0:00

Hey upstream listeners, today we're

0:02

releasing a conversation from turpentine

0:04

show, Complex Systems, between host

0:06

Patrick McKenzie and entrepreneur and

0:08

investor Azimazar. They discussed the

0:11

history of data centers, the landscape

0:13

of energy infrastructure for AI,

0:15

the coming LLLM boom, and more. Please enjoy.

0:32

Hi Dho everybody, my name is Patrick

0:34

McKenzie, better known as patio 11 on

0:37

the internet, and I'm here with my buddy

0:39

Azimazar. It is so great to be here.

0:41

I love the name of your podcast. Oh,

0:43

thank you very much. So Azim

0:45

runs a newsletter called Exponential View

0:47

and we're going to be talking

0:49

about the power economics specifically that

0:51

of data centers today. People might

0:53

have heard recently in the news

0:55

that Open AI at all are

0:58

building a multi-billion dollar Stargate facility

1:00

down in Texas. People might have

1:02

heard some... sort of hand-wavy or

1:04

more evidence calculations that data center

1:06

usage is going to be tens

1:08

of percent of all power usage

1:10

in the near future. And I

1:12

think for folks inside the industry,

1:14

these are eye-popping numbers, but they're

1:17

somewhat excellent at numbers. For people

1:19

who are outside the industry, this

1:22

is all just a little bit

1:24

wild. So let's take the very

1:26

long view. How does this compare

1:28

to other infrastructure rollouts over the

1:30

last day? couple of centuries, and then

1:33

go into the nitty gritty. But

1:35

speaking of things that are a couple

1:37

of centuries old, we've been around this

1:39

computer thing for a while, haven't

1:41

we? We certainly have. I still

1:43

have my first computer. It's a

1:45

Z80 processor, and the computer's called

1:47

the ZX81, and I got it in

1:50

1981. So it's 43 years old. It's

1:52

going to be 44 this year, and

1:54

probably older than many of the listeners.

1:56

The first modem I ever used was

1:58

300 bits per second. for those of

2:00

you who have had cell phones

2:03

for your entire life. Anyhow, so

2:05

let's talk about slightly more ancient

2:07

infrastructure. So as we're doing this

2:09

build-out of data centers and the

2:11

electricity, both generation and transmission apparatus,

2:13

that powers them, I'm sometimes put

2:15

in mind of other societal-wide infrastructure

2:17

build-outs. What's one that comes to

2:19

mind for you? You know we

2:21

see these infrastructure build-outs every 50

2:23

to 60 years roughly speaking and

2:25

the really really big ones in

2:28

the US and in Europe were

2:30

for the railways and they were

2:32

for electrification just over a hundred

2:34

years ago or a hundred to

2:36

80 years ago in the US.

2:38

One of the interesting facts around

2:40

all of these build-outs is just

2:42

how significant they were. If you

2:44

look at the... build out of

2:46

the railways in Great Britain in

2:48

the 1840s to 1860s, roughly 6

2:50

to 7% of GDP her annum

2:53

went into capital investment to build

2:55

out those those rail lines, which

2:57

in US terms today, given the

2:59

size of US economy, would be

3:01

approaching kind of a trillion, a

3:03

trillion, a half dollars a year.

3:05

Whenever we see a general-purpose technology

3:07

like this, like artificial intelligence, like

3:09

the internet, like telecons and electricity,

3:11

or the railways, infrastructure needs to

3:13

get built, and the thumbs of

3:15

money that go into it are

3:18

always eye-popping. And so too, by

3:20

the way, are the dynamics of

3:22

over-investment, and then, you know, at

3:24

some point there being a bubble

3:26

that pops, because it's so hard,

3:28

as you know, to... forecast exactly

3:30

when demand and where demand will

3:32

arise. Mm-hmm. Vern Hobart has a

3:34

book called Boom with Stripe Press

3:36

that lays out this theory with

3:38

the school author Tobias, I believe.

3:40

And essentially he says that infrastructure

3:43

overinvestment tends to happen with bubble

3:45

dynamics, but that the bubble is

3:47

actually somewhat positive in that it

3:49

causes a shelling point of... and

3:51

different people firms, government entities, etc.

3:53

that all have different pieces of

3:55

the puzzle that would not coordinate

3:57

to do a nationwide or even

3:59

worldwide infrastructure revolution, but for somewhat

4:01

bubble dynamics. And then... often bubbles

4:03

pop because as you mentioned it

4:06

is virtually impossible to get these

4:08

questions right a priori but the

4:10

demand sort of back fills over

4:12

the intervening decades after the popping

4:14

of the bubble. I honestly don't

4:16

know there's substantial technical uncertainty etc

4:18

etc etc etc. This might be

4:20

the bubble that doesn't pop and

4:22

always dangerous say this time is

4:24

different but what I'm saying is

4:26

like there is some possibility that

4:28

there is a discontinuity here with

4:31

with previous infrastructural build outs but

4:33

for people who are our generation

4:35

who remember very keenly the dot-com

4:37

bust happening. People remember the dot-com

4:39

bust as being about the application

4:41

layer about webvan and pets.com. We

4:43

had great.com domains back in those

4:45

days, but they remember that, but

4:47

by count of money invested, almost

4:49

all of it was doing coast-to-coast

4:51

fiber and copper rollouts, and that

4:53

substantially... enabled the development of the

4:56

modern internet and its use as

4:58

e-commerce, etc. platforms. Yeah, I agree

5:00

with that. And in fact, the

5:02

telecoms bubble is really salutary. The

5:04

total investment in telecoms infrastructure you

5:06

described at kind of coast-coast, but

5:08

there was stuff happening outside the

5:10

US as well, was, as I

5:12

recall, around $600 billion between 19

5:14

and 96 and 2001. and US

5:16

telecoms firms alone took on about

5:18

350 to 370 billion dollars of

5:21

debt in order to do this.

5:23

And they did it at a

5:25

time when the telecoms market was

5:27

growing at about 7% per. Anum.

5:29

And so what feels really different

5:31

about this AI market is that

5:33

firms are growing much, much faster

5:35

than that. I mean, 7% per

5:37

week is not unheard of right

5:39

now. And you're getting these early

5:41

stage startups that like Curser, which

5:43

are getting to $100, $150 million.

5:46

of annual recurring revenue within 18

5:48

months or so, which is an

5:50

absolutely dramatic, dramatic result. And in

5:52

fact, six months ago, we were

5:54

looking at data that showed that

5:56

AI-based software startups were growing three

5:58

times faster than fast-growing SAS startups

6:00

in the pre-AI era, and that's

6:02

compressed even further. And the reason

6:04

that's different to the telecoms infrastructure

6:06

build-out is that the revenues... are

6:08

probably ramping faster than the infrastructure

6:11

growth is ramping, whereas the reverse

6:13

was true back in the telecoms

6:15

bubble, which has been relabeled the

6:17

dot-com bubble by some historians. Yeah,

6:19

so I'm also hearing reports of

6:21

this anecdotally. Don't treat the following

6:23

as... a quote from an informed

6:25

market participant but just treated as

6:27

a bit of market color on

6:29

the grapevine. People are saying that

6:31

firms which are at about stage

6:34

for raising investment in Silicon Valley

6:36

and typically at that point you

6:38

continue to be quickly growing but

6:40

are sort of reinvesting in processes

6:42

to support the next 10X over

6:44

the next couple of years and

6:46

getting let's say a little less

6:48

cowboy about the operations. There are

6:50

people who are saying that they

6:52

are seeing firms at the C-round

6:54

stage that are growing as quickly

6:56

as Y-cominator companies that are, you

6:59

know, double-digit weeks old with 10%

7:01

week over wreath growth, etc. and

7:03

not growth in a vanity metric

7:05

or growth in eyeballs for reviewing

7:07

cap videos, but growth in enterprise

7:09

spend on AI capabilities. Which to

7:11

the extent that one trusts that

7:13

observation is mind-blowing? It's absolutely wild,

7:15

and I have heard similar things

7:17

as well, so presuming we're not...

7:19

hearing the same rumor sourced from

7:21

the same person. Let's assume that

7:24

these are coming from different places.

7:26

That seems to match somewhat with

7:28

what I hear as well. And

7:30

you can see it in from

7:32

other data sets. So there is

7:34

a platform called OpenRooter, which has

7:36

access to a whole bunch of

7:38

LM APIs. data, they, when I

7:40

looked at it, I saw an

7:42

8x growth in token usage, as

7:44

in tokens served by a set

7:46

of these models, over a somewhat

7:49

less than one year period. And

7:51

so we're going from, you know,

7:53

a few hundred million tokens per

7:55

month to, you know, several billion

7:57

tokens per month, and of course,

7:59

while token prices have come down,

8:01

that is showing that elasticity of

8:03

demand and the fact that the

8:05

demand exists. and the demand is

8:07

not being saturated at all. So

8:09

I sometimes feel that we are

8:11

the blindfolded men walking around this

8:14

elephant and we have to sort

8:16

of put the full picture of

8:18

what's happening in the market together.

8:20

And I don't really see many

8:22

of the feelers that suggest we're

8:24

not dealing with something that's big

8:26

and fast growing. I mean, typically

8:28

every conversation I have, sometimes a

8:30

bit skeptical about them. points to

8:32

the kind of dynamic that you

8:34

talked about, which is that the

8:36

growth is really, really fast. It

8:39

continues to be fast. And, you

8:41

know, as prices come down, spend

8:43

goes up because you can, you

8:45

know, basically make the economics work

8:47

on newer use cases. I also

8:49

think that people, there's a comparative

8:51

advantage in having seen the dynamics

8:53

of SAS companies scale in understanding

8:55

how this is going to scale,

8:57

because I think some of the

8:59

best informed people with regards to

9:02

what LLLM's can actually do these

9:04

days are folks that are playing

9:06

with them every day. But if

9:08

your view on an LLLM's capability

9:10

is I chat with Claude all

9:12

the time, he seems very emotionally

9:14

supportive. I've done this sort of

9:16

song generation and Suno, which is

9:18

a wonderful experience by the way.

