Digital Disruption

00;00;01;16 - 00;00;04;27

I think the economics of it is

is not very clear at this point.

00;00;04;27 - 00;00;08;09

So where the dust settles is

is to be seen.

00;00;08;11 - 00;00;10;16

But that is the challenge.

00;00;10;16 - 00;00;14;17

That will probably start

a domino effect of correction over time

00;00;14;17 - 00;00;18;04

as the power acquisition stalls.

00;00;18;11 - 00;00;23;10

As the, you know, the many states

in the US and other countries globally

00;00;23;20 - 00;00;27;01

push back on these mega-projects

to deploy data centers.

00;00;27;02 - 00;00;30;24

Hyperscalers will push out these $200

billion CapEx cycles out

00;00;30;25 - 00;00;32;27

versus

deploying all of that money each year,

00;00;32;27 - 00;00;36;14

and that will then have a trickle down

effect on the revenue, the bookings

00;00;36;14 - 00;00;40;22

and the quarterly performance, which will

then start the stock correction.

00;00;42;04 - 00;00;45;09

This is a show about the future of tech

and the future of work.

00;00;45;10 - 00;00;46;16

I'm Jeff Nielsen

00;00;46;16 - 00;00;50;05

and today we're diving into the physical

infrastructure shaping AI.

00;00;50;09 - 00;00;56;08

From compute and microchips to power to

robotics, edge AI and industrial systems.

00;00;56;09 - 00;01;01;11

My guest today is Anker Saxena, investment

director at TDK ventures.

00;01;01;15 - 00;01;05;22

Anker has well over a decade of experience

in venture capital investing in AI,

00;01;05;24 - 00;01;09;03

including a key role at Cisco Investments,

and has a deep

00;01;09;04 - 00;01;12;04

understanding of the entire AI ecosystem.

00;01;12;07 - 00;01;16;21

I've found that the layers of AI beyond

software are completely underexplored,

00;01;16;21 - 00;01;20;23

so I want Anker to take us on a deeper

dive through the whole stack.

00;01;21;00 - 00;01;24;07

How should we be thinking

about the technology holistically?

00;01;24;11 - 00;01;28;11

What are the implications on our current

AI boom or AI bubble?

00;01;28;11 - 00;01;32;12

And does it even matter at all for people

living outside of Big Tech?

00;01;32;24 - 00;01;34;01

Let's find out.

00;01;37;02 - 00;01;38;06

Thanks so much for joining us today.

00;01;38;06 - 00;01;39;12

Really excited to have you.

00;01;39;12 - 00;01;42;29

I'm curious, just before we kind of

dive into things from your perspective.

00;01;43;00 - 00;01;47;02

You know, I wanted to hear a little bit

about your outlook on the future of AI,

00;01;47;03 - 00;01;50;08

the future of technology

over the next handful of years.

00;01;50;08 - 00;01;54;04

Where do you see this space evolving to,

and what has you

00;01;54;04 - 00;01;57;09

most excited and concerned

about the current environment here?

00;01;57;28 - 00;01;59;23

Yeah, Jeff, happy to be here.

00;01;59;23 - 00;02;02;13

First of all, thank you for having me.

00;02;02;13 - 00;02;04;07

Future of AI is pretty bright.

00;02;04;07 - 00;02;08;01

I've been monitoring this market,

actively investing in this market

00;02;08;01 - 00;02;11;03

for more than ten years now, and we are

00;02;11;13 - 00;02;14;13

what seems to be

like the second innings of AI.

00;02;14;18 - 00;02;18;14

The first one was more around

the chips and infrastructure.

00;02;18;19 - 00;02;21;15

Now we are seeing a way more secular

00;02;21;15 - 00;02;26;10

adoption, and the conversation has shifted

from just startups talking about

00;02;26;11 - 00;02;30;22

AI and impact of AI to governments,

to enterprise customers, to consumers

00;02;30;22 - 00;02;33;23

like you and me

actively using AI in day to day life.

00;02;35;06 - 00;02;35;25

Having said

00;02;35;25 - 00;02;39;02

that, the current investment cycle

00;02;39;02 - 00;02;42;26

seems to be a little inflationary,

hyped up kind of cycle.

00;02;42;26 - 00;02;45;09

So we can we can talk about that.

00;02;45;09 - 00;02;50;01

And like a lot of other technologies,

humans tend to overestimate the impact

00;02;50;01 - 00;02;54;02

of a technology in the short term

but underestimated in the long run.

00;02;54;04 - 00;02;57;26

So we can maybe unpack both sides,

but I feel like it's

00;02;57;26 - 00;03;01;09

an exciting time to be investing in AI

and also to be as a consumer of.

00;03;01;10 - 00;03;03;26

I think it's a very, very exciting time.

00;03;03;26 - 00;03;07;01

I think we're seeing a significant

00;03;07;16 - 00;03;10;29

step up in productivity gains

leveraging AI.

00;03;12;02 - 00;03;14;07

So there's a lot of promise, and I'm

00;03;14;07 - 00;03;17;07

particularly very excited about the space.

00;03;17;13 - 00;03;20;03

I'm interested in.

00;03;20;03 - 00;03;22;27

You know, the fact that you're describing

the investment environment

00;03;22;27 - 00;03;25;22

as sort of inflationary here

and you know, that we're maybe

00;03;25;22 - 00;03;28;07

even over investing in it,

not because I disagree with you.

00;03;28;07 - 00;03;30;26

I think that's probably right,

because it's interesting to me

00;03;30;26 - 00;03;33;26

that even on the investment

side, being as close to it as you are,

00;03;34;03 - 00;03;35;20

that that you're noticing this.

00;03;35;20 - 00;03;38;28

And so, you know,

what are the ramifications of that.

00;03;38;29 - 00;03;42;05

Where does the investment environment

go from here?

00;03;42;06 - 00;03;45;25

Do you believe we're sort of over

invested in the space?

00;03;45;25 - 00;03;49;26

And what does that mean in terms

of what you look for in the investments

00;03;49;26 - 00;03;50;24

you're looking to make?

00;03;51;26 - 00;03;54;05

Yeah, I think

00;03;54;05 - 00;03;54;15

on the

00;03;54;15 - 00;03;58;18

positive side, it's actually good that

a lot of investment is going in this space

00;03;58;18 - 00;04;03;18

because we do need all the kind

of capital support that we can provide

00;04;03;20 - 00;04;07;19

to entrepreneurs who are really pioneering

innovation in key

00;04;07;19 - 00;04;11;18

building blocks of AI within software,

hardware or boats.

00;04;12;25 - 00;04;15;23

But there is a lot of crowding

that is happening.

00;04;15;23 - 00;04;16;00

Right?

00;04;16;00 - 00;04;21;14

And all of this is what never really

invest in deep tech areas or tech areas.

00;04;21;24 - 00;04;24;20

They are now investing in AI

because they don't

00;04;24;20 - 00;04;28;13

want to miss out on

what is the next gold rush, right?

00;04;28;14 - 00;04;30;03

They want to play an active part.

00;04;30;03 - 00;04;32;14

They want to have an investment exposure.

00;04;32;14 - 00;04;39;04

So this kind of uninformed

or less informed capital

00;04;39;09 - 00;04;44;27

that's in getting invested in the market,

that is the thing I have issued, right.

00;04;44;29 - 00;04;46;21

More capital is long term.

00;04;46;21 - 00;04;49;21

But what I'm also seeing is in

00;04;49;22 - 00;04;52;28

a King Kingman ship, kind of, you know,

00;04;52;29 - 00;04;58;06

effort that's going on from a non obsess,

where typically they do not less than AI

00;04;58;13 - 00;05;02;20

or at least in the AI infrastructure

layer, but now they are deploying

00;05;02;21 - 00;05;06;26

hundreds of millions of dollars of capital

in companies at a very early stage.

00;05;06;27 - 00;05;12;24

And the round sizes are just gotten

out of hand and they become really big.

00;05;12;25 - 00;05;14;29

And a lot of companies

are actually raising

00;05;14;29 - 00;05;18;23

significant capital

without actual product now,

00;05;20;05 - 00;05;21;11

let alone productivity.

00;05;21;11 - 00;05;25;09

And the value of also

just getting out of hand.

00;05;25;19 - 00;05;26;18

Right.

00;05;26;18 - 00;05;29;17

People think that

00;05;29;17 - 00;05;31;27

this invest AI investment supercycle

00;05;31;27 - 00;05;35;04

will lead to significant exit outcomes.

00;05;35;04 - 00;05;38;04

We're not talking about billions

or tens of billions anymore.

00;05;38;16 - 00;05;41;07

We're talking about hundreds of billions

or trillions of dollars

00;05;41;07 - 00;05;45;15

of of exit thesis being underwritten

by by a lot of investors.

00;05;45;15 - 00;05;49;16

So they are looking at these

as compounding platforms, your sort

00;05;49;16 - 00;05;53;29

of next generation Googles or Meta's

or Amazon's in the making.

00;05;53;29 - 00;05;59;18

And with that kind of lens in mind,

they are investing a lot of money

00;06;00;15 - 00;06;04;10

at very early stages, which is actually,

00;06;04;10 - 00;06;09;19

I would say, doing the contrary work

of instead of supporting the ecosystem,

00;06;09;19 - 00;06;13;26

it's actually crowding out

a lot of the other players, other,

00;06;13;26 - 00;06;15;17

other startups that may not have access

00;06;15;17 - 00;06;18;17

to the same capital

but might still be doing fundamentally

00;06;19;10 - 00;06;22;17

reasonable work

and might be actually waiting.

00;06;23;12 - 00;06;26;20

So this this we see

strategy of concentrating capital

00;06;26;20 - 00;06;30;05

in a handful of companies just because,

you know, they look promising.

00;06;30;05 - 00;06;33;24

Or maybe it's their second venture

or card venture

00;06;34;03 - 00;06;36;23

that is not great

for the overall ecosystem.

00;06;36;23 - 00;06;38;19

So I'm a little concerned about that.

00;06;38;19 - 00;06;41;27

And this inflationary environment again

has ramifications

00;06;41;27 - 00;06;44;27

in terms of how many such companies

can actually exist

00;06;44;29 - 00;06;48;15

in exit at tens of billions

or hundreds of billions of valuations.

00;06;48;20 - 00;06;49;05

Right.

00;06;49;05 - 00;06;52;28

Is the public

market, is the acquisition market really,

00;06;53;24 - 00;06;56;17

you know, prepared

for those kind of liquidity events?

00;06;56;17 - 00;06;58;21

Probably not. Right.

00;06;58;21 - 00;07;04;13

So over the next few years there will be

some sort of correction in the market.

00;07;04;13 - 00;07;07;18

It may start in the public markets. Right.

00;07;07;19 - 00;07;11;08

Which is another another, you know,

kind of discussion topic in itself.

00;07;11;09 - 00;07;15;22

I probably markets have also seen the

AI stocks, the memory stocks have

00;07;16;03 - 00;07;20;12

have you know, there

isn't a lot over the last 12 to 18 months.

00;07;20;17 - 00;07;21;02

Right.

00;07;21;02 - 00;07;25;04

On the hoax of this

multi-decade AI supercycle.

00;07;26;04 - 00;07;29;11

Is that all

really justified in the next few years,

00;07;29;12 - 00;07;34;14

or is there a lot of reform and hype

that's kind of driving these valuations?

00;07;34;14 - 00;07;39;03

That's that's another kind of point

that's, you know, debatable..

00;07;40;20 - 00;07;42;08

If you're a technology leader.

00;07;42;08 - 00;07;45;26

Infotech is a name you need to know

for getting ahead with AI.

00;07;46;07 - 00;07;49;18

Infotech helps leaders

deliver on top technology initiatives.

00;07;49;18 - 00;07;54;13

And one of our most popular programs

right now is AI prototyping.

00;07;54;16 - 00;07;58;15

We've helped hundreds of organizations

design and build a genetic

00;07;58;16 - 00;08;01;29

AI prototypes using our step by step

methodology.

00;08;02;02 - 00;08;05;19

Infotech road tested

approach gives you a trusted blueprint

00;08;05;19 - 00;08;08;19

to execute faster

and avoid costly mistakes.

00;08;08;19 - 00;08;13;04

And what you walk away with is a prototype

tied to real business outcomes,

00;08;13;04 - 00;08;17;25

integrated with your systems and built

with the right guardrails from day one.

