The World's Largest AI Data Center Has a Massive Problem

[00:00:00] Speaker 1: The largest AI training computer on Earth is running inside an old warehouse in South Memphis, and the power plant built to keep it alive sits in another state. Just over the Mississippi Line in South Haven, dozens of natural gas turbines run day and night to feed what Musk says is more than half a million NVIDIA chips. Getting those chips turned out to be the fast part. Keeping them switched on is where the real engineering fight is now, and Microsoft, Google, Amazon, and Meta are walking straight into the same wall. The machine is called Colossus 2, and it belongs to XAI, the artificial intelligence lab Elon Musk founded, and then folded into SpaceX in February. On January 17th, Musk posted that Colossus 2 was live and called it the first gigawatt training cluster in the world. The goal behind it is easy to say and brutal to build. XAI wants a million NVIDIA GPUs in the Memphis area, all wired together so they can train one AI model as if they were a single, enormous chip. That also means a hiccup anywhere in the power supply can stall the whole training run at once. From the street, it doesn't look like much. Colossus 2 lives inside a huge former warehouse on Tulane Road in the Whitehaven neighborhood of Memphis, with the word macrohard painted across its roof. By late September, Musk said that building was running 550,000 NVIDIA chips, more than five times what the original Colossus launched with. Musk says more batches are due to switch on before the end of this year. Those are the company's own numbers, so it helps to check them against something that can't be spun. Epoch AI, an independent research group, tracks Colossus 2 from space, counting cooling units and turbines in satellite photos. Its latest estimate puts the site at about 946 megawatts of server load. That's nearly a gigawatt flowing into computers alone, before a single cooling fan or light bulb in the building is counted. And that number is where this whole story turns. NVIDIA ships chips by the truckload when the money's there. The electricity to run them has to arrive through wires, every second of every day, without so much as a flicker. Finding that power, steadying it, and getting rid of the heat it leaves behind is the obstacle XAI ran into in Memphis. And it's the same obstacle now sitting in front of every company racing to build bigger AI. When XAI first came to Memphis in 2024, it moved the original Colossus into a shut-down Electrolux appliance factory near the Mississippi River. The substation serving that building was sized for an appliance factory. So even the first Colossus wanted far more electricity than its neighborhood could hand over. And the chips have only got hungrier since. The reason sits inside every server cabinet. A single NVIDIA Gigabyte 300 rack, the tall, liquid-cooled cabinet these chips now ship in, pulls around 135 kilowatts when it's working hard. That's more electricity than a hundred average American homes use, packed into a single cabinet. NVIDIA did that on purpose. Squeezing dozens of GPUs into one rack lets them talk to each other faster, which is exactly what training a giant model needs. The price of that speed is a power cable that would have looked absurd in an older data center. Now, line those cabinets up by the thousands. Fully built, Epoch expects Colossus 2 to need about 1.5 gigawatts for its servers alone, nearly double what the entire city of San Diego uses. To see how far off the normal scale that is, look just across the state line. The Tennessee Valley Authority already runs a gas power plant in South Haven rated at about 780 megawatts. Even if XAI could take every watt that plant makes, it still wouldn't cover the full build in Memphis. A city's worth of demand dropped onto a grid that was planned years earlier for factories and homes. Utilities add power plants and transmission lines over years. XAI was planning in months. So instead of waiting for Memphis light, gas, and water to catch up, Musk went shopping for a power plant of his own. In July 2025, Musk confirmed a claim that sounded like a joke. XAI had bought an entire power plant overseas and was shipping it to the United States because a new one couldn't be built here in time. Where it came from has never been made public. Buying an existing plant and moving it across an ocean was the faster way to get turbines spinning. A hint of a shortage this story comes back to later. The landing spot was just over the state line. An XAI affiliate called MZX Tech took over a former Duke Energy Power Station site on Stanton Road in South Haven, Mississippi, a short drive from the Colossus 2 building in Memphis. Then the turbines started showing up. These were mobile gas turbines, the kind that arrive on the back of trailers. And Mississippi's environmental regulator treated them as temporary equipment that didn't need an air permit for up to a year. Epoch's analysts have pointed out that putting the generators in Mississippi may have been about getting faster approval. The bigger move came next. In March, Mississippi's Environmental Permit Board voted unanimously to let XAI build 41 permanent gas turbines at the South Haven site. Built out, that permanent plant is designed to produce roughly 1.2 gigawatts, enough electricity for a mid-sized city, and it answers to an AI lab instead of a utility. And for the thing XAI