.

>> Yes. in Los Angeles, nothing. Right? New

York, maybe a little bit, right? Is the

bodega down the street using AI?

Probably not. So, there's a way to

really there's a way there's a ways to

go to really see it permeate in the rest

of society. And I think it will be a net

benefit if you if used in the right way

for the right purposes. What are you

following in terms of the frontier labs

and their monetization in terms of

anthropic in terms of open AI? What are

you seeing in terms of their revenue and

just how fast is that growth relative to

that revenue and how does that compare

to how their costs scale as well?

>> The the beauty of this business

obviously it's very capital intensive.

Almost think about it as it's almost

like a bakery type operation. You build

this huge bakery plant which is like the

data center and your inputs are flour

and water or in this case GPUs and

electricity and then your output the

tokens is the bread right so uh

>> but the flour and water cost a lot of

money and the biggest expense for

building a data center is GPUs the

biggest expense for running a data

center from a gap perspective is GPU

depreciation

>> I think it's GPU are actually lasting

longer than people expected. I think you

have H100s and A100s that were like over

3 years old, still running, and they're

still useful. Not so much for training,

but you can use them for inference. So,

useful life is uh is still very very

there. But the thing is because there's

so much money going into these training

runs, you make your money off the

inference and the gross margin on

inference with the API pricing seems

like it's trending towards, you know,

the high 70 maybe even 80% gross margin.

So this is like the new basically

software paradigm. The new SAS is

inference, right? The question is is if

you're a software company, are you

buying tokens wholesale and adding your

intelligence and selling retail? Do you

have enough margin to survive against

these frontier model companies that are

coming after your business who own the

whole stack who control their whole

their whole margin profile? I mean

that's why you've seen this huge

derating in software like software was

unassalable.

>> Yes.

>> Until now. I wonder why. But Ben, the

core point that software had very high

gross margins that you would spend a lot

of capital, hire a lot of expensive

software engineers to make pay software

um sellers a lot of money to make to to

sell the software, distribute the

software, but that once it was running,

it was very very high gross profit

margins. AI does delivering AI does have

lower margins.

>> Delivering AI does have lower margins. I

mean, not from what I've been reading.

The profitability on inference is there.

I think again going back to the bread or

oil analogy, I think if a company like

Anthropic is made truffle infused gold

flaked bread at a low cost and is

selling pieces, thousands of pieces at a

at a high price, then they're really

making good money and they're probably

heading towards cash flow positive at

that. I mean, the guys, he said in

February they they 8x their revenue

plan. I mean, that's crazy. It's like

the fastest growing company in the world

at at this point.

>> So, yeah, I just I just pulled up

Salesforce, a classic SAS business.

Yeah. Gross margins of like 72 to 77%.

You're So, you're saying that you've

read or heard that anthropics gross

margins for inference are are what you I

just want to be clear.

>> I think in the 70s or higher.

>> Yeah, that that um that is pretty pretty

high. What about OpenAI?

>> I haven't seen their numbers. I mean,

from a user perspective, I do use some

chat GPT and things like that, but for

deep commercial work, it's mostly Claude

and Fable and all this other stuff. I

think they've really won on the

enterprise, but I don't know what Open

AI is trying to do. Maybe 5.6

is it? I know they have way more compute

than than Anthropic, that's for sure. Do

you think that companies are happy with

the money that they're spending? I know

you watched Alex Karp on CNBC

just absolutely trashing um

I won't say trashing. He he was very

stern that the executives in corporate

America are displeased with the cost of

AI, the the value that it provides in

terms of outputs and also the the data

lake that like anthropic could be like

taking the data or the companies aren't

allowed to own their own data. Set

setting that aside and and then you know

the CNBC people were saying well what

are what are you saying and and and Alex

Karp said I am a vessel I am speaking

for corporate America. It was honestly

pretty pretty entertaining TV, but that

Yeah. What do you What do you make of

that argument that the companies are

like, "Whoa, I'm spending a billion

dollars. This is ridiculous."