00;08;18;02 - 00;08;21;04

What sets Infotech

apart is how we deliver.

00;08;21;09 - 00;08;25;11

Unlike traditional researcher

AI tools, we don't stop at guidance.

00;08;25;11 - 00;08;28;26

Infotech works with your team

to build the prototype.

00;08;29;00 - 00;08;33;05

You can do this through a focused one week

workshop, ongoing analyst support,

00;08;33;05 - 00;08;37;23

or by using our full methodology

and toolkit to get it done yourself.

00;08;37;28 - 00;08;42;07

The outcome is a working prototype

and a team that's ready to scale it.

00;08;42;13 - 00;08;45;09

To learn more about our agenda

AI prototyping.

00;08;45;09 - 00;08;49;02

Check out the link below

and don't forget to like and subscribe!

00;08;50;09 - 00;08;50;18

You know,

00;08;50;18 - 00;08;54;02

anchor from from my perspective,

one of the trends that we're seeing,

00;08;54;13 - 00;08;58;07

as you mentioned

is there's this sort of gold rush

00;08;58;08 - 00;09;01;15

around, this sense

that this is a once in a generation

00;09;01;16 - 00;09;06;01

opportunity to shift the environment

and create something of value.

00;09;06;01 - 00;09;09;24

And to your point about the public markets

and this sort of,

00;09;10;06 - 00;09;13;25

you know, for lack of a better word,

gold rush mentality, it feels like there's

00;09;13;26 - 00;09;17;27

there's sort of this greed here

as investors,

00;09;17;28 - 00;09;21;19

you know, unsophisticated investors

look for a place to park their money

00;09;21;19 - 00;09;23;20

where they think they're going

to get outsized returns.

00;09;23;20 - 00;09;30;05

And one of the themes I've seen is

it seems like as this market

00;09;30;05 - 00;09;34;15

and these technologies mature, it's

becoming obvious that some of the usual

00;09;34;15 - 00;09;39;00

suspects are no longer in a position

where we can really expect them

00;09;39;00 - 00;09;42;17

to have 100 X growth, and investors

00;09;42;17 - 00;09;46;29

are starting to very quickly and,

you know, almost with minimal due

00;09;46;29 - 00;09;50;20

diligence, turn over rocks and ask

like, well, what is the next big,

00;09;51;02 - 00;09;52;03

you know, scaler here?

00;09;52;03 - 00;09;56;24

And whether that's looking in semiconductors or microchips, hardware, software,

00;09;57;27 - 00;09;58;12

just kind of

00;09;58;12 - 00;10;02;01

this frantic search for crazy returns.

00;10;02;04 - 00;10;03;18

Is that your read as well?

00;10;03;18 - 00;10;06;02

And how do you see this

environment playing out?

00;10;06;02 - 00;10;09;08

Is there necessarily

going to be a correction here

00;10;09;10 - 00;10;13;14

as investors realize, okay, there's

probably not a silver bullet?

00;10;13;14 - 00;10;16;14

Or do you still believe

there are silver bullets to be found?

00;10;17;00 - 00;10;18;27

Yeah, there are no silver bullets.

00;10;18;27 - 00;10;21;27

First of all, to just kind of address

that point right away.

00;10;23;00 - 00;10;25;10

Fundamentally AI

00;10;25;10 - 00;10;28;08

market is a really big growth area.

00;10;28;08 - 00;10;31;13

But are the valuations justified

00;10;31;13 - 00;10;34;28

in terms of actual revenue

delivery over the next 3 to 5 years?

00;10;34;29 - 00;10;37;09

That is questionable, right.

00;10;37;09 - 00;10;41;06

And the kind of massive infrastructure

rollouts, the CapEx,

00;10;41;14 - 00;10;44;10

you know, expenditure of these,

some of these big hyperscalers

00;10;44;10 - 00;10;47;27

that are spending close to $200 billion

each year, building

00;10;48;02 - 00;10;51;12

data centers all over the world,

and particularly in the US.

00;10;52;06 - 00;10;56;24

Is all of that going to generate

economic returns for those hyperscalers?

00;10;56;24 - 00;10;58;19

That is questionable.

00;10;58;19 - 00;11;02;27

Having said that,

they're also in this prisoner's dilemma

00;11;02;29 - 00;11;08;15

of, you know, they have to actively invest

in the AI infrastructure.

00;11;08;16 - 00;11;11;27

If they lag, it will fall behind.

00;11;11;29 - 00;11;14;28

Then they just basically, you know,

00;11;14;28 - 00;11;19;00

fall behind significantly

and then they lose the market share,

00;11;19;08 - 00;11;23;06

then they lose the investment dollars

that are going in the stock.

00;11;23;09 - 00;11;25;05

Right.

So there is this prisoner's dilemma.

00;11;25;05 - 00;11;30;03

And everyone has to keep on betting more

and more capital expenditure money into

00;11;30;04 - 00;11;34;22

AI infrastructure, which is supporting

all these venture capital investments.

00;11;34;22 - 00;11;38;27

Because if Google, meta, Amazon

and others are going to continue

00;11;38;27 - 00;11;42;13

to deploy $200 billion of CapEx,

where does that CapEx going

00;11;42;14 - 00;11;46;24

that is going in buying power

infrastructure, semiconductor chips,

00;11;46;25 - 00;11;50;03

data center networking

and the whole sort of fiber rollout

00;11;50;26 - 00;11;54;07

and then the software layer

and all of that is getting funded

00;11;54;07 - 00;11;58;23

from these approximately $2 billion

per hyperscale or kind of annual CapEx.

00;11;58;24 - 00;11;59;10

Right?

00;11;59;10 - 00;12;04;29

So that is kind of the whole trickle down

effect of how this hyperscale investment

00;12;04;29 - 00;12;08;23

cycle is driving the venture investment

cycle in this whole gold rush team.

00;12;09;13 - 00;12;11;06

Now, where the disconnect is.

00;12;13;07 - 00;12;13;24

The physics

00;12;13;24 - 00;12;16;24

of it,

right, or the real estate of it, right.

00;12;16;24 - 00;12;22;14

The physics aspect is

we only think about ChatGPT Claude,

00;12;22;14 - 00;12;27;03

all all the AI investments

or AI innovations happening in software,

00;12;27;03 - 00;12;31;21

but where the real bottleneck

is, is, is the physics of it, right.

00;12;31;22 - 00;12;37;03

The current GPU infrastructure

is not the most our efficient, energy

00;12;37;03 - 00;12;42;26

efficient, cost efficient infrastructure

to be deployed at scale, right?

00;12;43;01 - 00;12;44;08

It is being deployed right now

00;12;44;08 - 00;12;47;24

because it is the only chip

that is available in the market at scale,

00;12;48;19 - 00;12;50;26

but it is not the most

00;12;50;26 - 00;12;55;07

economically efficient

infrastructure compute infrastructure.

00;12;55;09 - 00;12;57;13

Right. So the physics is the bottleneck.

00;12;57;13 - 00;13;00;19

And where that physical constraint

will come into

00;13;00;21 - 00;13;04;05

picture is the power acquisition,

the power delivery.

00;13;04;06 - 00;13;04;14

Right.

00;13;04;14 - 00;13;06;16

These hyperscalers,

wherever they're building

00;13;06;16 - 00;13;10;08

these big data centers, they need access

to cheap and sustainable power.

00;13;10;18 - 00;13;14;14

And a lot of regions in the US

have started opposing

00;13;15;07 - 00;13;19;28

the use of natural resources,

their power grid, to power data centers.

00;13;20;01 - 00;13;24;07

I think the economics of it

is, is not very clear at this point where

00;13;24;26 - 00;13;29;01

hyperscalers are claiming

that it will lead to economic boom,

00;13;29;01 - 00;13;33;02

more jobs in those regions,

but it will likely come at a cost of,

00;13;33;11 - 00;13;38;18

you know, more pollution of natural

resources, water and air, and also,

00;13;38;22 - 00;13;42;13

you know, increase in power costs for the

the local consumers there.

00;13;42;15 - 00;13;42;27

Right.

00;13;42;27 - 00;13;46;14

So where the dust settles is

is to be seen.

00;13;46;20 - 00;13;48;26

But that is the challenge

00;13;48;26 - 00;13;52;21

that will probably start

a domino effect of correction over time.

00;13;52;23 - 00;13;53;07

Right.

00;13;53;07 - 00;13;58;06

As the power acquisition stalls, as the,

00;13;58;09 - 00;14;02;10

you know, the many states in the US

and other countries globally

00;14;02;20 - 00;14;06;11

push back on these mega-projects

to deploy data centers.

00;14;06;17 - 00;14;10;10

You know, Nvidia's GPU chips

will you know, the demand for

00;14;10;10 - 00;14;11;25

it will reduce, right.

00;14;11;25 - 00;14;15;19

Hyperscalers will push out these $200

billion CapEx cycles out

00;14;15;19 - 00;14;17;22

versus

deploying all of that money each year.

00;14;17;22 - 00;14;21;08

And that will then have a trickle down

effect on their revenues, the bookings

00;14;21;08 - 00;14;25;16

and the quarterly performance, which will

then start the stock correction.

00;14;26;08 - 00;14;30;02

The other thing is the chip supply,

the memory supply that is also severely

00;14;30;02 - 00;14;30;25

constrained, right.

00;14;30;25 - 00;14;34;27

Memory prices

have more than quadrupled in the market.

00;14;35;16 - 00;14;38;04

You know, Apple

and others have have decided to pass

00;14;38;04 - 00;14;41;04

on some of those increased memory costs

to consumers

00;14;41;06 - 00;14;45;01

in the form of price increases for iPhones

and tablets and other products.

00;14;45;05 - 00;14;49;00

I think that is not very sustainable,

right?

00;14;49;03 - 00;14;52;05

The consumers will only absorb

so much cost, right?

00;14;52;06 - 00;14;55;23

So I think that the chip supply,

the memory supply,

00;14;55;24 - 00;15;00;09

all of all of these physical products,

you know, their supply is constrained,

00;15;00;10 - 00;15;03;25

their actual delivery is constrained,

their actual deployment is constrained.

00;15;03;25 - 00;15;09;16

That will bottleneck the

the hyped up growth plans of hyperscalers.

00;15;09;17 - 00;15;09;22

Right.

00;15;09;22 - 00;15;13;17

And that is how the whole connection

cycle religion at some point.

00;15;14;15 - 00;15;18;08

You mentioned, you know, at some point

feels like

00;15;18;10 - 00;15;19;18

it feels like a given these days.

00;15;19;18 - 00;15;21;02

And you mentioned that,

00;15;21;02 - 00;15;24;27

you know, you're not exactly sure

where the dust is going to settle.

00;15;24;29 - 00;15;25;29

But, you know,

00;15;25;29 - 00;15;29;15

can we maybe speculate a little bit about

what are some of the scenarios

00;15;29;15 - 00;15;33;04

about how this plane plays out,

regardless of exact timeline?

00;15;33;08 - 00;15;38;19

What could this look like in terms of the

the impact of that constraints

00;15;38;19 - 00;15;40;28

and just kind of the downstream impacts

00;15;40;28 - 00;15;44;14

in this entire technology

and investment ecosystem?

00;15;45;11 - 00;15;46;09

Yeah.

00;15;46;09 - 00;15;49;06

So think of AI infrastructure

00;15;49;06 - 00;15;52;22

or these data centers

as AI token factories, right?

00;15;52;23 - 00;15;56;01

These are modern factories

where raw inputs go in

00;15;56;01 - 00;15;59;20

and inference,

intelligent inference comes out.

00;16;00;08 - 00;16;00;20

Right.

00;16;00;20 - 00;16;04;12

And like I said earlier,

they are bound by the physics

00;16;04;12 - 00;16;07;21

of the chips, the power,

the thermals, the interconnects.

00;16;08;01 - 00;16;08;26

Right.

00;16;08;26 - 00;16;14;16

Deploying data centers at scale, running

these ChatGPT cloud models at scale,

00;16;14;16 - 00;16;17;22

where you're powering

millions of inference queries every second

00;16;18;22 - 00;16;20;09

is sustainable in the long run.

00;16;20;09 - 00;16;23;23

Only if the economics of this factory

works right.

00;16;23;24 - 00;16;27;24

And these factories are not running at

00;16;27;25 - 00;16;31;13

at the right kind of optimal level right

now, the

00;16;32;17 - 00;16;37;00

the hundreds and thousands of GPUs

that are deployed by the hyperscalers are,

00;16;37;19 - 00;16;41;00

you know, getting powered on, but they're

not actually getting used optimally.