9:20

You're probably not predicting what those

9:22

capabilities in an API plus a

9:24

two to five year enterprise integration

9:27

cycle looks like because after that

9:29

exists it's not going to be

9:31

you know you invoking one LLLM

9:33

at a time for a couple

9:35

hours per day it will be

9:37

everybody getting a staggering number of

9:39

LLLM invocations on their behalf every

9:41

day most in the background where

9:43

the rate model isn't a person

9:45

having a conversation with some, it's

9:47

more similar to what happens when

9:49

you open up the New York

9:52

Times and several hundred robots conduct

9:54

an instant auction for your attention

9:56

on your behalf. And there's an

9:58

entire ecosystem affirms mad tech that

10:00

make that happen and you will

10:02

never know most of their names.

10:04

Anyhow. Yeah, well, can I add

10:06

to that? I mean, we already

10:08

got that crossing point on web

10:10

traffic a couple of years ago

10:12

where 50% of traffic generated is

10:14

bots traffic doing various things like

10:17

that. And within. you know, processes

10:19

within the enterprise, there are so

10:21

many advantages in having LLLM's or

10:23

AI systems talk to each other,

10:25

I mean, fundamentally because they can

10:27

be much faster than we can.

10:29

And that's one of the things

10:31

that emerges when you see these

10:33

sort of optimized distilled LLLM's running

10:35

at a thousand tokens per second,

10:37

which is that is really, really

10:39

significant in terms of, you know,

10:42

ingesting information, making a decision on

10:44

that information, and sending a signal

10:46

back out to the next step

10:48

in the process at millisecond speed,

10:50

whereas humans work at, you know,

10:52

minute speed, and at a thousand

10:54

tokens a second rather than five

10:56

to eight tokens a second, which

10:58

may be where we might sit

11:00

at the very, very best of

11:02

times when we're reading something, let's

11:04

learn when we're writing something. And

11:07

that velocity, I think, in of

11:09

itself, will breed, will breed faster.

11:11

you know, faster velocity. And that's

11:13

why I love the idea of

11:15

this being a complex system and

11:17

the complex systems podcast, because that

11:19

is ultimately a complex system, right,

11:21

with all of these feed-forward loops

11:23

and flywheels. Yeah. And if I

11:25

can give people concrete examples of

11:27

what is going to happen very,

11:30

very quickly. If you're a kind

11:32

of sewer of the fictional experience

11:34

of watching rich people talk to

11:36

each other, a line that you

11:38

hear a lot in those... movies

11:40

and etc. is let's have your

11:42

people talk to my people where

11:44

the two principles are aloof from

11:46

their own calendar management but there's

11:48

some implicit team in the background

11:50

that takes care of that for

11:52

them. Let's have your alum talk

11:55

to my LLLM is already happening

11:57

and will certainly happen in the

11:59

future in every conceivable way. As

12:01

an example I recently a push

12:03

notification, email, and paper letter, all

12:05

from a bank, asking me, hey,

12:07

we haven't seen you update your

12:09

address with us in a while.

12:11

We need to be on top

12:13

of your address. And so do

12:15

you have an update for us?

12:17

If not, just tell us so.

12:20

And that is totally going to

12:22

be an alum-driven conversation in the

12:24

future. You can optimize out the

12:26

stamp entirely, the annoying interaction with

12:28

customer. And also optimize out probably

12:30

all the time on the bank

12:32

side. given the nature of the

12:34

current process. Yeah, can I give

12:36

you a really concrete example that

12:38

I use as well? Sure. One

12:40

of the tools that I use

12:42

is a network of LLLM. So

12:45

I'll have a single LLLM that

12:47

acts as a kind of orchestrator

12:49

and I'll have several other LLLM

12:51

that acts as members of a

12:53

focus group. And I will put

12:55

my question in to the orchestrator.

12:57

and the orchestrator will pass that

12:59

question which is sort of evaluate

13:01

this idea or evaluate this product

13:03

to the underlying agent lens. Typically

13:05

I'll use three or four. Each

13:07

one of them will have a

13:10

fairly detailed profile of a pen

13:12

portrait, right, a persona, a marketing

13:14

manager or an investor and you

13:16

know the last one might be

13:18

a recent college grad and they

13:20

will iterate and argue between themselves

13:22

about the merits of this particular

13:24

product. with a view to come

13:26

into some kind of distinct consensus.

13:28

And in that virtual focus group

13:30

where I run three or four

13:32

of these, they will go back

13:35

and forth and they will generate...

13:37

tens of thousands of tokens and

13:39

then ultimately the orchestrator will respond

13:41

to me. Now I do that

13:43

to help me sense-check ideas that

13:45

I might want to explore or

13:47

research or to kind of really

13:49

push them or to find a

13:51

very exacting through-line or weakness, set

13:53

of weaknesses in the idea. And

13:55

there are companies that are out

13:58

there electric twin is in the

14:00

UK that is one. that are

14:02

building panels of LLLM's where panels

14:04

of virtual synthetic personas that are

14:06

built within LLLM's where these

14:08

conversations will happen much more

14:10

rapidly and at much larger

14:12

scale in order to do in

14:15

silico what we might have had

14:17

to do quite slowly, you know,

14:19

sort of in vivo and much

14:21

more expensively with real focus groups.

14:23

And I think that those are

14:25

also examples that are beyond, you

14:27

know, AI agents interacting on

14:29

fixed processes and process flows

14:31

where they're much much more

14:33

open-ended and actually the the amount

14:36

of resource that could go into

14:38

those could be quite quite significant.

14:40

So stepping back for a moment,

14:42

I think that I've used as

14:44

part of my writing process for

14:46

many years is to either have

14:48

a formal review from other people

14:50

or more for more usually have

14:52

an informal review where I conjure

14:54

my mental model of a particular

14:56

person in my head. read it and then,

14:58

okay, you know, what would this person

15:00

in the industry think and think of

15:02

this piece right now? Patrick Ellison has

15:04

described using that for his own writing

15:07

process, but I think I stole the

15:09

idea from him. At any rate, the

15:11

one can increasingly do that with like

15:13

telling an LLLM role play as someone

15:15

who has a large corporate on the

15:17

internet, and you don't necessarily need them

15:19

to successfully anticipate more than, you know, 80,

15:21

90% of what the person would say, just as

15:23

a idea generation thing. I did it over

15:25

the weekend on something which was

15:28

extremely professionally significant, extremely enough professionally

15:30

significant, that I was always also

15:32

spending social karma and no small

15:34

amount of money with a number

15:36

of external professional advisors. And the

15:39

fact that the LLLM needs no

15:41

social karma at all and trivial

15:43

amounts of money to step through

15:45

like role-playing as say five different

15:47

potential audiences for a piece was

15:50

revelatory for me in terms of

15:52

the quality, speed of iteration,

15:54

etc., etc. etc. for the

15:56

advice I got. And if, you know, sort of

15:58

early adopters like us are using

16:00

this in production. Like this was

16:03

a this was a real thing

16:05

for me. A lot was riding

16:07

on the line of this, you

16:09

know, not a fun test just

16:11

to try out the new toy.

16:13

If the early adopters are using

16:16

this in production right now for

16:18

this, you can imagine when every

16:21

marketing team in many places

16:23

is running, as you said,

16:25

in silico panels. If I help

16:27

us kind of frame this as well, which

16:30

is that, you know, when Deep Seek was

16:32

released at the end of last year, sort

16:34

of both V3 and R1, and

16:36

then we had the flurry of

16:38

excitement in late January, you know,

16:40

the idea of Jeven's paradox surfaced,

16:42

right, which was Saty Nadella from

16:44

Microsoft said this, and, you know,

16:47

the point about Jeven's paradox was

16:49

that essentially, you know, if you've

16:51

got clog traffic around Austin, and

16:53

you build another freeway within a

16:55

few months you'll have more more

16:57

traffic jams because ultimately there's positive

16:59

elasticity of demand right you reduce the cost

17:01

and and demand goes up you know up until

17:03

a saturation point and one of the

17:05

questions is where is that saturation point

17:07

with with AI and The truth is,

17:10

I think we are nowhere near, and

17:12

by nowhere near I can't even find

17:14

the phrase for the size of the

17:16

fraction of being nowhere near. So much

17:18

of what we do in business and

17:21

often in our personal lives is fundamentally

17:23

gated by the fact that we don't

17:25

have enough time to think. through to

17:27

get to the optimal solution. So we just

17:29

use a heuristic and where we accept that

17:32

being a bit of slack and a bit

17:34

of loss. And essentially we're going to we're

17:36

not going to do that anymore, right? We're

17:38

going to let these AI agents do tons

17:40

and tons of thinking enormous amounts of it,

17:43

things which we, you know, we just did

17:45

out of habit, we will get them to

17:47

to improve, particularly in business. And so I

17:49

don't think there's any end to the

17:51

amount of demand that we could individually

17:54

generate as individuals either as

17:56

people at home or in

17:58

the workplace for thinking to

18:00

be done by these machines. And the second

18:02

lever of that is very, very few people,

18:04

few of us are currently doing that

18:07

and very few organizations are doing that.

18:09

I mean, I think Open AI has

18:11

125,000 people paying for the pro level

18:13

of chat GPT, which is absolutely

18:15

phenomenal. You know, and if you're in a

18:17

job that pays 100K a year or more,

18:19

you should be investing in that straight away

18:22

because your ROI will be incredible.

18:24

And so I think that these two, these two...

18:26

elements, which is one is how

18:28

deep does each personal organization want to

18:30

go and what new organizations emerge, and

18:33

how many of us are participating that

18:35

are both going to expand very, very

18:37

dramatically, and as prices for intelligence

18:39

or per token come down, the

18:41

break-even point will rise and will

18:43

accelerate the demand. And so that

18:45

gets us to the question of

18:47

like, what is the infrastructure that's

18:49

going to serve all of that?

18:51

So we'll get to that infrastructure

18:53

in one second, but to remind

18:55

people of a famous phrase in

18:58

the technology industry, there was a

19:00

point where intelligent people, well steeped

19:02

in the worldwide situation for demand,

19:04

said that there is a worldwide

19:06

market for perhaps five computers. And

19:08

it turns out that we can

19:10

deploy many more than five. We are,

19:12

I think, right now in the five

19:14

computer days of LLLM usage, where we've

19:16

applied it to the obvious applications. and

19:18

people who are only seeing

19:20

the obvious applications fail to appreciate

19:23

what will happen once it gets

19:25

injected into just about everything and

19:27

that. I feel pretty confident

19:29

is going to happen over the

19:32

course of the next 10-20

19:34

years with new developments on

19:36

a week-by-week-by-month basis. We'll continue

19:38

our interview in a moment after a

19:40

moment after a word from our sponsors.