>> Yeah, Alex is uh I love him. He's like

my spirit animals. I think he just says

whatever's on his mind and he doesn't

care. And I think that's the way people

should be. And I that's really how

things are being done. Even at Nvidia, I

posted a few times when I was there at

Marchant GTC, a VP of AI Abdullah

Hollik's like we use uh Opus to

basically orchestrate and we use Neotron

which is our own open- source model and

we can achieve frontier level results by

combining these two and we own our own

compute. So all those the grunt work of

of tokens is owned by us and we don't

pay anybody for it. So you so the idea

of using frontier models for everything

is that's like when you blow your token

allowance and you spend too much money

on wasted projects and or like what am

someone on Amazon spent like a million

dollars or something on tokens. It just

it it doesn't work. And I think that's

where enterprises are heading with this

like router approach where if it's

really hard, use the frontier. If it's

not, use a local model and have the the

frontier orchestrated. Does that make

sense?

>> Yeah, that that does make sense. And so

short term that may may lead to less

demand for frontier tokens, but longer

term you think it actually is is

bullish.

>> Yeah. Cuz at the end of the day, the

frontier is always going to be like

pushing the limits and it's sort of

maybe like the luxury

car. Maybe it's like the Ferrari of the

market. There's always going to be

buyers for that who want the speed and

they want to be the best, but maybe some

people don't want that or maybe some

people want to mix. I mean, that's what

I'm saying. Like the the AI market is

sort of maturing. Not one not one

sizefits-all. There's different ways to

use like for say you're like a CPG

business and you want to have like a

chatbot about like your your consumer

products. Just load all the data and

then train it with like a like a open-

source model with the weights and

everything and that's it. Like talk

about the product. It's not going to

like solve math or find new chemicals or

whatever for you. It's just going to be

talking about hey yeah the frosted

flakes has this much sugar in it. There

you go. Whatever. Like like the

different parts of your stack are going

to use different types of models. It's

overkill to use something like a

chainsaw for everything when you just

need like a a butter knife,

>> right? Yeah. You don't need Albert

Einstein for you doing data entry or

honestly even sending emails that aren't

that important. Ben, what do you make of

opensource? This is the another bare

thesis. Okay. Yes, AI is transformative,

but open-source models that charge so

much less, so so so much less are just

going to sap the pricing power of the

American Frontier models. What's your

reaction to that? Number one, and where

does Nvidia's Neatron come into this?

>> Yeah, I think you need both. It's just

not, like I said, again, it's not one or

the other. It's a combination of the

two. And what I've told people is like

with Nvidia you get like a free open AR

anthropic like Neotron is just as good.

Like they have like you should look at

the team of researchers like the guy's

like oh I just got hired from Meta Super

Intelligence because I'm working on

Neatron to solve to solve science and

hard math problems and enic workflows. I

was like whoa like these are the people

that Nvidia is hiring. Part of doing all

these things is to keep your customers

honest, right? Like what if a company

like Anthropic are so smart they're like

oh you know what we'll just make our own

Nvidia we'll make our own chips we'll

make our own models like we're just

going to outdesign you on everything. So

like if that happens like what what leg

does Nvidia have to stand on? they need

to have their own model, right? So, it's

a this is also a part business. It's

it's a business decision also to make

sure your your customers don't put you

out of business, right? And I think a

lot of the world wants open source. I

think like the parallel again going back

to Apple uh Frontier models are sort of

like iOS, right? And it's sort of like

closed and the the the open source is

like Android. It can be anywhere and you

can use it on anything. But the the cool

thing is is that Neotron in itself is

optimized on CUDA and CUDA with the

software layer is optimized for Nvidia

hardware. So like you have basically

developers all around the world on an

open source contributing figure out ways

for free to make this product better for

you. You don't have to do anything. You

can wake up like oh look the community

came up with this and look we we made

this a little bit better. Oh that's

awesome. like I get free software

updates. Oh, it's great because I bought

this and I have this. So that's like the

benefit of open source and and the

silicon that it comes with.

>> So that's good to know about Nvidia has

this ecosystem. I I didn't know this Neo

Neotron that is really good to know.

People should look into that and you

know if it's not obvious you're giant

Nvidia bull which has obviously worked

out well. But Ben just go back to open

source. I'm let's say there's a company

five people and they are going to do AI

agentic workflows for coding that really

are going to transform their business

and provide a million dollars worth of

value. The problem is that they're that

you know anthropic or openi is going to

charge them like four or 500k to do

that. So it's still a good deal but it's

really expensive. Then comes in Quen,

comes in Deepseek, comes in Jeep, all

these Chinese models that instead of

charging $500,000 or four $500,000 are

going to charge $6,000.

How is that number one are my priors

wrong? Like is that just not accurate?

Or number two, how is that not wildly

bearish for the Western AI things that

are closed source?

