welcome back to investing wizards we

have gavin baker

founding of founder of atreides

management and because

gavin works at a hedge fund i need to

read a disclosure first

so all opinions expressed by gavin baker

in this podcast

are solely his opinions and do not

necessarily reflect the opinion of a

tradies management lp

this podcast is for informational

purposes only

and should not be relied upon as a basis

for investment decisions

clients of atreides may maintain

positions in the securities discussed in

this podcast

atreides has invested in koifen as a

seed investor

all right now that we have that

disclosure out of the way

i wanted to introduce gavin gavin thanks

for joining us

awesome to be here rob uh great to have

you

so uh we we met on the internet uh

about a year ago and uh as i mentioned

you were

a seated investor in coiffin and and

very happy to have you

on board on our journey but today we're

gonna focus on

uh sort of public markets investing in

your background um so can you tell

viewers about your background

um and what school you went to and how

you got started in investing

sure well first i feel like i should say

um personally i do love coiffin

and that's um how we came to know each

other and

um i guess the tradies came to be an

investor in koi fan but in

um i had a year where i did not work as

a professional investor

and i extensively investigated all of

kind of the

free um investing tools available

on the internet and i thought that koi

finn

was by far um the best

um and i was happy to see you uh when

that twitter poll

uh run by uh patrick o'shaughnessy um

please please please go on yes that's

where you guys decisively

won but it really is yeah it really is

um it really is awesome and i do go to

uh even now having a you know

a lot of resources i do find myself

going to koi fin a lot for its uh

simplicity

ease of use elegance but um but it's

your question i grew up in houston texas

um i had zero interest in the stock

market growing up

uh but my entire life from some of my

very

uh earliest memories um

are being interested in history so i

loved um before i could barely read to

love uh

read almost you know you know picture

books about um

the greeks the romans the egyptians

ancient civilizations i

vividly remember in a second grade my

dad would

drive me to uh driving to school every

day and drop me off

and i'm sure you know my dad also loves

history

but but i'm sure he'd been studying

every night at the time i was kind of in

awe of his knowledge

but um every day for the entire school

year

we did the history of world war ii kind

of in a

chronological order um you know in

little 15-minute slices

and i loved it and um

you know as i grew up i read more and

more history and it was one of my

favorite things to do

you know when i was when i was 12 years

old there were these great time life

series on like the

the battle of britain the air war um you

know world war ii

um you know all sorts of really big

illustrated 20 book series on on major

historical events which i love

and then

[Music]

beginning freshman year of high school

i also um kind of developed an intense

interest in current events

so as a very um you know as a teenager

as a 14 year old

i started reading the houston chronicle

every day the new york times every day

i read every issue of the economist u.s

news time

newsweek and um

and then when i got to college i kind of

still had those interests

and um i um

i began to get really interested in

games of skill and chance like uh

poker and chess uh but in college the

most important thing in my life

by far um work outdoor sports like rock

climbing and skiing i love to rock climb

i love to ski um

my kind of stated intention was to um

i'd worked i'd worked in the ski

industry for one winter during college

i was a housekeeper at the gold miners

daughter in alta utah i cleaned a lot of

toilets

and my stated plan was i was going to

after graduating from college buy a

really old pickup truck with a

cap on the back put a bed in the back of

it

and kind of travel around the american

west i was going to work in the ski

industry in the winter

i was going to work on a river in the

summers and i was going to climb

full-time in the shoulder seasons

and um you know try and support myself

also in the shoulders even

seasons not only on the savings from

those from those jobs but on um

but by trying to you know to write write

articles maybe take some wildlife

photography um and you know try and

write uh

the next great american novel um i'm

certain i would have failed um writing

the next great american novel at least

um and i was an english and history

major and really no interest in the

stock market really no interest

in um investing economics anything

and uh my parents i was very lucky they

they paid for me to go to college

and they said hey gavin we really

support this plan

we think it's a great plan um and they

were kind of probably who i'd inherited

a love of the outdoors for because we

did a lot of hiking trips growing up

but um they said you know we really

support this but we think we would

really appreciate it if you just do one

professional internship but i thought

wow that's super reasonable

of course i'll do that the only church i

could get was in the stock market

it was working uh for donaldson lufkin

and ginret

um actually here in boston where i'm

based to this day um

and i would say i have a very um

obsessive addictive personality

like if i get interested in something i

can just go really really deep for a

long time

and that's what happened to me with the

stock market after two days i was

done it combined everything i'd ever

been interested in

it was this giant game where your

ability to win

was based upon having the um

the best the most extensive possible

knowledge of history

because while history does not repeat it

does rhyme intersecting

that with the most um

up-to-date uh understanding of the

present state of the world and future

events

and intersecting those two um

to form a differential opinion about the

future different from what was

discounted in the stock market because

the stock market is the world's biggest

period mutual system

where you know implied odds of different

outcomes are always changing and if you

could

by intersecting the knowledge of history

and current events

um get an edge it was kind of like

competing in the world's most

meritocratic game that was it was a

combination of poker

and chess where history and current

events drove your edge

and i was super addicted i would say i

read a book on investing probably every

two days on average

for those 90 days i lived in the

basement of a woman's house in

in cambridge um i mean i i never saw her

as

you know she was renting out a room um

and just it was intensive study probably

you know

six to eight hours a day i taught myself

accounting

um you know i read all warren buffett's

letters to his shareholders

um i read just all you know the peter

peter lynch books

and i was completely hooked i came i was

going to dartmouth um

at the time dartmouth college i came

back i shifted my major from english and

history

to history and economics

started what i termed an options trading

hedge fund

in college and um

yeah and it that started like a lifelong

love

and passion um you know i could still

remember

learning about um return on invested

capital

and like how amazing that that was to me

you know as a junior senior in college

and the concept of incremental returns

on incremental invested capital and

reinvestment ratios

um so so what year was this that you

were at dlj like where was the market at

the time

not not to date you but i believe

it was 1996.

