Full Transcript

Martin Shkreli Joins The All-In Crew To Speak On Quantum Computing

8:03853 summary words · ~4 min readEnglishBy Shkreli PlanetTranscribed Oct 13, 2026

Analyze another video with Pro30-day money-back guarantee

Summary

Quantum computing is fundamentally unsuited for commercial AI workloads due to architectural bottlenecks and algorithmic narrowness, whereas analog optical computing offers a superior physics-based paradigm for accelerating matrix multiplication.

Billions of dollars in enterprise CapEx and venture investment risk being misallocated toward quantum hardware that cannot outperform classical silicon on general computational workloads.

Section summaries

Opening Banter and Quantum Fraud Thesis

optional

The hosts welcome Martin Shkreli and immediately prompt him regarding his public claims that quantum computing commercialization is fraudulent and could lead to executive indictments. They jokingly inquire about survival advice for prospective corporate inmates based on his federal prison tenure. Shkreli establishes the context for why he took a public short stance against quantum computing startups.

Shkreli has publicly framed commercial quantum computing claims as fraudulent hype.
The discussion frames deep-tech corporate promotion through the lens of securities liability.

Primarily casual host introductions and prison humor before technical arguments begin.

Algorithmic Limits and Hardware Realities of Quantum

watch

Shkreli cites research insights from his wife Chelsea, an early OpenAI staffer who studied under quantum theorist Scott Aaronson at MIT. He explains that quantum computing originated around Peter Shor's theoretical work showing polynomial-time reductions for integer factorization and discrete logs, rather than universal acceleration. He contrasts this with basic tasks like sorting algorithms in CLRS, noting quantum hardware has slow clock rates in the hundreds of hertz, lacks quantum RAM, and cannot accept high-bandwidth data inputs.

Quantum computing is not an across-the-board accelerator; its advantage is mathematically constrained to integer factorization and discrete logarithm problems.
Commercial algorithms like sorting perform far worse on quantum systems than on classical CPUs.
Lack of quantum memory architectures eliminates compatibility with data-intensive AI models.

Core breakdown of the mathematical and architectural misconceptions surrounding quantum computing.

The Transistor Counterargument and Short-Seller Drawdowns

watch

The host pushes back by comparing current quantum hardware to the primordial transistor era, asking whether today's limitations are simply early engineering constraints. Shkreli responds with historical market precedents like vacuum tubes, punch cards, and mercury delay lines that went obsolete before yielding sustainable businesses. He notes that empirical testing on cloud systems from IBM and Google failed to compute simple double-digit products, though he acknowledges sustaining substantial financial losses shorting quantum equities during market irrationality.

Historical hardware paradigms often bankrupt their early investors before dominant silicon designs emerge.
Current public cloud quantum systems struggle to solve elementary two-variable arithmetic reliably.
Market timing risk can cause severe financial losses when shorting technically flawed concepts.

Crucial context on market mechanics, failed historical hardware analogs, and the pain of shorting hype cycles.

Matrix Multiplication Scaling and Analog Optical Computing

watch

Shkreli shifts the focus to modern artificial intelligence, defining modern AI accelerators purely as matrix multiplication engines operating under cubic time complexity. Digital GPUs execute a billion operations to multiply two thousand-element square matrices, consuming massive amounts of power. Shkreli introduces analog optical computing based on a Microsoft Nature paper, arguing that calculating matrix multiplications through light wavelengths occurs near-instantaneously without gate switching overhead, citing startups like Neurophos and Lumai.

GPU compute scaling for deep learning is strictly bounded by digital matrix multiplication latency and power draw.
Photonic systems execute linear matrix transformations passively through light wave interference.
Optical hardware startups represent a higher-probability paradigm shift for machine learning than quantum devices.

Delivers the central technological thesis comparing GPU matrix bottlenecks with passive photonic acceleration.

Early Trading Career and Wall Street Background

optional

The host pivots from compute hardware to Shkreli's formative years in finance, asking about his career origins and the controversial pharmaceutical acquisitions that drove his notoriety. Shkreli recounts landing an analyst role at Jim Kramer's hedge fund at age sixteen in Manhattan, describing Kramer's high-turnover discretionary trading approach during the late 1980s and 1990s that generated strong fee-adjusted annual returns.