>> But where is your compute? Is it are you

are you do you have your own compute and

you're running your models locally or

you're running these models hosted on

some something else and trying to get it

to work?

>> And the question is is will China block

open source models out of their country?

And then if you're building on a model

that's like blocked and you're halfway

through building in your business then

what happens? What kind of business risk

are you introducing to your company? So

the question is is like how do you build

your business in a way where you're the

CIO or the CTO and you're not so model

dependent on one company where tomorrow

you're building on some sort of volcano

and they erupt pull your access or

increase your pricing. What do you have

to fall back on by then? They have you

by the balls and then you're

>> Okay. So, so people are going to pay a

ton of money for Western models rather

than Chinese models because they're

worried about the Chinese government and

China taking their data. That's what

you're saying

>> that or at the same time you just don't

know where that's going to head. I think

with with anything with obvious China I

mean like their playbook has always been

let's deflate the cost out of anything

that we make and just try to export it

around the world. And what I've coined

in other segments, it's called inference

through influ influence through

inference. So the idea is like what

China's been doing with like companies

in south countries in Southeast Asia

like Frontier Africa is like like we

don't have any running water. We don't

have an airport. We have no

infrastructure. Don't worry, China will

come in. We'll build it for you and

we'll we'll lend you the money and you

know all these natural resources that

you get out of the ground and these rare

earths, we'll take it. Thanks. So, I

think the next thing China is going to

do is like, you know what? Oh, you're in

Zimbabwe. You want AI? No problem. We'll

build you a Huawei data center and we'll

give you the models to run on it. And

it's the uh GLM or DeepSeek. And by the

way, if you look up Tenement Square, it

never happened. And they basically they

basically influence this is like a this

is like a global policy perspective from

America and democracy, right? How the

world views America through the lens of

China in these models. So you're

influencing through inference and I

think that's another thing that America

can afford to lose on the race for AI of

the world. So your argument that China

is not going to displace the pricing

power of these western models is based

on geopolitics

rather than tech. I am not saying that

you're wrong and that totally like the

reason that people all around the world

in Europe use Microsoft as opposed to

like you know some technology developed

in Singapore is is precisely for that

reason or just that we have kind of

network effects in the American tech

stack. But don't you think with the vast

sums that are being spent and will be

spent by companies on AI that you know

if it really is so transformative

wouldn't you want to save $2 billion a

year like if you are I don't know uh

Coca-Cola or something and you know in

many years and you're spending let's say

$50 million a year as Coca-Cola wouldn't

you want to cut that to $15 million are

you heavily heavily incentivized to do

that

>> sometimes there's career risk with

putting all your eggs in one basket and

I do see somewhat of a contingent hybrid

approach when it comes to model

selection by CIOS and CEOs of companies.

You can't go all in on one. You have to

have like a a few that you're basically

routing around and then you'll have like

a blended lower cost hopefully.

>> Ben, what do you think is the most

overrated stock or seg segment of the

supply chain?

>> Overrated.

uh

>> like like like I don't I suspect that

you're not short any semiconductor

stocks, but if you were like running a

long short company hedge fund and you

had to be short, which ones would you be

and why? I think there's a lot of these

companies that are doing like optical

materials like AXTI or indium phosphate

like like again like it's just you're

just going after like a raw material

that's supposedly like very high in

demand that will have like boom and bust

and things like that or it's just to me

it's like I like to be in places that

have some sort of defensible moat as

good management uh that will last over

time and Those types of names to me are

just like very esoteric. And there

there's obviously there's some cartoon

characters on X who shall not be named

that have large followings who tell

people to to go into these names which

to their own risk uh they can do all

that stuff.

>> Yes. You know I understand that there

are some commodities that actually are

very rare like the the photo resist and

they're very hard to make and like two

to four Japanese companies make them.

But yeah, I mean like it that makes

sense to me that why if if if the

commodity is literally in the um you

know in the periodic table that sounds

like something that shouldn't have

pricing power in the long term.

>> Correct. I mean the whole idea is like

the idea of modes and how hard is it? I

think that's where uh you know I live

through it with LED lighting right? It

was like a commodity. It's like a diode

and we had competitors from China and

Korea and like everyone was doing it and

Amazon. So it's like if you're getting

squeezed from all different directions

like that's not a good feeling to be in

because everyone's chasing the same

market. You want to be in something

that's very hard and has definitive

layers against it. And I think when you

look at from a macro perspective, like I

said, all hardware eventually

commoditizes, right? The last scalable

consumer or electronic device that

hasn't commoditized is what you're

talking to right now, your MacBook, your

iPhone, right? How come it's just a

computer, right? It's just like a phone.

