Martin Shkreli: Big Pharma's Blind Spot, Betting Against Quantum & What Prison Taught Him

NVIDIA

Across the shows: 12 speakers in other episodes, 12 bullish and 4 bearish calls, past 30 days

• Martin Shkreli · Q/C TechnologiesBullish

  He is long semiconductors and NVIDIA, expecting Jensen to win despite efficiency gains.

  He says he is long semis and that Jensen always wins in the end, using a Jevons-paradox argument: cheaper compute gets consumed more. He also concedes risks. A thousand-times more efficient transformer would be bad for NVIDIA, and he doubts demand will grow a million-fold. He expects H100, B100 and B200 chips to be scrapped within about 20 years, and photonic chips could outrun NVIDIA. His long is therefore a qualified bet on demand outrunning efficiency gains.

  “I think that doesn't mean NVIDIA's not [a long]. And I'm long semis right now.”at 11:49

OpenAI

Across the shows: 12 speakers in other episodes, 3 bullish and 10 bearish calls, past 30 days

• Martin Shkreli · Q/C TechnologiesBullish

  He holds OpenAI and calls Sam Altman's hardware-plus-model strategy brilliant.

  He says he holds OpenAI through several vehicles. He thinks OpenAI's hardware push is brilliant. The company will likely always have the best model, but being able to deliver that model at scale could prove more valuable. He points to the Muse product work as evidence that OpenAI is opening new avenues beyond the model itself. He says we are still on day one of AI, and the chatbots of today will look primitive later.

  “So we have OpenAI obviously through a few different ways. And I think that Sam's hedge with hardware is really brilliant because... He'll always have the best model I think, but being able to deliver that model is probably gonna be more valuable”at 19:34

Pharma & biotech

• Martin Shkreli · Q/C Technologies

  Big Pharma sells drug assets far below their worth, so buyers of molecules capture large alpha.

  The market for molecules is inefficient. He bought a drug from Bristol Myers for $2 million, and that company is now worth $6 billion. Pfizer would sell drug rights for a million dollars when they were worth a billion, because it did not care. The bottleneck was vetting assets, which he tackled with an NLP company that tried to read all 40 million medical papers. His strategy of buying and developing undervalued molecules is the source of the alpha he describes.

  “Pfizer would sell you a drug for a million bucks that was worth a billion dollars. They didn't care.”at 9:09

• Martin Shkreli · Q/C Technologies

  AI won't meaningfully help drug discovery; he sees no AI-relevant drug in the next ten years.

  He has built drugs and gotten an FDA approval himself, and says it is hard to do. Computers have been used in drug discovery for decades; he built a Cray cluster to analyse drugs years ago. He disputes Dario's claim that AI would cure all diseases and says AI does not help drug discovery in any meaningful way. He is confident that no AI-relevant drug will arrive in the next ten years.

  “He also said that we'd cure all diseases. I don't think there'll be a single AI relevant drug in the next ten years. ... AI does not help drug discovery.”at 16:05

Quantum computing

• Martin Shkreli · Q/C Technologies

  Quantum computers won't beat GPUs for business use; he is short quantum computing companies.

  Quantum machines run at hundreds or thousands of hertz and are single-threaded, so they are not faster general computers. Their only real advantage is a narrow factoring-type speedup (integer factorization and elliptic-curve discrete log), which matters to nerds, not businesses. Quantum has no storage, so problems must have tiny inputs, and they can never beat even one NVIDIA machine. He and his wife tested every quantum computer available and could not multiply 300 by 200 on any of them. Google and IBM spend hundreds of millions a year on a speculative payoff. He has been short quantum computing companies and says the trade lost millions, but the insight was worth billions.

  “So they'll never beat an NVIDIA machine. They'll never beat a million NVIDIA machines.”at 3:59

AI chips & semiconductors

• Martin Shkreli · Q/C Technologies

  Photonic matrix-multiply chips could run circles around NVIDIA GPUs, and several startups are building them.

  GPUs spend about 99.9% of their time on matrix multiplication, which costs n-cubed operations. A Microsoft analog optical computing paper showed matrix multiplication in light is essentially instant. Startups including Neurophos, Opticore and Lumai are pursuing photonic matmul cores that could outperform NVIDIA. He says Sam Altman has also publicly said photonic computing could be the future. Nonlinearity, storage and other details remain unsolved, and he is only investigating whether to start a company.

  “There's a couple of startups like Neurophos, Opticore, Lumai that wanna make these photonic MapML cores, which would run circles around what NVIDIA does.”at 7:26

AI infrastructure & data centers

• Martin Shkreli · Q/C Technologies

  Covering the earth with GPU data centers is unwise when photonic computing could be faster.

  He questions the logic of the current buildout. If photonic computing can deliver a faster machine, building more GPU capacity across the planet looks wasteful. He also says it is insane to be thinking about space and nuclear power for compute. The argument rests on the expectation that photonic architectures will displace GPU scaling, and he gives no capex numbers.

  “So I'm just wondering why we think this is a sane idea to cover the earth with computers when we can make a faster one.”at 11:49

AI regulation & safety

• Martin Shkreli · Q/C Technologies

  AI existential-risk fear-mongering is ridiculous; panic is premature because AI is just software.

  He says the fear-mongering is ridiculous and that there will be a time to panic, but not now. AI is just software, and you can turn off the computer. It is the most important software ever made, but it will not kill everyone. He thinks nerds get carried away when they want to draw attention to their own work, and he accuses effective altruists and the EA movement of narcissistic fear-mongering. He points to Meta and the White House as pushing back against the same narrative.

  “I think the fear mongering is a little ridiculous and there'll be a time to panic. It's just not now.”at 15:09

AI & jobs

• Martin Shkreli · Q/C Technologies

  Dario's full-unemployment prediction is hype; demand for intelligence and work is effectively infinite.

  Dario Amodei said about two years ago that people would be fully unemployed within one to five years, and Martin says he has three years left to prove it. He argues that productivity gains come from infinite demand for intelligence and infinite demand for work. Limiting the total addressable market to knowledge workers, as Brad did, undercuts the opportunity. He also says there is no ceiling on intelligence, so AI expands the pie rather than simply replacing people.

  “I think that what you mentioned earlier, productivity, I mean, there is infinite demand for intelligence and infinite demand for work.”at 16:57

AI bubble & tech valuations

• Martin Shkreli · Q/C Technologies

  AI capital spending isn't a valuation trap; enterprises will pay heavily for tokens.

  Asked whether roughly $400 billion of VC money redirected into AI amounts to LPs buying tokens and a valuation trap, he says no. The opportunity is boundless because intelligence has no ceiling. As an example, McDonald's could easily spend $10 billion on 10 million GPUs to decide its next sandwich, and he calls that NPV-positive. The argument only fails once more tokens stop improving answers, which he says has not happened yet.

  “So I think that the opportunities are boundless. And one last thing is that McDonald's can afford to spend $10,000,000,000 and say, OpenAI, give me 10,000,000 GPUs... That's an NPV positive.”at 17:58

AI: overall outlook

• Martin Shkreli · Q/C Technologies

  AI should resemble the Internet, with value flowing mainly to consumers and businesses, not model makers.

  His mental model for AI is the Internet. The only big winners would be consumers and businesses, who get a tax-free highway of value. Cisco and a few others capture some of it, but the main benefit is free intelligence in everyone's pocket that is already here. He is unsure who will over-earn and for how long, which implies that model makers and infrastructure providers will not capture the lion's share.

  “It has to be like the Internet... where the only winners would be consumers and businesses. They get this tax free kind of Internet. Cisco and some other guys will get a little bit out of it”at 18:52