cared about most, the gamble paid off. Land clearing on Tulane Road started in February of last year. By this summer, the site was running close to a gigawatt of servers on power it had largely built for itself. That's a timeline an ordinary grid connection doesn't allow. And XAI hit it by turning itself into a utility on the other side of a state line. SpaceX's own filing for its stock market debut shows how fast that meter climbs. It says the first Colossus II cluster drew around 210 megawatts for about 110,000 chips. That works out to close to 2 kilowatts for every chip once you count the server and networking gear wrapped around it. That's more than a household space heater on its highest setting for every single chip in the building around the clock. And every phase since has stacked more chips onto that same math, which is why the power plant in South Haven keeps growing alongside the data halls in Memphis. But generating the power turned out to be only half of the problem. The chips inside Colossus II have a habit that no gas turbine was ever designed to put up with. Training a giant AI model draws power in pulses. Hundreds of thousands of GPUs crunch numbers in lockstep, then pause together while they swap results across the network, then all fire up again at the same instant. So the power draw of the whole building can lurch up and down in milliseconds. With ordinary web traffic, millions of separate requests smooth each other out. A training run moves like one giant animal breathing in and out because every chip in the building breathes at the same moment. Grid operators have good reason to fear big loads that move all at once. On a July evening in 2024, a single faulty part on a transmission line in Northern Virginia caused 60 data centers to flip to their backup power in the same instant. About 1,500 megawatts of demand vanished from the grid in a blink. More than one and a half times everything Colossus II's servers draw today. PJM, the grid operator that runs that part of the country, had to scramble to cut power plant output and keep the region from tripping offline. Those data centers weren't even doing anything unusual that night. Now picture a single customer the size of a city whose demand swings up and down by design. XAI found this out on the original Colossus. Its engineers saw that those millisecond spikes and drops were rough enough on the power equipment that they put Tesla Megapack batteries, shipping container-sized battery units, in the middle, between the generators and the computers. The generators charge the batteries. The batteries feed the cluster. And the turbines see a smooth load even when the chips don't. At Colossus II, that fix turned into something enormous. Satellite photos from this July showed 720 Megapack containers lined up on the site. In August, a company energy developer told the board of the Tennessee Valley Authority that the battery holds enough to run all of Memphis for about two hours. That's SpaceX's own figure. And if it holds up, it's bigger than any battery operating anywhere in the United States today. The money for all that storage moved from one Musk company to another. XAI spent $430 million on Tesla Megapacks last year. And the combined SpaceX company spent another $405 million in the first half of this one. Then came the twist. The same day it heard that pitch, the Tennessee Valley Authority Board approved connecting the site directly to its grid. The batteries now double as backup for the moments the data center has to curtail, which means stepping back from the grid when power runs tight. So the battery that started life as a shock absorber for the chips became Colossus II's ticket onto the public grid. On a scorching Memphis afternoon, when the region needs every megawatt it can find, the data center can lean on its own batteries instead. That matters to every other customer on the system because a giant load that can step aside on demand is a load the utility doesn't have to build an extra power plant for. But every one of those megawatts, whether it comes from a turbine, a battery, or the grid, ends up as the same thing inside the building: heat. Each GB300 chip runs so hot that these racks are cooled with liquid instead of air. Every chip wears a cold plate, a slab of metal with liquid running through it pressed right onto the silicon, and pipes carry the heat out of the rack. The trouble starts once that heat leaves the rack. Somewhere outside, all of it has to be dumped into the air. And that's where Colossus II fell behind its own announcement. When Musk declared a gigawatt cluster in January, Epoch satellite analysts found only about 350 megawatts of cooling installed on the site. That's roughly a third of what a full gigawatt of chips would need. Without that cooling, the extra chips couldn't run at full power. So that winter, the real limit on Colossus II was the fans. XAI kept building them. Satellite images of Tulane Road now show long rows of dry coolers, giant radiator walls with fans across the top, beside rows of chiller units pushing heat straight into the Memphis air. By June, the cooling had finally caught up with the power. Those dry coolers were a deliberate trade. A closed-loop cooled by air barely uses any water, but it burns more electricity running fans and compressors, while cooling towers save power by evaporating water away. Memphis has good reason to care which one gets picked. The