00;16;41;01 - 00;16;44;26

You know, I've read that utilization rates

of a lot of these chips

00;16;44;26 - 00;16;47;27

that are deployed

in data centers is below 30%.

00;16;48;05 - 00;16;51;04

So for the remaining 70%,

they're consuming a lot of power.

00;16;51;04 - 00;16;55;00

They're getting heated up, but they're not

really optimally driving in inference.

00;16;55;00 - 00;16;58;00

So it's like you're running a factory

00;16;58;04 - 00;17;01;08

at only 30% utilization, right?

00;17;01;17 - 00;17;05;20

Over time, investors,

whether it's public market investors

00;17;05;20 - 00;17;06;22

or in private markets,

00;17;06;22 - 00;17;10;15

you know, VCs will start demanding

that there be a sustainable way

00;17;10;20 - 00;17;14;09

where, you know, people will start talking

about what are the margins on these data

00;17;14;10 - 00;17;15;10

centers, right.

00;17;15;10 - 00;17;19;18

Just the money going in and deploying

these data centers is not enough.

00;17;19;20 - 00;17;21;10

What is the payback? Right.

00;17;21;10 - 00;17;23;23

What are the margins

of operating this at scale?

00;17;24;24 - 00;17;26;21

These conversations will start to

00;17;26;21 - 00;17;30;26

to permeate the public market

and the private markets more and more.

00;17;30;27 - 00;17;34;26

And that is when I think this brute force

scaling of just deploy

00;17;34;26 - 00;17;38;24

more GPUs at whatever cost, that brute

force scaling will have to stop.

00;17;38;24 - 00;17;42;24

And people will think about the real

economics of running these factories.

00;17;42;25 - 00;17;43;14

Right?

00;17;43;14 - 00;17;46;02

This is like

the steel mills of 100 years ago, right?

00;17;46;02 - 00;17;50;13

Steel mills were started

near coal mines and rivers

00;17;50;13 - 00;17;54;28

because they wanted access

to water to coal and railroads.

00;17;55;05 - 00;17;58;09

Data centers are also getting started

around the power sources

00;17;58;09 - 00;18;02;05

and and around,

you know, chip availability.

00;18;02;07 - 00;18;04;17

Right. But how do you actually operate

at that scale?

00;18;04;17 - 00;18;08;18

There was only one US steel or, you know,

few steel companies that survived,

00;18;08;19 - 00;18;13;14

right out of maybe 100 others

that were started a century ago.

00;18;14;18 - 00;18;17;21

Economics

was the the factor that basically

00;18;19;02 - 00;18;21;25

was forcing the survival of the fittest.

00;18;21;25 - 00;18;22;09

Right?

00;18;22;09 - 00;18;25;16

And that economics is not something

that's actively debated

00;18;25;16 - 00;18;28;16

right now or actively analyzed right now.

00;18;28;23 - 00;18;31;00

I've started to think about the economics

00;18;31;00 - 00;18;34;05

of this token factory

in the form of five piece.

00;18;35;01 - 00;18;38;02

These are performance, power,

00;18;38;17 - 00;18;42;11

price, privacy, and programmability.

00;18;42;28 - 00;18;46;11

People have so far these hyperscalers,

everyone in the chip

00;18;46;11 - 00;18;49;11

industry mostly focused on performance.

00;18;49;12 - 00;18;49;18

Right.

00;18;49;18 - 00;18;52;25

I can drive so much inference per

second or per watt.

00;18;53;03 - 00;18;56;13

Power is the next gating item

that people have started to realize.

00;18;56;13 - 00;19;00;16

And it's a boardroom conversation

right now in a lot of public companies

00;19;00;16 - 00;19;02;29

that are actively deploying

AI infrastructure, right.

00;19;02;29 - 00;19;06;25

How do we get access to power

cheaply, affordably, and sustainability?

00;19;07;01 - 00;19;10;04

But what is also not talked

about is at what price

00;19;10;04 - 00;19;16;07

and are we impacting the privacy

and programmability refers to?

00;19;16;08 - 00;19;19;02

How do we actually orchestrate this?

00;19;19;02 - 00;19;23;18

This whole AI token factory so that

we drive up the economics, the margins.

00;19;23;19 - 00;19;23;26

Right.

00;19;23;26 - 00;19;27;05

So the other three P's

will will get talked about more and more.

00;19;27;11 - 00;19;31;07

And those are the factors

that will be used to scrutinize

00;19;31;09 - 00;19;36;06

the operations of these big infrastructure

vendors at scale over time.

00;19;36;08 - 00;19;39;06

And I think once they start to falter

on some of these,

00;19;39;06 - 00;19;42;12

that's when I think people will start

to schedule or the markets will start

00;19;42;12 - 00;19;46;20

to settle into more sort of rational view

on valuation,

00;19;46;24 - 00;19;50;28

long term growth

prospects and long term CapEx cycles.

00;19;52;11 - 00;19;53;21

I really like that model.

00;19;53;21 - 00;19;55;25

And I like the

00;19;55;25 - 00;19;58;14

I like kind of the economic lens

on all of this as well.

00;19;58;14 - 00;20;02;14

And so keeping with the theme of

economics, and we can keep it kind of 101

00;20;02;15 - 00;20;02;20

here.

00;20;02;20 - 00;20;02;26

You know,

00;20;02;26 - 00;20;06;25

we've got a supply constraint and

we've got a significant amount of demand.

00;20;07;17 - 00;20;10;13

I'm curious, anchor,

from your perspective, you know,

00;20;10;13 - 00;20;13;17

clearly we've got something that's

unsustainable from a pricing level here.

00;20;13;18 - 00;20;13;24

Right.

00;20;13;24 - 00;20;16;24

Because there's price subsidization

going on.

00;20;16;26 - 00;20;20;09

You know, there's performance increases,

but there's, you know,

00;20;20;10 - 00;20;23;10

a limited amount of supply

00;20;23;17 - 00;20;26;05

in terms of like the next stages here.

00;20;26;05 - 00;20;30;16

Is demand in your mind

going to stay steady

00;20;30;16 - 00;20;35;23

or increased to a degree where it forces

an increase in the amount of supply.

00;20;35;24 - 00;20;38;22

Is demand going to soften?

00;20;38;22 - 00;20;40;12

And, you know,

one of the interesting things

00;20;40;12 - 00;20;45;02

about supply here, and you know a lot more

about this than I do, I'm sure, is,

00;20;45;13 - 00;20;49;19

you know, it's not that easy

to just spin up a new AI infrastructure,

00;20;49;20 - 00;20;50;28

right, to come up with something.

00;20;50;28 - 00;20;54;28

The the lag time on that and the

the upfront investment is significant.

00;20;54;28 - 00;21;00;06

So using that economics lens,

how do you see that playing out.

00;21;01;01 - 00;21;01;23

Yeah.

00;21;03;09 - 00;21;03;20

Yeah.

00;21;03;20 - 00;21;09;00

First of all, I agree that it's not easy

to spin up a data center, right?

00;21;09;01 - 00;21;12;03

That is why a lot of these forward looking

00;21;12;26 - 00;21;15;07

investments are happening

for hyperscalers.

00;21;15;07 - 00;21;17;27

Because again, back to the prisoner's

dilemma point.

00;21;17;27 - 00;21;21;05

They don't want to be the ones

whose, you know, inference

00;21;21;20 - 00;21;25;16

latency is several milliseconds

more than their competitors.

00;21;25;17 - 00;21;26;04

Right.

00;21;26;04 - 00;21;29;09

They want to serve

all these inference queries at the best

00;21;29;09 - 00;21;32;10

possible

throughput and and other criteria.

00;21;32;11 - 00;21;34;14

That's

why they're investing ahead of the curve.

00;21;34;14 - 00;21;35;19

Not talking about the curve.

00;21;35;19 - 00;21;40;07

The demand curve I feel like demand curve

will continue to grow

00;21;40;27 - 00;21;44;24

for the time, you know, for

the foreseeable time, because right now

00;21;45;10 - 00;21;50;09

it is not constrained by costs as much.

00;21;50;12 - 00;21;53;25

Right, because the costs

are being subsidized by the vendor,

00;21;53;25 - 00;21;55;21

because they are in a line graph

situation.

00;21;55;21 - 00;21;58;16

Right. OpenAI.

Now we've done the challenge.

00;21;58;16 - 00;22;02;10

Sorry about infrastructure vendors, but

let's talk about the model vendors, right.

00;22;02;11 - 00;22;06;24

Model vendors are in a land

grab situation themselves.

00;22;07;25 - 00;22;09;05

They cannot differentiate

00;22;09;05 - 00;22;12;20

purely based on the frontier models

that they're building.

00;22;12;27 - 00;22;16;16

They are going to sustain their roads

00;22;16;27 - 00;22;20;11

scale through retention

and through the strong value

00;22;20;11 - 00;22;23;15

proposition of their models,

which will really come with

00;22;24;08 - 00;22;29;17

metrics like can they serve these models

at a lowest latency or the cheapest cost,

00;22;29;20 - 00;22;32;21

which also has infrastructure

ramifications, but then they're

00;22;32;21 - 00;22;36;16

also going to move up the application

stack, right?

00;22;37;04 - 00;22;39;06

So far,

00;22;39;06 - 00;22;42;05

I've started to see enterprise

customers spend

00;22;42;05 - 00;22;45;08

millions of dollars each month on running

00;22;45;08 - 00;22;49;03

inference workloads that I think

00;22;49;03 - 00;22;53;08

is already significant expenditure

for a lot of enterprise customers.

00;22;53;08 - 00;22;57;00

But we've barely touched

the tip of the iceberg here, right?

00;22;57;01 - 00;23;01;23

I think there is so much more potential

first in the digital realm

00;23;01;23 - 00;23;05;12

and then over time in the physical world,

the physical realm of how

00;23;05;13 - 00;23;09;12

AI can actually impact

real physical world interactions.

00;23;09;13 - 00;23;12;13

So I feel like that is the

00;23;13;26 - 00;23;15;09

optimism that's driving

00;23;15;09 - 00;23;18;09

all these investments

by both public market

00;23;18;13 - 00;23;22;04

public companies

as well as venture capital firms.

00;23;22;06 - 00;23;24;21

Right. The demand is

what's keeping all of this going, right?

00;23;24;21 - 00;23;29;07

People are seeing companies

built like never before, right.

00;23;29;19 - 00;23;32;21

Companies in the coding space,

they were attained

00;23;32;21 - 00;23;38;10

billions of dollars of RR when it took 7

to 10 years for the the the poster

00;23;38;10 - 00;23;42;24

children of the SaaS era to attain

billion dollar AR in 7 to 10 years.

00;23;42;24 - 00;23;46;11

Right cursor like companies

have attained it in less than a year.

00;23;46;24 - 00;23;50;13

So there is something

substantially different in this

00;23;50;14 - 00;23;53;19

AI cycle compared to the SaaS cycle

of a decade ago.

00;23;53;27 - 00;23;56;27

That is what is prompting these.

00;23;58;01 - 00;23;58;27

Investments, right?

00;23;58;27 - 00;24;03;12

These super investments in companies

where $100 million rounds are

00;24;03;12 - 00;24;07;10

are so common these days,

when it was so uncommon in the in the SaaS

00;24;07;10 - 00;24;09;16

key, right, was a special thing.

00;24;09;16 - 00;24;14;11

Nowadays, almost every day of the week,

you hear about 100 or 200 million,

00;24;14;12 - 00;24;17;11

maybe even $500 million fundraisers.

00;24;19;12 - 00;24;22;05

Do you think, are we at a pace here

where you see that

00;24;22;05 - 00;24;26;21

continuing indefinitely into the future,

at least for the medium term?

00;24;26;21 - 00;24;29;06

Or are we coming

close to a reckoning point?

00;24;29;06 - 00;24;33;02

And I guess the, you know, getting back

to the underlying sort of financial

00;24;33;02 - 00;24;37;08

physics of that, are those valuations too

00;24;37;08 - 00;24;40;08

rich, just based on investors

00;24;40;09 - 00;24;43;09

wanting a taste of this gold rush?

00;24;44;12 - 00;24;46;10

I think at some point

00;24;46;10 - 00;24;49;17

the valuations will revert to

00;24;49;18 - 00;24;52;20

more reasonable SaaS like valuations.

