Welcome back to Data Driven, the podcast where we talk about how data

and AI are changing the world. And sometimes we

even understand it. Today's guest is the brilliant Carmen

Lee, CEO of Silicon Data and former Bloomberg brainiac

who's now on a mission to bring financial grade transparency to the wild west

of GPU compute markets. If you've ever wondered how to hedge

your AI infrastructure costs the way airlines hedge fuel, or what

a futures market for GPUs even looks like, you're in for a

treat. Carmen's turning raw compute into a tradable

commodity, normalizing chaos, and possibly building the

Bloomberg terminal for AI infrastructure. Minus the beige

keyboard, we cover everything from tokenomics and TSMC

to why your AI startup's margins are flatter than the earth in a

conspiracy forum. Oh, and there's a used GPU car

lot somewhere in Virginia. Stick around. This one's a data

geek's fever dream in the best way.

Hello and welcome back to Data Driven, the podcast where we explore the

emerging field of data science, artificial intelligence, and

this crazy AI world we live in. But it's all underpinned by data

engineering. And with me, as always, is my favoritest data

engineer in the world. Even my dog is barking, giving you a shout out.

Andy Leonard. How's it going, Andy? It's going well, Frank. How are you?

I'm doing well, I'm doing well. I'm keeping busy.

We were talking about other podcasts that we have and

the other one is Impact Quantum. So go to impactquantum.com

definitely check it out. And had a very fascinating

conversation with our guest in the virtual green room. So without

further ado, let's welcome Carmen Lee to the show. She's

the CEO of Silicon Data and she is driven by a

passion for developing and delivering cutting

edge derivative products and data solutions that

provide essential data, intelligence and efficiency to compute

markets worldwide. Her company's vision is to

revolutionize these markets through unparalleled data transparency

and financial innovation. Welcome to the show, Carmen.

Thank you. You deliver up my tagline so well I might want to

hire you to do the whatever. Thank you. This is like.

Thank you. This is like I was looking the other day. This is almost our

400th show, so I do have a face for radio and

apparent thankfully. But a voice for radio. So good for me.

This is great. And speaking of radio, we were geeking out because

I started my career in New York in finance

and Bloomberg. Having a Bloomberg terminal on your desk was

a status symbol. There were the ones who had it and the ones who didn't

and the ones who wanted it. And you know radio,

right? Bloomberg radio, which we also get here in dc. And you used to work

for Bloomberg, so that's really cool. That's right. I had a great time

working for Bloomberg and my team was part of the

data team I thought is

one of the most cutting edge data company especially in the

financial services industry. Back then I cover all content,

all product data integrations with any third

party ecosystems. So think about any training

cycles from Fedmin back offices, think about any

cloud providers and database Systems and

even AI, LLMs, whatever you call them,

different use cases, real time

reference data, aesthetic data, anything. It's

really fascinating. I learned a lot my background before that I

was in all financial services and I don't know if I bore your audience at

this point. I started my career in trading, high frequency trading

in Chicago. So to me transparency,

efficiency and free market is sort of in my blood.

100% brainwashed at this point in life. So one of

the things I noticed when I was a Bloomberg is there's a

lot of interesting ecosystem

platform came up last year, right? So they all leveraging gen

AI. You're the first few adopters which is good for them

and their client basis sometimes can be financial institutions. So boom, client

basis. So one of the things I noticed is it was a really fascinating conversation.

So those startups, they're gaining a lot of tractions. Good for them. So

obviously I was like oh you're doing so well. And they will complain to me

saying that they were sassed, right? They were 100% SaaS

revenue so static and then it's pivoting to

AI driven SaaS. So their cost, think about last year The

GPU per GPU per hour was like $9 or

6, 7, 9. Back to like 3 if

you own interruptible instances, right? So the swing is like

300% within the same day but then their revenue

is static, right? So their margin like

positive 40% to negative, 60 to

positive and there's no way for them to manage it. And also

same time it's not like they bring on more clients. They

can enjoy the scalability. It's like again

same thing, the margin is uncontrollable and they have this problem say how

do they actually coming out a cash flow plan for next year and then

they obviously complain. Totally strikes me to be

hey, this industry needs financial

infrastructure layer, right? It's almost like talking to American

Airlines. Say hey airline, you cannot hatch your oil prices

fluctuation. How are they Going to price their tickets. They can't, right?

And it's not like American Airline cost OPEX in like give me five year long

contract. They don't do that. Every single of those commodities

pricing discovery and hedging happens in divers market. So

futures options because there's a few reasons, right? Number one is it's just

efficient. Number two is cheap. Both is flexible and then you and me,

we can do the same thing. We have oil exposures, we don't have to be

American airline. But today if you are crowing for

hyperscalers, you can go to those, you know,

whoever, right? Produce chips, right? And get a long term contract.