19:42

Even if you think it's a bit

19:45

overhyped, AI is suddenly everywhere. from

19:47

self-driving cars to molecular medicine to business

19:49

efficiency. If it's not in your industry

19:51

yet, it's coming fast. But AI needs

19:54

a lot of speed and computing power.

19:56

So how do you compete without costs

19:58

spiraling out of control? Time to upgrade

20:01

to the next generation of the

20:03

cloud. Oracle Cloud Infrastructure, or OCI.

20:05

OCI is a blazing fast and

20:07

secure platform for your infrastructure, database,

20:10

application development, plus all your AI

20:12

and machine learning workloads. OCI costs

20:14

50% less for compute and 80%

20:17

less for networking. So you're saving

20:19

a pile of money. Thousands of

20:21

businesses have already upgraded to OCI,

20:24

including Vodafone, Thompson Reuters, and Suno

20:26

AI. Right now, Oracle is offering

20:28

to cut your current cloud

20:30

bill in half if

20:33

you move to OCI

20:35

for new US customers

20:37

with minimum financial commitment.

20:39

Offer ends March 31st.

20:42

See if your company

20:44

qualifies for this special

20:46

offer at Oracle.com/Turpentine.

20:50

That's Oracle.com

20:52

slash Turpentine. And I bet I'm not

20:55

alone. In a world of endless scrolling

20:57

and ad blockers, sometimes the most powerful

20:59

way to reach people is in the

21:01

real world. That's where Adquick comes

21:03

in. The easiest way to book out-of-home

21:05

ads, like billboards, vehicle wraps, and airport

21:07

displays, the same way you would order

21:10

an Uber. Started by Insticard

21:12

Alums, Adquick was born from their

21:14

own personal struggles, to book and measure

21:16

ad success. Adquick audience targeting.

21:19

campaign management and analytics.

21:21

Bring the precision and

21:23

efficiency of digital to the

21:26

real world. Ready to get

21:28

your brand the attention it

21:30

deserves. Visit adquick.com today to

21:32

start reaching your customers in

21:35

the real world. Again, that's AD,

21:37

QUICK.com. Let's talk about that

21:39

infrastructure. So. Data centers. We have

21:41

a very diverse listener group for this

21:44

podcast and so some people are intimately

21:46

familiar with walking to a data center,

21:48

some people less so. Let's see first.

21:50

You walk into a data center, look

21:52

to your left, look to your right, what

21:54

do you see? It looks like that scene

21:57

in The Matrix when Neo and Trinity asked

21:59

for guns and you're in a white

22:01

room and just racks and racks of

22:03

guns show up. And that's what you,

22:05

you know, what you see in a

22:07

data center, you see very large numbers

22:09

of racks that have within them, you

22:11

know, pizza box-sized computers stacked up and

22:14

cooling systems at the back in power

22:16

delivery. And what is, you know, how,

22:18

I mean, in a sense, what you

22:20

see today is not too different in

22:23

my view to what we saw 20

22:25

years ago, but in reality. everything

22:27

is much denser, right? The

22:29

networking bandwidth interconnects are running

22:31

100 times faster than they

22:33

used to be. The power

22:35

demand is much, much higher. I

22:37

mean, I think that the most

22:40

recent, most recent sort of H100

22:42

racks, which are the H100 is

22:44

a big big invidious, a GPU

22:46

that people like Meta use, they'll

22:48

be running at 130 kilowatts per rack,

22:50

which is... 70 Kettles is how

22:52

I think about it. You know,

22:54

to put that in the context,

22:57

the first servers that I

22:59

ran to serve websites back in

23:01

1996 ran to 200 watts each,

23:03

and I had four of them.

23:06

So I had 800 watts sitting

23:08

there, roughly, including the sort of

23:11

little Ethernet switch that was

23:13

connecting them to the internet.

23:15

So we've gone from 800

23:17

watts to 130 kilowatts, her

23:19

rack. and that's the power demand

23:21

but of course you then have to

23:24

to cool this as well and you

23:26

have you have sort of storage requirements

23:28

as well and I think roughly

23:31

speaking about 40 to 45% of

23:33

this goes into the actually doing

23:35

the thinking right the the flops

23:37

and the processing about 40% goes

23:39

into the cooling and then the

23:42

rest is you know networking redundancy.

23:44

The cooling often surprises people who aren't

23:46

specialists in this, but it basically comes

23:48

down from, well, data centers don't get

23:50

to cheat the laws of physics. If

23:52

you pump X amount of energy into

23:54

them, that energy has to go somewhere.

23:56

And what you typically need to do

23:58

in most locations. is active cooling to

24:01

remove the energy that you've pumped

24:03

in to the external environment in

24:05

some locations that are very cold at

24:07

certain points of the year you can

24:09

use passive cooling but we're putting data

24:11

centers all over the place. Yeah and so

24:14

active cooling basically means we have to pump

24:16

a liquid in and we have to have

24:18

a heat exchanger somewhere else on the other

24:20

side and you know I think in Colossus

24:23

which is Elon Musk's data center there's a

24:25

fantastic video that shows the size of the

24:27

final cooling pipes and you know that they

24:29

come up to your shoulder, they're pretty

24:31

phenomenal. So one final bit of color

24:33

before we go into the recent developments

24:35

in this, but density drives so much

24:38

of both the economics and the operational

24:40

concerns of data centers and as we've

24:42

heard over the last couple of decades

24:44

they're getting more dense for square centimeter

24:46

cubic centimeter I suppose said because height

24:49

is a material thing we stack these

24:51

boxes on top of each other. Why

24:53

does density matter fundamentally for the operator?

24:55

Because a data center is fundamentally a

24:57

real estate business with some value add

25:00

to being the power cooling and on-site

25:02

technical services. But an example of

25:04

a counterintuitive thing with respect to

25:06

what the density and intensification does

25:08

for you, data centers because they

25:10

have huge amounts of electricity running

25:12

through them operating at high temperatures

25:14

as high as you can do

25:16

without damaging the chips. are not

25:18

the safest places in the world

25:21

to be, particularly during some failure

25:23

modes. And so when I first

25:25

got the badge that would allow

25:27

me into a data center when

25:29

I was working at a Japanese

25:31

system and grader back in the day, I

25:33

had to be given a safety briefing before

25:35

I got the badge, because there is a

25:37

device on the wall called colloquially a big

25:40

red button. And there are many genres

25:42

of big red button in the world. A

25:44

thing you really want your young engineer to

25:46

understand before they walk into that room is,

25:48

is this the big red button that just

25:51

drops all the power in the room? Or

25:53

is this the big red button that you

25:55

have 60 seconds after you press it before

25:57

every living thing in the room dies? Yes.

26:00

the Halon, the Halon gas, right,

26:02

comes out to extinguish an electrical

26:04

fire. That's one of the things

26:06

your on-site safety engineer will be

26:08

really, really interested in making sure

26:11

that everyone taking even a guided

26:13

tour of that room understands. Anyhow.

26:15

So, now you know what it looks like.

26:17

in the data center. So let's take

26:19

a look around the... Can I just

26:22

add to this point, right? So, you

26:24

know, one of the things that you've

26:26

described, which is this, that increasing density,

26:29

also has an impact on the physical

26:31

real estate asset. So many data centers

26:33

that exist today, and, you know, if

26:36

we've driven down freeways in parts of

26:38

the US, you'll have seen these buildings

26:40

that look like warehouses, but are not

26:43

warehouses, turns out... you can't upgrade them

26:45

to these modern AI data centers because

26:47

they actually can't maintain the power delivery

26:49

and the cooling delivery that the new

26:52

chips require. So you know, as the

26:54

chips get more and more dense, they

26:56

get hotter, they need better cooling, they

26:58

need more reliable power, and in fact

27:00

you need... different physical architectures. You physically

27:03

need new buildings as well. And there's

27:05

a kind of unintended consequence, I guess,

27:07

of Moore's Law and Wang's Law and

27:09

whatever else has sort of replaced those

27:11

laws to that make the chips more

27:14

power efficient and kind of more dense for

27:16

flops per cubic centimeter is that

27:18

they need different buildings. Yeah, one

27:20

of the more mind-blowing things in

27:22

my career as a system is

27:24

engineer. Systems engineer build combination software

27:27

and hardware systems. I was definitely more

27:29

on the software side than the hardware

27:31

side. If I never own another server

27:34

in my life, I will be very

27:36

satisfied with that. But there existed a

27:38

data center and the physical amount of

27:41

weight of the server racks was over

27:43

the physical capacity of the floor that

27:45

the server racks were on. And you

27:47

can't simply run a command on your

27:50

terminal to make the floor stronger than

27:52

the architect designed it to be. And

27:54

at the point where you are saying,

27:57

okay. We'd like to replace not just the

27:59

thin metal shell that is on top

28:01

of the floor, but no, actually, we

28:03

need the structural floor replaced. Then you

28:05

start thinking, okay, it might be time

28:07

to build a new building. Right. I

28:09

think that what you just described there

28:12

is one of the things that's most

28:14

misunderstood about the, you know, the nature of

28:16

this particular game, because for, you know,

28:18

the bulk of us, our experience with

28:20

super computers are things that weigh, you

28:23

know, five ounces, right? It's our cell

28:25

phones. and they've always got smaller and

28:27

they've largely got lighter and that's always

28:29

been the way they've been sold to

28:31

us and you know that's true about

28:34

laptops as well and we think about

28:36

our monitors they get smaller and

28:38

smaller and smaller on our desks and

28:40

the in a way the the miniaturization

28:42

the the packing of more transistors onto

28:44

every square centimeter or square inch of

28:47

a dye has the reverse effect on

28:49

physical architecture and it's a really important

28:51

notion and I think it makes it

28:53

really complex when you start to think

28:55

about the you know we tend to depreciate

28:58

buildings over a many many multi-decade period

29:00

you don't depreciate computer hardware over that period

29:02

and it used to be four years now

29:04

Google and you know Amazon or alphabet and

29:06

Amazon have moved that to six years but

29:09

you have this sort of difference in in

29:11

the kind of tenor on the you know

29:13

the financing side and the depreciation side it's

29:15

much much much more complex than just upgrading

29:17

your iPhone your iPhone every three years. We'll

29:20

continue our interview in a moment after

29:22

a word from our sponsors. In an age of

29:24

AI and big data, your personal information has

29:26

become one of the most valuable commodities.