and um i

am i actually kind of worked for that i

was an intern for the guy who ran the

office

and i did a lot of things as an intern

but my my

my favorite activity um

was writing up i wrote up one stock a

week and then he would sit down and

talk to me about it for an hour and it

was it was it's something that i've

always remembered

so um i wrote i did so much work and

wrote this

um you know enormous 15-page report on

cisco

who was so into the idea that it was the

arms dealer for the internet and was

going to grow with the internet

and it was all you know i was spending a

lot of uh time on the internet dartmouth

was

very early to email it had something

called blitz mail there wasn't such a

thing as email at the time they called

it blitz mail

um and i was so excited about cisco

but you know the stock had been a really

good stock even as of um

maybe it's 97 it was either 96 or 97 but

it had been a really good stock

um and maybe i don't know what the

market cap was

it was a 10 billion dollar market cap i

don't we could go back and look

um and the head of the office you know

just looked at me and said you know

gavin i

you know there's all these great things

about it but how much more can it go up

and i was just like geez i think we're

really early in the internet i think it

could go up a lot

um you know if you look at other you

know other technologies and

you know capital expenditures as a

percentage of global gdp

you know and i'd gone back and i looked

at you know canals and railroads and all

these things were kind of foundational

technologies and i was like wow i think

we're very early

um and he but he didn't love it and he

was

a great man and a great investor i

learned a lot from him but i do always

think of that

um just because you know cisco you know

i think it went up 20x

in the next four years um

that was probably a pretty uh

foundational experience for me as an

investor

yeah you you uh this this inner thing is

going to be big

and and and so that was kind of like um

it sounds like that was the basis of

where

uh you learned fundamental investing and

i know from

our previous conversations um you use

fundamental investing

and technical investing or you look at

technical analysis

where was the point where you started

thinking about uh using technical

analysis

and was it weird to you when you kind of

like someone said to you the concept of

hey looking at historical prices may

have an impact or does have

some predictive power of the future yeah

well

so i do tend to have a very um

academic approach to things and

so i had done a lot of work on kind of

behavioral finance inefficiencies as

part of my economics

major and i actually wrote my um

senior thesis on reconciling the two

most dominant statistical patterns in

the stock market

are actually long-term mean reversion

and

short-term momentum so short-term return

continuation

so in other words over very short time

frames

statistically if a stock has been going

up it is likely to keep going up and if

it's been going down it's likely to keep

going

going down that works really not

um not much longer than 12 months and

nine months is

probably generally the sweet spot in a

purely quantitative sense this is not in

a technical or chart sense

um and then the reason reason there's a

long-term mean reversion pattern in the

stock market is because of capitalism

just you know an industry that's

underperformed for a long time where the

stocks have been going down probably

capacity is leaving

um whereas you know an industry

that's been going up you know it

attracts competition

it attracts capital and that's why

historically there has been mean

reversion

in the stock market and um

you know basically the two most powerful

statistical factors in the market are

actually valuation and momentum

um and momentum is almost as powerful as

valuation as a factor

so i was aware just from all of that

academic work i had done in writing my

senior thesis

um that there actually was

information um contained in

historical prices that was relevant for

future predicting kind of future prices

and you can do a lot of things you know

if you intersect momentum with with

volume

it makes it even more powerful and

fundamentally all

quantitative investors all they are

doing is intersecting momentum with

valuation

so i came into um

probably investing sympathetic to

um technical research um but

because of that and then you know i read

the um is it is it terence odin who

wrote the paper on prospect theory i

believe believe it was

i read the prospect theory paper um

which really kind of helped me

understand why technical analysis works

prospect theory is basically human

beings um

are feel the pain of losses twice as

intensely as they feel the

joy of gains you know it's kind of like

you know there's a well under

uh understood thing and kind of

management that you you almost have to

give somebody ten compliments for every

one piece of criticism they give them

because we're all so much more sensitive

to being wrong

to negativity um and you know

fundamentally that is why there is

support and resistance because of

regret and people feeling acute regret

that they

didn't sell and so when they get the

chance to sell there again

they sell and that's why there's

resistance and feel acute regret that

they didn't buy support levels

um and basically to me prospect theory

explains a lot of technical analysis so

it's very sympathetic to it

and then i would say broadly speaking

while i am

i will never buy a stock just because of

its chart

it is an important input into its into

my process

and by the way i will buy stocks with

terrible choice i

i'm actually um you know hopelessly

addicted to the 52 week low list

and you know even though that that

sounds smart it's

actually not smart statistically

you know statistically the smart place

to work off of is the 52e kylo

552 week hi list but i find myself drawn

to the 52 week

low list um so i will buy

something on the fundamentals alone even

if it doesn't have a grade chart

um but

i use um yeah i think there was

i i forget who it was i think it was

bruce kovner in market wizards he said

that you know he just thought

you know technical analysis was part of

his toolkit you know it's like if

you know doctor had a thermometer

technical analysis was his thermometer

you just kind of check the health of the

patient

but i'd say the simplest and crudest way

that i use

charts is but i do this less and less

but particularly when i was a newer fund

manager

like i just start by looking at a

20-year chart if

you know there was something that was

new to me the

quickest way to determine if something

is a good business an average business

or a mediocre business

is to look at the long-term chart good

businesses

the charts are usually up and to the

right you know who

average businesses the charts are you

know they're

um you know maybe more in line with the

stock market you know you want to look

at both the app you know the chart of

the equity itself and then first the

market

and then you know charts are bad

businesses they're often long-term

um underperformers so there's um

a 20-year chart can tell you a lot and

then i actually further

intersect this and this is a view that

they um that we

that i use um in koi fin with earnings

um

is over the long term what drives almost

all stocks is just growth in either

earnings per share of free cash flow per