00;24;52;24 - 00;24;56;00

But again, at some point is is indecisive.

00;24;56;00 - 00;24;59;25

So I would say at least in the next

few years I expect this to happen, right?

00;24;59;26 - 00;25;01;15

Not ten years away.

00;25;01;15 - 00;25;03;27

Next few years.

00;25;03;27 - 00;25;08;15

Having said that,

like Cloud anthropic and OpenAI,

00;25;08;16 - 00;25;11;27

their scale that they have attained

in the last 2 or 3 years is just

00;25;12;01 - 00;25;16;07

completely unimaginable, right from wasas

era, right?

00;25;16;10 - 00;25;19;10

Nobody was growing from

00;25;19;20 - 00;25;22;17

2 billion to 10 billion in a year, right?

00;25;22;17 - 00;25;26;01

Very few companies attained

even 10 billion scale after ten years.

00;25;26;02 - 00;25;26;19

Right.

00;25;26;19 - 00;25;31;05

But these companies have attended

in two years or three years of operation.

00;25;31;28 - 00;25;34;28

So they do deserve more.

00;25;35;03 - 00;25;39;27

I mean, a rich valuation,

but over time, the underlying physics

00;25;39;27 - 00;25;43;28

and the economics

will bring their growth rates down.

00;25;44;16 - 00;25;49;18

There will be extra scrutiny or not

just the growth rates, but on the margins.

00;25;49;20 - 00;25;51;08

Are you actually delivering these

00;25;51;08 - 00;25;52;07

astronomical growth

00;25;52;07 - 00;25;55;28

rates at sustainable margins, or

are you actually just funding your growth

00;25;55;28 - 00;25;59;19

rate through these losses and you're

just raising money for external investor?

00;25;59;19 - 00;26;03;19

I think that that will be the

the reality check that

00;26;03;19 - 00;26;07;15

these companies

will get in the next couple of years. I.

00;26;08;09 - 00;26;11;11

I'm inclined to agree, and it's

going to be very interesting to see

00;26;11;11 - 00;26;15;21

how that plays out and how, you know,

sustainable or not sustainable, that is.

00;26;15;22 - 00;26;19;29

And yeah, I think a lot of enterprises

are in for a rude awakening once

00;26;20;01 - 00;26;23;16

you know, the token subsidization,

you know,

00;26;23;27 - 00;26;26;28

if that starts to go away,

there really going to have to think hard

00;26;26;28 - 00;26;30;17

about, you know, their inference costs

and their inference usage.

00;26;30;20 - 00;26;33;13

I'm curious though, Anker,

with that in mind,

00;26;33;13 - 00;26;35;00

we've talked a little bit

about this demand.

00;26;35;00 - 00;26;37;14

I want to unpack a little bit more,

00;26;37;14 - 00;26;39;21

I guess, the shape of this ecosystem.

00;26;39;21 - 00;26;44;08

So specifically thinking about the AI infrastructure that we've discussed so far,

00;26;45;20 - 00;26;46;13

you've you've

00;26;46;13 - 00;26;49;26

talked about the idea of enterprises

00;26;49;26 - 00;26;52;26

using inference,

being concerned about power,

00;26;53;20 - 00;26;57;28

the the infrastructure sides behind

the scenes.

00;26;58;08 - 00;26;59;27

Is the demand really just coming

00;26;59;27 - 00;27;03;23

from a handful of hyperscalers,

or is this something that most enterprises

00;27;03;24 - 00;27;07;21

need to have on their radar

in a more fulsome way?

00;27;10;01 - 00;27;11;22

You mean demand for infrastructure,

00;27;11;22 - 00;27;14;22

the data center, physical infrastructure,

right?

00;27;14;27 - 00;27;15;10

Yeah.

00;27;15;10 - 00;27;21;09

So this is in sharp

contrast to the pre-Code

00;27;21;10 - 00;27;25;18

era when everything was run on on

premise servers, right?

00;27;25;19 - 00;27;29;18

All the big companies

had these big closets of servers

00;27;29;18 - 00;27;31;18

that were running in their campus

somewhere

00;27;31;18 - 00;27;34;20

or in their enterprise buildings

managed by it.

00;27;35;05 - 00;27;37;26

Right then the cloud basically

00;27;37;26 - 00;27;41;25

took a big portion of that and,

you know, basically made

00;27;41;25 - 00;27;45;06

it available as a public utility

for anybody to rent that capacity

00;27;45;07 - 00;27;48;24

versus managing all these big closets

of compute themselves.

00;27;49;08 - 00;27;52;20

That could still manageable, because

00;27;52;23 - 00;27;56;27

the underlying chips in those servers

were all of the same architecture,

00;27;56;27 - 00;28;01;02

the XT6 architecture that Intel and Dell

and others built.

00;28;01;05 - 00;28;06;00

But now, over the last couple of years,

we have entered a heterogeneous compute

00;28;06;01 - 00;28;09;18

era where it's not just CPUs, there's GPUs

00;28;09;18 - 00;28;13;20

and there's grok

and other type of chips and accelerators.

00;28;13;20 - 00;28;18;03

And frankly, any enterprise,

it is not prepared.

00;28;18;03 - 00;28;21;15

They just don't know how to manage

these different chips, let alone

00;28;21;16 - 00;28;25;02

the whole sort of cooling infrastructure,

the power infrastructure,

00;28;26;06 - 00;28;27;08

you know, some of these data

00;28;27;08 - 00;28;32;04

centers that run GPUs, their temperatures

spike up into several hundreds.

00;28;32;07 - 00;28;33;22

It's like a furnace in there.

00;28;33;22 - 00;28;36;25

How can you run that furnace in the bottom

of the basement of a building?

00;28;36;25 - 00;28;40;04

You cannot write the kind of

cooling infrastructure that is needed.

00;28;40;05 - 00;28;44;12

You can't do that out of a basement

in a big city, right?

00;28;44;13 - 00;28;49;07

So it's just not possible

to do big hyperscale

00;28;49;08 - 00;28;52;08

kind of data centers

in enterprise campuses.

00;28;52;20 - 00;28;57;20

Having said that,

even the hyperscalers are not fully adept.

00;28;57;21 - 00;28;59;14

They are

00;28;59;14 - 00;29;01;15

gradually acquiring

the right kind of talent

00;29;01;15 - 00;29;04;15

and the technology

to run these data centers at scale.

00;29;04;17 - 00;29;06;01

But like I said before,

00;29;06;01 - 00;29;10;05

their utilization rates of these massive

rollouts are less than 30%.

00;29;10;05 - 00;29;11;19

So they've just deployed

00;29;11;19 - 00;29;15;08

massive data centers, but they're not even

optimally using themselves.

00;29;15;17 - 00;29;19;01

And I've not even talked

about the software inference optimization

00;29;19;01 - 00;29;22;07

that's needed on top of these,

these physical chips.

00;29;22;08 - 00;29;23;08

Right.

00;29;23;08 - 00;29;27;17

You need an air traffic controller

kind of software orchestrator

00;29;27;17 - 00;29;30;25

to think about based on the right

workload.

00;29;31;01 - 00;29;36;09

What is the right rack or right data

center to run this specific workload on?

00;29;36;10 - 00;29;40;10

Right, like an air traffic controller

decides amongst a couple of flights

00;29;40;10 - 00;29;43;16

that are all trying to learn

in the next five minutes,

00;29;43;16 - 00;29;47;13

how do I prioritize who lands at

which runway and at what time?

00;29;47;13 - 00;29;48;17

That is the kind of role

00;29;48;17 - 00;29;52;24

that software needs to play,

and that software stack is in the works.

00;29;53;03 - 00;29;56;03

The hyperscalers are also figuring it out.

00;29;56;14 - 00;29;59;09

What they built in the cloud era

00;29;59;09 - 00;30;03;02

using Kubernetes is not applicable

for AI stack.

00;30;03;15 - 00;30;06;13

Everything has to be built

from from scratch.

00;30;06;13 - 00;30;10;26

Almost nothing from the cloud era can be

borrowed for the infrastructure layer.

00;30;10;27 - 00;30;11;20

You need

00;30;12;23 - 00;30;15;28

specific chips, specific cooling,

specific interconnects.

00;30;15;28 - 00;30;19;24

Then you need it's specific tailored

software on top to do all the optimization

00;30;19;24 - 00;30;24;18

and drive that less than 30% utilization

to more than 70% utilization,

00;30;24;19 - 00;30;28;12

while achieving the kind of margins

that you were able to achieve

00;30;28;12 - 00;30;29;24

during the cloud layer.

00;30;29;24 - 00;30;33;11

If you fail to achieve that over

the next five years as a hyperscalers,

00;30;33;19 - 00;30;35;27

you know your stock will suffer, right?

00;30;35;27 - 00;30;39;26

So so that's that's kind of why

I think that

00;30;41;06 - 00;30;44;07

most of the demand for AI infrastructure

00;30;44;09 - 00;30;47;09

would be.

00;30;48;11 - 00;30;50;16

Concentrated in hyperscalers market.

00;30;50;16 - 00;30;53;02

But then there are also these neo

clouds, right.

00;30;53;02 - 00;30;55;05

Like Lambda and Corvi.

00;30;55;05 - 00;30;59;21

When others that were started in the

AI era again because they did

00;30;59;21 - 00;31;04;13

not have the, the, the baggage,

if you will, of the hyperscalers.

00;31;04;13 - 00;31;05;02

Cloudera. Right.

00;31;05;02 - 00;31;07;14

They had no cloud before those new clouds.

00;31;07;14 - 00;31;10;27

They built these GPU clouds

and the software

00;31;10;27 - 00;31;14;16

stack on top of it

to exclusively serve the AI market.

00;31;14;25 - 00;31;18;03

These new clouds don't serve your

your other non eye workloads right.

00;31;18;03 - 00;31;21;03

They don't

you can't rent capacity for non I work

00;31;21;08 - 00;31;23;22

you use them for AI workloads.

00;31;23;22 - 00;31;25;25

That's a big emerging market.

00;31;25;25 - 00;31;30;10

And then there are

these massive capitalized

00;31;30;19 - 00;31;34;18

AI native companies like OpenAI and Tropic

and other companies

00;31;34;18 - 00;31;37;04

that are also building

their own data centers

00;31;37;04 - 00;31;41;06

because they realize that they are

spending so much money on on their,

00;31;41;07 - 00;31;45;08

you know, hyperscale bills that they

might as well just do this themselves.

00;31;45;08 - 00;31;48;10

So they are in market for chips,

for talent,

00;31;48;10 - 00;31;52;14

for power, for someone

who can run this whole operation at scale.

00;31;52;15 - 00;31;54;08

Right. And they're doing this themselves.

00;31;54;08 - 00;31;57;16

And a lot of the chip companies

are now working directly with OpenAI

00;31;57;16 - 00;32;03;07

or anthropic versus working with meta

or Amazon or or Nvidia or other companies.

00;32;04;25 - 00;32;08;03

So, you know, you've described fairly

clearly that there's

00;32;08;03 - 00;32;12;18

this huge opportunity for optimization

here, right?

00;32;12;20 - 00;32;15;21

You call it air traffic control,

sort of a, you know, if I can call it

00;32;15;21 - 00;32;20;08

an infrastructure orchestration layer,

I guess, between the chips

00;32;20;08 - 00;32;25;15

and the actual inference requests here,

what are the implications there?

00;32;25;16 - 00;32;28;10

Who's best position

to take advantage of that?

00;32;28;10 - 00;32;29;29

Is it going to be that chip companies.

00;32;29;29 - 00;32;32;02

Is it going to be companies?

00;32;32;02 - 00;32;33;01

Is it neo clouds?

00;32;33;01 - 00;32;37;24

Is it organizations like OpenAI

or you know, in your view, is that field

00;32;37;25 - 00;32;42;08

sort of wide

open for disruption by new entrants?

00;32;43;11 - 00;32;45;26

Yeah, I feel it's the latter.

00;32;45;26 - 00;32;48;26

The latter, the part that,

00;32;49;01 - 00;32;51;18

you know, I feel that

00;32;51;18 - 00;32;55;02

if this kind of inference control

or inference optimization software

00;32;55;03 - 00;32;57;09

were to come from one of the hyperscalers,

00;32;57;09 - 00;33;01;01

they would naturally be biased towards

their own chips in their own cloud.