But you, if you and me start Neo Cloud, guess what? We don't have access

to kind of pricing. It's not good. We you have a

few players who have the pricing, who have that way to hedge it. But

then the smaller Prius just couldn't get in the game, right? It's really not good

for the ecosystem's health performance and

the risk management. So that's really struck my core.

Last year I was like man, someone needs to do the

index, the pricing, the benchmarking layer of the

GPU compute as human resource

I feel like will be the biggest human resource in the next few years.

Surpass all energy combined, right? So that's why I left Bloomberg right

away. Super passionate. I think we can bring so much transparency to

ecosystem will benefit everybody, right? Not only benefiting other people,

needs compute benefiting like you know, the end consumers. Because think about

the whole funnel, right? You had finance and gpu the

actual clusters cost, right? So

if the banks don't have enough information or hedging for the

banks then they have to charge you high interest, they have no other way. Or

you have to look for alternative capitals which traditionally

they're more expensive, right? Because they're not banks. Banks are cheap as a

cost of capital, right? So then the cost from you

know, stage zero is high. Then think about the second stage, third

stage and then people like you and me using Sora with OpenAI

everything will be more expensive because of that, right? So fix the problem

with transparency from this from Gecko is really really

critical and then their benchmarkings and encourage the

secondary markets and all those flexibility and then

availability will be really incredible to benefit the whole

ecosystem. Interesting. So is it fair to say you've built basically a

futures market for GPU. Compute I building a

benchmark index layer. We are working with future exchange,

right? So I'm not a futures exchange so that would be something we

will think about S and P. Right. So they license the index with a. Right,

right, right, right, right. That's what we do. Right. Well, we will index

to an exchange and they will have futures options on top of that and other

financial products. That's a fascinating concept because like

you're right, we need that because the scarcity

of GPU compute is a real issue. It comes up.

And if, if, if Amazon, the rate. Of volatility, how do

you. With, with. With like 40,

60% fluctuation every daily volatility and then it's

just not a, a very transparent market

which is. Breeds inefficiency. Right.

Absolutely. So for those of. Oh, sorry. Go ahead,

Andy. Okay. I was just going to ask. So are you tracking

features and functionality and all of that? That, that would be the. How you value

the GPU itself and compare that to the price and

you're coming up with some ratio. Exactly. So

compute is not like. Unfortunately it's not as easy as electricity

or even oil have different grid. Right. So even 100

has different configurations. Right. They all, it's not the same. Right.

Different CPUs, different RAMs and geolocation matters.

Right. So a lot of things. So normalization become very critical

component to financially settle index.

Right now we have H100A100 indexes published at Bloomberg and Refinitiv.

So the way we do it is we have a base case and all

the factors normalize to the base case. And the way we normalize

historical data, what factor is actually important to the users,

the CPU matter? How much does it matter? What's the wave, whatever it

contributes. How often do we calibrate? Maybe it matters

today, maybe tomorrow. This, this particular. Whatever

inputs value more. Right. So we do calibration,

period of calibration as well.

Interesting. Yeah, it's fascinating to kind of see because I mean

it always seemed like there's something missing around

the GPU market. Right. Because it's just. And I also think too

it's been a while since we had any kind of compute limitations on what we

wanted to do. Right. Like that CPU is like. Yeah, it's cheap

and you can get what you want and it's not supply demand kind of shifting.

Yeah, I agree. Right. So I didn't really think of like,

you know, kind of this, this market kind of response to

it, which I think is, is an interesting approach and I think, I think,

I think it's fascinating. Yeah. Even if you think about

AI SaaS company. Right. I don't know if you heard the saying that

SAS is 80% margin AI SaaS is 0% Mar.

So I mean it depends on how you run your workflow. If

you are not being thoughtful, right.

You just dump everything, everything you need to do into the most

expensive closed source model. And you're not

optimizing your thinking tokens, your input tokens,

output tokens. It can get very pricey very

quickly, right? Not batching it, you're not doing all the right things.

And even you do all the right things, it's gonna be such a meaningful

percentage of your cost. And then all those companies not ready for it. Right.

Because in before what's the raw material cost?

Electricity. Like really nothing. Right. But now

every company becomes, you know, a which is great company

but then their cost structure is shifting from zero cost

to. To 40%, 60%, any percent to token

or to GPU at the end. Right? Right. So how do you think about hedging

that kind of cost component? Can you control that? Can you optimize for it? Can

you monitor it, can you benchmark it? You know, can you hedging it? So

no, that's a good point. So do you think

there's multiple, I guess, inputs and levers to this? Right. Because it doesn't seem like

this would be a straight thing. So what's, you know, Andy mentioned that you were

tracking certain benchmarks. Like what benchmarks are you tracking? Because I'm very curious about

this. Right? So there's a few things, depends on

your position at least this can change

every single day. Like our ecosystem is so nuts, right? So it

depends your

positioning, the whole workflow, right. So think about if you are new

clouds, you are selling token, right?

The cost for you is a gpu, right. So then your margin

becomes the diff between the margin and the GPU cost. And that's

the way we calculate it, right. Which is different units.