share

it's amazing if you look at a 20 30 40

year chart

um whether it's microsoft whether it's

procter gamble whether it's coca-cola

um you know maybe less coca-cola and png

recently

you know any business you know high

returns on invested capital

that's able to reinvest those good rates

um you can see it just goes

up and lockstep with the earnings line

and if you scale them

so it's at about 20 times earnings 15 to

20 times earnings

the lines are almost on top of each

other for 20 to 30 to 40

um your time frames um

the compounders right the like yeah and

by the way it's like you know

there's loads of ways to make money and

you know you know

like i i you know that you know someday

you know there's

these names that are just kind of

wigglers instead of compounders they

just

bounce around up and down and they're

very cyclical and over the long term

they destroy value but you make a lot of

money

um buying them at the bottom and selling

them at the top i think it takes a high

degree of skill and a lot of knowledge

um but there's many ways to make money

in the stock market yeah

absolutely and i think it's it's

important for people to identify which

is the way that

they're good at or what fits their style

and really to focus on that and not try

and

be someone else oh yeah i think the

secret

i think the most important thing in

investing is there's many ways to

succeed as an investor

and ultimately investing is all about

finding the right

balance between humility and conviction

and you have to find an investment style

that fits your own

personal emotional makeup because i

think the most important factor in

investing is the ability to be rational

when wrong

so you have to file and find an

investing philosophy or style

where you really believe in it you're

really comfortable

you're going to stay with it

but that helps you be rational when

you're wrong because you're going to be

wrong a lot and it's really hard for

most people to be rational and wrong

so i'm i always think it's very strange

that people have such

strong views um you know this is

this is not science it's not like you

know there's there's you know there's

only one theory of relativity um you

know even though the theory of special

relativity doesn't actually explain

everything

um you know there's um

there's only you know there's there's

only one theory of gravity you know the

sun

does rise in the east and set in the

west um

because of the direction that the earth

rotates you know

this is not science where there are

established facts or where theories are

so well established it's so well proven

that they're basically facts

um that's important remember all of

those are actually just hypotheses that

you know are proven every millisecond um

investing there's many many ways you

know to win there's wasted many ways to

succeed

you could succeed as a value investor

you could see it as a growth investor

you could see it as a

technical investor you could succeed as

a momentum investor

by the way momentum investing growth

investing are often conflated they're

not the same

you know so i find it so strange that

people have such strong

views that their way of investing is the

only way of investing

there are many ways to succeed you just

have to find the way that works for you

as an investor your own particular

emotional makeup

and you have to stick with that and

is there and so what you're talking

about in terms of being flexible

and just being able to uh to change

your view how do you balance that with

just having conviction in something and

if something's going against you you

know having the conviction to see it

through some

temporary setbacks like is there

anything that you've learned throughout

your career have seen others do that

help you balance that

um

yeah i mean i think this is um you know

michael steinhardt

um and and market wizard said

um investing was all about finding the

right balance

between um having the courage of your

convictions relative to the flexibility

to admit when you're wrong

you know so your largest position is

down 50

you've been wrong you know a lot of

people have this attitude oh

i'm not wrong just a little work you

don't yeah you just need to wait if

you're down 50

you were wrong um

you know a lot of people say oh surely

um you know a

former fidelity colleague of mine my

name george vanderheid

george vader hayden was fond of saying

being too early is the same thing as

being wrong

this is a very competitive demanding

pursuit

um so it's down 50 you've been wrong

can you be rational take all the new

information that led to the stock being

down 50

and make the correct decision which may

be to double your position

and maybe to do nothing because hey the

story is changed and

now your position's half what it was and

that's the right position size or maybe

it's to sell it all because the new

information was fundamentally thesis

changing

that's really hard to do and that's why

i think you have to have a philosophy

that works for your own emotional makeup

you know it's not um

you know some of the investors i know

with the strongest stomachs

um are actually growth investors um

you know you you you know and there's a

lot of deep value investors who have

very strong stomachs you know there's

there's some great research done

um showing that um great value investors

you know differentiated themselves on

the entry

great growth investors differentiated

themselves on the exit um

but um yeah i i

you have to find what works for you for

me it is feeling like and i talked about

this on patrick o'shaughnessy's podcast

so

you know i wrote um i wrote a medium

article on it called investing in the

name of the rose but for me

feeling like i have a very high

knowledge level

um helps me make good decisions when i

am wrong

because even when i have a high

knowledge level i'm wrong a lot um

but when i feel like i have a high

knowledge level and i've and i'm

wrong because of kind of a risk that i

thought about

um i find that i make better decisions

um so that's what works for me

but everybody has to find what works for

them yeah

and and and another concept related to

this which you mentioned on patrick's

podcast

uh also from the western wizards is that

boulder on top of the mountain and the

boulder community oh yeah

either way and you really have to

imagine what are the different scenarios

where the boulder goes

any which of the 360 degrees

uh especially in today's market where

the boulder could literally go anywhere

and you could have a lot of different

permutations yeah if you said to

yourself

you know here are the seven big risks

here are the ones that

you know i'm comfortable with and here

are the ones that if they materialize

would material would really change the

thesis

um you know it makes it a lot easier you

you almost you know this is the whole

concept of

having a pre-mortem written um

for a position um you know in other

words if we were wrong

these are the five seven ten reasons

that we're likely to be wrong

it's actually very funny you know

there's gpt-3 this new ai algorithm

yeah that um you know if you feed it

anything

um it can spit out what it's often

really impressive

text afterwards i am tempted to

experiment

because one of the biggest things in

investing are unknown unknowns

and gpt3 was you know trained on you

know

an enormous corpus of written material

and i am