00;33;01;03 - 00;33;06;08

We need an unbiased,

a Switzerland of orchestrator

00;33;06;08 - 00;33;08;08

that can orchestrate workloads

across different

00;33;08;08 - 00;33;12;00

clouds, different chips,

different data centers and racks.

00;33;12;05 - 00;33;13;18

Right?

00;33;13;18 - 00;33;16;00

Basically, whenever there's

a big architectural shift,

00;33;16;00 - 00;33;19;00

there is a new control plane

that emerges in the software layer.

00;33;19;01 - 00;33;19;06

Right?

00;33;19;06 - 00;33;22;12

So there was Kubernetes for the cloud era

00;33;22;13 - 00;33;27;10

that virtualized the underlying cloud

environments or the online environments.

00;33;27;11 - 00;33;32;16

There was the VMware hypervisor, right,

that virtualized the bare metal.

00;33;32;22 - 00;33;36;09

There was also the

the software defined networking that

00;33;36;10 - 00;33;39;25

that virtualized the underlying networking

from Cisco and Juniper and others.

00;33;39;27 - 00;33;43;06

Now, in this AI infrastructure,

we need the next

00;33;43;07 - 00;33;46;07

VMware to emerge, right?

00;33;46;07 - 00;33;51;29

We want a unbiased inference control

vendor to emerge in the software layer

00;33;52;00 - 00;33;56;03

that's able to support a variety

of workloads, variety of chips,

00;33;56;03 - 00;33;59;12

and all the underlying hyperscalers

and data center racks.

00;33;59;13 - 00;33;59;21

Right.

00;33;59;21 - 00;34;03;24

So it's it's a Nvidia software like this.

00;34;03;26 - 00;34;07;00

But again,

in reading this, a biased party in this,

00;34;07;05 - 00;34;10;18

and I think right now

for the lack of a better alternative,

00;34;11;04 - 00;34;13;26

their software works

and their software is being used.

00;34;13;26 - 00;34;18;00

But over time we will see the next VMware

in this space emerge.

00;34;18;04 - 00;34;21;03

Big company in this space independently.

00;34;21;19 - 00;34;23;19

So with that in mind,

I want to come back to something

00;34;23;19 - 00;34;25;26

you said earlier,

which is that you believe

00;34;25;26 - 00;34;27;09

there's too much investment

00;34;27;09 - 00;34;30;00

potentially being concentrated

in a few big players here.

00;34;30;00 - 00;34;33;04

And there's, you know,

quite a large number of startups that are,

00;34;33;07 - 00;34;36;21

you know, could be very well positioned

to disrupt the market

00;34;36;21 - 00;34;41;03

and add real value, but that, you know,

investors aren't necessarily seeing that.

00;34;41;03 - 00;34;45;05

So with that framing in mind,

where do you see the investment

00;34;45;06 - 00;34;49;01

opportunities and what do you look for

when you're looking for,

00;34;49;03 - 00;34;52;11

you know, the

next AI startup to invest in?

00;34;54;17 - 00;34;57;01

So each investor has its own

00;34;57;01 - 00;35;01;19

set of constraints fund size, timelines,

risk appetite and all that.

00;35;01;19 - 00;35;03;26

We are all stage deep tech investors.

00;35;03;26 - 00;35;07;17

We invest when the companies are just

starting out with their product idea.

00;35;07;18 - 00;35;09;27

The teams are getting together.

00;35;09;27 - 00;35;11;28

They have a vision, they have a plan.

00;35;11;28 - 00;35;13;27

Hopefully they have some validation

00;35;13;27 - 00;35;17;16

or key industry insight

that sets them apart, right.

00;35;17;21 - 00;35;20;19

And they have to be deep tech,

which means,

00;35;20;19 - 00;35;23;24

you know, we're not investing purely

in AI application companies.

00;35;23;24 - 00;35;25;02

We're investing in the underlying

00;35;25;02 - 00;35;28;02

physics, the underlying

material science innovation companies.

00;35;28;02 - 00;35;31;20

And whenever there is a gold rush,

you invest in the picks

00;35;31;20 - 00;35;34;20

and shovels that help people, you know,

00;35;35;22 - 00;35;39;10

filter the sand out of the water

and find the gold gets there, right.

00;35;39;10 - 00;35;43;14

So basically,

we have invested in the whole building,

00;35;43;21 - 00;35;45;23

set of building blocks

of the modern data center.

00;35;45;23 - 00;35;49;10

And it's the next generation power

delivery generation cooling,

00;35;49;15 - 00;35;52;15

next generation set of chips

and different architectures again

00;35;52;15 - 00;35;55;16

there, which are more power efficient

than Nvidia GPUs.

00;35;55;24 - 00;35;59;13

We've been measured in the the fabric

or the railroads of the steel mills.

00;35;59;13 - 00;35;59;18

Right.

00;35;59;18 - 00;36;02;20

The connecting connectivity

fabric companies.

00;36;03;03 - 00;36;06;03

So basically these kind of companies

versus

00;36;06;07 - 00;36;10;25

a big sort of data center developer,

new cloud, which is a very capital

00;36;10;25 - 00;36;13;05

intensive play and something that

we are not comfortable on

00;36;14;06 - 00;36;15;29

as an investment opportunity.

00;36;15;29 - 00;36;19;23

So that's that's

what we have done on the data center side.

00;36;19;24 - 00;36;22;18

Now, the world of the eye is broader,

right?

00;36;22;18 - 00;36;27;17

There is the physical component as well,

which is how you actually embody this

00;36;27;18 - 00;36;28;26

AI in the real world,

00;36;28;26 - 00;36;32;21

whether it's in the plan of robots

or self-driving cars or drones

00;36;32;21 - 00;36;36;06

or augmented reality

or other forms of physical interaction,

00;36;36;17 - 00;36;39;08

that's another key area of focus for us.

00;36;39;08 - 00;36;42;02

We've invested in several robotics

companies.

00;36;42;02 - 00;36;47;21

We've invested in enablers of robotics

in the form of compute companies

00;36;47;21 - 00;36;49;01

whose chips can actually go

00;36;49;01 - 00;36;53;18

in these robots and other physical

AI machines and help them do

00;36;53;20 - 00;36;57;24

inference on the device itself, versus

depend on connectivity to the cloud.

00;36;58;27 - 00;37;01;07

And we have invested in

00;37;01;07 - 00;37;04;03

also the the enabling manufacturing

and supply

00;37;04;03 - 00;37;05;26

chain companies

and the material science side

00;37;05;26 - 00;37;10;22

that will enable physical AI and

and the AI infrastructure to to grow.

00;37;12;03 - 00;37;13;23

So all of these are

00;37;13;23 - 00;37;17;21

deep tech investments

fundamentally rooted in scientific

00;37;17;21 - 00;37;21;03

innovations started by teams

with deep domain experience.

00;37;21;03 - 00;37;25;04

These are not college dropout

kind of teams like this.

00;37;25;06 - 00;37;30;11

This deep space needs people who have

significant experience, the connections

00;37;30;11 - 00;37;34;16

and the insights to to bring these hard

technologies to market.

00;37;35;26 - 00;37;40;15

So under the banner of deep tech,

we've talked about that kind of hard

00;37;40;17 - 00;37;43;15

AI infrastructure

for most of this conversation,

00;37;43;15 - 00;37;47;15

but we've also now sort of backed

into another category of deep tech,

00;37;47;15 - 00;37;51;03

which I don't know if you would

describe it as edge AI or edge computing,

00;37;51;03 - 00;37;55;01

but, you know, you just you described the,

you know, the computation

00;37;55;01 - 00;37;58;26

or the AI calculations

happening on the device itself.

00;37;59;05 - 00;38;02;25

You know, for listeners who don't

necessarily have a good understanding

00;38;02;25 - 00;38;06;29

of that space, can you talk a little more

about what that looks like,

00;38;06;29 - 00;38;10;02

what the advantages are

and how advanced that space is right now?

00;38;10;26 - 00;38;11;18

Yeah.

00;38;12;26 - 00;38;14;01

So edge AI

00;38;14;01 - 00;38;17;24

is basically running

AI inference on the device itself.

00;38;17;24 - 00;38;20;29

And those devices could be smartphones,

watches,

00;38;21;05 - 00;38;24;19

satellites, robots, industrial machinery.

00;38;25;05 - 00;38;30;26

A lot of times these physical devices

have to take decisions in real time.

00;38;30;27 - 00;38;34;29

And interaction with cloud,

that round trip of signal going

00;38;35;00 - 00;38;38;23

to the cloud for computation in the cloud,

and then coming back with a decision.

00;38;38;23 - 00;38;41;23

That round trip is

00;38;42;00 - 00;38;45;06

very latency heavy, and real time

00;38;45;06 - 00;38;48;17

decisioning is not possible

if you're depending on the cloud.

00;38;48;17 - 00;38;51;27

So you need to add

in real time on the edge itself.

00;38;51;27 - 00;38;55;26

And that's why all the AI related compute

has to happen on the device itself.

00;38;55;27 - 00;38;58;27

Now you can't put a big and media GPU,

00;38;59;11 - 00;39;02;11

let alone a server on a

on a watch or a phone.

00;39;02;11 - 00;39;06;19

So you need a toned down

version of the compute

00;39;06;19 - 00;39;10;16

that is exclusively

tailor made size wise, power wise,

00;39;10;16 - 00;39;15;19

price wise, to be fit into a watch

or a phone or similar device.

00;39;15;20 - 00;39;15;27

Right?

00;39;15;27 - 00;39;20;10

And that's a market

that is a specialty market where you do

00;39;20;10 - 00;39;24;09

have incumbents like in Qualcomm

and other companies.

00;39;24;14 - 00;39;29;02

But these companies have not kept up

with the innovations in the edge AI space.

00;39;29;02 - 00;39;33;15

And that is creating a very clear wedge

for startups

00;39;33;15 - 00;39;37;03

to build

very focused set of chips for the market.

00;39;37;12 - 00;39;38;10

Right.

00;39;38;10 - 00;39;42;09

And the space is actually heating up a lot

00;39;42;10 - 00;39;46;07

with the advent of physical

AI and how you actually enable machines

00;39;46;07 - 00;39;48;28

to interact with humans

and also do work in real time

00;39;48;28 - 00;39;53;15

without depending on cloud,

so that space is heating up.

00;39;53;15 - 00;39;57;18

And AGI compute is actually one of the

key enablers of physical AI.

00;39;59;24 - 00;40;00;25

I was just sort of

00;40;00;25 - 00;40;04;28

smiling there at the phrasing

of heating up, because where my mind went

00;40;05;00 - 00;40;11;09

was you described on the infrastructure

side, these challenges where doing these,

00;40;11;10 - 00;40;14;23

these

AI calculations can actually physically,

00;40;14;24 - 00;40;17;29

you know, heat up these data centers

to hundreds of degrees.

00;40;17;29 - 00;40;19;12

And I have to imagine that.

00;40;19;12 - 00;40;22;04

I mean, obviously

nobody wants that on their wrist,

00;40;22;04 - 00;40;26;20

but that some of the challenges

you have with that infrastructure

00;40;26;22 - 00;40;31;28

or even more pressing, if we're talking

about local computation there.

00;40;31;28 - 00;40;36;08

So the need for optimization

is probably more important than ever.

00;40;36;09 - 00;40;39;06

How how does that sort of play out

in practice?

00;40;39;06 - 00;40;42;18

And, you know,

where do you see the space going?

00;40;43;22 - 00;40;45;15

Yeah.

00;40;45;15 - 00;40;50;22

So edge computing as a space has existed

for more than a decade, right?

00;40;50;23 - 00;40;55;29

People have talked about digital twins,

edge computing, both on the telecom side

00;40;55;29 - 00;41;00;13

and on the industrial side, and also

on the the variables and devices side.

00;41;01;21 - 00;41;04;29

But edge

computing as a space has never taken off.

00;41;05;08 - 00;41;09;14

But AI at the edge is the keyword

application that has now,

00;41;11;00 - 00;41;14;14

you know, driven a lot of interest,

both from enterprise customers

00;41;14;14 - 00;41;17;11

but also from the ministers

to invest in this market.

00;41;17;11 - 00;41;20;06

Now we startup can stand out in

this space.

00;41;20;06 - 00;41;23;21

Is the deep software hardware integration.

00;41;23;22 - 00;41;28;16

How the code is an edge is an area

which is probably fragmented.