29:28

Data brokers are the silent players in

29:30

this new economy, building detailed profiles of

29:33

your digital life and selling them to

29:35

the highest bidder. But data breaches happening

29:37

frequently, these profiles aren't just fueling spam.

29:40

They're enabling identity theft and putting your

29:42

financial security at risk. But what if

29:44

there's a way you could opt out

29:47

of that system entirely? Incogne is a

29:49

service that ensures your personal data remains

29:51

private and secure. They handle everything from

29:54

sending removal requests to managing any pushback

29:56

from data brokers. What I trust most about

29:58

Incogne is how it provides with ongoing

30:00

protection. Once you're set up,

30:02

it continuously monitors and ensures

30:04

your information stays off the

30:06

market. Want to extend that

30:08

peace of mind to your

30:10

loved ones? Incogney's family and

30:13

friends plan lets you protect

30:15

up to four additional people

30:17

under one subscription. Take control

30:19

of your digital privacy today.

30:21

Use code upstream at the

30:23

link below and get 60%

30:25

of an annual plan. Again, that's

30:27

Incogney.com slash upstream. Eric here. In

30:30

this environment, founders need to become

30:32

profitable faster and do more with

30:34

smaller teams, especially when it comes

30:36

to engineering. That's why Sean Lanahan

30:39

started Squad, a specialized global talent

30:41

firm for top engineers that will

30:43

seamlessly integrate with your org. Squad

30:46

offers rigorously vetted top 1% talent

30:48

that will actually work hard for

30:50

you every day. Their engineers work

30:52

in your time zone, follow your

30:54

processes, and use your tools. Squad

30:56

has front-end engineers excelling in typescript

30:58

and react, and NextJS, ready to

31:00

on-board to your team today. For

31:02

back-end, Squad engineers are experts at

31:05

NoJS, Python, Java, and a range

31:07

of other languages and frameworks. While

31:09

it may cost more than the

31:11

freelancer on Upwork billing you for

31:13

40 hours, but working only two,

31:15

Squad offers premium quality at

31:17

a fraction of the typical

31:19

cost, without the headache of

31:22

assessing for skills and culture

31:24

fit. Increase your velocity without

31:26

amping up burn. Visit choose

31:28

squad.com and mention turpentine to

31:30

skip the wait list. So, both in

31:33

the historical perspective and in

31:35

the near future perspective, where

31:37

did we build data centers

31:40

in the physical universe? Well, we

31:42

started, I mean, the very first

31:44

data centers, I mean, to go

31:46

back to that, tended to be

31:48

in cheap bits of land that were...

31:51

reasonably close to where

31:53

customers were. So in the

31:55

US, it would be in

31:57

Reston, Virginia, as one example.

31:59

because you had the bold baronet in

32:02

Newman which was one of the architects

32:04

of the the the sort of NSF

32:06

net kind of precursor to the commercial

32:08

internet was based over there and you

32:10

know I think that was one of

32:12

the reasons why AOL ended up being

32:14

there in the United Kingdom there was

32:16

a place called Docklands where which was

32:18

very very cheap light industrial land that

32:20

was not too far from the city

32:22

of London where people those banks were

32:24

among the first users of you know

32:26

high-speed cabling so you were really you

32:28

really thought about you know that particular

32:30

dimension which is like relative

32:32

proximity to customers but also

32:34

still being quite you know quite cheap

32:37

for for the land that dynamic

32:39

has has of course continued and

32:41

we know that you know the

32:43

northeast bit of the US is

32:45

really really big for for data

32:47

centers but I think there are

32:49

now considerations around the accessibility to

32:51

fundamentally to electricity, right? Is there

32:53

sufficient electricity for the work that

32:55

we need to get done? Before

32:57

we get into the electricity point,

33:00

some fun color, why do we

33:02

put data centers close to the

33:04

customers? Back in the day, it

33:06

was more about putting them close

33:08

to employees slash skilled technicians. And

33:10

so if your server breaks down,

33:12

you might need your... system administrator

33:14

to drive out from Chicago to

33:16

one of the suburbs to reboot

33:18

it. But say in the late 90s network

33:20

latency was not that huge of a

33:22

consideration because who in the late 90s

33:25

was doing anything where you could tell

33:27

a difference between an 800 millisecond ping

33:29

time and a two second ping time.

33:31

Fast forward to today network latency

33:33

is a primary consideration for where

33:35

these go. And so there are

33:38

worldwide networks of data centers at

33:40

the largest firms in capitalism and also

33:42

out of firms selling to the rest

33:44

of the economy, well the largest firms

33:46

to capitalism also sells to the rest

33:48

of the economy. We'll talk about that

33:50

in a moment to optimize essentially for

33:52

network latency. And then this new power

33:54

constraint is sort of new. The data

33:57

center usage up until recently in

33:59

the United States was probably single

34:01

digit percentage of all the national

34:03

electricity demand. So no small amount,

34:05

but we get no small amount

34:08

of value out of computers. So

34:10

that's fine. But with the densification,

34:12

with the notion of having entire

34:14

buildings full of H100s running on

34:17

training and inference, we start to

34:19

have real constraints about can we

34:21

physically pump as many electrons through

34:23

the grid as we need to call

34:25

back to last week's episode. Anyhow.

34:28

Yeah, but can I can I also put

34:30

some history on this as well because I

34:32

You know AI is this a technology which

34:34

I think is incredibly important

34:37

It's also turned out to

34:39

be very divisive in debates

34:41

both within the industry and

34:43

outside the industry and I

34:45

can't really remember technology But

34:47

you know triggered such a

34:49

split in in people's perspectives. Oh,

34:51

can I can I can I get

34:53

one? Yes, please cloud was a in

34:55

say the 2000 late 2000 to early

34:58

2010s, depending on where exactly you were

35:00

in the world, there were huge debates

35:02

within both the engineering community and in

35:05

the ones who were specifically hands on

35:07

the middle for most of their careers.

35:09

Will big businesses ever consent to use

35:11

somebody else's server and somebody else's building

35:14

for their most private customer data? But

35:16

okay. Yes, there used to be this

35:18

ad by an on-prem company, I forget

35:21

which it was an advert, and

35:23

it said, it's someone else's computer.

35:25

as a way of saying something negative

35:27

and derogatory about the cloud.

35:29

I remember working at a Japanese

35:31

system integrator where we had a

35:33

multi-year debate with the customer base

35:36

which were mostly universities in my

35:38

part and this would have been

35:40

in the late 2000s and the

35:43

universities would say we really really

35:45

want to have All of our student

35:47

information in a location that we

35:50

control where it will be safe,

35:52

not in some building somewhere where

35:54

we have no visibility. And the

35:56

true engineering fact of the matter was

35:58

the location they could was literally an

36:00

unlocked broom closet on campus where anyone

36:03

could walk in under the influence of

36:05

a hangover or similar and walk out

36:07

with all the data. And there was

36:09

a bit of an adoption curve in

36:12

the Japanese enterprise, but the

36:14

somewhat stodier members of the Japanese

36:16

enterprise did get there a few

36:18

years after the American enterprise

36:20

and similar did. But this was a

36:22

live issue back in, you know, as recently

36:25

as 10 years ago. The power issue

36:27

has also been a longer

36:29

term issue than generative AI

36:31

large language models in the

36:33

chat gPT moment. One of

36:35

the reasons I think this

36:37

has become so present in

36:39

people's minds has been that there's

36:42

a lot of skepticism about

36:44

the value that AI brings.

36:46

But before chat gPT in November

36:49

2022 we were all and anyone

36:51

knowing this was going to be

36:53

a thing. We'd already started to

36:55

see places like Singapore which host

36:57

a lot of data centers and

36:59

Ireland and a number of cities

37:02

start to say we can't provision

37:04

any more data centers and those are

37:06

data centers were just for sort

37:08

of pre into pre-A.I. uses like

37:10

just moving customer data becoming a

37:12

digital digital business and if you

37:14

look at the CAPX of a

37:16

firm like I'm just going to

37:18

look at Microsoft for example. In

37:21

2021, to that year, Microsoft was

37:23

going to spend 20 billion dollars

37:25

on CAPX. Only three years earlier,

37:27

it was at 10 billion. So

37:29

it was doubling in three years.

37:31

And this was well before the

37:33

open AI deal had manifested itself.

37:35

It was well before chat GPT.

37:37

So one of the things I

37:39

think we need to also contextualize

37:41

was that even before the before

37:43

AI and before this Gen AI

37:45

thing. data center demand was growing

37:48

really really significantly partly because of

37:50

your the point you made about

37:52

the cloud right companies want customers

37:54

data close to customers and they're

37:57

moving all of it off-prem and you know

37:59

we are now seeing that accelerate, but

38:01

it's not first and foremost in

38:03

my view something that is just

38:05

about, you know, just about AI.

38:07

We've certainly concentrated it, but it

38:09

hasn't, you know, been the sole

38:12

spark for it at all. Yep. And if

38:14

I can give a shout out to Leopold's

38:16

paper here. situational awareness. These are

38:18

the sort of things which were obvious to

38:20

some people back in the day, but they

38:23

were not obvious to, you know, extremely informed

38:25

planners of electricity demand for metropolitan areas and

38:27

nations. They weren't obvious to hedge funds that

38:29

were following the space, etc., etc. They were,

38:32

you know, conversations at dinner parties in San

38:34

Francisco that said, hey, we might need a

38:36

trillion dollars worth of new power built out

38:38

in the next couple of years. Trillion with

38:41

a teeth, that's kind of wild, huh, what

38:43

could we do with that? a trillion of

38:45

anything is a lot, but let's also

38:47

just look at US overall

38:49

electricity demand. I mean electricity

38:51

or energy in general is

38:53

wealth and energy is prosperity and

38:56

energy is health. There are no

38:58

countries with good outcomes for their

39:00

people broadly defined that don't have

39:03

high levels of energy consumption per

39:05

person whether you're efficient about it

39:07

or not. The thing about the US is that

39:10

as of 2020, 2020, 2021, electricity

39:12

usage was pretty much at the

39:14

level of 20 years earlier. Now

39:17

that is a really really important

39:19

thing to look at. Now of

39:21

course you see energy efficiency,

39:23

right? The switch to

39:25

LED light bulbs is

39:27

tremendous. You see environmental

39:30

standards emerging and those...

