Martin Shkreli: Big Pharma's Blind Spot, Betting Against Quantum & What Prison Taught Him
Summary
Shkreli’s quantum view is both technological and business-model bearish: today’s machines accelerate one narrow class of problems while losing badly at ordinary computing. He argues that quantum’s real advantage is moving integer factorization and elliptic-curve discrete logarithms from exponential to polynomial time—not sorting or general computation. Current systems run at hundreds or thousands of hertz, lack quantum RAM, and, in his testing, “can’t multiply 300 by 200.” He says companies spending hundreds of millions on the field are betting that “something might come of this someday.”
The host’s strongest counterargument—that quantum may simply be at its transistor-era beginning—does not change Shkreli’s investment conclusion. He concedes the technology could eventually work, then asks whether investors would have wanted punch cards, vacuum tubes, or ferrite-core memory: important stages whose owners largely disappeared. He has lost millions shorting quantum-computing companies but jokes that he “gained billions in insights,” conceding that market timing is the hard part.
His preferred compute disruption is analog optical computing, because AI workloads spend “99.9%” of their time multiplying matrices. Shkreli says a 1,000-by-1,000 matrix multiplication takes roughly one billion operations on an Nvidia machine, whereas multiplication in light can occur instantly. He is exploring photonic cores and predicts that within 20 years H100s, B100s, and B200s will “end up in the trash.”
That technological bearishness does not translate into a clean Nvidia short: Shkreli owns semiconductor stocks and is still positioned for AI upside. Jevons’ paradox may absorb efficiency gains, but he insists a transformer that is 1,000 times more efficient would still be “bad for Nvidia”; whether demand then expands one million-fold is unresolved. His warning is historical: in 2000, he probably would also have bought Yahoo and Cisco.
Big Pharma’s blind spot, in Shkreli’s telling, is the mispricing of discarded molecules rather than a shortage of scientific ideas. He bought a Bristol-Myers drug for $2 million and says the same company is now public and worth $6 billion; sparsentan was redirected from hypertension to FSGS, eventually giving roughly 30,000 people who had never had medication access to an approved treatment. “I can be a good person too,” he says, while calling this “a great way to make a billion dollars.”
Shkreli rejects near-term AI drug-discovery hype but remains extremely bullish on the market for machine intelligence. He calls claims that AI will develop drugs within five years “complete nonsense” and says he does not think there will be a single AI-related drug in the next 10 years, yet sees effectively unlimited demand for intelligence and labor. His endpoint is not merely chatbots but platforms building whole businesses—“an entire S&P 500” inside OpenAI, Anthropic, or their successors.
Prison changed Shkreli’s account of his own failure from overidentification with achievement to a more grounded view of himself. He says refusing to admit he was not the world’s best hedge-fund manager—even though he was a strong entrepreneur—“got me to prison.” His broader lesson is that gifted people can turn a felt obligation to achieve something enormous into self-destruction, while prison revealed the ordinary needs shared beneath status: missed friends and loved ones, and the wish to be a good parent.
Deep dive
1. Quantum’s scientific magic trick does not make it a business
Shkreli’s core distinction is between faster computing and one specialized speedup: Peter Shore’s thought experiment makes factoring and elliptic-curve discrete logarithms polynomial-time problems, but it does not accelerate ordinary algorithms such as merge sort. “Unless you’re factoring numbers, there’s nothing to get excited about.”
Today’s quantum machines, he argues, operate at hundreds or thousands of hertz rather than gigahertz, behave as single-core, single-threaded computers on conventional tasks, and have no quantum RAM. His categorical conclusion: “They will never outperform an Nvidia machine,” let alone a million Nvidia machines.
He also frames the problem as a business bet. Google, IBM, and Microsoft, he says, spend hundreds of millions of dollars a year because “something might come of this someday,” which he does not consider a very good business.
The host’s pushback—worth keeping—is that the industry may be where classical computing was at the transistor stage. Shkreli agrees that progress is possible but asks whether anyone should have invested in punch cards, vacuum tubes, or ferrite-core memory merely because they were necessary precursors.
His empirical test is deliberately mundane: after he and his wife Chelsea tested every quantum computer in the world, “We can’t multiply 300 by 200 on any of them.” He has lost millions shorting quantum-computing companies, acknowledging that “guessing the timing in the markets” is not easy while joking that he gained billions in insights.
