uh so first we want to thank fi and our

our friends in the Kingdom of s for

inviting us here to speak with all of

you thank you very much Gavin and I are

here together because we collaborate

very closely on arti intelligence

Investments because we share the same

thesis uh we have our relationship goes

back more than a decade um started with

uh with Tesla has become very close over

time and I'm going to ask Gavin to start

with our thesis Gavin you tell people

what we think here sure uh so wanted to

Echo all of Antonio's thoughts of

gratitude for towards fi and all of our

hosts for having us here

um I think we share the thesis that when

there is a fundamental platform shift

from the mini computer um to the PC and

the client server um to the mobile phone

to the cloud and now to AI the safest

and best way to invest at the beginning

of that fundamental platform shift is at

the infrastructure layer which I'm going

to talk about and I'll let you talk

about the verticalized apps that come L

later um and the infrastructure layer is

particularly important with AI

because a GPU which two years ago not

many people have heard of um I'm a gamer

so I had heard of it and I I was the

Nvidia analyst back in the year 2000 so

I've heard of it for a long time GPU is

about half the size of this iPhone it is

made from Sand um Taiwan Simi does

amazing things to that sand and sells it

to Nvidia for $7 to $800 that die Nvidia

then sells that for $50,000 companies

like Microsoft and

Amazon and Google um and that is the

probably the biggest markup in human

history significantly bigger than the

birken bag for instance and um selling

drugs yes like narcotics yes for nerds

narcotics for nerds is a great way to

describe it exactly um and if you think

of Jensen likes to call Data Centers AI

factories I think maybe an easier more

fasile analogy is to think of them as

restaurants and the GPU that

$50,000 uh chip half the size of an

iPhone is a star chef and it has gotten

50 times faster in the last four years

which is a big part of why we're living

through this AI

Revolution the problem is you know it's

it's got a convection oven you know all

sorts of things it's much faster all the

other ingredients in the restaurant all

the other parts in that um supply chain

have only gotten two or three times f

faster so the food needs to be delivered

the SF needs to prepare it um and

because the GPU has gotten so much

faster it is doing nothing most of the

time 70% of the time it's doing nothing

it's just drawing power so our kind of

shared core infrastructure level thesis

is we're investing in things that

increase the GPU utilization rate

because if you can take that GPU from uh

being utilized 30% of the time to 60% of

the time you've doubled the output of

that restaurant or that AI Factory and

this is particularly important for AI

because we've gotten used to a world

where compute and computation is

essentially free everybody in this room

is an investor has heard that software

has zero marginal cost and it has almost

no marginal cost to produce even in the

cloud this is not true of AI every time

you ask chat GPT or Gro a question a lot

of comput is happening and it's

extremely expensive um and the

bottleneck is about to shift from GPU

availability to power availability it it

is possible that we will face a global

power shortage in the next few years

because of ai's ever increasing demands

for power and a great way to deal with

that is right now that GPU

is consuming full power 100% of the time

doing nothing 70% of the time and if you

can drive the utilization up that's

going to be great and so we've have a

lot of shared investments in kind of

storage networking and memory

technologies that improve the

utilization rate of those gpus Tonio

maybe you want to talk about the other

part of our I mean Gavin's background

stemies has been super useful to us and

you know we have a long background our

intelligence States back Deep Mind out

I'm going to go quickly because we're

going to run out of time otherwise um so

our our thesis is inform what we do and

what we don't do so we are investing in

we call verticalized applications these

are applications that have uh very

definitive data modes and reinforcement

learning that either exist because the

software is being used in a way that

reinforces the the data or it's piece of

hard or robotics that reinforces that

learning as well we think that the only

valuable Assets in the entire stack are

going to ultimately that have margin

that can be uh that can create we call

it a data mode that creates large margin

going Downstream is uh the data and the

reinforcement learning those are the

most important parts the models will

basically be to zero and because the

models are free so to Gavin's point the

data center is a commodity energy you

know might be a shortage for a while but

is also a commodity uh the models are

free in that entire supply chain all we

really see from end to end that has real

value is going to be in the data mode

and reinforcement learning we're

investing companies in that space so

this is about $2 companies so Gavin so

people want to know what we think what

you think about training our company

I'll do a public since you're in the

public business why don't you to do a

couple of privates and and we can trade

back and forth sure um well based on the

framework you just outline outlined I do

think if you're a foundation model

company and you do not are not attached

to a major internet platform for

distribution and reinforcement learning

and you don't have proprietary data you

your value is going to ASM toote to zero

very quickly but if you're a foundation

model company attached to a giant

internet platform with proprietary data

I think you're going to be immensely

valuable there are only a few of these

um there's Google um but it's already a

multi-trillion dollar company there's

Microsoft and open AI multi trillion

dollar Plus 90 billion and meta has

chosen to make their AI available um for

free open source the only one left is

xai and I submit to you with I think

it's a reasonably high confidence um

hypothesis that the sum of xai and X

will be worth more than a trillion

dollars because it has some of the most

valuable proprietary data in the world

for training these models and a lot of

distribution um to help them to help

them learn uh as far as a public company

I agree with you Gavin by the way um

public company I'm going at the risk of

being accused of bias and I have a large

position this in this in this company so

I'll say that up front I I think Tesla

will be the next public company you see

p $2 Enterprise value for AR

intelligence as an AI company not as an

EV company but as or a Mobility company

but as an AI company because uh people

underestimate the impact of FSD version