interesting to see and i have no idea if

it'll work we haven't done it yet

um but um i'm interested to see

that if um i were to feed an investment

thesis into gpt3

and then you know here here's you know

here's here's the hypotheses that go

into this

um you know here is you know an expected

irr

and a base case in a bowl case over the

next five years here's here's here's the

risk

and then say you know here are the risks

um

you know that i really see and then you

say then you ask gbt to generate

gpt3 to generate other risks i'm curious

if it will generate good

unknown unknowns and thereby kind of

convert which are actually if we're

being very precise

um uncertainties rather than risks you

know uncertainty and risk for different

things

and a set of bounded outcomes is risk

when the when outcomes are unbounded

it's uncertainty and so they're always

definitionally uncertainties in the

stock market because

almost anything can happen but can you

turn a few unknown unknowns and then

known unknowns you know turn a few

uncertainties into risks that

well you probably can't bound them

precisely you can begin to frame them

yeah sounds like you want to automate

away the analyst function

which i hope i hope the analysts on your

team are not watching this

no no i'm a huge believer

um you know you know in some ways the um

our iphones and our android devices are

already personal ais for us

right you know they know a lot about us

they know you know they effectively know

everything about us they know where we

are

they know who we're with um they know

when we're sleeping

um and you know they can answer almost

any question

but i do believe you know human beings

working in concert with artificial

intelligence or computers

are are going to be able to produce

superior outcomes that are the humans

alone or computers alone for many years

to come so

i do not believe that analysts are going

to be automated away anytime soon

yeah and i think that probably goes for

for a lot of industries that people

assume are going to be automated away

um cool so um you know

super interesting background and thanks

for sharing kind of how uh

your investing style has evolved i

wanted to turn it over

to maybe some specific uh investment

ideas or investment themes that that

you're thinking about

and let me share koi finn um and i'll

kind of pull up charts as you're talking

about different things

awesome um so i would say what i am most

excited about

right now is the um

rising is the higher semiconductor

intensity

of artificial intelligence relative to

software written

by human beings um

and i think this may be one of the kind

of most important

investment themes over the next 10

20 30 years um

and so kind of you know

how did we get here why is this

important well you know mark andreessen

famously wrote this um

wall street journal op-ed i believe

eight years ago entitled software is

eating the world

basically saying that every company was

going to become a software company every

industry is going to be dominated by

software

and i think if you were to rewrite that

essay today

who'd probably say artificial

intelligence is eating the world

um because artificial intelligence is

just another

is an is another kind of software our ai

is almost software written

by software software that was trained on

a big data set

and you could almost think of

software written by human beings as

almost being an expert system you know

it's hand coded by a human

being if this happens then that happens

um and it's interesting ai has been

around

um for you know roughly 60 years now

and people were really excited about it

in the 60s 70s and 80s

and then i think they quickly you know

neural networks were invented long ago

and i think they quickly limited quickly

realized that hey the neural networks

um we're kind of limited by the amount

of

compute power available um

and then ai went through what was called

an ai winter

um where there was no progress and

you know the uh the people who who

believed in it were um

you know a box clementis and deserto

um you know to reference my alma mater

alma mater a voice

um crying out in the wilderness um yes

here's the chart of nvidia which has

certainly been central to

artificial intelligence and i and i can

explain um

and i and i could talk about that um

and um you know there's all sorts of

stuff in symbolic systems

and you know just tried to teach an ai

everything about the world

and then um in

in 2006 a computer

scientist a woman named faith hayley um

who is now a luminary of artificial

intelligence

had the idea to create an enormous

database of labeled images that she

called imagenet

and they actually had to use amazon's

mechanical turk service

um to label all the images but this

meant that you had

an enormous data set that you could test

various types of computer vision

software on

and there are many different ways to do

computer vision and

they started having an imagenet com uh

competition

and um deep learning was

ported to over to gpus um nvidia wrote

wrote uh

software that they call cuda um it's

it's

you can think of it as a programming

language or programming environment a

tool chain

um for doing deep learning

on gpus and gpus

are it stands for graphics processing

unit

which are used uh to generate the 3d

graphics that you see when you play a

computer game

or you see any you know really kind of

sophisticated

animation um

and fundamentally 3d graphics are all

about doing

um enormous amounts of

very simple math really quickly

and you could almost think of the

difference between a cpu and a gpu

the cpu processes

instructions serially and gpus do them

in parallel

and i think of cpu as being like a

checkout register

that can check out any item in the world

um so it's like you can picture the

world's most sophisticated checkout

system and it can check out anything

ranging from a cow to a car

to a fighter jet to a computer you know

to an apple

to you know a piece of meat to a book to

a magazine they can check anything out

one after another so fast boom

whereas the gpu can only check out like

100 different things

but with that cpu you really only have

one checkout register and you can have

hundreds of checkout registers in a row

if you're a gpu so if you're a store

that only sells a few things it's much

faster to check out through gpus and in

this analogy

the software programs or the stores and

the items that are being checked out

and the gpus and the cpus um

are their kind of checkout registers um

so

in 2012 um

a team led by jeffrey hinton who is

another

um luminary of artificial intelligence

today

entered a deep neural network called

alexnet

in the um imagenet

competition and they absolutely

um annihilated

um the saw the computer vision software

that had been written by human beings

and this was you know tens of thousands

of hours of man hours of programming

that had got into the best computer

vision software humans could write

and it got annihilated um by a deep

neural network

that had been sped up um on gpus

and that really actually kicked off the

current

kind of deep learning um revolution

um an artificial intelligence revolution

um

you know jeff jeff um

uh jeff bezos um you know i mean it just

it was a massive win

um alex that won by an 11

margin which doesn't sound like a lot

but it was 41

better it's kind of like you know it's

like you know nobody had broken the four