city drinks from the Memphis Sand Aquifer, some of the purest groundwater in the country, and the original Colossus has been cooled with that same drinking water. XAI's answer was an $80 million recycling plant that would clean up treated city wastewater for cooling, up to roughly 49 million liters a day. The plan was for that recycled water to serve the Tennessee Valley Authority as well, so two of the biggest industrial users in South Memphis could stop pulling from the aquifer. In April, construction stopped. Musk said the team needed to finish Colossus II and make it extremely stable first, and the city now says the company has committed to restart work no later than early 2027. So, the heat got handled the same way the power did, by spending more electricity, while the water promise slid down the list behind the chips. If you want the real numbers on who can actually build and power the chips everything runs on, subscribe, because that's the only story this channel covers. And Memphis is only the most extreme version of a problem every AI giant now shares. In late 2025, Microsoft chief executive Satya Nadella said on a podcast that Microsoft could be left with stacks of chips and nowhere to plug them in. What he was short of, he said, was warm shells, finished buildings with the power already hooked up and waiting for racks. That's the same Wall-X AI hit, now standing in front of one of the richest companies on Earth. The grid isn't coming to the rescue quickly either. In America's busiest data center hubs, like Northern Virginia, the wait for a big new grid connection now routinely runs five to seven years. So the giants are starting to copy the Memphis playbook and build their own power. And that runs them straight into the next bottleneck. GE Vernova, one of the world's biggest makers of the heavy gas turbines that power plants are built around, is sold out for years. A big turbine ordered from it today won't arrive until around 2031. NVIDIA, meanwhile, now brings out a new flagship chip design every year. So a turbine ordered alongside a rack of today's Blackwell chips will show up after those chips have been replaced at the top of NVIDIA's lineup several times over. That's why the smaller, truck-mounted turbines XAI leaned on have turned into a hot commodity. CRUSO, the company building OpenAI's Stargate campus in Abilene, Texas, ordered a fleet of GE Vernova's mobile turbines worth about one gigawatt, specifically for AI data centers. For the biggest labs, the wait for turbines now runs longer than the wait for chips. And the racks themselves aren't getting any easier to feed. At this year's Open Compute Summit, a yearly gathering of data center hardware makers, NVIDIA shared new open rack designs built to carry up to 250 kilowatts each. That's nearly double the power of the cabinets filling Colossus 2 today. Every new generation of NVIDIA hardware packs more computing into the same footprint, and the power cable grows right along with it. So, Memphis is an early look at where the whole chip industry is heading. The other answer is nuclear power, which is cleaner and steadier, and slower still. Microsoft agreed to buy 20 years of power from Constellation Energy, which is bringing back a reactor at Three Mile Island in Pennsylvania, now renamed the Crane Clean Energy Center. Constellation is still aiming to switch it back on in 2027, even after the regional grid operator's own analysis warned its connection could slip into the next decade. Google has backed small modular reactors, compact factory-built nuclear units from Kairos Power, and Amazon has lined up electricity from the Susquehanna nuclear plant, also in Pennsylvania. Then, in January, Meta signed nuclear deals with Vistra, Terrapower, and Oklo that could support up to 6.6 gigawatts by 2035. A typical nuclear plant puts out about 1 gigawatt, so that's 6 or 7 nuclear plants worth of power lined up for a single social media company. Now look at the calendar. The brand new reactors in those deals start arriving around the end of this decade at the earliest. Colossus 2 went from land clearing to nearly a gigawatt of servers in about a year and a half. That gap is why the fastest-built AI cluster on the planet runs on gas and batteries today. The cleanest power the industry is paid for won't show up for years. And speed like that came with a bill that landed on people who never signed up for any of it. In July, a Reuters investigation found that XAI had installed 59 natural gas turbines for Colossus 2 without federal clean air permits. That's roughly double the 27 turbines the company had said it was running as of January. 