39:32

making people think much more about

39:34

their energy efficiency. It's also good

39:36

business because electricity costs money and

39:38

you know if you can do

39:40

the same commercial output with less

39:43

sort of cost of inputs that's

39:45

more profit for you. But at the

39:47

same time for it to be flat says

39:49

that there is something about the

39:51

the kind of collective agreement by

39:54

power providers to invest in

39:56

in capacity. And you know this was

39:58

off the back of essential. a doubling

40:00

of electricity consumption between 1975 and about

40:02

the year 2000. So to it suddenly

40:04

going flat. And at that time we

40:06

also started to see the electrification certainly

40:08

of certain types of passenger car transport,

40:10

right? The Tesla's show up and so

40:12

on. So I think that when we

40:14

start to diagnose this we have to

40:16

also go a little bit further back

40:18

and I'm not a historian of the

40:20

US sort of electricity system in great

40:23

detail. you know, it's kind of odd

40:25

that it flat lines 20 years ago

40:27

and we don't start to make, we

40:29

don't start to make those investments, frankly

40:31

either in the US or in many

40:33

parts of Western Europe. I'm also myself

40:35

not exactly an energy economist. I would

40:38

say one thing which probably contributes to

40:40

it was the US has undergone some

40:42

structural economic changes over the course of

40:45

less over decades and there was a

40:47

bit of a substitution between manufacturing output

40:49

in the United States as the size

40:51

that's ever been. manufacturing employment

40:54

is slower, people sometimes confuse those

40:56

two. But we largely shifted from

40:58

a manufacturing focused economy to a

41:00

services focused economy and per dollar

41:03

value of output services, use various

41:05

resources, and inclusive electricity less intensely.

41:07

But I do agree that there

41:10

was a failure to anticipate future

41:12

needs, and also I think in

41:14

the United States in many places

41:16

in Western Europe and many places

41:18

near and dear to the hearts of

41:20

many listeners of this podcast. There's

41:23

been a real reluctance to

41:25

build things in the physical

41:27

world. It's almost like we have

41:29

lost either that will, the knowledge, the

41:31

capacity to do so in

41:33

some places in ways that

41:36

seem absolutely mind-boggling. And when

41:38

we... I live 20 years in Japan and

41:40

Japan has many problems but refusal to

41:42

be able to build buildings is not

41:44

one of them. Right. And then, you

41:46

know, look over the ocean over to

41:49

China. China certainly has not forgotten how

41:51

to, you know, do solar deployments, for

41:53

example. And I think one of the

41:55

most crucial things in this sort of

41:57

moment we find ourselves in is rediscovering.

42:00

the complex system that will

42:02

allow us to actually build

42:04

the infrastructure there are future

42:06

economic needs to depend on.

42:08

Yeah, I mean, I absolutely

42:11

agree with that. I mean, I

42:13

think with China, we see a

42:15

willingness, a desire at sort

42:17

of senior levels of government,

42:19

but also as the sort

42:21

of acceptance amongst people that

42:23

infrastructure is really really valuable

42:26

and it's not just solar manufacturing

42:28

capacity it's also solar solar deployment

42:30

and it's deployment of solar at utility

42:32

scale and on rooftops it's about

42:34

the deployment and build out of

42:36

nuclear power stations very very rapidly

42:38

it's about high-speed rail it's also about

42:40

transmission one of the things that

42:42

of course is really challenging in

42:44

in in the US a lot of which is

42:47

to do with market structure and regulation

42:49

is building transmission lines but you know

42:51

China has 34 ultra-high voltage transmission

42:53

lines that you know very very kind of

42:56

energy efficient and don't leak a lot of

42:58

have a lot of energy loss over those

43:00

long long distances but totaling tens of thousands

43:02

of miles right and one of the things

43:05

that that does especially when you deal with

43:07

intermittent resources like solar and wind is it

43:09

allows you to move the electrons to where

43:11

they need to be you know consumed you

43:14

know if it's sunny in the place and

43:16

they're not being consumed locally you can move

43:18

them to where they're needed. We had discussion

43:20

about this a few weeks ago with Travis

43:23

DeWaltter on the changing needs of

43:25

transmission lines in the United States,

43:27

and if I can elaborate just

43:29

slightly more on what you've said,

43:31

I think one of the most

43:34

important facts of energy economics has

43:36

been the extreme performance of the

43:38

learning curve for solar power versus

43:40

cost over the course of the

43:42

last 25 years. I remember at

43:44

the course. At the time where I

43:47

graduated university about 2004, it

43:49

did not look likely that

43:51

solar was ever going to

43:53

be economical against coal, for

43:55

example, absent huge subsidies

43:57

for social reasons. And it turns

43:59

out... that not only did we

44:01

continue down the cost curve we

44:04

actually bent that curve the learning

44:06

accelerated as there were you know

44:08

multi-billion tens of billions hundreds

44:10

of dollars of investment

44:12

into solar deployment and so

44:15

the vacation of the energy grid is

44:17

one of probably the most central aspects

44:19

of the coming infrastructure wave, but solar is

44:21

not the only power generation thing that is

44:23

going to shake up in the course of

44:26

the next decade or two. You mentioned that

44:28

China has been doing large-scale nuclear bolt-outs, which

44:30

I kind of feel a little bit jealous of,

44:32

but do you want to say a few words

44:34

about the hottest blum-pum, new nuclear technology that we

44:36

might be collocating with data centers in the near

44:39

future? Well, in the near future, let

44:41

us talk about China's electrical capacity.

44:43

They added about 335 gigawatts of

44:45

capacity in 2023, and it was

44:47

29 gigawatts in the US. So

44:49

that's a scale of where we've

44:51

got to. And I think the

44:53

point about the learning curves with

44:55

solar is that they really start actually

44:58

back much, much further back. Back in

45:00

73 or 74, there was a James

45:02

Bond film called The Man of the

45:04

Golden Gun, where this sort of British

45:07

spy has to steal. back solar technology.

45:09

It was so importantly sending the best

45:11

secret service agent in the world to

45:14

get it. And now solar panels are

45:16

so cheap that in Germany they've

45:18

actually fallen below the price of

45:20

fence panel and you're starting to

45:22

see people build out vertical balcony

45:24

fences and fences between them and

45:26

their neighbors, which don't, you know,

45:28

they don't catch as much sunlight,

45:30

but it's cheaper and it generates

45:32

some electricity for you. I think

45:34

a lot is, you know, we're

45:36

hoping for... quite a lot from

45:38

nuclear and in particular a small

45:40

modular nuclear reactor. So the idea

45:42

of a small modular reactor is

45:45

that it's all of those things.

45:47

It's meant to be small and

45:49

it's meant to be modular. What's

45:51

the benefit of that? The benefit

45:53

of that is that you tend to

45:55

see better learning effects when you

45:58

make more of something and so... and

46:00

you get density better learning effects when

46:02

those things are modular, rather than built

46:04

as products, rather than as projects. And

46:06

so one reason why solar has had

46:09

these amazing learning curves is that, is

46:11

that, you know, the panels, whether on

46:13

my rooftop or in a solar field

46:15

in Texas, are essentially the same. And

46:17

of course, the implementation is slightly different,

46:19

because, you know, ones on a roof,

46:22

the others on sort of flatish ground

46:24

with, you know, mounted in particular ways.

46:26

Nuclear reactors have been built. and

46:28

in many cases are sort of n

46:30

of one. You have to start from

46:33

the beginning. A multi-decade bespoke engineering process

46:35

where we get very, very little learning

46:37

between the nth and nth and nth

46:39

plus one iteration of it. Right, absolutely. And

46:41

the idea between the small modular reactor

46:43

is that you can build these things

46:46

in a modular fashion, so you can

46:48

get learning effects. Because they are small,

46:50

you scale out rather than sort of

46:52

by kind of magnifying the scale of

46:54

things. So if you want more capacity,

46:56

you buy more of them. And frankly,

46:58

that's what we've done in the computer

47:00

industry. If you need a super powerful

47:02

computer, you don't go off and get

47:04

a massive mainframe with huge chips. You

47:06

go and get 10 H100s and stick

47:08

them together. you know, a thousand H100

47:10

and sticking together. This also works

47:13

well against the nature of the

47:15

demand for data center electricity because

47:17

for a large scale nuclear power

47:19

plant that would produce enough electricity

47:21

for a large fraction of a city, and

47:23

you don't have full control in sighting

47:25

where that plant is, you probably can't

47:27

justify putting one directly next to the

47:30

newest data center that you popped up

47:32

by the freeway, but for a... a small

47:34

modular nuclear reactor that might fit in

47:36

a footprint that is about the size

47:38

of standard-sized shipping container? Sure, put one

47:40

right next to every data center. Put

47:43

two if you want. Put two if

47:45

you want. Right. And I think the

47:47

other thing about the small modular reactors

47:49

is that if theoretically they are safe

47:52

by physics, by the laws of physics,

47:54

as opposed to safe by layers and

47:56

layers of containment systems and safety

47:58

systems. So in some... sense they are

48:01

much more appealing. I guess the issue

48:03

around the SMR that we have to

48:05

recognize is there's a sort

48:07

of TRL technology readiness level risk

48:09

at least in the West. So

48:12

there are some SMR units

48:14

operational in China and Russia. I'm

48:16

not sure how quickly we're going

48:18

to sort of import them into

48:21

the US or into Europe and

48:23

there are lots of companies who

48:26

are building new designs with reference

48:28

designs and you know, we are hoping to

48:30

see them take off. And, you know,

48:32

in the senses, if there's a

48:34

sort of tailwind of demand and

48:37

capital that's available, you could potentially

48:39

scale these out much, much faster than

48:41

we have scaled out, you know, certainly

48:44

nuclear plants, but I think there is

48:46

a recognition also that you need the

48:48

electricity provisioning. today, which is why, you

48:51

know, we're starting to see gas generation

48:53

on some of these bigger data, you

48:55

know, data centers, whether it's Metas or

48:58

its XAIs or Colossus. So you mentioned

49:00

T.L.R. there. Can you say a few

49:02

more words for the benefit of the

49:05

audience? Oh, TRLs. Yeah, technology readiness level.