Shkreli started working on Wall Street at age 16 under Jim Cramer.
Cramer's early fund generated roughly 23% net annual returns through high-frequency discretionary churn.

Historical personal narrative and trading history rather than compute or economic analysis.

Key points

Narrow Algorithmic Scope of Quantum Advantage — Quantum computers do not provide general-purpose acceleration; their mathematical advantage is almost exclusively restricted to Shor's algorithm for prime factorization and elliptic curve discrete logarithms. Conventional computer science operations like sorting or basic arithmetic run vastly slower than on commodity silicon.
Architectural Bottlenecks: Memory Absence and Clock Latency — Existing quantum machines function at single-core clock speeds measured in hundreds of hertz rather than gigahertz and possess zero persistent quantum random-access memory (RAM). This architectural deficit requires extremely small input vectors, prohibiting integration with high-throughput data pipelines.
Analog Optical Computing for Matrix Multiplication — Modern GPU compute scaling is constrained by the O(N^3) time complexity of digital matrix multiplication across billions of operations. Optical systems exploit the passive wave mechanics of photons to compute linear algebraic matrix multiplications essentially instantaneously with minimal thermal dissipation.

“The modern quantum computers run at a clock speed of hundreds of hertz or thousands of hertz, not at the gigahertz that we're used to.” — Martin Shkreli

“Chelsea and I have tried every single quantum computing uh quantum computer in the world. We cannot multiply 300 by 200 on any of these.” — Martin Shkreli

AI-generated from the transcript. May contain errors.

0:00

Please welcome Martinrelli.

0:06

>> Good.

0:07

>> Welcome, Martin. Thanks for coming out.

0:09

[laughter]

0:09

>> Nice to meet you.

0:10

>> You wore Brad's outfit today.

0:11

>> Exactly.

0:12

>> All right. Welcome, Martin. Um, we got

0:15

so many topics to go into with you, but

0:17

I want to start with one. Um, you've

0:20

[snorts]

0:22

basically taken down quantum computing.

0:24

You said you'd think that this is a

0:26

fraud uh that people might be going to

0:29

jail. So twopart question. Number one,

0:33

uh maybe you could give us your thesis

0:35

on on why this is a fraud or fugazi or

0:38

fugizi whatever it is. Um

0:41

>> discretiad

0:42

>> and a discard and then two your best

0:45

advice for those CEOs when they do go to

0:48

jail. You've got a little experience.

0:50

>> Yeah. [laughter]

0:52

best advice in terms of you don't seem

0:55

to be the kind of guy who would thrive

0:57

in prison, but from what I understand,

0:59

Martin, you shot,

1:02

you turned into Lex Luthther and you

1:05

were running the yard, I understand. So,

1:07

break us through the quantum trade and,

1:09

you know, best practices when you go to

1:11

the Huscow.

1:12

>> Yeah. So, I I have to shout out my my

1:14

wife Chelsea. So Chelsea um very few

1:18

people know her but she's uh I think a

1:20

big deal in technology. She was one of

1:21

the first people at OpenAI and her

1:24

father believe it or not was one of the

1:25

first people at Nvidia. In my family I

1:27

was one of the first people in prison.

1:29

So [laughter]

1:30

I thought I'd switch the gene pool you

1:32

know shifted over a little bit but

1:35

Chelsea t um studied MIT and taught uh

1:38

quantum computing with Scott Aronson the

1:39

the pre-minent guy in quantum.

1:42

you know, we we uh she she got to teach

1:44

me over many hours, you know, uh

1:47

penetrated my thick head, you know, what

1:48

exactly quantum computing even is. And I

1:50

think a lot of people sort of synonymize

1:51

it with faster computing, which sounds

1:53

great. That's really not what it is at

1:55

all. No.