minute mile

um and you know alex net ran it in three

minutes and the number two finisher was

at four and a half minutes that's that's

how big that margin of

victory was and then in 2014 um

it was over 2013 the human beings were

like hey we're going to come back

um it was like that old story of you

know the steam engine versus

who's the guy splitting wood um

paul bunyan yes exactly yes whatever it

was

um you know the steam engine eventually

obviously went very decisively

um and um

you know the human said oh we're gonna

do a lot better 2013 we're

really going to get focused they got

annihilated again and in 2014

every submission uh was a deep neural

network

um and then you know and this is why

jeff bezos you know in 2013 2014

basically he observed there's been no

change in algorithms the algorithms we

have been using

um our um

our really old algorithms it just turned

out for them to be

accurate we needed way more data and we

needed way more compute we had way more

data

basically because smartphones are

gathering so much unstructured data

about the world

um and then we had more compute because

of cloud computing so cloud computing

was a very foundational

um invention and innovation enabling

artificial intelligence but i think the

fundamental we are going to come to

stocks but the fundamental thing

i think to um understand

artificial intelligence and this may be

i think the single most important fact

to understand the world we're living in

is that ai quality generally doubles

with every

10x increase in the amount of data used

to train the algorithm

um and um

this led to and this really um

so so if you want to double ai quality

you need to spend a lot more compute

and compute means cpus gpus memory

everything

um networking all of it um

it's all scaling you need them in a

roughly fixed ratio if anything the

amount of memory

um needed it may be growing even more

faster than

even faster than the compute that's

debatable we've just

we've actually just had a lot more

architectural progress in compute than

we have

had in memory i mean there's there's

been this explosion of you know gpus

getting better

tpus from google um lots of startups

just an explosion and architectural

innovation

but i read an article in wired um

in early 2017 and it was called building

an ai chips uh save

google from building a dozen new data

centers um

and basically what happened was they had

realized that these saved

deep neural networks that were so

effective at image recognition

were incredibly good at speech

recognition

google was a big believer in voice

search and voice is a new

um ui and that their new deep neural

network

um for speech recognition they realized

and google at this point had probably

50 billion dollars of capex in the

ground in terms of data centers

um so in fact check that number but

order it i certainly did not look it up

beforehand but in 2017

google had a lot of invested capital and

data centers and their engineers

realized

that if every android phone on this

planet

used the new google search algorithm

for just three minutes it google

did not have enough capacity and they

not only that they needed to

double the amount of compute capacity

they had globally

um and so that's actually why they um

you know why they designed the tpu

although i think the gpu progress has

probably been a lot more rapid than they

would have expected

but just you're seeing incredible um

truly incredible um

progress like that alex net 2012 was

trained on two gpus

over five days in 2019

um it took a thousand chips

six days and each one of those chips was

probably

10 to 20 x faster than those original

gpus

you know so so the training thought so

the training time

you know had gone out you know the

training compute had gone up by kind of

a

you know like a factor of probably ten

to twenty thousand

um uh there's a

english to french machine translation

algorithm

and if you were to

do the training necessary to take its

accuracy from 50

to 10 percent so in other words well

fifty percent error rate to ten percent

error error rate so from fifty percent

to ninety percent

you would require a billion billion

times as much

um computation power you know stuff like

this is why a lot of futurists believe

in dyson spheres

you know that eventually the compute

needs of humanity

will be so great that the raw energy

required to run them

will require a dyson dyson sphere which

means you will enclose

the entire solar system with a bunch of

reflectors that will gather solar energy

to power compute

um but anyways so this is

very exciting to me and just you know

what it means

in real terms you know like to bring it

back to you so there's nvidia

almost all ai in the world today is

trained on

um nvidia gpus uh

a lot of inferencing happens on you know

intel and amd cpus then

nvidia and amd um gpus but you know

nvidia along with google

really um

from a factual market share position of

ai training workloads they

they they they do the majority of them

today

there's loads of startups coming for

nvidia um

jensen is a truly exceptional ceo

um and an interesting kind of

observation about

digital processor um digital processor

markets

is you've never really seen a number one

unseated

so in other words um intel's always been

number one in x86 cpus um

nvidia's always been you know number one

in in gpus

um you know qualcomm has always been

number one in base bands

um and this is just you know nvidia had

an effort

you know loads of people have tried to

compete with intel and cpus loads of

people have tried to compete with nvidia

gpus

nvidia and intel both try to compete

with qualcomm and basebands

digital processor markets the barriers

to entry

are really underestimated code

gets optimized um for you know a sp

you know a specific flavor of cpu gpu or

base band

and that gives um you know somebody said

that it was an exorbitant privilege that

the united states could borrow in its

own currency

you know that the number one digital

processor player

enjoys the exorbitant privilege of

having code optimized for their own

specific

architectures um

but you know it's not just you know and

by the way

ai training and inference it just it is

much

more it's more cpu intensive it's more

gpu intensive it's more fpga intensive

it's more ai accelerator intensive

it's more memory intensive it's more

networking intensive

than software written by human beings

and that's really important to me as an

investor because if

i believe that 40 or 50 years from now

it's probably going to be illegal for

human beings to write software

ai will kind of run the world um

anything that uh you know a um

one of these pioneers of ai research you

know i i retweeted him but he said i

found it mildly scary anything a human

being can do

a deep neural network can do better if

it is provided with sufficient data

and enough um compute training capacity

um you know so ai will drive cars fly

the planes

steer boats control traffic systems

um you know they'll everything will be

priced by ai

advertising will be driven by ai you

know there's almost nothing that will

not be

touched by ai um but that was

that goes back to the analyst comment

before do you think that that's going to

be exclusively driven by

ai or there's no so so the so the humans

the role of human the computer

scientists

they will they will almost become

hypothesis generators

they will uh frame problems

