Thousands of new datacenters filled with millions of NVIDIA chips made with billions of transistors are being built to make trillions of dollars — doing multiplication.
1,000,000,000,000,000,000,000,000,000,000,000,000 multiplications.
You’ve probably never head the word undecillion, but that’s how many multiplications they will do all-in.
Every AI you’ve ever talked to is a pile of numbers getting multiplied together — billions of multiplications for every word it spits out. We got the idea in the 1950s by trying to imitate neurons in the brain, and for fifty years it barely worked because computers were too slow. Then video games happened. GPUs were built to draw explosions in Quake, and by coincidence, the math for drawing explosions is the same math for simulating neurons. With AI, neural networks took over the world, not because multiplication is the right way to think, but because we had warehouses full of chips that were good at it. This coincidence got expensive, now we’re multiplying numbers that never needed to be multiplied.
EvoChip went around the whole mess. They compile training data straight into logic, chains of the operations transistors do in a single clock cycle, essentially for free. Nothing gets pruned, quantized, or compressed, because there is no neural network to shrink — and there’s hardly any multiplying left to do. The model is the circuit.
SidePath, a Dell partner, tested EvoChip against the fastest TensorFlow configurations they could build, across seven public datasets. EvoChip came out 13× to 41× faster on ordinary CPUs, with 35× to 301× fewer parameters, at the same accuracy. A $200 FPGA cranks out over 3 billion inferences per second. An Arduino, the computer from your kid’s science fair project, ran a model in 521 bytes of RAM.
Speed isn’t even the interesting part. There’s a reason your plane doesn’t have an AI copilot: nobody can prove what a neural network will do next, including the people who made it, and you can’t certify a black box for jobs that kill people when they fail. An EvoChip model is a logic circuit. Same input, same answer, same number of clock cycles, every time. You can hand the schematic to a regulator. If AI is ever going to fly planes or run pacemakers, this is how it will have to work.
When we invested in EvoChip in 2024, this was 1000× speedups in the lab. Now the benchmarks are public, the products are launching, there are 100+ patents on the core, and LLMs are next. If you know people trying to fit AI into something that runs on a coin cell battery, certify it for something that can kill you, or who just can’t get their hands on a zillion H100s, send them to EvoChip.