1:56

>> Um and I think the the the basic point

1:58

of quantum computing is that there's one

1:59

algorithm. In fact, the whole field was

2:01

invented because of Peter Shor's thought

2:03

experiment. Peter uh considered this

2:05

idea that there are kinds of algorithms

2:07

that you can run faster in different

2:09

media. That was a revolutionary idea in

2:12

computing but it's not really a

2:13

revolutionary idea for business because

2:15

and this is my thesis of course uh

2:18

because there's one algorithm that runs

2:20

faster in a quantum computer and so if

2:22

you ask somebody which computer can sort

2:25

a deck of cards if anybody know the

2:26

famous algorithms book CLRS you know the

2:29

first five pages are about sorting

2:31

algorithms like bubble sort and merge

2:32

sort and things like that and I asked

2:34

quantum fans well which one can sort it

2:36

faster and they said well of course the

2:37

quantum computer and I said of course

2:39

That's not the case. The modern quantum

2:41

computers run at a clock speed of

2:43

hundreds of hertz or thousands of hertz,

2:45

not at the gigahertz that we're used to.

2:47

They're single core and they're single

2:49

threaded when it comes to doing a job

2:51

like a merge sword or something like

2:52

that. So the only thing it's really good

2:54

at which indeed is a magic trick is

2:56

going from exponential time algorithm

2:57

which is integer factorization or

2:59

elliptic curve discrete log problem

3:01

which I want to mention a little later.

3:03

Shout out to the ECDLP problem. The

3:05

quantum computers can do those in

3:07

polinomial time, which is remarkable.

3:09

And to a nerd, that's the greatest

3:11

coolest magic trick ever. But to a

3:13

business person, unless you were in the

3:15

business of factoring numbers, there's

3:16

there's not a whole lot uh to be excited

3:18

about here. So the other algorithms that

3:20

quantum could do are not much faster. So

3:23

they'll never beat an Nvidia machine.

3:24

They'll never beat a million Nvidia

3:26

machines. And quantum doesn't have

3:28

storage. There's no such thing as

3:29

quantum RAM. Uh so you have to design a

3:32

problem that has a really really small

3:33

input. uh you cannot put the world's

3:35

compute into it. Every time somebody

3:37

pitches open AI from what I've heard

3:39

about a quantum computer, they they say

3:40

no, we're not interested.

3:41

>> I I think I'm closer to you in my

3:45

perspectives on quantum. But let me take

3:47

the other side. Let's just take the

3:48

other side. The other side would say,

3:49

look, we're in this very early

3:51

primordial phase of actually, you know,

3:54

constructing the transistor. And from

3:56

the transistor, what we're going to do

3:57

is like put a bunch of these things

3:58

together. We're going to figure out

3:59

gates. We're going to figure out logic.

4:01

We'll then be able to build some

4:03

reasonable operating system on top of

4:05

it. We'll have some reasonable, you

4:06

know, as like uh mechanisms to write

4:09

software to these things and none of

4:10

that really exists. So, do you buy that

4:14

we're just at an early phase and this is

4:16

why we're frustrated that there's a a

4:18

computational limitation except to a

4:20

very narrow use case? Like do you

4:21

believe that that could be possible?

4:23

>> Yeah. No, of course. Uh I think the

4:24

issue though is do you buy would you go

4:26

long punch cards and vacuum tubes,

4:28

right? uh mercury delay lines, core

4:30

memory, these are all things that

4:31

disappeared and went bankrupt. Those are

4:33

those were stocks at one point. Uh CTR,

4:36

you know, uh computer tabulator machine

4:38

got incorporated into IBM.

4:38

>> But okay, so but your critique is is

4:41

business model and timing or is it

4:43

technological?

4:43

>> I think it's both. You know, I think

4:45

that if you when you talk to Google or

4:46

IBM or Microsoft now, why why are you

4:48

spending hundreds of millions a year on

4:49

this? They say, well, someday there

4:52

might be something that comes out of it.

4:53

That's not a great business to me.

4:54

>> Well, it is very shocking. I mean, you

4:55

you can probably tell the story, but how

4:57

many machines actually how how many

4:58

machines does IBM and Google actually

5:00

have?

5:00

>> Uh, very few. So, Google has a few

5:02

prototypes. IBM has a little cloud.