you know ai is is brilliant but narrow

today it may become more robust if

transfer learning ever works

but for now it's brilliant but narrow

and

um so the skill in human beings

will almost be in

choosing the data sets to train the a

ion

understanding framing the problem is in

the right ai

so there will be a role for human beings

but it will be very different

but this matters for semiconductors

because definitionally

if all this software all software runs

on semiconductors today

definitionally and if ai

consumes more compute than software

written by human beings

it consumes 6x the memory um if you you

know

consume 6x the memory uh much more much

more

much more compute much more networking

then the semiconductor

intensity of software is going to go up

significantly by 4 to six x so this

means

to me as an investor over the very long

term that whatever the

current um

um you know semiconductor share of

global gdp is it should go up by at

least a factor of four or

five or six over the next 10 15 20 years

as it takes share

and then that's even more exciting so

that's the demand side on the supply

side

by the way there's you could pull up the

sox the smh

you know this is going to apply for

memory makers networking semiconductor

makers you know a

wide variety you know fpgas this is

going to touch on a lot of

semiconductors

um you know just if you want tickers you

know

you know intel and amd make cpus xilinx

makes fpgas

um you know micron and samsung make

memory

um you know liam research and asml

make the equipment that you use for all

of this lrcx and asml

taiwan semi makes a lot of these

so this is going to touch almost all of

semiconductors

from a significant upward inflection

in demand and if you if you think about

it this is important because the

semiconductor industry

has absorbed two massive negative demand

shocks

um over the last

uh call it 15 years

um the first one was the iphone you

would think that the iphone was a

was positive for demand it was actually

negative because what the iphone did is

it replaced pcs

um you know people used to you used to

be able to confidently project

that you know billions of people were

going to have laptops well

they don't and the reason is they have

iphones or android phones

and even though there's a lot of

semiconductors in an iphone

a lot of semiconductor dollar content

it's lower than a laptop

so having iphones replace laptops was a

negative demand shock

when you saw that you saw the

replacement cycle for laptops extend

and then it had to absorb the negative

demand shock of both virtualization

and cloud computing which really

increased

the utilization of the world's existing

semiconductor infrastructure that was a

negative demand shock

we've kind of worked through both of

those and now we're going to have a

positive demand shock

from the ai intensity and we're going to

have that just

has almost every semiconductor market

has really consolidated

um all of these markets whether it's

semiconductor equipment

whether it's memory cpu gpu fpga

they've become incredibly consolidated

with extremely high barriers to entry um

you know brad slingerland who's an old

friend

um wrote this up um and went back and

forth on twitter

um and i am going to write a lot of this

up in a medium post at some point rob

but he basically and i was talking about

how high the barriers to entry were

you know that modern semiconductors are

the closest thing to magic

in the real world today you know just

that where

you know you are manipulating wires you

know that are

just you know almost

incredibly small these wires they're so

thin that the human being

the human mind cannot cannot really

apprehend how small these wires are

and you know the the thousands of miles

and wire

of wiring that are inside you know a

chip that's you know the size of

maybe two thumbnails put together um

as it is so thin um it's the closest

thing to magic and uh brad

brad uh tweeted at me magic has high

barriers to entry

and that is going to be the title of my

medium post on this which i'm writing

but um so you have this positive demand

shock in the form

of ai intensity and then you've got a

very consolidated industry and look a

lot of technology investors have been

very focused on software and internet

for the last 10 years and so

semiconductor investing is a little bit

of a lost art

i was a semiconductor analyst back in

the year 2000 so i actually love it

um it is pull up asml um

rob so asml is a

uh is it a dutch company so i'm not yeah

it's yeah

it does it does it's a lithography

company

and it is just amazing it's like almost

all human progress depends on this

company

okay um because to keep moore's law

going which is

really important look all these

companies you know all these processor

companies whether it's intel amd

um nvidia cerebros

you know rock you know mythic ai

whoever you want to um talk about

um yeah i should declare disclose

tradies management as an investor in

mythic ai um

a lot of some of these other names but

um they're getting a lot better

architecture it's squeezing more

performance out of each generation of

wars law

but to keep moore's law going asml has

to make

progress and humanity we really need

another five to seven x turns of moore's

law

um to unlock really amazing stuff and

health care

and material science um

you know really keep technological

progress going and then after you know

after moore's law there's all sorts of

cool stuff we can do

um you know silicon photonics

um using light instead of electricity

quantum computing is

is super cool and beginning to work

super cool no pun intended um you know

um super cooling i mean i mean just you

cooling is a problem with with quantum

computing um

but we're gonna um chill

um conventional chips down to extremely

low levels

um as a way to solve this there's a lot

of neat stuff coming

but asml if you could pick one company

on this planet

you know if kind of nvidia porting

um

artificial intelligence to gpus via cuda

and by the way the whole thing about

parallel computing is

artificial intelligence deep neural

networks all they are is doing very

simple math over and over again at

massive scale

just like graphics and that is why gpus

are good at artificial intelligence

sorry to come back to that i think maybe

i've forgotten to make that point

but you know kind of nvidia and their

architectural work and their software

work has kind of

really underpinned the modern revolution

and artificial intelligence

um and you know i think really made a

lasting contribution to kind of science

and humanity um

and look now they're facing a massive

wall of venture-funded

competition and vc uh evidently told the

ceo jensen that uh jensen had

single-handedly brought brought back

semiconductor uh

uh venture capital investing because

people were so excited to fund

competitors to nvidia

um and um you know lots of those are

being acquired by intel

um you know has their own big big

efforts but asml

lithography they need to keep

making progress um to keep moore's law

going

please rob so i i was gonna so uh

lithography again is is more of like a

semi-cap equipment concept

in the supply chain but i was gonna ask

you uh a couple of things based on what

you said so

um yeah so the competitors the

competitors that are coming

online to compete with nvidia uh you

know before you mentioned

as a student of history that that