2. Photonics could overturn Nvidia’s matrix-multiplication regime
Shkreli translates AI into “matmul”: GPUs spend “99.9%” of their time multiplying matrices, with conventional multiplication scaling at N cubed. Two square matrices of 1,000 elements therefore demand roughly one billion operations.
His alternative emerged from a Microsoft article on AOC, or analog optical computing: matrix multiplication in light is “free” and happens instantly. He cites Neurophos, OptiCore, and Lumai as startups pursuing photonic cores, while conceding that nonlinearity and storage leave substantial engineering work.
The long-range call is stark—H100s, B100s, and B200s could be trash within 20 years—but the portfolio stance is not. He owns semiconductor companies and is “playing for the upside,” even while warning that a transformer 1,000 times more efficient would plainly hurt Nvidia.
3. Neglected molecules were Shkreli’s highest-alpha market
Shkreli began at Jim Cramer’s hedge fund at 16; he credits Cramer with earning 23% annually after commissions but discovered that he himself was “a much better entrepreneur and private-equity specialist than a trader.” Public-market trading offered less alpha than buying overlooked pharmaceutical assets.
His defining example is a Bristol-Myers drug purchased for $2 million; he says the same company is now public and worth $6 billion. He says Pfizer and other drugmakers could sell rights to assets for roughly $1 million even when those assets might support billion-dollar products because the assets were immaterial to them.
Sparsentan carried the model from financial arbitrage into clinical repositioning: originally developed for hypertension, it was redirected toward FSGS, a disease for which no drug had been approved. After FDA approval, Shkreli said that about 30,000 people who had never had medication now had it, calling the result both useful and “a great way to make a billion dollars.”
The binding constraint was information processing. His team needed people to read papers and patents across universities and drug companies; in 2015 he started an NLP company to confront a corpus of 40 million medical articles after finding Watson “wasn’t great.”
4. AI drug hype is premature, but intelligence demand may be boundless
Shkreli’s disagreement with Dario’s forecast is categorical: AI developing drugs within five years is “complete nonsense,” and he does not expect a single AI-related drug in the next 10 years. Computers and GPUs help analysis, he says, but drug development and FDA approval remain difficult and AI cannot really develop drugs.
When the host suggests that roughly $400 billion of venture capital may be funneling investor money into buying tokens, Shkreli rejects the resulting valuation-trap argument. He sees “an infinite demand for intelligence and an infinite demand for labor,” extending far beyond knowledge workers.
He cites a friend who left OpenAI and started Core Automation, whom he says invented the reasoning model at OpenAI, then imagines AI platforms creating domain businesses such as OpenAI Pharma, OpenAI Motors, and OpenAI Aviation—“an entire S&P 500” inside OpenAI, Anthropic, or comparable platforms.
The McDonald’s example supplies the economic test: spending $10 million on “10 million GPU brains” to select the next sandwich makes sense whenever the resulting intelligence has positive net present value. The thesis breaks only if more tokens eventually stop improving the answer.
5. Distribution captures value; ego determines who survives it
Shkreli’s market analogue is the internet: consumers and businesses receive something like “tax-free” infrastructure, while suppliers may earn only temporary excess returns. He favors Meta for distribution through Messenger and product execution, while calling OpenAI’s hardware strategy “brilliant” and holding OpenAI through multiple avenues. He points to io and Jony Ive’s creative product work as evidence of a new path.
On AI safety, he sides with effective accelerationism—“move fast and break stereotypes”—and dismisses present existential rhetoric: the system is still software, “you can turn off the computer,” and “it won’t kill us all.” The host’s proposed peer-review compromise does not soften that view.
Asked about Sam Bankman-Fried, whom he says he advised, Shkreli notes that SBF was early in Anthropic and that his 2021–22 portfolio looked like genius in retrospect. He links unconventional brilliance with dangerous self-belief: gifted people may feel accountable to some enormous destiny, while speed and refusal to admit failure can convert that burden into misconduct.
His self-indictment is the episode’s closing turn: “It was hard for me to admit that I wasn’t the best hedge fund manager in the world, but I was a great entrepreneur.” Prison then stripped away status, revealing how consistently people miss friends and loved ones and want to be good parents.