49:07

So it's a kind of standard level

49:09

of technology readiness that runs from whether

49:12

something is, you know, really, really, really

49:14

at the high-risk scoping-scoping-scoping we know exactly

49:16

how to build it, how to price

49:18

it, and how to implement it, what

49:21

its kind of total life cycle looks

49:23

like. And, you know, in a sense,

49:25

small modular reactors are sort of

49:27

lower down that scale, probably in, I'm

49:30

guessing, I'm kind of extenprising slightly, but

49:32

certainly in the fours and fives and

49:34

fixties rather than the nines and tens

49:36

and that creates a certain degree of

49:39

risk and uncertainty of what the outcome,

49:41

you know, looks like. There's also,

49:43

of course, a regulatory slash political will

49:45

issue about nuclear reactors where I'm going

49:47

to make a terrible pun, but I

49:50

am a dad. I get to do

49:52

dad jokes. They were politically radioactive for

49:54

a number of years in many Western

49:56

democracies, and I think we are in

49:58

a moment the last. couple of years

50:01

where we can, partly through a

50:03

combination of engineering fact, the new

50:05

technology is simply safer than existing

50:07

technologies, but partly because we had

50:10

good substitutes for base load demand,

50:12

or acceptable substitutes for base low

50:14

demand, liquid natural gas, etc., etc.,

50:17

etc., for many of the last

50:19

couple of decades. And things have

50:21

changed. One is the climate issue,

50:23

of course. We would strongly

50:25

prefer to avoiding using

50:28

combustion of hydrocarbons. and then

50:30

the geopolitics of energy usage have

50:32

changed quite radically over the course of

50:34

the last 10 years or so to a point

50:36

where say much of Europe where

50:39

there's say a relatively extreme level

50:41

of political engagement around environmental issues

50:43

it's like well you can choose

50:46

either fulfilling all of one's preferences

50:48

with respect to domestic constituencies that are

50:50

vociferously anti-nuclear, or you can choose to

50:52

be warm during the winter. And when

50:54

push comes to shove, many of our

50:56

truest and dearest friends over there will

50:59

probably choose to be warm during the

51:01

winter. Yeah, coming back to small modular

51:03

reactors though, you know, you know, Google

51:05

has this deal with chyros energy for,

51:07

I think it's six, maybe it's seven

51:09

small modular reactors, and... I think 2030 is

51:12

a delivery time, so we're talking five

51:14

or six years out. A couple of

51:16

interesting things about this is that given

51:18

that it's seven, it is, we are going

51:20

beyond first for kind, so first of a

51:22

kind tends to be the really expensive one,

51:24

and I did write about this a few

51:27

months ago saying this is kind of a

51:29

Google gift to humanity, because the learning

51:31

curves will be shared, learning experience will

51:33

be shared by all of us. and

51:35

they're the ones who are paying the

51:37

price to bring these things out at

51:39

a high cost. But look at the

51:41

timing, it's six years, and that's only

51:43

seven reactors, and these are any

51:46

small reactors. And so the electricity

51:48

requirements across the US economy are

51:50

really, really enormous, and we have

51:52

to ask how quickly can this

51:55

actually fundamentally scale up. But what

51:57

Google did was they addressed something

51:59

that... this complex system has

52:01

as a roadblock, which is

52:03

that mezzanine financing, which

52:05

is not venture capital level

52:07

risk, but nor is it

52:09

the low-risk guaranteed return of

52:11

asset finance. And that has been an

52:14

issue with a lot of these

52:16

energy and sort of electrification

52:18

hard technologies, which is

52:20

that when you're developing the

52:22

IP, the international property around

52:24

which the equity value in

52:27

the extreme return comes, venture

52:29

capitalists are willing to take

52:31

that risk. But venture capitalists

52:33

are a really small asset

52:35

class. They can't fund infrastructure

52:37

projects. But once they've built the first

52:40

one, you still have risk. You have

52:42

a lot of deployment risk. You have

52:44

all of the learning effects before it

52:46

turns into something that is steady and

52:49

stable, like a solar farm or a

52:51

wind farm where infrastructure and

52:53

investors come in and they ask for

52:56

very very steady state returns with very

52:58

little chance for extreme upside which is

53:00

what the the venture capitalists are after.

53:02

So you have this middle period which

53:05

is kind of ends of a kind

53:07

financing risk which has been really really

53:09

difficult to address. It's known as amongst

53:12

sort of people in climate tech as

53:14

of one of the valleys of death.

53:16

There are many valleys of death and

53:19

you know it's a struggle to cross

53:21

it. fortuitous topic for you to bring

53:23

up because we didn't plan this in

53:26

advance but I've actually spent a good

53:28

portion of my professional cycles the last

53:30

two years volunteering with a focused focused

53:32

research organization that is attempting to popularize

53:35

next generation geothermal and it is exactly

53:37

that problem and there are VCs

53:39

who are willing to write checks

53:42

into the hopefully defensible IP for

53:44

power generation using next generation geothermal

53:46

but every time you do an experiment in

53:48

the field, you need to spend 20 to

53:50

60 million dollars to ask Caliburton to provide

53:52

you professional services and what is the thing

53:54

they're going to do for you, they're going

53:57

to dig a really deep hole and 60

53:59

million dollars a whole. Let's go. As you

54:01

said, that is challenging in VC land.

54:03

In a world where all the

54:05

technology risk has been shaken out,

54:08

there are virtually unlimited amounts of

54:10

capital available to do this. So

54:12

the oil and gas industry in

54:15

the United States, for example, it's

54:17

the same people digging functionally the

54:19

same holes. But can you go to

54:21

a bank or other sources of capital and

54:24

get the marginal gas will financed? Absolutely.

54:26

You know, there are people who do

54:28

that every day, like, give us your

54:30

number, give us your engineer's numbers, I'll

54:32

put in my spray tree, do, do,

54:35

green, we go, you'll have your word

54:37

tomorrow. And so a lot of the

54:39

last two years from me has been

54:41

attempting to cheat the value of death on

54:43

behalf of NGO. Can I ask

54:45

about your experience there that's super

54:48

interesting to me? And what is

54:50

next generation geothermal rather than geothermal?

54:52

Sure. So the brief version is

54:54

that. The geothermal that most people think

54:57

of is places where heat energy from

54:59

the earth bubbles up so close to

55:01

the surface that you can physically perceive

55:03

it in some cases. So hot springs,

55:05

geysers, etc., etc. When you think of the

55:07

places in the world that are the largest

55:10

geothermal energy producers currently, you think

55:12

of places like Iceland. and they

55:14

have been blessed by nature with

55:17

the particular subsurface formations give them

55:19

abundant access to geothermal. Most places

55:22

in the world are not similarly

55:24

blessed by nature. However, due to

55:26

the, particularly the fracking boom in

55:29

the United States, we've gotten much,

55:31

much better at drilling to depths that

55:33

were not economically drillable before. And if

55:35

geothermal can only tap energy that is

55:37

available at the surface of the earth,

55:39

you have to be blessed by nature

55:42

to do it. If, on the other

55:44

hand, you can go down, say, I

55:46

don't know, 6 to 10 kilometers, then

55:48

essentially everywhere is blessed by nature. And

55:50

to 90 plus percent of the continental

55:52

United States is an emir that I've

55:55

heard thrown around. And thus, there is

55:57

still technology risk, you know, digging the hole,

55:59

sure, but... But you need to figure

56:01

out what the generation station is

56:03

that you put on the top

56:06

of the well and what the

56:08

curve looks like for heat in

56:10

the immediate vicinity of the well

56:12

that you have fracked is a

56:14

limited resource. So you tend to

56:16

get a trailing off of the

56:19

generation over a sometimes scale. And

56:21

so we're really looking for those

56:23

next like one to 20, 40, 100

56:25

wells to see what do those curves

56:27

look like? And then it's just a

56:29

numbers game. Like in one in one

56:32

version of fiscal reality, this is

56:34

not cost competitive with other forms

56:36

of heat or electricity generation. And

56:39

in another version of physical reality,

56:41

we have free clean abundant energy available

56:43

in large portions of the world. And

56:46

so ask me in 10 years which

56:48

reality we're living in. I will have

56:50

a very confident answer in 10 years,

56:52

but I don't currently. And what's the

56:54

price that you think is cost competitive?

56:56

the number that I had cashed

56:59

off the top of my head

57:01

because I didn't know we were

57:03

going to be talking it. Interestingly,

57:05

just to say a few more

57:07

words on the why fracking matters,

57:09

fundamentally fracking. The oil and gas

57:11

people love to call it subsurface

57:13

engineering, but we had a prior

57:15

episode about tracking. I will link

57:17

it in the show notes, but drill

57:19

the hole, pump a working liquid down

57:21

the hole, and use that to break

57:24

rock around the vicinity of the vicinity

57:26

of the hole. thing here, sorry folks,

57:28

but in one version of the

57:30

system you pump water or some

57:33

other working liquid which filters into

57:35

the cracks that you've made. And

57:37

because the surface area in

57:39

those cracks is, the cracks look

57:42

fractal in nature and the surface

57:44

area is absurd relative to the

57:46

diameter of the hole. And so

57:48

you can pull heat from the

57:51

surrounding rocks for a very long

57:53

time, hopefully, until the... the rate of

57:55

heat moving into the vicinity of the

57:57

rocks that your water is touching is

58:00

no longer sufficient to sustain the rate

58:02

of heat that you're extracting from the

58:04

top of the hall. And yeah, that's long story

58:06

short for people who want to learn

58:08

more about this field, they'll drop some

58:10

links in the show notes. I mean,

58:12

geothermal is, I think, really interesting and

58:14

potential technology, and it speaks to the fact

58:16

that the energy system feels like it's

58:19

going to continue to be heterogeneous. You

58:21

know, what happened with computing is that...

