OpenAI Close to Landing $500 Billion Data Center With Nvidia's Backing

When I saw that OpenAI is building a $500 billion data center, I had to know what cost half a trillion dollars to build a data center. And why Nvidia is backing half the financing for it. I spent weeks researching and talking to people who build data centers. How do you build a $500 billion data center? First, they need some land in Southern Ohio. There's an old uranium enrichment site that's already owned by the federal government. They'll be leasing. We don't know terms yet, but it's 3700 acres, which comes out to between $1-300 million. Next is power. America's grid dates back to the 1970s, and there's not enough spare power near this site to feed a 10 gigawatt data center. So they're building the world's second largest gas-fired power plant on site, aiming for 9.2 gigawatts. For that, they need gas turbines, which already have a 5+ year wait time, plus new transmission lines to the data center. All this, $37 billion. But you can't just plug a data center into a power plant. The voltage gets stepped down through substations, and power goes out then back in. So you need backup diesel generators, another 3 year wait list. So they need fuel cells to bridge the gap. One ceramic fuel cell gives you about 25 watts. A stack of stacks combined into 300 kilowatt blocks. And companies like Bloom Energy make arrays of blocks giving you megawatts in months. All another $20 billion. Now we actually have to build the thing. You'll need engineers, construction workers, plus concrete and steel to create what's effectively a hurricane and tornado proof box. A powered concrete shell, all the technology inside it, us. That's about $55 billion. Next, cooling. A huge source of controversy. The way it used to work, heat was pumped from the servers into a cooling tower outside. Evaporation carried the heat away, and new cold water was pumped through the servers. 25 billion gallons each year. Used it here, as much water as the city of San Francisco. The new system is more like a car radiator, where the water gets filled once, and that's it. A cooling plate gets attached to each chip, which is then a heat exchanger. Racks of chips now face each other with a cold alley in between. Some heat going out the back. Pipes pull the rest of the heat off the chips, carry that hot water past fans, cool it down, and send cold water back through the alley between the racks. Instead of the whole city of San Francisco, you only use as much water as a city block. The new closed loop system costs about $35 billion. Next up, the network. There's millions of chips in here, but they only work if they can talk to each other inside the racks. Chips communicate through Nvidia's NVLink, 2 miles of copper cabling in each rack. The language they speak is CUDA, Nvidia's coding language. To then link the racks together into one coherent machine, you need fiber optic cables and switches carrying data around at the speed of light. All in $30 billion. Last but not least, the chips. Each unit in a rack consists of a high bandwidth memory system with a logic chip. We explain all this in our other videos. Now do that 4.5 million times. And you get $350 billion. Plus a lot of the networking spend is going straight to Nvidia. Sum it all up, and this project costs $527 billion. But more importantly, you now understand why Nvidia is willing to backstop $250 billion of this project's debt. They're guaranteeing themselves almost $400 billion of revenue by backstopping the loans that pay for it. Sounds like circular financing. We'll cover that next.