And then your worry is okay, so for

token survey for the tokens, how much money can I get rate

from one particular gpu? The flops, right? How can I optimize for that? And

what if I'm doing even hosting open source models?

And how do I make sure people using that open source model, should I

shifting it? What's the pricing for that? Think about that strategy and GPU said

okay, am I renting GPUs? I'm like outright

purchase those GPUs and put on my books and depreciate it. How

long can I depreciate it for? How do I let's say if

everyone's the latest and greatest, I'm selling the GPU after second, third year,

what's the terminal value for the GPUs? Who should validate that? Which bank should

depreciate the asset classes. So it's a lot of things coming to the new

cloud space. If you think about your inferencing infrastructure,

right? So let's say you're

AI tech company, right? Then your revenue is token,

right? Ideally they're paying you based on token

use cases as well. And then your cost is token which is

easier but same time for you is thinking through okay so

right now open source tokens, the price

they do move up and down. For example

if you look at Deep Seq, even Deep Seq, they host their own servicing but

then the price changes, they have the off peak hours and that change all the

time. Or you can do closed source which the price is pretty

static. The way I think about it is again it's extremely

free market approach, right? Is how can we

make sure especially open source ones, the token prices

is driven by the market demand supply curve,

right? Let's say if everyone, if I have like 100 GPUs

right now and obviously let's say I

choose to host only one llama open source

model and then I know I can produce X amount of tokens,

both input output tokens, right? And I can just auction off

and you guys and you can buy a million token and one day he's like

I'm not going to use it, why do I sell it to Frank? Can this

be some market where right now you are stuck

with it, right? In

my mind, unfortunately I'm very brainwashed to free market. I feel like you have to

give people option. The more option you give people and

any have flexibility, franchise flexibility and people more

willing to participate because they know they can get out. Because right now you're stuck

with hyperscaler GPUs or any tokens, you're stuck with it

and then you're less likely to commit because you know you

can get out or you get fined even worse, right? You know those cases, you

get fined millions of dollars when you back out on cloud deals.

That's one of the things I really think I should encourage people thinking about tokens

and GPUs as a main cost structure. How can we drive

efficiency so people can commit and then get out if

they need to and then swap out and everyone gets more value

and efficiency from those transactions. So is it

more like an exchange or an auction?

What's the mechanism? Right? So from token GPU side

obviously there's Spot exchange already like compute

exchange, where you can actually tell them, hey, I need this

configuration how many nodes? And then they will

say okay, let's do an auction. And then the

best price, best quality, whatever combination wins. Right?

Yeah. You can potentially do other asset class as well. Right. So we're.

Siliconita is a data company. So think about us as the Bloomberg and there's the

Nices, NASDAQqs and everybody, right. This spot, right,

you can actually get GPUs. You, you can actually get stocks from those exchanges.

And the FAST is we collect data from those exchanges like

Bloomberg. Right. And then we'll produce financial products on top of that. Right.

So that's right, there's spot, which is the

nasdaq. Right. You can buy and sell, get actual physical

deliveries, all the compute or token you need. And there's

data side which is making data the Bloomberg. Right. And then FAST

is structurally the financial products layer. Right, data layer. And

then we're agnostic, meaning we look agnostic of chips,

agnostic of spot markets, agnostic of everything. Right.

And it's a future exchange which they license

our indexes to create futures product. Ideally we're

settling to spa. Maybe some of them will sell at spa. Right. So it's pretty

standard practices. So

would the currency or the coin of the realm be tokens

or compute time or compute seconds? Things

change. It's, it's making my life really fun

and you know, also different. Yeah, all the time.

And then you, you mentioned you have this quantum thing, right? Right.

It's a lot. We track all compute. So it doesn't for us

what chips and what, what architecture framework

and you know, we don't really care. We benchmark the performances and the data

inside. And everything we don't know for us is getting

ready for everything. So we want to create product

that's actually going to be helpful to the marketplaces, not just creating

things like gambling table. People bet on binary things. Right. For

us, how can we make it useful for the people who actually

naturally long compute? So the Neo clouds everybody else,

they need product to hedge their revenue fluctuation. Right.

So they issue short futures and whoever naturally short compute.

So you need computer and for you is a cost management.

So I want to make sure my product is usable by them. It

depends on how they pay. Right. If they pay tokens,

nothing to create token products. You're very right now paying people paying

per GP power and you create product for that. If they pay

things all right, then it's different contracts for that.

So it really depends on how people using it today and tomorrow.

And then, you know, we. We hyped to create products that may

not be the S&P 500, which live forever. We probably create financial

products live for next five to 10 years. Because guess what? Chips

what our style, right? The A100 people still using it, but

like L4s, people are using it, but like other chips like the V's,

the, you know, probably not as much. Right. Then similarly, my

financial products associated with that underlying asset

probably will, you know, retire, be retired. Right? Which is fine.