industries are mean reverting and that

success begets competition and then

failure uh destroys competition and it's

a natural cycle

and then you also mentioned kind of this

this um

concept of leadership in semiconductors

and kind of leadership uh having a

competitive edge

over a long period of time whether uh

it's intel or 18c or whatever it is

um so how how do you think about nvidia

being a leader and

all these startups nipping uh at its

heels is

will we be at a point where someone will

succeed

well just to just to address your first

point about kind of comp you know mean

reversion

one of the most fascinating things is

until maybe 15 years ago

roics were highly mean reverting so

companies with high roics would

generally see those roics mean revert

over a period of time

that stopped being true

roughly 15 years ago um

and this is kind of well established

it's been all over twitter

i bel i forget which

um which consul it might have been bane

that did the work

um but they've done this work and it's

been extensively replicated

and you can see it i mean it just i mean

anybody with access to financial

data can do it but really high roic

companies are maintaining those roics

with no sign of those roics falling

and i think this goes back to you know

brian arthur who

um kind of wrote a similar paper i

believe in 1996 on increasing returns

you know there are in technology there

are you know increasing returns to scale

network effects that lead to these roics

being much more resistant to competition

than in traditional industries so for

instance um

you know the biggest risk to google is

that somebody invents an algorithm

that actually changes this relationship

between

um data quantity and ai quality

because the quality of google search

is really driven by the fact that they

have more data than everyone else

they do more searches than everyone else

on a daily basis so their data mode is

only growing

and so as long as data is the

determining

factor for the quality of their product

you actually if you wanted to go spend

200 billion dollars to compete with them

you still couldn't you know amazon tried

to compete with them

they failed um you know they had a

startup i think

they had a search engine um i think they

called it

um alexa um

no a123 is what they called it i believe

um

they failed microsoft has largely failed

um it is you know these these internet

businesses

are natural monopolies the likes of

which the world is never seen

because of this relationship between

data quantity and ai quality

and because of these increasing returns

to scale and network effects

you know like what do i mean by that um

you know railroads and cable or natural

monopolies

because and utilities they're natural

monopolies because it only makes sense

to have

one and it's so expensive to put it in

so let's put it in and then let's

regulate them and in the case of cable

it's not regulated because it was done

with private capital

um and and that was largely enabled by

the creation of the high-yield debt

markets

but these internet businesses um they're

natural monopolies because

once you get a dominant position

you have more data to train your

algorithms so your algorithms are always

going to be better

and unless they hit some ceiling where

quality no longer matters and we have

yet to find that

you can't spend any amount of money to

compete with them like

you know so for you know if you wanted

to compete with a utility or a cable

company or a telecom company

or a railroad and you had an infinite

amount of capital you could

for these internet businesses you can't

um and so i think and how you regulate

them is going to be

very interesting and beyond algorithmic

innovation which i think is the most

important risk for any

internet business where their moat is

kind of fundamentally the um

you know the data quantity driving ai

quality leading to a better product

because they have a better product they

get more users and more usage which

gives them more data which makes their

product

better in that flywheel and that

virtuous cycle spinning

um you know apart from algorithmic

innovation regulation is what really

matters for all of these companies you

know i do think probably the most

logical way to regulate them is actually

to break them up

um because you know fundamentally they

are kind of giving

you know you know most of them are

either free to the user or cheaper than

alternatives

but anyways so

the statement i made about mean

reversion has actually not been

true for kind of 15 years and it is

easy to show that in the roics in the

top decile they are no longer mean

reverting

uh but nvidia look um you know the um

the head of the head of ai at um at

nvidia was

was uh was saying look everybody's

trying to eat our lunch

but we're trying to be a uh we're trying

to keep our lunch moving pretty fast

um and it's a nvidia is a moving target

with what they just unveiled with ampere

was i think pretty

pretty astonishing um and look you know

cerebrus what they have done with wafer

scale computing that is a

truly kind of fundamental you know

scientific engineering technological

breakthrough you know they're going

about

about it in a very different way

intel bought habana

which you know by the way this is why a

video about melanox because um

gpus stop scaling linearly once you get

i think to like 50 link gpus

um and if you could put um i mean we

don't have to go into details but if you

can embed the networking on the chip

they can

scale linearly so that's kind of you

know habana's big advantage and intel

bought them

and then you've got um you know grok

which was started by the um

by the tp by the team that built the

first tpu for google you have google

itself with its tpus

um you have a startup called sambanova

um you have a

huge array of very

well funded venture companies going

after nvidia

and then you have amd who you know lisa

sue

um has done a great has done a great job

she's an exceptional executive

um and they have both you know cpus and

gpus

intel is trying to make a gpu um

of their own um this will be their kind

of

um they they had a first try called

knightsbridge which never really saw the

light of day

um this will be their second try they're

very serious

um they've hired um the former head of

architecture at amd to

work with them on their gpu architecture

so we'll see so look

there's a massive amount of competition

coming for nvidia

um and so they're gonna have to run fast

to keep their leadership position

but i do think it is interesting that

the history of these digital processor

markets

and that exorbitant privilege um

kind of conferred by having software

optimized you know it is it's it's a

powerful competitive advantage for

any dominant um uh

processor company yeah

and and and and some of the kind of like

secular trends that

that you're mentioning are evident in

this market cap chart of

of amd versus uh nvidia and let me just

put this on

on a linear scale to kind of show just

the

uh the total magnitude but obviously log

scale you could see the percent change a

little bit better

uh you know amd four years ago went from

uh 1 billion

market cap to 70 billion today

uh and still trailing nvidia by a decent

amount

um how do you how do you think about

kind of like amd

versus nvidia going forward both are in

the gpu market

um like how how would you think about

these companies strategically over the