12 when you use this product you will

see it drives like a human and that will

blow you away and then at some point

you've seen videos of Optimus and you I

think people sort of they they under

they undervalue or don't believe quite

what's happening with Optimus when you

see the videos but I've actually seen

the robot itself operate it is

extraordinary and imagine um when you

have the most complex thing a human does

is drive a car so we have a model at at

Tesla knows how to drive a car and does

that well like a human you have a robot

that operates like a human uh you put

those things together you get a humanoid

Stout robot that then you can deploy

into a factory and get reinforcement

learning so that that lock in between

data data mode which is UN by the way

does unable data mode in humans driving

a car plus the lockin of a robot working

in a factory doing all the things we do

in a factory right what's the torque on

the bolt how do I not run the gabin

while I'm walking by him where's the

bathroom whatever robots do right the

those things um are going to inform our

training data by reinforcement learning

of the robot itself that will blow

people away and I think you'll see that

you know I don't know exactly when but

let's say in the next year or two uh

this will become more obvious and it

will accelerate the company people will

really see it as an AI company which

which is what it is that's my view Ai Ai

and Robotics yes and Robotics yes yeah

um great so the future of uh of AI what

what do you think happens next uh um

well I listened to this discussion this

morning about regulation with interest

um and I think this is a very very

dangerous topic um what we want as

humans is a world where there are many

many different AIS um we want a

multipolarity of AIS we don't want to

live in a world where there's one AI

superpower are two AI superpowers so I

do think in addition to those kind of

you know those three the grock uh xai

Google um open AI Microsoft it's great

that meta has open sourced their models

it's great that MRA exists it's great

that Google just open sourced a model

the most dangerous world for humans and

I would just say you know if you run any

survey of the top thousand people

working on AI and they call it P Doom

probability of Doom probability that AI

is an extinction level event for

Humanity they can consistently pull it

around it's a 15% chance so that's

pretty high now there's an 85% chance

it's it's amazing um but the most

important thing to lower the odds of the

15% and improve the odds of the 85%

being awesome are that we have many many

AIS which increases the odds that at

least some of them are friendly to us um

has humans um and I would say I do think

in terms of kind of the impact of AI to

build on some of Antonio's comments

about um Tesla's Optimus robot and

Robotics in general there has been a lot

of focus on the impact of AI on

knowledge work and white collar work I

think when we see AI um and llm dropped

into a very functional humanoid robot

there may be an immense impact on blueco

collar work manual labor um and this may

you know it may seem very funny to

historians that anyone was worried about

inflation in front of this kind of tidal

wave of kind of uh deflation I also

think um having embodied AI in a robot

may be very um important to them

progressing becoming sentient becoming

conscious um there there's a lot of

thought that it is hard for an

intelligence to become conscious without

being embodied but Antonio what are your

what are your thoughts I mean I I I

agree with you with h with a lower

probability set I hope on on extinction

level event um you know I had these red

socks now if you notice these are my

Star Trek socks okay because I believe

in a Star Trek future um as opposed to

DET Terminator future and I think we are

here to determine that that happens

events like this are very important

because we're having these dialogues

this morning you know Michael Dale

mentioned this which is the models we

build as humans will be imbued with the

values of their creators and you know I

started my career um wanting to be a

development Economist particularly in

Latin America and worked at Goldman

Sachs in that area and it's um the

questions about global South and Global

North that we have here at F5 I think

are very important this is a period of

time where this particular technology

can enable the rest of us the people

that are live in the global South to

equalize I think their living standard

in a way that doesn't destroy the planet

because productivity will fundamentally

change um it also is a time when llms

that are the powerful models you're

talking about uh they will be created in

individual languages so every country

represented here whether it's Arabic um

if you're speaking Spanish whatever

language we're in Miami I'll say Spanish

but Spanish and Arabic the two languages

there uh the the models we train in that

in those languages with respect for your

cultures respect for for the local

cultures local values and if those

models are rational they should

cooperate like in Star Trek they should

actually cooperate if they cooperate and

they respect local cultures local

languages then I think we end up in a

Star Trek future we end in a place where

humans are better better off as a group

in total and the the parts of the world

been left behind they can access

technology it's very easy these miles

are basically free and you can train

them in local languages right we Falcon

and Arabic is a good example of this um

this is why I think it's mostly going to

be benign I think people are overvaluing

the possibility of a very negative

future and Gavin is right the key to a

positive future is multimodel I call

multipolarity so that each country and

each individual group can train their

models in their ways with their values

to protect them and also to cooperate

other models that maybe trans with other

values and if they're told to cooperate

the machines will cooperate that's what

they told to do and you know the

ultimate multipolarity is where each

human has its own AI that is you know

trained on our own personality on our

own values and is definitionally

friendly to us and I do

think AI may eventually be mediated

through brain computer interfaces um and

we may co-evolve rather than having kind

of you know AI embodied in robots is

probably a first step um and and that's

a good step I think on balance for the

world but I think a future where we

co-evolve as a species with AIS and we

all have our own personal individual AI

that would be an awesome future and that

would that would that would be the

ultimate Star Trek future we're going

for Star Trek we're going for Star Tre

all right gin thank you thanks thanks

guys thanks thanks everybody thanks

brother