58:23

we tend to have these sort of

58:25

winner-take halls, although it's a little bit

58:27

more heterogenous than it looks at the

58:29

surface because kind of an arm chip

58:31

is different to an Intel chip is

58:33

different to a GPU. But I do

58:35

see a kind of world of different

58:37

energy technologies. Definitely. And as we heard

58:40

in the episode with Travis Duelter a

58:42

few weeks ago, the heterogeneity of power

58:44

generation makes the grid more stable because

58:46

there are... different physical aspects to

58:48

be different power generation technologies, nuclear

58:50

geothermal, etc. are stable base load

58:52

power, and then solar seems to

58:55

be scalable to the moon, oh

58:57

man, dad joke number two, but

58:59

solar is of course only available

59:01

during particular hours of the day.

59:03

Well, but I think, I think it's

59:05

worth asking a question about how far

59:08

you can actually go with solar, because

59:10

I suspect that it's further than most

59:12

models take it, and just hear my

59:14

case for that. Because solar is highly

59:17

modular, the market expands significantly and that

59:19

means that homes and small businesses as

59:21

well as large scale utility providers can

59:23

get into solar and we've seen this

59:25

happen in large parts of the US

59:28

with community solar but of course rooftop

59:30

solar in China is absolutely enormous and

59:32

Pakistan has a great example where businesses

59:34

got sick and tired of the grid

59:36

failing and so just went off and

59:39

bought loads of solar panels so at

59:41

least 10 hours a day they could

59:43

run their business. the cost curves are

59:45

really really in their in their favor and

59:47

even though we've had 50 years of learning

59:50

it's not clear that solar price decline panel

59:52

press declines are going to stop you know

59:54

sort of any time in the next five

59:56

to ten years and there are new technologies

59:58

bubbling in the wings. So even though it's far

1:00:01

from perfect, the total system cost

1:00:03

is something that's quite dynamic and

1:00:05

the other aspects of the total system

1:00:07

cost will be whatever happens with batteries

1:00:10

and other forms of storage, and batteries

1:00:12

are really early in their learning effects,

1:00:14

I mean the cost of kind of

1:00:16

prestige batteries, right, the lithium-ion battery,

1:00:18

have declined from... $1,200 per kilowatt

1:00:20

hour to about $40 per kilowatt

1:00:22

hour since 2011-2012 to the beginning

1:00:25

of 2025. And there's probably still

1:00:27

some room to run and there

1:00:29

are cheaper technologies with different physical

1:00:31

characteristics like sodium iron air batteries

1:00:33

that are in the wings. And

1:00:35

then you have to think about

1:00:37

how do you manage your distribution

1:00:39

because that becomes an important part

1:00:41

because as you say, right, it's

1:00:44

sunny in one place and it's

1:00:46

not sunny somewhere else. And there are

1:00:48

a couple of really interesting projects that

1:00:50

are going on at the moment. One

1:00:52

is built in Australia to take power

1:00:54

all the way up to Singapore, another

1:00:57

in Morocco, to take power to the

1:00:59

United Kingdom with these subsea high voltage

1:01:01

direct current cables that are being

1:01:03

built out. There's also some other

1:01:05

interesting second, third, maybe even seventh

1:01:07

order effects for some of these

1:01:10

technologies where Casey Hand of our

1:01:12

previous podcast, his company Terraform Industries,

1:01:14

I believe. Yeah, is attempting to

1:01:16

do a direct capture of carbon

1:01:18

dioxide to turn into hydrocarbons using,

1:01:21

quote, alien science, end quote, which

1:01:23

is, of course, heavily energy intense

1:01:25

because you can't cheat the laws of

1:01:27

thermodynamics. Right. But given that you have

1:01:29

huge amounts of solar generation in one

1:01:31

part of a nation, if hypothetically you

1:01:33

can do local generation of hydrocarbons, then

1:01:36

you can ship extremely energy dense hydrocarbons

1:01:38

to wherever you want to put them

1:01:40

in the world, can bust them there,

1:01:43

and then just, you know. the carbon

1:01:45

goes back into the atmosphere, you suck

1:01:47

it right back down and turn into

1:01:50

hydrocarbons again. So Casey is amazing, but

1:01:52

and let's talk about, let's just go

1:01:54

through that cycle again. If we take

1:01:56

carbon dioxide out of the atmosphere and

1:01:59

we push it over... the second or

1:02:01

third thermodynamics hump and we

1:02:03

turn it into methane and we

1:02:05

combust that methane, we're net

1:02:08

zero, right? We've not put

1:02:10

any additional CO2 into the

1:02:12

atmosphere and the energy density

1:02:14

of gasoline or methane or

1:02:17

kerosene is absolutely staggering and we

1:02:19

have a whole load of systems

1:02:21

that already know how to use

1:02:23

that. And I think that's a

1:02:25

great example of why it may

1:02:28

be that. solar could end up being,

1:02:30

and I think it will be the

1:02:32

dominant supply of, you know, first-party energy

1:02:34

and electrons into, you know, into

1:02:37

the system. And there are other

1:02:39

things that you can start to

1:02:41

do with the system, like demand

1:02:43

response. So that's where you kind

1:02:45

of affect and create incentives for

1:02:47

people's behavior to change. There is

1:02:49

a, there are lots of air

1:02:52

conditioners and heat systems, heating and

1:02:54

cooling systems in homes in Texas.

1:02:56

whose behavior is actually managed

1:02:58

by the energy provider to

1:03:00

respond to, to respond to, to,

1:03:03

to, to, by minute, an

1:03:05

hour-by-hour changes in the electricity

1:03:07

demand and pricing. And that

1:03:09

can also extend to how

1:03:11

we might shift workloads for compute around

1:03:13

data centers at different times a day.

1:03:15

Not everything needs to be right on

1:03:17

the front end. You know, your Akamai

1:03:19

servers, serving up live video, need to

1:03:21

be close to the customer the whole

1:03:24

time. I'm not the world expert on

1:03:26

this, but we seem to be sort

1:03:28

of in land grab mode at the

1:03:30

moment with respect to training new AI

1:03:32

models. But one can imagine a future

1:03:34

in which, for those of you who aren't

1:03:36

familiar, these wonderful AI models that we are

1:03:39

using these days, have typically two phases. There

1:03:41

is a training phase and then an inference

1:03:43

phase. The training phase is the months of

1:03:45

hard work that... open AI or anthropic or

1:03:47

another lab put into putting out one of

1:03:49

their new numbered releases of a model. And

1:03:52

then the inference phase is what happens in

1:03:54

the few seconds between when you ask a

1:03:56

question and when an answer comes back to

1:03:58

you. My guess, finger to the wind. without your

1:04:00

amounts of inside knowledge is that

1:04:02

the chips that have been doing

1:04:04

training have been running hot essentially

1:04:06

24 hours a day, seven days

1:04:08

a week for the last while.

1:04:10

However, one can imagine future iterations

1:04:12

of this technology where there is

1:04:14

actually some sort of cost benefits

1:04:16

curve associated with it. And so

1:04:18

at times where times and places

1:04:20

where electricity is particularly expensive, just

1:04:23

stop training for a while. and

1:04:25

continue doing the inference on an

1:04:27

on-to-band basis. Or, again, currently, you

1:04:29

know, we are the dominant public

1:04:31

deployment of AI as with a

1:04:33

user sitting up the keyboard typing

1:04:35

into a computer, but that will

1:04:37

not be the case for forever.

1:04:39

It might make sense to... provision

1:04:41

more intelligence when electricity is cheap

1:04:43

to do these sort of quote-unquote

1:04:45

offline calculations on behalf of industry

1:04:47

than to simply continue running inference

1:04:49

at the same levels everywhere in

1:04:51

the 24-hour clock. Or we could

1:04:53

end up in a world where

1:04:55

cognition is just so stupidly valuable

1:04:57

that why would you ever turn

1:04:59

it off just to save on

1:05:01

electricity bills? Well and what I

1:05:03

love about this is these are

1:05:05

these are a few scenarios and

1:05:07

let's throw out some other scenarios.

1:05:09

One is the algorithmic optimizations. of

1:05:11

the functional level of intelligence that

1:05:13

we want at a given at

1:05:15

a given time. And, you know,

1:05:17

we think about Moore's Law being

1:05:19

this remarkable thing, right? 60% cost

1:05:21

declines on price performance every year

1:05:23

for decades and decades. But software

1:05:25

optimizations can be orders of magnitude

1:05:27

of improvement instantly, like a phosphoria

1:05:29

transform or, you know, doing something

1:05:31

with a bloom filter rather than,

1:05:33

you know, mechanically walking through bloom

1:05:35

filters that brings me back. That

1:05:37

takes you. Yeah, sorry. I'm just

1:05:39

an old dude. I can't I

1:05:41

can't help it. And so this

1:05:43

there's one thing we have to

1:05:45

think about which is which is

1:05:47

like software optimizations. There are also,

1:05:49

you know, novel novel architectures. So

1:05:51

I invested in probably the world's

1:05:53

first reversible computing company. So what

1:05:55

reversible computing does is it has

1:05:58

a different way of processing information

1:06:00

and the reason why Envidia chips

1:06:02

give off so much information is

1:06:04

that they're irreversible. So they increase

1:06:06

entropy and you destroy information and

1:06:08

that is, appears as heat loss.

1:06:10

If you don't do that and

1:06:12

you have reversible processes, you actually

1:06:14

can be a couple of orders

1:06:16

of magnitude in theory, more energy

1:06:18

efficient and it comes at a

1:06:20

kind of certain cost of sort

1:06:22

of complexity for building the gates

1:06:24

up that are required in a

1:06:26

chip. But you know, you could

1:06:28

see, you know, 10, 2050X. improvement

1:06:30

in energy efficiency, which is faster

1:06:32

than the wonderful improvements of energy

1:06:34

efficiency we've seen over the last

1:06:36

30 years in computing since we

1:06:38

started to move towards laptops and

1:06:40

then cellular devices. But none of

1:06:42

that actually, that is all extremely

1:06:44

helpful and the market will surely

1:06:46

move to more energy efficient systems

1:06:48

because electricity will always cost something.

1:06:50

but we still have the specter

1:06:52

of Jevan's paradox, which is I

1:06:54

think your observation, which is we

1:06:56

never know how useful cognition will

1:06:58

be. And as we said earlier

1:07:00

in our discussion, I think we're

1:07:02

barely scratching the surface. You said

1:07:04

we're at the five computers stage

1:07:06

of LLLMs. And so, so. You

1:07:08

know, net net, net, I'm sure

1:07:10

all of this stuff is going

1:07:12

to become orders of magnitude more

1:07:14

efficient, right? Sort of digital IQ

1:07:16

points per watt. We'll get far,

1:07:18

far better than it is today.

