AI WARS: IS GOOGLE COMING FOR NVIDIA?

Blackstone just announced they are financing a $35bn deal for Anthropic to buy Google's proprietary AI chips. Wall Street made NVIDIA the most valuable company on Earth for selling THEIR chips that power AI. But nobody actually explains how all these chips work or what makes them different from one another. If you want to understand semiconductors, hyperscalers, and the rest of the AI buildout, you have to have a working definition of all the parts, so we challenged ourselves to explain just th Silicon is an element. Some elements conduct electricity, others block it. Silicon is in between, which is why it's called a semiconductor. Chip makers like TSMC in Taiwan, who produce chips for NVIDIA, take a bunch of sand, melt it, purify it to like 99.999% silicon, grow it into a giant crystal, and then slice it into these razor thin, mirror polished discs. They then basically take a stencil with some holes, shine a light through the stencil to shrink the pattern to a fraction of its size and effectively print tiny on/off switches onto that silicon wafer. On means one, and off means zero. Those are the only two letters in a computer's alphabet. These switches are so tiny, thousands of them fit across a single strand of human hair. You're gonna print billions of these switches onto this one slice, then connect them. So switches flip other switches, send electricity into it, and voila, a computer, the thinking machine behind AI. But not all chips are created equal. AI requires one key thing: parsing massive amounts of information into an answer. People often explain it like a restaurant trying to feed brunch to 10,000 people all at once. We need different chips to do this. Three kinds of chips... 1. CPU. That's the chip inside every laptop. In our restaurant, the head chef. He's brilliant, he can cook anything from soufflé to fish, but he works one step, then the next, one plate at a time from start to finish. But AI isn't one complicated order, repeated billions of times at once. Our chef can't do that by himself. Two options to help him. The first is a graphics processing unit or GPU. That's the kind that NVIDIA is famous for. Think of an army of thousands of line cooks all chopping at the same time. Unlike our head chef working sequentially, our line cooks work in parallel. GPUs were originally built to color in every pixel of a video game screen at once until people like Jensen Huang realized they could just run AI math in parallel instead. The third type is an application specific integrated circuit or ASIC. A chip that does exactly one thing perfectly. In our kitchen, it's a gadget like a waffle iron. Big guys like Google each have their own slightly different ones, designed to avoid paying a fortune to NVIDIA. These chips do AI math and nothing else. You can't buy one, you can only rent time on it. Pair your CPU with an army of GPUs or ASICs and you get a trained AI model. Hyper scalers like Meta and Microsoft acquire millions of chips and wire them together in one giant AI kitchen, a data center. Power and inside CPUs, GPUs or ASICs are directing teams of memory and network. And network warehouse is working like a giant brain. The byproduct of all that work is heat. And the dish you get when you ask Claude or ChatGPT a question. Follow for part 2! These data centers, energy and supply chains are keeping Elon Musk up at night.