That's cool. I'm sorry, go ahead, Andy.

I was just thinking about it and a couple of ideas popped into

my head as you were describing that, Carmen. One is

capacity. It sounds like you're literally selling

compute capacity, GPU capacity, time, just

whatever. But it kind of falls into that bucket under one hand.

But then on the other hand, it seems like that

it almost creates this utility market.

Is that fair or am I missing something, right? No,

you're right. But two pieces. So one is a compute exchange part, right? This is

where you can actually get either depends on what people,

the mode of people preferences. You can get GPUs or get

tokens, whatever, right. Physically delivered, you do you. You

don't have to touch any financial products, right. It is literally like you going to

a store buying stuff. And then the more option based, right.

You can actually get instances. And the silicon data is. You

cannot actually getting any compute. Right? Like you cannot

get any stocks from Bloomberg. Well, you can get this data.

What asset is trading, what prices? So that informal decision, ideally

in your spot market be like, hey, I think everyone, you know,

the H100 price is a little too high, in my opinion. I'm not going to.

Right. Right now, like, forget about this. And I can totally use a

100. Right. It's fine. So this data is data

layer, which is liquid data, right? So those are those the

sort of two pieces to I guess resolve the

workflow equation. So it's kind of like when you go to the supermarket. I'm

sorry, Andy. When you go to. That's okay, go ahead. When you go to the

supermarket, you buy the beef, you buy the pork, but you don't think about the

pork belly futures and stuff like that. It's kind of abstracted away from you.

Exactly. The farmers will think about this, right? Yeah, farmers think about it.

Yeah. They need to hatch the corn futures, right? But if

you are farmer, you still say you were someone to eat the

corn. You go supermarket, you don't think about, hey,

Right. So you may have covered this already, but how does

or does fungibility come into play?

It's a great question. So I went through so many different iterations about this.

Initially I was like, okay, why don't I just normalize across flops? And I was

like, nope, can't do that because there's

just, there's so many things wrong with this approach. But obviously

we can dig into details, but we're not going to do that. And then secondly

is okay, why don't we do like inferencing

chips? Like just make a pot and then we realize, okay, how can.

So again back to the initial question. I want to make product actually going

to help people hedging. Right. If you

do a combination of different chips, then if you

are and you know we're using of a lot of people, are you going to

really use that to hatch? How would some correlation look like. Right.

Maybe you just rather have different chip types and then just hatch accordingly

because the correlation will be much higher than the combination of indexes.

Maybe the composition of indexes is good for just tracking

general, but not for actually financial products. So we have, we have,

we can have all. Some of them will be tradable. Some of them. Well, right.

For us is if people start, if, if we move to the world

where it's not going to be Nvidia only kind of play

in the like amd. We can eventually,

it'll probably end eventually. Well, we'll see when, right? We'll

see quantum happens first or everyone catching up first. I have no

idea. Right. So if it's like a more vibrant

ecosystems. Right. And then maybe we're thinking about, hey, maybe we can do

like doing some of the chips. Even different firms would normalize it and then we

do something like a inferencing chips, chaining chips. I don't

know. So that's another thing. Or like token, token indexes. Right.

So can we do open source ones? Multimodality? Is

multimodality going to be a thing in a few years? Everything going to go back

to one model only? Because right now with different models. But maybe it's the interim

stage. Right. We. I don't know. So it's one of the things we have to

keep like looking and thinking and just moving things

forward. Yeah, I was thinking too about, you

know, the, the amortization that people

do in their heads at least when they buy a new car. Yeah. So

that's the math is you drive it off the lot, it's worth what, a 75,

80% of what you pay for.

So we need a Carfax for GPUs, right? So that's what we do too

for silicon Mark. So what we do is okay, everything. Well

at least right now or before Last year or T minus 1, everything

is brand new. So okay, we'll take whatever the

number they published and tdbs, the flops, we all know

there's like haircut to that number.

That's funny, right? And then a year later, right, A year later I

say, Andy, you're growing great in great data centers. Your

thermal cooling was doing great. I'm old data

center, I don't have the latest cooling. Obviously my chip

is after year. You can argue they own different curves,

decay curve. And are we treating the same prices even

though same configuration? Probably we shouldn't. Should it be reflection of

the actual quality? So that's something Mark does.

And then we do things even more basic than that. So number one is

when you tell me you have H100 like 100 nodes, each node has

say 8 GPUs, right? Yeah. Is that true? Can I

number one verify the UID of that? And you see, it's all the CPUs

and this operation systems on all

the nodes, they all live connected. Number one, can we just

verify are they connected? What's the latency? So that's very

basic things, right? So we do that piece at least, you know,

are they truly UIDs and CPUs? The machine, is the machine ever

changed? Because we do mesh IDs based on

CPU changes. We know something changes, right? And then the UID of every

chip. So we do the decay curve for the individual chips and also the machine

level and then thermal staggeration, everything. So we do

that and then we do validation. Almost like Bloomberg Validate fixing

compound. Because you have to understand the issuers and it's

a bridge and it's a school and with cash flows and all those stuff. So

we do that for GPUs. The geolocation. If you build data

centering somewhere in North Korea,

it's great, but no one going to use it, right?

