Martin Shkreli explains why light could replace Nvidia chips. 🤯

Shkreli's framing is blunt. When people say AI, he hears MatMul. A GPU spends almost all of its time multiplying lists of numbers, and it does it in n cubed time, so two 1000 by 1000 matrices means about a billion operations. Then he gets to the part worth checking. Microsoft published a paper in Nature last September on an analog optical computer, and the pitch is that doing those multiplications in light is close to free. The paper is real. A small prototype ran image classification and non-linear regression in 20 nanosecond cycles, and the team claimed more than 100 times the efficiency of GPU-based vector processors. The startups are real too. Neurophos raised $110 million in January, led by Gates Frontier with Microsoft's own venture arm in the round. It has no shipping product and no revenue, and its first pilot is planned for 2027. Here's the catch he skips. Photonic matrix multiplication currently tops out around 4 bits of precision. Inference generally wants at least 8. Light makes the multiply almost free. Making it accurate is the expensive part. 🔎 We fact-checked 10 expert AI predictions so you know whose calls are worth listening to. It's free in bio. Follow @HyperbolicAI for more content on AI and technology. Media: All-In Podcast #ai #nvidia #photonics #chips #technologynews