1:07:20

And boy, are we going to

1:07:22

demand a lot of it. Points

1:07:24

on the scale. It's, hmm. IQ

1:07:26

is a useful obstruction to, like,

1:07:28

bad around casually about it. I

1:07:30

think we'll probably have more powerful

1:07:33

abstractions in the next while to

1:07:35

describe something that it like, you

1:07:37

know, this OLLM is... extremely limited

1:07:39

with respect to its capabilities, but

1:07:41

doesn't have to be super genius

1:07:43

to successfully route a package from

1:07:45

point A to point B that,

1:07:47

you know, we will have others

1:07:49

that are assisting people in doing

1:07:51

cutting out of scientific research, and

1:07:53

every time there's a new model

1:07:55

released every six years, the laboratory

1:07:57

assistance gets, sorry, six years. Oh,

1:07:59

yeah, exactly. That was a verbal

1:08:01

disfluency, rather than a prediction of

1:08:03

immediate cratering of the learning curve.

1:08:05

The research assistance will be getting

1:08:07

shockingly more capable over very compressed

1:08:09

times fans. Well, I think what

1:08:11

you've described there, though, that ecology

1:08:13

is so important for everyone, for

1:08:15

us to understand. You don't always

1:08:17

need, you know, a... PhD in

1:08:19

negotiation to help you decide how

1:08:21

much to pay for the pair

1:08:23

of socks in Walmart that you're

1:08:25

about to buy. That's the price

1:08:27

you just pay for it and

1:08:29

you're done. And I think that

1:08:31

will also be true for the

1:08:33

way in which we embed intelligence

1:08:35

in our in our system. But

1:08:37

we've only really started to sketch

1:08:39

that I mean if you think

1:08:41

about humanoid robots that you know

1:08:43

they're down to a few thousand

1:08:45

dollars in from unitry in China,

1:08:47

how much intelligence or whatever proxies

1:08:49

for common sense do we do

1:08:51

we want, I would say that

1:08:53

it's got to be at least

1:08:55

a GPT-4 level. I mean you

1:08:57

wouldn't want a robot like that

1:08:59

understanding the world as well as

1:09:01

GPT-2 did, which was random sentence

1:09:03

fragments and then going off anywhere.

1:09:05

And you'd certainly want more controllability.

1:09:07

I'm not saying you could just,

1:09:10

you know, lump one of these

1:09:12

GPT-4 class open source models into

1:09:14

a unitry robot and say, go

1:09:16

and look after my kids. I'm

1:09:18

saying that that's kind of surely

1:09:20

the surely the surely the baseline,

1:09:22

surely the baseline. As you say,

1:09:24

you don't necessarily need it to

1:09:26

make a Einstein-level discovery while it's

1:09:28

loading your dishwasher. And these will

1:09:30

be sub-components of larger engineer systems.

1:09:32

I think people extensively underrate that.

1:09:34

We've had robots for a very

1:09:36

long time. There's many folks that

1:09:38

science fiction officinadoes that can quote

1:09:40

Asimov's three laws of robotics and

1:09:42

talk about a day in which

1:09:44

a robot could kill someone for

1:09:46

the first time. And the... worked

1:09:48

in Central Japan, System Engineer and

1:09:50

Me says, oh, that day actually

1:09:52

happened in the 1970s, an industrial

1:09:54

accident. But, you know, we can

1:09:56

do things in factories, like saying,

1:09:58

okay, it is possible. that a

1:10:00

human factor system might not be

1:10:02

sufficiently intelligent to lock in an

1:10:04

uncontrolled environment right now. So all

1:10:06

right, let's cheat all those assumptions.

1:10:08

One, it won't be human factor,

1:10:10

it will be just a grabbing

1:10:12

arm. Two, we're going to put

1:10:14

it in the middle of a

1:10:16

factory where we control everything around

1:10:18

it and put in some factories

1:10:20

yellow hazard tape describing the physical

1:10:22

like... the physical maximum extent that

1:10:24

the arm can move. And so

1:10:26

you can guarantee the robot system,

1:10:28

the invariant, will never be a

1:10:30

human skull inside this physical hazard

1:10:32

tape during your operation and therefore

1:10:34

you cannot crush anyone like an

1:10:36

egg. And anyhow, and then there's

1:10:38

sort of a, we will not

1:10:40

be in a stable equilibrium as

1:10:42

the software gets smarter, as the

1:10:45

LOMs get smarter as they unlock

1:10:47

additional fun toys for us on

1:10:49

the software side. we'll find new

1:10:51

ways to build hardware to take

1:10:53

advantage of those new capabilities to

1:10:55

build out larger engineered systems to

1:10:57

put the smaller hardware systems in

1:10:59

such that they can produce even

1:11:01

more value at scale. So wild

1:11:03

time to be alive. It's a

1:11:05

wild time to be alive and

1:11:07

what we have to do is

1:11:09

fix our are mental models around

1:11:11

how these technologies emerge. And this

1:11:13

intersection between electricity and computing or

1:11:15

AI is a really, it's bringing

1:11:17

together two very different worlds. So

1:11:19

what's happened in the computer industry

1:11:21

since the, you know, the 1960s

1:11:23

and the 1970s is that we

1:11:25

have tried to bring the computing

1:11:27

closer and closer to the end

1:11:29

user, from the mainframe to the

1:11:31

mini time series, then you know,

1:11:33

personal computers and then, and then,

1:11:35

you know, even smaller smart devices,

1:11:37

I'm wearing a smart ring. on

1:11:39

my hand right now. And we've

1:11:41

also then moved into this a

1:11:43

hybrid environment where there are certain

1:11:45

tasks that I do locally, like

1:11:47

my sleep tracking, and there are

1:11:49

other tasks that I push out

1:11:51

where there's lots and lots of

1:11:53

computers, lots of storage, and I

1:11:55

just need to get them done

1:11:57

in the cloud. The energy system

1:11:59

was never like that. The energy

1:12:01

system was all main frames. There's

1:12:03

three mile island, there's... the Hoover

1:12:05

Dam, there's a huge coal station,

1:12:07

some power station, and then you

1:12:09

pipe that power over and we're

1:12:11

all consumers. And so all of

1:12:13

our mental models around this are

1:12:15

based around those types of ideas.

1:12:17

And I think a really good

1:12:20

analogy for how the system changes

1:12:22

is what happened with telecoms moving

1:12:24

from that world, which is what

1:12:26

the old telephony pre-internet telephony system

1:12:28

look like, to what the new

1:12:30

internet. looks like. We went through

1:12:32

a period of time when the

1:12:34

new internet was sort of highly

1:12:36

decentralized point-to-point, but all the servers

1:12:38

were back in, you know, Reston

1:12:40

Virginia or Telly House in Docklands

1:12:42

or Palo Alto Internet Exchange, and

1:12:44

then we've started to hybridize this.

1:12:46

So quite often, when you access

1:12:48

a resource that is a resource

1:12:50

in another country, that's actually served

1:12:52

quite local to you, perhaps only

1:12:54

20 miles away from a front-end

1:12:56

cashing system like an Akamai edge

1:12:58

server or a cloudfare server. That

1:13:00

is a tiered topology. And I

1:13:02

don't see any reason why that's

1:13:04

not what AI infrastructure ends up

1:13:06

looking like. But the bit that's

1:13:08

a real sort of mind... shock

1:13:10

for people is that's what the

1:13:12

energy system might end up looking

1:13:14

like that supports this with localized

1:13:16

generation and vehicle to grid and

1:13:18

community batteries part of the mix.

1:13:20

Yeah, to analogize directly to the

1:13:22

mainframe, love that analogy, you know,

1:13:24

the large scale nuclear plant as

1:13:26

a mainframe, we had a few

1:13:28

decades of usage of like the

1:13:30

local server room with particularly for

1:13:32

industrial uses co-located perhaps behind the

1:13:34

meter small typically combustion-based electricity generation

1:13:36

and then we've had rooftop solar

1:13:38

the last couple of years I'm

1:13:40

somewhat barefsh on rooftop solar but

1:13:42

that's neither here nor there at

1:13:44

least in I think there are

1:13:46

places in the world where it

1:13:48

makes a lot of sense where

1:13:50

for example grids are less reliable

1:13:52

I think it like California An

1:13:55

interesting rabbit hole for people to

1:13:57

go down. California and Texas made

1:13:59

very different bets with respect to

1:14:01

both loving solar. California thought we

1:14:03

really really want rooftop generation. That

1:14:05

seems extremely incentive compatible. It will

1:14:07

be green, etc., etc. But it

1:14:09

won't spoil our beautiful landscapes. And

1:14:11

Texas said, the desert is free.

1:14:13

We are going all in on

1:14:15

utility scale generation. And that experiment

1:14:17

was run. The results are in.

1:14:19

Texas won by a lot. And

1:14:21

so yeah. Anyhow. So. But there's

1:14:23

chicken in every pot. There might

1:14:25

be a battery in every garage

1:14:27

in the very near future. Certainly,

1:14:29

Elon Musk would love to make

1:14:31

wave a magic wand and make

1:14:33

that happen. And community scale batteries

1:14:35

operating at material scale might very

1:14:37

much be a thing in the

1:14:39

next couple of years. It'll be

1:14:41

wild times. So I feel like

1:14:43

we could continue having this discussion

1:14:45

for a very long time, but

1:14:47

I do want to be respectful

1:14:49

of your time and the audience's

1:14:51

attention as well. Where can people

1:14:53

find you on the internet to

1:14:55

see? The best way to find

1:14:57

me is at exponentialview, which is

1:14:59

exponentialview.com or just put it into

1:15:01

your search engine of choice and

1:15:03

it'll show up and you can

1:15:05

sign up to my newsletter there

1:15:07

and then all of my other

1:15:09

links will sort of leaf off

1:15:11

that. Awesome. Thank you very much

1:15:13

for coming out today and for

1:15:15

the audience, thank you very much

1:15:17

for joining us again. We'll be

1:15:19

back next week. Thank you.