We took all those in considerations when we created those data models. So then

we figured out, okay, so based on the setup and

we run a benchmark on specific GPUs, this is our grade and then

this is our validation. Obviously you can do whatever you want. And then you can

say hey, screw that, I believe this is much higher price. You can do that

as well, Right? But this is our valuation. So almost like a scoring system.

That's interesting. So My mind immediately went

to, when, when we started talking about cars, my

mind immediately went to, you know, the used GPR lot

some guy in bib overhauls out here in Farmville, Virginia

kicking the tires. What's it going to take to get you into this

gpu?

Yep. See, there we go. And network them together. Right. Like I think there's also,

you know, maybe, you know, I don't know if

you've been tracking the, the DGX Spark device

that Nvidia has, but apparently they have ports

in them so you can network I think up to four together. I'm not sure

but yeah, I'm sorry I

cut you off but like. No, no, no. Nvidia we

definitely leveraging a lot of. So we do the container within

container and we do integrate with Nvidia DGX

benchmarking. So they have open sourced some of their LM

benchmarking based on GPUs and we do streamline their products so

you can test lms. So Nvidia Digitex testing

through system data. The benefit is if you do it all right yourself,

number one, you can

obviously people want but people can just change up the, the

benchmark results themselves. Right? It's open source but through us it's data Oracle. You

can't really change results. Number two things is more streamlined. It takes a few hours

to run versus take weeks because you've download a bunch of things you may or

may not need. You may or may not need.

Well, I also think too like, you know, how does this, you know, you

mentioned you, you kind of skirted around the location thing with sovereign

AI, right? So like if I'm okay with using Google

Services, right. And I can, I have access to TPUs, right. I have a lot

more access to whatever Amazon's chip. Microsoft I think is

working on something. Custom that's on prices too, right. The Geolocation

they have different prices and different carbon footprint. We haven't even touched that.

Right, right, right. We do track that as well based

on local grid power grid information. We do track the carbon cost associated with

different AI workflows. I think it's important, I think so

for me is let me at least surfacing the number to you and

you decide what to do with it. Right. So I think that's a good idea

or you know, maybe it turns out that you know,

this type of model of GPU is you know, depending on what your

core. I think it's, I think it's great because I think one of the things

that I've heard And I didn't Peter

Drucker. What gets measured gets managed, right? So you're, what you're doing is you're providing

ways to measure GPUs and GPU performance. Right.

So if I don't care. One of the things I heard about and I'm sure

you have some thoughts on this is like cloud providers that are

starting up and they're just doing

GPUs, right. They're just doing kind of training loads. Right.

And they don't need to be located anywhere special. Right. Like they don't

need to be in the northeast corridor. They could be in the middle of

nowhere as long as they have power. Right. And

because you're going to run a load, right, you're going to run a load on

the thing, it's going to take 72 hours say to run. You don't really care

if the latency is, you know, 150 milliseconds versus

3. Right. It doesn't really matter. Yes.

That's why you see a lot of us get built up in like Iceland, Finland,

the users can be in Americas, can be in Asia. Right,

right. For them is can they get the capacity

looking for and you hard deal if you're thermal powered

data centers, cheap electricity. Yeah.

And then it's cleaner supposedly. Right. As

long as you're not on the volcano belt.

Right. As long as it's not going to blow up. Yeah.

But yeah, so we definitely see that trend and a lot of energies, you

know, what do we call it oversupply sometimes can

be in Spain because overbuilt and the grid couldn't handle it. And

then they need to get data center up and running like now to take over

the power. But then

it takes a lot to make the racks start running. Right.

More than just the GPU itself, you need the connectivities and network

and that could be in shortage. So you need to solve a lot of different

pieces to actually deliver the actual computer.

But that's why it's fascinating industry for us because

we see things from dsml, tsmc, side.

So anything supply demand shifting will have

an impact on the whole ecosystem. And then this industry is winner takes off

from LTSMC to a solution level.

You have to be the solution. Your alternative solution just not

going to work. So. So every single piece is so critical to

the whole chain packaging. Right. You have to work,

right. If you don't know how to do it, then you just can't do it.

It's not like you can buy a cheaper pair of socks or whatever

so we do. We're from end to end, right. From the SM of production,

tsmc. So we're official TSMC partners are going to be actually

TSMC conferences to this

November. Very cool. It is really cool. I

kicked out by those stuff very quickly. And all the way to

the model A, the token layer. Right. Agentic layer. So

we sort of see things all the way. Which

I think my brain get overclocked every single.

