Key Takeaways

  • Martin Shkreli predicts Nvidia flagships like the H100 and B200 will sit in the scrap heap within twenty years, replaced by optical architectures.
  • Artificial intelligence workloads spend 99.9% of their execution time on matrix multiplication, a math operation that light can process instantly with near-zero energy.
  • Silicon Valley plans to power data centers with dedicated nuclear reactors or orbital arrays miss the coming hardware transition entirely.
  • Scale AI founder Alexandr Wang told Shkreli that Jevons paradox will expand aggregate compute demand a million-fold as costs drop.
  • Software algorithms that deliver 1,000x efficiency gains on the transformer architecture threaten to crush GPU maker pricing power long before silicon hits physical scaling limits.

The Expensive Illusion of Silicon Matrix Multiplication

Tech companies are currently scrambling to secure nuclear power plants and designing orbital data centers to feed hundreds of thousands of screaming silicon boards. Shkreli views this capital frenzy as an architectural dead end.

“Every time somebody uses the word AI, I actually think of the word matmo,” Shkreli observed. “These are matt mole machines... multiplying two lists of numbers is what a GPU spends 99.9% of its time doing.”

Right now, companies pay tens of thousands of dollars per chip to force electrons through dense copper interconnects, generating heat that requires massive cooling infrastructure. Shkreli points out that physics offers a far cleaner shortcut through photonic computing. Photons passing through an optical core complete matrix multiplications instantaneously without drawing continuous resistive power.

“What I saw in the corner of my eye was that Matt Moles in light are free,” Shkreli said. “They happen in instant time, which is a really remarkable thing.” When the math itself costs nothing in terms of heat and latency, the logic behind burning gigawatts of thermal power on traditional chips collapses.

When 1,000x Algorithmic Jumps Meet Jevons Paradox

The market assumes that data center operators will keep writing blank checks for clusters indefinitely. Shkreli disagrees. While he remains long semiconductors over the broad horizon, he warns that today's capital expenditure models rely on brittle unit economics.

“I do think in 20 years these H100s are in the scrap heap. B100s are in the scrap heap, B200s in the scrap heap,” Shkreli said. “I think it's insane that we're thinking about space. I think it's insane we're thinking about nuclear.”

Hardware margins face pressure from two opposing forces: algorithmic efficiency and total demand volume. First, researchers are already rewriting model foundations. “There are guys in the valley talking about better transformers that are a thousand times better,” Shkreli noted. “If there's a thousand times more efficient transformer, that's bad for Nvidia. There's no doubt about that.” If a startup needs one-thousandth of the silicon footprint to run equivalent inference, supplier pricing power erodes overnight.

Second, Jevons paradox complicates the fallout. When a resource becomes cheaper to use, consumption rises instead of falls. Shkreli recalls discussions with Alexandr Wang on this dynamic: “Alex Wang told me Jevon's going to take care of all this, will be a millionx compute improvement.” Total compute consumption will balloon across the economy, but the specific silicon chips minted during the 2024 to 2026 buildout will depreciate to zero faster than cloud providers can amortize them.

What to Do With This

Audit your infrastructure roadmap this week to identify every multi-year GPU reservation or hardware lease your engineering team signed. If your company locked in long-term compute contracts at fixed 2024 pricing, draft a renegotiation timeline for when model weights compress or inference pricing drops another 50%. Shift your technical budget from reserved hardware capacity into model optimization and quantization pipelines that cut your raw token costs immediately.