5:04

Chelsea and I have tried every single

5:05

quantum computing uh quantum computer in

5:07

the world. We cannot multiply 300 by 200

5:10

on any of these.

5:11

>> And you're shorting them?

5:12

>> Yeah, I've been short them. Uh,

5:14

unfortunately, we lost millions of

5:15

dollars shorting quantum computing

5:17

companies, but I told my wife that we

5:19

gained billions in insight. [laughter]

5:22

>> Right. It's uh it's it's not easy to

5:25

time the markets. Yes, definitely.

5:28

>> So, where where is the natural

5:30

computational edge? I mean, some people

5:31

would say it's going to be towards

5:33

encryption and defeating encryption

5:35

because now we can actually use what

5:36

it's actually good at to essentially

5:38

brute force solutions to things that we

5:40

can't solve today.

5:41

>> I I think it's crazy to even think about

5:42

it because we have a computational

5:44

problem that we're trying to solve. And

5:46

every time somebody uses the word AI, I

5:49

actually think of the word Matt Mole.

5:50

These are Matt Mole machines. And I

5:53

think that as we

5:54

>> matrix multiplication is what he's

5:56

referring

5:56

>> right. So so multiplying two lists of

5:58

numbers is what a GPU spends 99.9% of

6:00

its time doing

6:01

>> but it does it in n cubed time. Now

6:03

what's funny I think things like

6:04

complexity log algorithms are going to

6:06

be parlance uh you know in the next 5 to

6:09

10 years as we

6:11

move towards a society. So n cubed means

6:13

if you want to multiply a thousand

6:15

element square matrix versus another

6:17

thousand element square matrix on an

6:18

nvidia machine it takes a billion

6:20

operations. That's a lot of operations.

6:22

As we learned about this computing stuff

6:24

with through through my lovely wife, I

6:26

was flipping through nature, which is

6:27

what we do as a hobby, and I saw a paper

6:30

from Microsoft that was called AOC,

6:33

analog optical computing. In New York,

6:35

that means something different, but AOC

6:37

in this case was

6:38

>> Yeah.

6:38

>> Uh, and what I saw in the corner of my

6:40

eye was that Matt Moles in light are

6:43

free.

6:43

>> Yeah.

6:44

>> They happen in instant time, which is a

6:46

really remarkable thing. So we're

6:47

investigating maybe we'll start a

6:48

company uh or something like that but

6:51

the there's a couple of startups like

6:52

Neurofas Opticore Lumi that want to make

6:55

these photonic matal cores which would

6:58

run circles around what Nvidia does. Now

7:00

of course that's not the whole story

7:02

there's a lot of details and

7:03

nonlinearity and storage

7:05

>> just be beyond the the hatred for

7:08

quantum and your raging short that may

7:10

or may not be winning. You were a hedge

7:13

fund manager, correct? And the famous

7:15

episode um just to catch everybody up

7:17

was uh you bought a drug and then you

7:21

inflated the price of the drug. Is that

7:24

what happened? Can you just give us the

7:25

just the

7:26

>> Yes. [snorts]

7:27

the history and what takes you now to

7:29

>> Yeah. I started at 16 at Jim Kramer's

7:32

hedge fund. Kramer was actually a

7:33

tremendously uh effective and skilled

7:35

trader. Uh but Kramer's idea of

7:38

investing was to you know buy and sell

7:40

all day every day everything and he

7:42

somehow squeezed profits out of it. That

7:44

was the way you do it sort of in the 80s

7:46

and 90s and he actually made 23% a year

7:48

uh net net of fees.

7:49

>> And so you worked for him as a 16-y old.

7:51

>> I worked for him at 16.

7:52

>> How did you get that job?

7:53

>> Uh just being in Manhattan having a big

7:55

mouth and you know it's a great city. Uh

7:57

and uh loving stocks, loving investing.

8:00

Bill Gates is my hero.

Continue with YouTLDR

Analyze another video with Pro

Process a new video, search every timestamp, compare sources, and keep the result in your library.

Get Pro — $12/month30-day money-back guarantee

More transcripts

Explore other videos transcribed with YouTLDR.