I know what you mean because I get till the time of like

2:33pm and I'm like, I can't take any more input. Like

and the muscle, my brain muscle just dead. I know. How

do you do that? How do you get a roller in my brain, just like

relax my brain muscles? I. I found going for a walk

is a. Is a good way to do it. Right.

No, like. And a co worker of mine calls it everything turns to

hieroglyphic hieroglyphics when he's like

looking at like stuff. And I was like, yeah, that's a good way to put

it. Because it's just kind of like, yeah, I can't. I don't want to have

time by a daughter. So I usually spend time with my daughters. I feel like

they've been silly. And I would tell them, I'm so stressed out. When my daughter

was like, me too. I was like, what are you stressed about one last donut

than the other guy. I was like, that's very important thing. I agree with that.

That's very stressful. I will be really upset if I get one less

donut. So. Yeah, so definitely put things in

perspective. Yeah, that's cool.

I think one of the best things. Any other questions? No, plenty,

plenty. Like, I'm just fascinated by this. I know

we're kind of short on time, but one of the things that you mentioned was

tcmc. Tsmc.

So for those who don't know who they are and how important they are to

the global economy, could you explain for those folks

and why I was so excited that you're going to one of their conferences? I

didn't know they had conferences, so. I don't think I would do the justice

of explaining how important TSMC is. All right, how about I explain it and

then you tell me where I'm wrong. I'm sure you'll do a better job

than I can. So. Tsmc. Taiwan

Semiconductor Manufacturing Company. That's right.

They are based in Taiwan. And

the reason why. Nvidia. There's a fascinating

story in the book called the Nvidia Way. I Don't know if you've listened to

that or read that book. Really awesome book. But basically

one of the advantages Nvidia had early on and arguably

now was that they off they outsourced their chip

manufacturing to this company tsmc. I'll get it right that

time. They are basically what they call a fab.

And you could, I mean not

now they're so busy like you know, you kind of the you in general. Right.

Like I couldn't call them up and be like hey, I have some prints for

you. I have some chip designs I want you to make for me. Can you

send me. They're not at that scale but

so they're a fab. And so what happens is people like Nvidia, companies like

Nvidia, a few other companies too will go and they will, they

will design their chips and then they'll, they'll basically

not drop ship but effectively kind of print to order

chips. Which frees up a company like Nvidia

from having to build their own fabs. Kind of like intel does. Is that a

good description? 100 so I usually call

on Nvidia and AMD like design houses and then sometimes

confused with people who's like oh, are they like Louis Vuitton was like no,

Right, right. Or like graphic designers? Yeah, yeah. So they're design

houses and then they are Fabless. Right. And intel,

which is interesting because they do both. Right? Yeah, yeah.

Intel like as I was saying that intel doesn't. Yeah, they do both. Yeah.

Right. And then it could be a great strategy. Could work

or. Well, depends on many things. Right then anyways,

so TSMC is like the, as I said before, this

industry, I don't know if it's good or bad but it's a winner takes

all market. Right. So TNC is definitely

the winner for a lot of different

reasons. I think for the leadership, self

and technical team for the whole supply chain ecosystem. The

gravity, all the years, the hard work they've put in.

So it's a position where I don't think anyone

can seriously challenge them

in a meaningful way in the next whatever

years. So they're very critical. And then the good

thing interesting about them, they're the agnostic of design houses,

right. So they have great relationship with Nvidia for sure and I'm sure with

them, with everybody, right. It's their job to

produce those chips and then it's

interesting enough it's aligned with mine. Silicon Data. Because

I'm agnostic of chips, right. So

obviously I want to create products that's most important to the

ecosystem. So right now people care a few chips and

those chips happen to be from one design houses. But let's say

if another design house start picking up a lot of momentum. For me, it's

like, how can I help everybody in ecosystem

compare, contrast hashing, right? Use them benchmarking, normalize

it in a meaningful way. So it's my job to work with all the design

houses. It's their job to produce chips that can be usable for

defunding the houses too. So we're very aligned in that sense. And

anything they do, right? So think about, they are

future looking because they're not thinking about next year or next quarter. They think

about 20 years, 10 years. It takes them five, six

years to build a fab, right? And then they need a fab to

be utilized. And they have a threshold, right? If you're

building a fabric and that's not utilized by year eight,

they plan right now by year a year 10, they are

losing a lot of money. A lot like billions of dollars,

right? Like can you make sure the fab will be utilized, the demand

will be there by year 10. Forecasting from today.

It's very, very, very hard job to do. And it's not

like it's not like a new reim, you know,

like what are minings and all things that you can hedge it, right?

Like there's a way to hatch the future curve. But like it's not like they

can forecast, forecast and do a swap on that because

the market is so concentrated and then very

binary and a huge size. Who's taking the other side?

I don't know. It's very hard over the concentrate to

do so for them is to get clarity supply demand curve in 10

years. I mean they do also edge computing chips as well, not just data

center chips. Right? But how do they think through that? I think that's

really challenging. I think will be really challenging for me

for sure. I'm sure they have way smarter people there to think through those problems.