Like, why hasn't it why is it $1,000?

How are they getting these? Like,

because the software layer matters. The

code design matters, right? And I would

say brand which I think in in the

semiconductor world brand kind of from a

consumer perspect doesn't really matter

like no one is like oh I'm going to buy

this phone because it has this

particular type of NPU.

>> It's not brand but you the thing is you

know an iPhone 9 has an operation of

49s. It's going to work 99.99%

of the time up time versus like these

like weird phones that were coming out.

you're like, I you had to reset it a

bunch of times and it was like a iPhone

wannabe. Like, you know, this phone

works, dude. So, like that's like you're

paying for reliability. The same thing

with Nvidia. You're gonna get four 9s

and if something goes wrong, someone's

there to like fix it. You're paying for

that. There's you're paying for that

certainty. But if you want to take the

risk and make your own chip or buy from

AMD or one of these startups, go for it.

Be, you know, be a guinea pig. See where

that takes you. I mean, you'll save

you'll save money on the front end, but

you're going to pay on the back end.

I've I lived through this many times,

you know, buying the cheapest thing and

on the back end I I pay for it. So, I

like I'm done. I rather just, hey, you

want to make this margin, okay, I'll pay

for it and then just move on. Ben,

obviously I'm not going to ask you to

give, you know, all of your alpha away

for free, but could you give us one name

or sector that you think is

underappreciated in the semiconductor

supply chain right now, as was perhaps

like, you know, you wrote about Bloom

Energy in in the winter of of last year.

>> Yeah, I mean, I wrote about this and

people are going to give me a lot of

crap. This company called Super Micro, I

think it was like a Catrini favorite. I

think

>> Shout out. Yeah.

>> Yeah. went up to like 600 bucks and then

like they managed to burn through all

their cash and then smuggle like GPUs to

China. I don't know why the guy was

doing that, but like it it's a real

company in San Jose. Like they have

people working building these server

racks and like like they're Nvidia needs

them to to survive. AMD needs them to

survive because they make a good liquid

cooled product. They can make AI racks

and hardware, right? Dell. Obviously, I

love Michael Dell. Like, he's like the

American He's like the American dream

guy out of college, out of his dorm

room. I I love him, right? And you have

companies like HP doing the same thing.

I think Super Micro just had maybe some

bad family members that were doing wrong

things, but I think from a valuation

standpoint and the IP that they have and

obviously they just raised about $7

billion of diluted funding recently. I

mean JP Morgan led it. If people are

putting $7 billion into this company and

it's probably we'll figure out what

institutions did then there's an order

book behind it and if SpaceX just raised

$85 billion they're going to be adding

more capacity to Colossus in Memphis and

Super Micro is going to get that call. I

don't know what their gross margin is

going to be. is going to be low, but

hopefully they can squeeze a few points

out of it and really rebuild this

business of what was and get back to

their glory days and probably not $600 a

share, but you know, something

reasonable. And I think that from a

fallen angel governance perspective uh

needs to be, you know, looked at again.

>> That is interesting. I remember when

when uh talks about it on my show and it

had an accounting scandal then and then

it you had huge demand from AI. It's

went up a tremendous amount and it's c

capital expensive. There was some bad

thing that people should look into. Um

but yeah. Okay, that's interesting. Ben,

one question I wanted to ask you. What

do you make of the bare case for the

software in the semiconductor supply

chain? Not talking about Nvidia, I

guess, although I kind of am maybe, but

I'm in particular the two companies are

Cadence and that they to explain for our

audience. They designed the software

that the engineers are going to use when

they make the and design the chips and

this has been the ultimately like high

power software and like a lot of

software super high retention and then

they they get you know net dollar

expansion upsells yada yada yada so

great business but I mean if AI really

is so transformational then maybe the AI

companies are going to design their own

software and save money on the uh you

know hundreds of millions of dollars

that it probably costs to to license

this software. So, you know, this is

probably the the stocks that are up the

least this year, Cadence and Synopsis.

What's your outlook on this on these uh

on these two names and the overall the

bare case on semisoftware that like you

know this AI AI can make software if it

cost software is zero?

>> Yeah, I mean that's I think when open AI

released the jalapeno chip they said