next five years

yeah so i would say amd has made you

know there's this legendary

architect named jim keller um

and he's one of

there's only a few truly great um

processor architects in the world or

kind of processor teams in the world

and there are by the way a lot of the

other dynamic we haven't talked about is

arm coming for x86 which you know is

certainly a big cross current for both

amd

and intel um and you know the

the you know apple is switching from

switch

switching the macbooks from x86 intel

chips to their own internally developed

arm based chips and these are all cpus

and you know arm is coming to the data

center and there's some really

interesting

um arm uh

arm based server cpu companies but um in

terms of nvidia versus amd look

there's this amd hired jim keller

um and working with a great design team

there

he kind of introduced this chiplet arc

architecture

um to amd

um that has let them get a

architectural advantage over intel in

the server cpu market

for really the first time since kind of

the 0.304 period

when amd had something called 030405 amd

had something

called the opteron architecture

basically in the

you know 20 years ago intel um

that really big on oh gosh what was it

called

um um

i don't remember titanium and a new

instruct kind of almost a new

instruction set

yep um and amd said hey we're just gonna

kind of

stick to the existing one do it better

they

kind of were a little faster they had

multi-cores but back then

um amd was still they still made their

own chips they were not fabulous

and um so they literally could not

fulfill demand

and intel used um their kind of

manufacturing muscle

and limitless capacity they kind of

slashed prices

and basically said hey amd can't fill

your demand

anyways but if you go to them instead of

staying exclusively on us we're never

going to remember it we're never going

to forget it

and you know the next time you know

there's a shortage we're not going to

give you chips

um it's they were able to kind of fight

off amd

um intel kind of

really has gone through kind of a period

in the wilderness where um

you know they made some silly

acquisitions they bought altera

they made a lot of silly acquisitions

they defocused

um and they took their eye off the ball

in manufacturing

um and so for this current

transition um into intel's 10 nanometer

is roughly

taiwan 77 nanometer and intel

for 50 years has had a manufacturing

advantage i'd say on average they're

nine to 18 months ahead of anyone else

to each node

and given moore's law you know we all

understand moore's law what it does for

price performance

but so if you're 9 to 18 months ahead to

each node you have a

massive embedded advantage um

and so intel had this incredible

advantage built in

and manufacturing advantages and

semiconductors

um they're very um you know one of my

favorite phrases

is from a british historian called

arnold i believe it's arnold toynbee and

he said empires are the lesson of

history is

empires are hard won and easily lost and

that goes for

many other things like investment

performance it also goes from

uh you know semiconductor manufacturing

so intel

um they're in the lead and being in the

lead is actually

it's um it does give you a big advantage

because

it is like baking a cake and you kind of

need to try out recipes

and so if you're ahead you have longer

to try out the recipe so you get there

faster

but intel they they didn't use asml's

euv tool

taiwan semi did that was an arrogant

decision by intel

and they tried to um

increase the transistor density too much

which

compounded the mistake so they have lost

their manufacturing lead

so now we have amd they have an

architectural advantage because of this

chiplet architecture

from jim keller jim and by the way jim

keller was also at um

tesla and he helped them develop their

own

um asic that now is in every tesla that

does the infancy for autopilot and

actually he did a stint at intel and he

helped them develop new chips that i'm

you know that we will see from intel

over the next two years um

and he is truly a one-of-a-kind

architect

um and it's like almost wherever he goes

um you know he kind of leaves magic

behind but um

so amd has an architectural advantage

but because of taiwan semi

they also have a manufacturing process

technology advantage

and then because taiwan semi has lots of

capacity

know they are theoretically going to be

able to meet more demand than the last

time

um they had an architectural advantage

nearly 15 years ago

so we'll have to see what happens but

intel they do have that exorbitant

privilege

of software has been optimized for intel

processors they're all these code

libraries

you'll code library for almost any kind

of software

you know ranging from artificial

intelligence to high performance

computing you know to you know

scientific programs to

anything it's been optimized for intel

cpus all those code libraries

um so amd has to be really good but look

they're starting to take share you can

see that in the um

you know in in in the kind of uh you

know in the amd verse intel chart there

but um you know we'll

we will see what happens going forward

intel

uh mary we'll call it i would say three

to four years later on 10 nanometer

they need to execute on 10 nanometer

they're gonna they're gonna actually

they're gonna show their tiger lake

um client cpus which are kind of their

mainstream tiger lake cpus for laptops

they're going to show those i think

highly likely in early september

but then the real test will be can they

run their 10 nanometer process

on cert cpus because basically

server cpus are always bigger than

desktop and laptop cpus

and as a chip gets

bigger yields become more important

because you can have you have to have a

lower defect density so

you know you make these chips on a wafer

they're 300 millimeter you know they

look like um

you know pieces of copper or gold and

they're they're circles and they're 12

inches in diameter

and if you're slicing that circle up

into lots of different little dies that

are going to be the cpus

you can have a higher defect density

because

you know let's say you're you're you're

you're slicing that circle up into a

thousand different chips

um and there's a defect

on five percent of that circle that

means you're only going to be throwing

away

50 chips you're gonna have 950 working

chips well if you're slicing that circle

up into only

you know i don't know 20 or 30 or 50

chips

um and those defects are randomly

distributed

you know you might be only yielding 50

for that

you know for that defect density um

so the real test of intel's 10 nanometer

process is going to come

um when they attempt um

to launch server cpus which are much

bigger chips with bigger dies

um on that 10 nanometer process and you

know

we'll see that's going to happen at the

end of this year

and either it will happen or it won't

and if it happens you know intel will be

on

a better competitive footing versus amd

if it

if it doesn't happen for intel then amd

is going to have a really big advantage

and you know we're going to know that

over the next six to nine months that's

on the cpu side of things now amd and

nvidia both compete head-to-head in gpus

and intel is about to come out with what

looks like a very interesting gpu

and their gpu was actually designed by

the former kind of head

you know another kind of processor