00;41;28;18 - 00;41;31;24

There are so many different device

types, right?

00;41;31;25 - 00;41;36;26

You have dozens of variable brands, dozens

of smartphone brands, values of robotics

00;41;36;26 - 00;41;42;13

companies, and even more number

of industrial machinery companies.

00;41;42;13 - 00;41;45;29

You cannot provide the same chip

to all these companies,

00;41;46;04 - 00;41;52;03

like you need a chip or a family of chips

that is tightly controlled by software

00;41;52;12 - 00;41;55;23

to be exclusively tailored

and just do the kind of functions

00;41;55;23 - 00;41;58;16

that are needed

for that particular device.

00;41;58;16 - 00;41;59;10

Right.

00;41;59;10 - 00;42;02;26

And there is a privacy element

to this as well, right?

00;42;02;28 - 00;42;06;04

You cannot run some of these computations

in the cloud

00;42;06;04 - 00;42;08;12

or offload

some of these computations per cloud.

00;42;08;12 - 00;42;11;15

How do you actually do

minimization of the latency

00;42;11;16 - 00;42;14;20

or compute minimization of the power draw.

00;42;14;20 - 00;42;18;00

Because most of these devices are battery

powered to write the thing,

00;42;18;01 - 00;42;21;15

you don't plug your phone and your watch

continuously to power, right?

00;42;21;17 - 00;42;24;02

Otherwise you use them.

00;42;24;02 - 00;42;27;15

They're supposed to be learned

and and be used on the go.

00;42;27;23 - 00;42;28;00

Right.

00;42;28;00 - 00;42;31;05

So so Fowler

constraints are really critical here.

00;42;31;06 - 00;42;33;03

Right?

You're drawing power from the battery.

00;42;33;03 - 00;42;37;01

So so there's a certain envelope of power

you can operate in.

00;42;37;01 - 00;42;38;29

You cannot go beyond that.

00;42;38;29 - 00;42;42;29

Otherwise you'll be in the battery

and make the device pretty useless.

00;42;43;00 - 00;42;43;20

Right.

00;42;45;07 - 00;42;46;09

So software is

00;42;46;09 - 00;42;49;11

the one that does

a lot of this optimization in the form of

00;42;49;11 - 00;42;52;15

when I'm not using the compute power

down the chip,

00;42;52;18 - 00;42;57;00

when I'm using the chip for a certain

workload, only light up

00;42;57;00 - 00;43;00;15

certain areas of the chip

so that the power draw is optimized versus

00;43;00;15 - 00;43;03;29

lighting up the whole chip

and drawing more power.

00;43;04;00 - 00;43;05;04

Right.

00;43;05;04 - 00;43;09;22

And then there's other optimizations

that are done in terms of making the

00;43;09;24 - 00;43;13;13

the really complex

AI models tinier in their footprint

00;43;13;13 - 00;43;17;06

so that they can be run on a lighter

footprint chip.

00;43;17;08 - 00;43;17;14

Right.

00;43;17;14 - 00;43;20;16

You can't run a 60 billion parameter

00;43;20;16 - 00;43;23;16

model LM model on a watch, right?

00;43;23;16 - 00;43;26;16

You need to to tone it down

to less than 10 billion

00;43;26;17 - 00;43;30;02

parameters to be useful

and to be run in a power efficient way

00;43;30;03 - 00;43;33;24

on a chip,

on a watch or a phone or a robot.

00;43;33;29 - 00;43;34;06

Right.

00;43;34;06 - 00;43;38;00

So those all those optimizations

are done in software.

00;43;38;12 - 00;43;42;16

But what I've seen in

the market is a lot of chip companies,

00;43;42;25 - 00;43;47;08

they don't think about hardware software

design from day one to them,

00;43;47;10 - 00;43;50;10

given their legacy in semiconductor

chip manufacturing

00;43;50;10 - 00;43;53;25

and their hardware legacy,

they focus on building the chip first.

00;43;53;25 - 00;43;56;02

And software is like an afterthought.

00;43;56;02 - 00;43;59;28

But the real differentiator is designing

the software and hardware together

00;43;59;28 - 00;44;04;02

from day one,

and that is part of our investment thesis.

00;44;04;02 - 00;44;06;19

We support companies

that think about software

00;44;06;19 - 00;44;11;06

and hardware innovations and core designs

together from day one, and that actually

00;44;11;06 - 00;44;14;28

has the potential to unify this fragmented

landscape through software.

00;44;14;28 - 00;44;15;17

Again.

00;44;17;10 - 00;44;18;14

It's really interesting.

00;44;18;14 - 00;44;19;16

And to me,

00;44;19;16 - 00;44;24;09

I can imagine at least a few different

competing design philosophies there.

00;44;24;11 - 00;44;28;04

One of them is, you know,

you described earlier with your five piece

00;44;28;06 - 00;44;33;08

like programmability and just creating

chips that are a little bit more flexible.

00;44;33;09 - 00;44;37;15

The other one is continuing

the fragmentation, but just having these

00;44;37;15 - 00;44;42;24

really narrow use cases on

on types of devices or types of workloads,

00;44;43;04 - 00;44;47;01

do you envision a world

where there is more general usage

00;44;47;01 - 00;44;52;16

and programmability of these chips

and of this compute, or do you think it's

00;44;52;16 - 00;44;55;16

going to be just more sort of,

I guess, like

00;44;55;16 - 00;44;58;16

vertical fragments based on use case?

00;44;59;11 - 00;45;02;13

I think the real value and lock would be

00;45;03;17 - 00;45;05;07

the programmability aspect.

00;45;05;07 - 00;45;09;02

If if a edge AI compute company

can deliver

00;45;09;03 - 00;45;12;06

a really rich software platform

00;45;12;06 - 00;45;18;01

for for running models on their chip

and maybe even support other other chips,

00;45;18;01 - 00;45;21;18

they have the real potential to become

the operating system for the Aja.

00;45;21;25 - 00;45;22;05

Right?

00;45;22;05 - 00;45;26;02

So imagine if a company can can let

a developer and app developer who's

00;45;26;02 - 00;45;29;23

building some sort of personalization

app for your watch.

00;45;29;29 - 00;45;35;00

If the developer can actually go down

to the lower level of optimizing the chip,

00;45;35;00 - 00;45;39;03

and not just this company ship,

but maybe a media chips or Qualcomm chips

00;45;39;03 - 00;45;43;21

or some other chips, then that software

ecosystem will be really rich.

00;45;43;21 - 00;45;46;27

And that is actually the power

of the Cuda ecosystem for me.

00;45;47;06 - 00;45;47;15

Right?

00;45;47;15 - 00;45;51;26

They invested in the in media Cuda

ecosystem 25 years ago,

00;45;51;29 - 00;45;55;09

probably not thinking that

that could be one of their modes,

00;45;55;09 - 00;45;59;06

but it is actually the most important mode

for Nvidia right now.

00;45;59;15 - 00;46;02;18

People can trump their chips,

but they cannot trump the

00;46;02;26 - 00;46;06;29

the huge ecosystem of developers

and the applications

00;46;06;29 - 00;46;09;29

and scientific libraries

that have been built on top of Cuda.

00;46;10;17 - 00;46;12;26

That is missing from the

AI in mind, right?

00;46;12;26 - 00;46;14;22

There is no Cuda in engineering,

00;46;14;22 - 00;46;18;17

and there is a potential for a startup

to build something like that.

00;46;18;17 - 00;46;22;22

And that will be,

I think, a significant value and lock.

00;46;22;22 - 00;46;25;05

If anybody can do that.

00;46;25;05 - 00;46;27;22

It's an area

where reading is not actively focused.

00;46;27;22 - 00;46;30;20

It's an area where Qualcomm

is not actively focused or Intel

00;46;30;20 - 00;46;34;23

or any other companies,

mostly because of the fragmentation.

00;46;34;23 - 00;46;36;17

But it's also a chicken and egg problem.

00;46;36;17 - 00;46;41;04

You only build a developer platform

if they're not developers, right?

00;46;41;08 - 00;46;44;08

We haven't seen a lot of developers

focus on,

00;46;45;00 - 00;46;48;17

so if you wait for it to come,

you don't build

00;46;48;19 - 00;46;52;10

the developer ecosystem,

the SDK and the platform, and then

00;46;53;15 - 00;46;57;02

there's

not a bit set of, you know, library.

00;46;57;02 - 00;46;59;15

So it's like a chicken and egg.

00;46;59;15 - 00;47;03;17

But I think if

if there is a visionary founder

00;47;03;18 - 00;47;08;04

or founding team that can build something

here, this is a key investment area.

00;47;08;06 - 00;47;12;16

We have a company in this space called

Edge Cortex that actually has the software

00;47;12;18 - 00;47;16;02

hardware co-design

methodology, embedded costs.

00;47;16;19 - 00;47;17;05

They were actually

00;47;17;05 - 00;47;20;28

a software company from day one,

and then they build their own high chip,

00;47;20;28 - 00;47;24;26

and they are trying to be

that operating system for for AGI compute.

00;47;25;09 - 00;47;27;28

But it's a massive undertaking.

00;47;27;28 - 00;47;31;12

We believe in them and average them,

you know, all the good luck.

00;47;31;12 - 00;47;34;12

But the thing began in.

00;47;35;29 - 00;47;38;11

I'm just I'm going

00;47;38;11 - 00;47;42;13

to dramatically oversimplify so much of

what you said here by just wrapping it

00;47;42;13 - 00;47;47;07

at under the banner of a lot of this,

especially in the context of Ajay.

00;47;47;08 - 00;47;48;19

Sounds really hard, right?

00;47;48;19 - 00;47;51;10

Like these are just

these are big, meaty problems

00;47;51;10 - 00;47;54;13

and you have to get a handful of things

right for it to work.

00;47;54;13 - 00;47;54;27

Well,

00;47;56;07 - 00;47;57;15

from a,

00;47;57;15 - 00;48;01;20

I guess from

a device or a robotics manufacturer side.

00;48;03;10 - 00;48;04;25

If you can avoid

00;48;04;25 - 00;48;07;25

Ajay,

you may try to to just say, okay, well,

00;48;07;25 - 00;48;10;25

we're going to try to avoid that

because it's too hard to get right.

00;48;10;25 - 00;48;13;25

So we're going to use more

of the traditional models.

00;48;14;04 - 00;48;18;02

Where do you see the use cases where edge

AI is really going

00;48;18;02 - 00;48;21;13

to be the biggest differentiator from,

I guess, like a product

00;48;21;13 - 00;48;25;09

or a device perspective,

like what can we unlock?

00;48;25;10 - 00;48;29;03

And if we get that right,

and what are the manufacturers

00;48;29;03 - 00;48;32;02

that are most likely going

to be looking at that?

00;48;32;02 - 00;48;33;06

Yeah.

00;48;33;06 - 00;48;38;27

So self-driving cars, they cannot depend

on cloud connectivity, right.

00;48;39;17 - 00;48;42;18

The round trip to the cloud

to make a decision of a,

00;48;42;20 - 00;48;44;14

you know,

whether you want to yield to the traffic

00;48;44;14 - 00;48;48;13

or actually go into the traffic,

that is you you can't depend on the cloud,

00;48;48;14 - 00;48;48;21

right?

00;48;48;21 - 00;48;52;09

You have to make that decision in the car

itself.

00;48;52;11 - 00;48;55;28

So these self-driving cars have Nvidia

GPUs because

00;48;55;28 - 00;48;59;25

they obviously have much bigger batteries

and they can power the GPUs and all that.

00;48;59;25 - 00;49;04;13

But if you take that

same kind of immediacy of the real time

00;49;04;21 - 00;49;07;23

aspect of decisioning,

there are several other use cases

00;49;07;23 - 00;49;12;20

where you cannot depend

on a wired connection to a big server

00;49;12;22 - 00;49;16;19

or a, you know, a wireless connection

to the cloud.

00;49;16;20 - 00;49;19;23

Things like drones saying

things like robots

00;49;19;23 - 00;49;23;27

that are operating in environments

where they cannot be tempered to a server.

00;49;24;06 - 00;49;24;12

Right.

00;49;24;12 - 00;49;27;10

So when creating a robot,

a lot of the companies actually

00;49;27;10 - 00;49;28;26

tether them to a server.

00;49;28;26 - 00;49;32;17

And while they're loading

the inference model on the robot.