57 of them sit in South Haven. The whole fight comes down to one word: temporary. The Federal Environmental Protection Agency said in January that even temporary turbines need permits once their emissions cross a certain threshold. Mississippi's regulator still says portable, temporary turbines don't need one at all. A university researcher told Reuters the nitrogen oxide output alone, the gas behind smog, would rank the site alongside some of the heaviest polluting gas plants in the country. The people living next door don't need a spreadsheet to feel it. Residents of South Haven have complained about turbine noise running around the clock. And the neighborhoods closest to the plant are predominantly black communities that, by Reuters' reading of government data, already suffer higher rates of lung disease. When Mississippi held its only public hearing on the permanent plant in February, the room in South Haven was packed, and the people who spoke were united against the draft permit. The final vote happened weeks later in Jackson, hours away from South Haven, on an election day. Not everyone in town is against it. Some South Haven residents have told local reporters to let XAI go ahead and build. And Mississippi's regulators say the permanent plant met every legal and technical requirement. State officials have pitched the project as jobs, investment, and a foothold in the AI economy. In April, the NAACP sued XAI under the Clean Air Act to stop the turbines. In June, the U.S. Justice Department moved to step into the case and have it thrown out, arguing it threatened national and energy security, including the military's growing use of AI. So, the hardest obstacle left in this story is permission. From neighbors, from courts, and from regulators who can't agree on what a temporary turbine is. And XAI has started answering it. At the end of July, SpaceX agreed with Mississippi regulators to pull the temporary turbines out by July 2027, as the permanent, permitted plant comes online. Whether that deadline holds while Colossus 2 keeps growing is the open question hanging over the whole site. The cost doesn't stop at the fence line in South Haven, either. Data centers are now pushing up the price of electricity for people who will never set foot inside one. PJM, which serves 13 states from Illinois to New Jersey, runs an auction that pays power plants to promise they'll be there on the hottest days of the year. In its latest auction, held this July, about $6.3 billion of the total traced directly to data center demand. That's close to two out of every $5 in the whole auction. And those charges flow straight into ordinary household power bills. Even at those prices, the auction came up about 6.8 gigawatts short of what PJM says it needs to keep the lights reliably on. The money is there, but the new power plants aren't arriving fast enough to take it. And this summer, homes and businesses across PJM's territory set a new all-time record for electricity demand. The industry's main answer is to do what XAI did and bring its own power plants. That takes some strain off household bills, and it moves the exhaust stacks into somebody else's neighborhood. Which is exactly the fight playing out in South Haven. Being first to a powered gigawatt has turned into a business of its own. In May, Anthropic agreed to rent essentially all of the computing capacity at the original Colossus. And SpaceX also rents capacity in Memphis to other labs, including Cursor. Then, in June, a regulatory filing showed Google agreeing to pay SpaceX $30 billion for computing on about 110,000 NVIDIA GPUs in SpaceX's data centers. The deal runs through the middle of 2029, which works out to roughly $920 million a month. Google designs its own AI chips and runs some of the biggest data centers on the planet. It still chose to pay a rocket company for compute that was already plugged in and running. What Google is buying is compute that's already switched on. And the biggest number in this story is still to come. Epic's satellite model projects that by early 2027, Colossus 2 will have the power and cooling for about 1.5 gigawatts of servers. The total bill for the site by then comes to roughly $58 billion. $58 billion is not far below what TSMC, the company that actually manufactures NVIDIA's chips, now plans to spend on new factories and equipment this entire year. Most of that Memphis money buys NVIDIA hardware, but the power and cooling under it decide when any of it starts earning. And XAI has already bought a third building in South Haven. A former warehouse Musk calls "MacroHard", meant to push the whole Memphis campus further still. On paper, the chips are the easy part. By Musk's own schedule, another batch was due in early October, a second in November, and, with some luck, a third in late December, which would carry Colossus 2 past the million GPU goal it was built around. Whether those chips actually run depends on a permanent gas plant, a yard of batteries, and a grid hookup that didn't exist a few months ago. That changes what the chip industry gets judged on. When electricity is the scarce input, the chip that wins is the one that squeezes the most training out of every megawatt a site can get, because the site can't get any more. NVIDIA already sells each new generation on how much more work it squeezes out of every watt. And in Memphis, you can see why that pitch is landing. In Memphis, the chip race is now being scored in megawatts. Right now, XAI leads on the one thing this race rewards most: speed to power, with the first gigawatt-class training cluster actually running. A court ruling that forces those unpermitted turbines offline before the permanent plant is ready could change that fast, and so could a rival's nuclear restart landing on schedule. The date to watch is July 2027, when the temporary turbines are supposed to vanish from South Haven, while Colossus II keeps growing. And if AI's next leap depends on who builds electricity fastest, look at China's northern deserts, where solar farms cover the sand on a scale nothing else matches. Whether that power is as close to free as it looks is what this video answers, so click on it next. It's on your screen right now.