But yeah, it's an interesting problem to have.

That's why TSMC and I, for example, they sell to

their clients who are in the vds of the world. So they have that kind

of transparency. But what they don't have, which

may be a different indicator for the supply demand curve in

10 years is end users

pricing volatility. Right? And then you know, okay, so if

every single chip, every single chip I produced, right, Data center

quality chips, one dying price, right.

Is the indicator for supply demand shifting. Maybe it

is Maybe it's not right. At least you have some, some data points which your

immediate sales and revenues which is T0

won't give you because then a few degrees removed from

end user experiences you give Nvidia and Nvidia packages it to

AWS and GCP and end users and you and me. Right.

So that's something that for them to think through as well.

Interesting. One of the stories I heard and I

wonder if it's true, was that part of the

reason why there was part of the reason

why Nvidia was able to really capitalize on this. There's a lot of

reasons, but one of them was the fact that in the

crypto craze, the run up to get chips for that Nvidia

had purchased. Now what you said makes a lot more sense now. Nvidia had purchased

the. They basically purchased a certain amount of capacity at TSMC

for like three to four years, something like that. And then that happened to

coincide with the AI boom. Is that, is that true? And

that. I guess that's a market too, right? Like you know, like hey,

so I wasn't. I know so 7 so I'm not following all ASICS

so they have a specific for. For the, for. For the mining

chips. That could be true. So I think

not because I'm straight, I mean and a girl can dream. I'm

strapped to be like, you know, to really

help the industry and then be, you know, like

the company the team hopefully can propel the industry

move forward. Right. I'm strive to point zero over percent people

and then competency is very important. Obviously execution, your

hard work is important. Not a big piece is you have to be

really, really lucky. That is also everyone's control.

And then Nvidia puts so much time effort into everything they do. You can argue

they were really great company even before the AI boom and

everything. But the lock piece and how do you control that? How do

you. How do you know quota gonna be like the piece

that's needed? Right. Well, some.

Someone said that, you know, Jensen Wang is like the epitome of,

you know, the better you, the harder you work, the more luck you have.

True. Like there's a lot to that and I know it's

complicated but like I'm just, I just. It's interesting how the crypto kind

of boom and bust really kind of also

propel us into the AI. Not, not all by

itself but it definitely I think gave. There was some momentum where

no momentum was expected, if that makes sense. Right. Yeah, I agree,

I agree Timing is so interesting, but

we just have to two point like the heart of your world. You have to

do everything you can with the environment. Right? That's

cool. That's cool. All data. So we'll see happens what

I mean. That'S the importance of data. Right. Like, you know, people don't realize that.

And I go calling back to Bloomberg. So I'm referring to Michael Bloomberg,

former mayor of New York. But before he was mayor he

basically started a company called Bloomberg. And

he was not the only factor but like

a big part of, you know, people getting into, you know, his

philosophy. As I understood it, if there's a good, if there's a good biography book

on him, I totally would want to listen to it. But basically getting

the traders access to data gave them an advantage. Right. And it was

really, he was really early on in the idea of that data is

not just something that's created as a byproduct of

transactions, but can actually be, you know, monetized

and arguably weaponized. Right. Like so.

And you know, Bloomberg terminals

before, you know, it was interesting because he basically sold these custom terminals so you'd

not to rely on like local ID who were still struggling with like, you

know, just keeping the network up and running, you know, these separate

devices that became status symbols. And ultimately he, that's become like

this media empire that, you know, I can watch Bloomberg on my

tv, I can listen to it, you know, whether it's a satellite radio or the

app or you know, FM or AM radio

stations. You know, I think it's in San Francisco, New York and

D.C. they have a big office in D.C. they always have an

interesting show called Political Capital. I think that plays

at 5pm every day. I listen to it because it's kind of the

policy side of finance and kind of what's going on in the world around.

And AI has come up a lot digital sovereignty. So it's interesting

how all of these worlds, I like your thoughts on this,

right. The worlds of finance, the worlds of tech and the worlds of policy,

politics and dare I say war. Right. They're all kind of like

crashing together in this giant thing. And

it's kind of cool, kind of scary.

I think it can be. I mean, sometimes I'm scared I was like,

you know, because you see a few things, it's like, whoa.

There's a lot I feel like for people born post Covid,

not born, but grew up post Covid, I would call Jen the second

Gen Z Gen Alpha. Yes. I think Gen

Z's apparently now like I'm all confused. But for

them it's like, of course they should. They should. My AI should be my

boyfriend, girlfriend. Right. Like whatever. And then for me it's like,

this is not comfortable at all. Weird.

Yeah, yeah, yeah. For me it's not. I have no idea what's going on.

Like, I just so creeped out by this. But for lot of people it's like,

of course you do that. Of course you tell AI all your secrets.

Of course they can. My phone can record my conversation. Of course

you can train, you know, your AI model. My

model use my all my Gmail content information.