00;49;32;19 - 00;49;36;22

But when it comes to inference,

which is the use of that model

00;49;36;28 - 00;49;41;09

in real time decisioning,

they you know, these robotics companies

00;49;41;09 - 00;49;44;28

have to depend on the on board chip

to make the decision,

00;49;45;07 - 00;49;48;12

because if the robot is tethered,

when it can't really do much right,

00;49;48;13 - 00;49;51;11

if it can't move around much,

I can't really do a lot of work.

00;49;51;11 - 00;49;55;05

And a lot of times these environments are

very, you know, hazardous.

00;49;55;05 - 00;49;59;03

So you can't really operate a server

and the wiring and all that.

00;49;59;03 - 00;50;03;04

So yeah, I mean,

and think about satellites.

00;50;03;07 - 00;50;06;09

There is no concept of wires convicted.

00;50;06;28 - 00;50;08;10

You're a satellite.

00;50;08;10 - 00;50;11;10

So satellite spaceships, drones,

00;50;11;10 - 00;50;14;20

robots, self-driving cars, variables.

00;50;14;26 - 00;50;15;05

Right.

00;50;15;05 - 00;50;19;29

Again you need to do inference on the

the watch itself, like next

00;50;19;29 - 00;50;24;11

generation of unlock for Apple Watches

will not come from fitting more sensors.

00;50;24;11 - 00;50;28;00

But how do you actually make sense

of the data that that sensor is,

00;50;28;18 - 00;50;32;19

you know, measuring

and emitting as data stream?

00;50;32;28 - 00;50;36;07

You have to embed a local model

that a chip

00;50;36;07 - 00;50;39;20

will run on the watch

itself to give you real time

00;50;40;26 - 00;50;41;14

measure

00;50;41;14 - 00;50;44;20

of a bunch of bio variables

and things like that.

00;50;44;21 - 00;50;44;27

Right.

00;50;44;27 - 00;50;47;29

So I think the the unlock will come

00;50;47;29 - 00;50;51;22

from embedding compute

next to the sensors.

00;50;51;26 - 00;50;56;08

And for that there will be innovation

both on the software side as models become

00;50;56;09 - 00;50;57;18

tinier and tinier.

00;50;57;18 - 00;51;00;20

And they basically filter out

00;51;01;01 - 00;51;04;24

what is not necessary for

for a core inference operation to run.

00;51;04;24 - 00;51;08;07

And they basically there's

a method called distillation.

00;51;08;08 - 00;51;11;26

So you start with a much bigger model,

20,000,000,030 billion parameter model.

00;51;11;26 - 00;51;14;27

You distill it down to 5 billion

or 1 billion parameters

00;51;14;28 - 00;51;19;06

by stripping away things or variables

that are not important for adults.

00;51;19;08 - 00;51;19;13

Right.

00;51;19;13 - 00;51;24;14

So there is software related optimization

that will be done in the model there.

00;51;24;17 - 00;51;25;17

Then there will be software

00;51;25;17 - 00;51;29;00

level optimization done in terms of

how you actually run a model on the chip.

00;51;29;00 - 00;51;32;04

And there will be chip

level optimization done in terms of

00;51;32;04 - 00;51;35;04

how do you actually light up

only a certain portion of the chip?

00;51;35;04 - 00;51;36;24

How do you miniaturized the chip?

00;51;36;24 - 00;51;39;29

How do you kind of density

the compute with more transistors

00;51;39;29 - 00;51;43;03

and another optimizations around memory?

00;51;43;05 - 00;51;48;10

So the bunch of innovations happening

here, and I think it's inevitable

00;51;48;27 - 00;51;51;00

for for physical AI, for AGI

00;51;51;00 - 00;51;54;27

to really work, the compute part has to

be done on the device itself.

00;51;56;23 - 00;51;57;20

Got it.

00;51;57;20 - 00;51;58;16

It's exciting.

00;51;58;16 - 00;52;01;22

And that gives me a nice sort of framework

for understanding,

00;52;01;29 - 00;52;05;08

you know what what that might look like

and the types of devices.

00;52;05;20 - 00;52;07;03

I want to shift gears a little bit.

00;52;07;03 - 00;52;08;15

And I know you talked about deep tech.

00;52;08;15 - 00;52;12;09

I understand one of the other areas,

you know, you look in and you're planning

00;52;12;09 - 00;52;15;13

for investment wise is quantum computing,

00;52;15;24 - 00;52;18;15

which has been sort of a hot topic

for a number of years.

00;52;18;15 - 00;52;21;19

And I guess just to set the table

a little bit.

00;52;21;20 - 00;52;25;05

I'm most used to hearing

about the impact of quantum

00;52;25;13 - 00;52;29;03

from a cyber security perspective

and post-quantum cryptography,

00;52;29;03 - 00;52;33;27

and what happens when quantum suddenly

breaks all your passwords and codes

00;52;33;27 - 00;52;37;01

and you know, you're

living in a non secured world.

00;52;37;01 - 00;52;41;00

So maybe I can start by asking,

what do you see as being

00;52;41;00 - 00;52;46;10

the biggest impacts that quantum will

have, you know at its onset.

00;52;47;20 - 00;52;48;26

Yeah I

00;52;48;26 - 00;52;54;07

think I would like to first clarify that

quantum computing is only starting to see

00;52;54;07 - 00;53;00;13

some real traction now in terms of actual

deployments and actual applications.

00;53;00;15 - 00;53;02;10

Security is definitely one area.

00;53;02;10 - 00;53;09;05

Will national governments of several

countries have made it a national policy

00;53;09;12 - 00;53;14;06

to to mandate the use of post quantum

cryptography or quantum key distribution

00;53;14;15 - 00;53;17;23

QCD in China, for example, and EQC

in the US

00;53;18;05 - 00;53;20;24

and Japan, as an example,

00;53;20;24 - 00;53;23;24

as kind of protective measures for

00;53;24;16 - 00;53;28;13

banks and other government institutions

to deploy before we hit the Q de

00;53;28;14 - 00;53;31;28

Q days, when,

you know, it's like the D-Day, right, when

00;53;32;14 - 00;53;36;04

quantum will be able to break

these existing encryption algorithms.

00;53;36;04 - 00;53;38;03

Right?

00;53;38;03 - 00;53;41;04

So so far,

we've seen a lot of these deployments

00;53;41;04 - 00;53;45;12

mandated

by governments are being proactively taken

00;53;45;12 - 00;53;49;08

as, as you know, you know, protective

steps by banks and other government

00;53;49;08 - 00;53;53;18

institutions like defense institutions

other than quantum security.

00;53;53;29 - 00;53;57;03

Quantum sensing is another area

where we've seen active

00;53;57;03 - 00;54;00;05

deployments of quantum sensors.

00;54;00;09 - 00;54;03;02

And I would also like to clarify

that people think

00;54;03;02 - 00;54;06;02

quantum

will will basically completely replace AI.

00;54;06;02 - 00;54;08;21

And the classical compute.

That's not going to happen.

00;54;08;21 - 00;54;12;11

Quantum is just another modality,

another compute paradigm.

00;54;12;11 - 00;54;14;19

In the overall computing toolbox.

00;54;14;19 - 00;54;15;26

It's not going to replace.

00;54;15;26 - 00;54;20;18

It's going to work in tandem

with your classical compute of AI,

00;54;21;06 - 00;54;24;27

GPU, CPUs and other, you know, chips

that are like you talk about.

00;54;24;29 - 00;54;29;15

And everyone all these computing

modalities will co-exist.

00;54;29;18 - 00;54;34;29

Quantum will only get deployed for certain

scientific discovery, chemistry, weather

00;54;34;29 - 00;54;39;16

forecasting and some other security

related problems or applications.

00;54;39;16 - 00;54;40;23

It's not going to take over the world.

00;54;40;23 - 00;54;42;23

It's not going to replace everything.

00;54;44;15 - 00;54;45;04

I think that's

00;54;45;04 - 00;54;48;05

a helpful framing device

and a helpful reminder.

00;54;48;05 - 00;54;52;22

So, you know,

I tend generally to shy away from quantum

00;54;52;22 - 00;54;54;28

because as you said, it's

a little bit more niche

00;54;54;28 - 00;54;58;23

and it's a little bit of a crystal ball

exercise to try and figure out

00;54;58;23 - 00;55;02;23

exactly when Q day is going to happen

and you know, when organizations

00;55;02;23 - 00;55;05;21

will actually be dealing with this

in a practical way.

00;55;05;21 - 00;55;07;17

But I've heard one of the phrases

you've used in

00;55;07;17 - 00;55;11;08

this space is enabling technologies

and starting to look at enabling

00;55;11;09 - 00;55;15;16

technologies around quantum

and some preparation there.

00;55;16;05 - 00;55;20;00

What in your mind does that look like,

and what sorts of organizations

00;55;20;00 - 00;55;24;10

should be starting to investigate

and adopt enabling technologies?

00;55;25;04 - 00;55;28;18

Yeah,

whenever there is a big architectural

00;55;28;19 - 00;55;33;09

paradigm shift,

there is an unbundling of the tech stack.

00;55;33;21 - 00;55;36;16

That's what we saw in the cloud era.

00;55;36;16 - 00;55;40;29

We are seeing now in the AI infrastructure

era where startups are starting

00;55;40;29 - 00;55;41;19

the are funding.

00;55;41;19 - 00;55;45;20

Other investors are also funding specific

building blocks of the infrastructure.

00;55;46;05 - 00;55;50;12

In quantum, we are kind of entering

the second innings of quantum computing,

00;55;50;12 - 00;55;54;25

where the first innings was all about

heavily capitalized, full stack place

00;55;54;25 - 00;55;58;29

where IBM, Google, Microsoft

and some of the heavily

00;55;59;00 - 00;56;01;16

capitalized startups were doing everything

in-house,

00;56;01;16 - 00;56;05;03

whether it's building

the actual quantum computer

00;56;05;09 - 00;56;10;03

or the cooling infrastructure to software

to control their like error qubit

00;56;11;07 - 00;56;12;22

correction, error correction and all that.

00;56;12;22 - 00;56;16;16

Everything was done in house,

but now the market is starting to grow,

00;56;16;16 - 00;56;19;16

and that gives enough

00;56;20;17 - 00;56;24;07

affiliate

to startups to start and focus on specific

00;56;24;09 - 00;56;30;06

tooling or specific enablers

of the quantum computing in our ecosystem.

00;56;30;06 - 00;56;34;18

So again,

like like the AI infrastructure goldrush,

00;56;34;19 - 00;56;38;09

at some point there will be a gold

rush in quantum computing where

00;56;38;11 - 00;56;41;09

selling the building blocks of the picks

and shovels of the gold rush

00;56;41;09 - 00;56;45;01

movement

would be a viable investment strategy.

00;56;45;01 - 00;56;49;23

So I have started to become really excited

about specific enabling

00;56;49;25 - 00;56;53;04

technologies like cryogenic,

like the coolers that actually,

00;56;53;07 - 00;56;56;06

you know, help these quantum

computer companies

00;56;56;18 - 00;56;59;22

operate

at these really low Kelvin temperatures.

00;56;59;23 - 00;57;00;27

Right? So that's one area.

00;57;00;27 - 00;57;04;20

Another area is the whole control circuits

00;57;05;24 - 00;57;07;10

like data center quantum

00;57;07;10 - 00;57;11;04

computers will also need these control

circuits and control electronics.

00;57;11;04 - 00;57;14;17

And a lot of the

AI infrastructure control electronics

00;57;14;17 - 00;57;17;17

can actually work with quantum computing

00;57;18;19 - 00;57;19;15

at room temperature.

00;57;19;15 - 00;57;20;12

But if they have to operate

00;57;20;12 - 00;57;23;12

at really low Kelvin temperature,

then there is a lot of engineering

00;57;23;12 - 00;57;24;20

problems there to solve for.

00;57;24;20 - 00;57;25;14

So that's some

00;57;25;14 - 00;57;27;02

there's some innovation there

00;57;27;02 - 00;57;30;09

that startups can actually focus on

and derive value from.

00;57;30;16 - 00;57;33;17

Then there is the whole error correction

piece, right?

00;57;33;18 - 00;57;37;11

The initial innings of quantum

computing was generating qubits.

00;57;37;22 - 00;57;40;21

But there is a lot of errors

in those in those qubits.

