On Thursday, we saw the biggest tech IPO of the year.

On Thursday, we saw the biggest tech IPO of the year. It's Cerebras. It's a chip company most people haven't heard of, but the CEO went to bed $3 billion richer, and it was up 68%. Let's unpack why. Cerebras had such a buzzy IPO because of their technology, a major innovation in microchip design. The typical process to build a microchip starts with a silicon wafer, a circle the size of a vinyl record. Making this wafer is really technical, so defects arise. To combat this, you dice it up into hundreds of tiny chips to get rid of the defects. Then you put them on top of a printed circuit board and wire them up together with cables and protocols so they can talk to each other. That's been the chip process for 70-plus years. But now that AI has come along, we've run into a new snag. Each individual chip is super powerful, but there's a lag in speed of communication between each chip, like the speed of data passing through telephone wires. That's called latency. And when you're running a large AI model, you're moving billions of numbers back and forth between memory and processor on every single response. For inference, the part where a model responds to you in real time, that delay is especially meaningful. The user is literally sitting there waiting. Every millisecond is noticed. Cerebras said, "But what if we didn't dice up the chip? Instead, let's make one big chip and solve for defects by creating redundancies inside the wafer. If something breaks, there's another duplicate to pick it back up." This innovation lets their chip be up to 21 times faster on inference, hold 2,600 times more memory bandwidth. The founder, Andrew Feldman, had already sold his previous company, also a chip company, to AMD in 2012 with the same founding team. At Cerebras, their big chip idea was relatively straightforward, but the supply chain to build it was complicated. The entire semiconductor supply chain was already designed for tiny chips. Cerebras had to build their own process from scratch. It was so brutally expensive that they almost folded in 2019. But once that infrastructure was built, it became a moat. Their process is proprietary, so every competitor would have to build their own. By 2025, Cerebras was doing $510 million in revenue. There was another wrinkle, though. 62% of that revenue came from one single customer, a university in Abu Dhabi. Another 24% from a second customer, also in Abu Dhabi. The company actually tried to go public once before, but the government blocked it in 2019. Too much revenue concentration in a Gulf state and a potential link their biggest customer and Huawei, a Chinese tech giant that US regulators considered a national security risk. After that, the company restructured and raised another billion, totaling $8.1 billion in venture capital. Then earlier this year, OpenAI struck a huge deal, $20 billion, a multi-year deal and 750 megawatts of inference. The deal isn't converted yet, but it was enough to respark an IPO. On opening day last week, it was priced at $185, opened at $350, and closed up 68% on day one. Only time will tell if OpenAI converts and if the big chip moment has really arrived. But Cerebras was early to see where AI was headed and survived long enough to start to cash in.