All edge computing. I have my own AI model. Of course you can wear,

you know, glasses and then record everything you and me talk

about. And how secure is everything

right now? Right.

The hardware level encryption

is only available on a very specific few chips.

TPU can do that. You rely on software encryption.

No, it's true. And software encryption that is vulnerable to a quantum

attack which is not that far away. We are not the

software and use cases moving so quickly. The hardware hasn't been able to cut

up. And it's expensive to do hardware encryption. It takes

longer and it's more expensive. That's why sometimes the hyperscaler charging

higher premium for that reason. Right. Are you willing to spend a

token and time and effort to do so? Some use cases, you can argue.

Yes, yes, absolutely. No edge computing

chips can do that kind of hardware level encryption.

And it's happening like now. Right, Right.

I was talking to a startup called Quantum Knight. Nate claimed to have a solution

that is a low, low compute kind of post

quantum ready thing. So I can send you their

link and information. Yeah, we, we track quantum

computing prices as well. Very different than GPU pricing and like, you know, like

a thousand per second per minute pricing versus hourly. Right. This is like

different cycles you run. And then GPU become like error correction component to the whole

thing. But for us it's like, okay, so

computers compute now, GPU and tpu, whatever, pu. And then

it becomes like quantum. How we think through that? I don't

know. My brain just like, you know. Yeah, I know. At some point it just

becomes like. I'm not smart enough

right now to, to. To. To figure that out. I tell you, like I go

through like quantum stuff and like I always joke with Andy, like I'd be like

15 minutes, I get a migraine, which is basically like my brain's version of

blue screening. And like, just like, okay, I can stop. I can get

to about. I can get to about 45 minutes now, which is, you know, an

improvement. But this is actually a good book.

And he was actually a guest recently on the Quantum Computing podcast.

It's a thick book. It's a thick book. But I'll tell you this.

The, the, the, the first three chapters, introducing the concepts

are probably the single best introduction to the

concept I have ever read. Yeah, I will send you the link. Yeah,

yeah. Dancing with Cubits.

Really interesting book. Super nice author too. He's a, he's a trip.

But it, it. No,

you're right. Like, these are. The thing that really worries me is I kind of

think about this like we built our entire economy and we're, we're

on a house of sand. Can we start on this? That's

another thing. We'll have to have you back on the show for

a second one. But like other countries

where they lay off hundreds and thousands of people, not. Not just by American

companies. Right?

Yeah. Don't even get me s on that. Well, like, and like, you know, we're

all based on. And, and the other thing, the elephant in the room, right, is

the fact that TC the, the T in

TSMC stands for Taiwan. Right. Kind of.

I know, I know it's very dangerous to talk about this, but, but like. It'S

kind of like, shoot. So I won't say much, but I'll just say it's

contested real estate. How about that? Right. That's a pretty safe way to say it.

Right? It's contested. Right. And you know,

the entire world effectively revolves around the kind of modern

civilization revolves around the manufacturing that happens there. And

God forbid, like, you know, whether it's man made or a tsunami or a bad

earthquake, like, I mean, our world, I mean, we, we get sent back

to the 1700s pretty quickly. You know, 1700 is not,

you know, there are still people, they're still human beings in the

hundreds. That could be worse than that. That's true. It could be way worse than

that. That is a good point. I was trying to keep it. I was trying

to end it on a positive. And I know you're traveling there

like no humans. Well, no, I mean, like,

I mean, there's a lot of ways that the, you know, this apocalypse could go,

so to speak. Right. It could be, you know, but like, it's a very. And

like, just from an infrastructure point of view and supply chain point of view, like,

you know, we, we. We've really championed

globalism and kind of all of these extended supply

chains for, you know, there were reasons there's always reasons, but like

at the cost of resilience. Right, right. That's kind of scary.

I assume you've read Taleb, right? The like anti fragile.

I'm so sorry. No, that's fine. That's fine. But I really appreciate you taking the

time. Where can folks find out more about you? Silicon

Data.com Silicon Data.com awesome. And we'd love to have you back on the show.

And you can tell us what these conferences were like. The. The

ts. Let's see how much I can understand

first. Right, right, right, right, right. That wasn't a good question.

That's why you got to be like the kids today and record all your conversations

so you can talk to the transcript later. All right,

nice seeing you guys. All right, thank you. And we'll let our AI finish the

show. And that wraps up another episode of Data Driven, the podcast

where we ponder the future of AI data and occasionally

the fate of humanity if we don't get GPU pricing under control.

Big thanks to Carmen Lee for joining us and blowing our minds with

compute market mechanics, financial innovation, and just a

touch of economic existentialism. Be sure to check out

silicondata.com to learn more. Just don't try to day trade

H1 hundreds after midnight. If you liked what you heard,

subscribe, leave a review, or send us compute credits.

Until next time, stay curious, stay caffeinated,

and remember, in a world of exponential AI, transparency

might just be the killer app.