00;57;40;26 - 00;57;44;02

How do you correct for those errors

to meet quantum computing

00;57;44;03 - 00;57;48;01

actually reliable

and be reliably deployed at scale?

00;57;48;01 - 00;57;49;08

That is the next gear.

00;57;49;08 - 00;57;51;21

So we're kind of moving

from the scientific

00;57;51;21 - 00;57;55;08

era of quantum computing to engineering

era of quantum computing,

00;57;55;08 - 00;57;59;05

and making these lab projects

be deployable at scale.

00;57;59;05 - 00;58;01;04

So now engineers

come into the picture, right?

00;58;01;04 - 00;58;02;18

And you have these specific tools

00;58;02;18 - 00;58;07;04

and technologies that these quantum

computing companies will need,

00;58;07;04 - 00;58;13;23

which will be a good target market

for all these picks and shovels companies.

00;58;13;23 - 00;58;18;19

So that's where I'm focusing my time

going forward in quantum computing.

00;58;18;19 - 00;58;22;01

I'll be I'll be looking at these fictions

of shovels investments.

00;58;22;15 - 00;58;23;23

It's it's very cool.

00;58;23;23 - 00;58;27;27

And it's it's fascinating that, you know,

to follow the picks and shovels analogy.

00;58;27;27 - 00;58;30;28

They're, they're very different

picks and shovels even from the,

00;58;31;03 - 00;58;32;24

you know, the, the AI era,

00;58;32;24 - 00;58;36;01

it seems like it's completely different

and somehow feels even more,

00;58;36;06 - 00;58;39;13

you know, science fiction than anything

that's that's come before it.

00;58;39;19 - 00;58;43;19

I do want to switch gears

a little bit though, anchor and,

00;58;44;11 - 00;58;48;00

you know, put on my, you know, business

leader or technology leader hat.

00;58;48;00 - 00;58;52;17

So for, for leaders

in kind of the long tail of organizations,

00;58;52;17 - 00;58;55;16

not in a hyperscale or not in the,

00;58;56;13 - 00;59;00;11

you know, frontier Labs,

we're probably developing and,

00;59;00;23 - 00;59;03;29

you know, actually engineering

the quantum tools themselves

00;59;04;04 - 00;59;09;25

in the entire space of what you describe

deep tech, what do they need to know?

00;59;09;26 - 00;59;12;06

What what should they be focused on?

00;59;12;06 - 00;59;15;11

And what, if anything, should

they be doing differently

00;59;15;11 - 00;59;19;25

or looking at adopting

that's going to help them get ahead?

00;59;20;24 - 00;59;22;09

Yeah, I think

00;59;22;09 - 00;59;26;22

first thing is to to stop worrying

about AI replacing them.

00;59;26;24 - 00;59;27;03

Right.

00;59;27;03 - 00;59;29;07

That's probably not going to happen

00;59;29;07 - 00;59;32;11

for most jobs,

at least for the foreseeable future.

00;59;32;11 - 00;59;38;04

I think the best way to to get ahead of

this, I is to start using these AI tools

00;59;38;05 - 00;59;42;04

actively, right, whether they're mandated

by your organization or not.

00;59;42;04 - 00;59;45;16

I think being curious, being active

00;59;45;16 - 00;59;49;08

in using these

AI tools is a worthwhile exercise

00;59;49;10 - 00;59;52;10

to kind of stay updated

on how these technologies work,

00;59;53;05 - 00;59;57;19

and then figuring out ways to actually

use them to to amplify your productivity.

00;59;57;20 - 01;00;00;00

I think that's that's really important.

01;00;00;00 - 01;00;03;12

I think that has already started

in the software world, where

01;00;03;13 - 01;00;08;19

a lot of programing and the documentation

and the show work is not AI driven.

01;00;08;25 - 01;00;09;04

Right.

01;00;09;04 - 01;00;13;14

And there are other pockets

of other functions like legal work or,

01;00;13;26 - 01;00;17;16

or resume scanning

or things like that, like content writing.

01;00;17;22 - 01;00;20;02

A lot of that is starting to leverage AI.

01;00;20;02 - 01;00;21;21

But when you talk about,

01;00;22;25 - 01;00;25;18

again, I'm kind of shifting gears

from digital AI to physical AI.

01;00;25;18 - 01;00;29;07

When you talk about a plant operator,

he or she was running a plant

01;00;29;22 - 01;00;33;23

manufacturing

or some other kind of physical operation.

01;00;34;01 - 01;00;35;25

How should they think about AI?

01;00;35;25 - 01;00;39;00

I think them, again, it should be outcomes

based.

01;00;39;01 - 01;00;43;22

Can I help me get to my outcomes

in a more efficient manner?

01;00;43;22 - 01;00;48;05

And by efficient

I mean, I mean I've invested in companies

01;00;48;11 - 01;00;52;03

that cut down

analytics times for plant managers

01;00;52;03 - 01;00;55;04

from three months to three days, right.

01;00;55;05 - 01;00;58;05

If you can, if you can only only analyze

01;00;58;14 - 01;01;02;18

one gigabytes of data in an Excel file

that is severely limited.

01;01;02;19 - 01;01;05;26

These plants, these refineries,

these manufacturing operations generate

01;01;05;27 - 01;01;08;28

gigabytes of data every second, right?

01;01;09;00 - 01;01;13;04

If you can use cloud and AI tools

to actually analyze and correlate data

01;01;13;04 - 01;01;16;29

from different sensors

and manufacturing lines, and that helps

01;01;16;29 - 01;01;20;23

you become more efficient as a plant,

why not go for it, right?

01;01;20;24 - 01;01;22;17

Like do it.

01;01;22;17 - 01;01;24;05

It's not going to replace anybody's job.

01;01;24;05 - 01;01;26;08

It's actually

going to make the plant more efficient

01;01;26;08 - 01;01;29;24

and more worthwhile as an investment

for the overall organization to continue

01;01;29;24 - 01;01;31;12

investing in the plant. Right. So

01;01;32;18 - 01;01;34;20

yeah, I think

01;01;34;20 - 01;01;36;18

using AI, diving into AI,

01;01;36;18 - 01;01;39;23

I think that's that's the kind of advice

I can give at this point.

01;01;39;23 - 01;01;44;12

And then obviously AI will have different,

different kind of timelines

01;01;44;12 - 01;01;48;12

up and back on different departments,

different teams, functions.

01;01;49;14 - 01;01;50;16

That's great.

01;01;50;16 - 01;01;52;00

Just just anchor.

01;01;52;00 - 01;01;52;25

Before we move to wrap

01;01;52;25 - 01;01;56;01

up, is there anything we didn't cover

in this conversation that you were hoping

01;01;56;01 - 01;01;59;11

that we could talk about, or

are you good to start the wrap up piece?

01;02;02;08 - 01;02;03;06

In the role of

01;02;03;06 - 01;02;06;13

humans are basically the the interaction

01;02;06;13 - 01;02;10;09

of humans with AI, and I think that is

maybe not talked about enough.

01;02;10;09 - 01;02;13;26

In general,

I feel like humans feel threatened.

01;02;13;26 - 01;02;16;26

There is all this hyperbole about,

you know,

01;02;17;16 - 01;02;20;07

jobs are going

to be replaced by, you know,

01;02;22;01 - 01;02;25;01

you know,

graduating classes from Stanford,

01;02;25;16 - 01;02;29;19

MIT and other places found difficult

to learn jobs in computer science.

01;02;29;21 - 01;02;30;11

People are

01;02;30;11 - 01;02;34;11

giving up computer science or programing

because I think we can can be automated.

01;02;34;11 - 01;02;37;09

We don't need these individual

undergrad people.

01;02;37;09 - 01;02;40;18

I think that just to do a proportion,

I think.

01;02;43;01 - 01;02;46;10

AI, like any other technology

paradigm, will

01;02;47;10 - 01;02;49;29

disrupt the human computer interaction.

01;02;49;29 - 01;02;52;11

It's not going to

01;02;52;11 - 01;02;53;23

remove the role of humans.

01;02;53;23 - 01;02;55;12

It's going to change the criminals.

01;02;55;12 - 01;02;56;26

Right? Instead of humans

01;02;56;26 - 01;03;01;01

completely out of the loop,

it will be humans on the loop, right?

01;03;01;10 - 01;03;03;17

Ultimately, AI,

01;03;03;17 - 01;03;07;11

you know, the the starting points

and the endpoints of AI operation will end

01;03;07;11 - 01;03;11;04

with an interaction with the human right,

whether it's physical AI,

01;03;11;05 - 01;03;14;03

whether it is digital

and you can of that sort.

01;03;14;03 - 01;03;20;06

You have to have a human input and a human

output or an output to a human right.

01;03;20;07 - 01;03;23;16

And then also there is there

is this thing where,

01;03;24;12 - 01;03;24;19

you know,

01;03;24;19 - 01;03;26;14

people think

that we can completely automate

01;03;26;14 - 01;03;29;22

all the processes, whether it's

in factories or even in digital

01;03;29;23 - 01;03;33;22

like enterprise processes

like HR, marketing, sales and all that.

01;03;34;21 - 01;03;37;21

It does not factor into.

01;03;39;05 - 01;03;41;01

The scene, the accountability of it.

01;03;41;01 - 01;03;41;12

Right.

01;03;41;12 - 01;03;44;24

And a lot of times when AI fails,

AI is not perfect, right?

01;03;44;25 - 01;03;48;00

AI is only as good as the

you know, the data is trained now.

01;03;48;00 - 01;03;50;06

It's not perfect.

It cannot cover everything.

01;03;50;06 - 01;03;53;22

But the accountability

still rests with somebody human.

01;03;53;29 - 01;03;55;18

You cannot jail AI, right?

01;03;55;18 - 01;03;57;23

You cannot choke AI's neck.

01;03;57;23 - 01;04;00;17

You still need a human

to be accountable for it. So?

01;04;00;17 - 01;04;02;12

So the role of humans will evolve.

01;04;02;12 - 01;04;04;22

It will not completely go away.

01;04;04;22 - 01;04;09;04

It'll it'll be more of a supervisory role

where they will still be accountable

01;04;09;04 - 01;04;12;04

for everything

that's going on in the AI realm.

01;04;13;03 - 01;04;18;13

So I think if folks can kind of figure out

a way to to not just work with AI,

01;04;18;14 - 01;04;21;13

but kind of figure out how their rules

for the world with AI,

01;04;21;13 - 01;04;24;19

I think they need to start doing that

and start preparing towards

01;04;24;19 - 01;04;28;25

that kind of eventuality, where

they'll be on the lookout in the loop.

01;04;28;26 - 01;04;30;26

They will be not completely off the loop

either.

01;04;30;26 - 01;04;34;01

They'll be in a supervisory role at

they'll still be accountable.

01;04;34;02 - 01;04;36;28

They may have to go

to jail. If AI now functions.

01;04;38;07 - 01;04;39;26

I think that's well said.

01;04;39;26 - 01;04;42;14

And an interesting and thoughtful

note to end on anchor.

01;04;42;14 - 01;04;44;10

I want to say a big

thank you for coming on today.

01;04;44;10 - 01;04;47;17

It's been a really interesting

and insightful conversation.

01;04;48;10 - 01;04;49;23

Thank you so much for having me.

01;04;49;23 - 01;04;51;17

It was great to have this chat.

01;04;51;17 - 01;04;53;28

Thank you.

01;04;53;28 - 01;04;57;14

Most viewers don't know this,

but Digital Disruption is developed by

01;04;57;14 - 01;05;02;10

Infotech Research Group, a leading advisor

to technology leaders around the world.

01;05;02;10 - 01;05;04;22

If that's not you, you don't need to care.

01;05;04;22 - 01;05;06;23

So skip ahead and enjoy our content.

01;05;06;23 - 01;05;09;23

But if you are a technology

leader, Infotech helps

01;05;09;24 - 01;05;14;00

IT teams get projects done faster, better,

and at a lower cost.

01;05;14;04 - 01;05;18;09

Infotech provides unlimited access

to practical tools and expert guidance.

01;05;18;09 - 01;05;22;16

You need to execute at a fraction

of the cost of traditional consulting.

01;05;22;19 - 01;05;26;11

No matter the project, from AI strategy

to cybersecurity

01;05;26;12 - 01;05;29;20

to vendor negotiation, Infotech

has you covered.

01;05;29;21 - 01;05;33;10

Check it out at the link below

and don't forget to like and subscribe!