Why AI Data Centers Are Becoming Impossible to Build — Transcript
Full transcript
- 0:00In early 2023, Meta tore down a data
- 0:03center it had half-built in Temple
- 0:05[music], Texas. You'd assume something
- 0:07went wrong on the job site, but the
- 0:09building was out of date before it was
- 0:11even finished because the AI chips it
- 0:14was meant to hold had already outgrown
- 0:16it. It was one of the first signs of the
- 0:18strangest problem in the AI race. The
- 0:21richest companies on the planet can
- 0:23afford all the chips they want, and they
- 0:25still can't get the buildings, the
- 0:27cooling, and the power ready fast enough
- 0:30to switch them on.
- 0:31Meta is the company to follow here
- 0:33[music] because it has hit every wall in
- 0:35this story first.
- 0:37Its goal sounds simple. Build AI
- 0:40supercomputers measured in gigawatts
- 0:41[music], each one drawing as much
- 0:43electricity as a city, and build them
- 0:46fast enough to keep up with Nvidia's
- 0:48chips [music]. The obstacle is
- 0:50everything those chips have to plug
- 0:51into, from the cooling in the walls to
- 0:54the power lines outside.
- 0:56Temple is where that obstacle first
- 0:58showed up in steel and concrete.
- 1:01Satellite images of the site, out on the
- 1:03open land between Austin and Dallas,
- 1:05tell the story frame by frame.
- 1:08Early that year, a long H-shaped Meta
- 1:11data center stands there, half-finished,
- 1:14built the same way Meta had built its
- 1:15server farms for years.
- 1:17A couple of months later, the building
- 1:19is gone, and the ground [music] has been
- 1:21scraped clear.
- 1:23Publicly, Meta talked about a pause and
- 1:26a redesign.
- 1:27Shortly before the teardown [music], it
- 1:29had stopped work on the Temple site, put
- 1:31building projects in Idaho and Alabama
- 1:34on hold, and canceled an expansion in
- 1:36Odense, Denmark [music] outright.
- 1:38Its explanation was that running AI at
- 1:41scale needs a different kind of data
- 1:43center from the ones it had been
- 1:44building to serve Facebook and
- 1:45Instagram.
- 1:47The satellite pictures show how far that
- 1:49redesign went. The problem was heat.
- 1:53The old design was planned around
- 1:54ordinary server chips, and GPU chips
- 1:57pull far more power and throw off far
- 2:00more heat.
- 2:01Meta's engineers said that forced a full
- 2:03rethink with liquid cooling brought
- 2:06right up to the chips.
- 2:07The networking had to change, too,
- 2:09because the GPU chips in an AI cluster
- 2:12trade data with each other constantly.
- 2:15Meta's own planning documents, reported
- 2:16[music] at the time, called for 24 to 32
- 2:20times the networking capacity of its old
- 2:22halls.
- 2:23Finishing the old structure would have
- 2:25meant opening a data center that was
- 2:26obsolete on day one. So, it came down.
- 2:30Work restarted on the new design later
- 2:32that year.
- 2:34Meta claimed the new buildings would
- 2:35cost 31%less than the old ones and go up
- 2:39much faster.
- 2:40That's Meta's own number, so treat it as
- 2:42a target until someone outside the
- 2:44company checks it.
- 2:46The rebuild took a while.
- 2:48This past July, the Temple data center
- 2:51finally started serving traffic more
- 2:53than 4 years after Meta first shared
- 2:55plans for it. Meta says it's the first
- 2:58building to go live on its new AI
- 3:00design, cooled by a closed liquid loop
- 3:03that the company says [music] needs no
- 3:04water for most of the year.
- 3:064 years is a long time in this business.
- 3:10In that window [music], Nvidia brought
- 3:12out two whole generations of data center
- 3:14GPUs, [music] Blackwell and then
- 3:17Blackwell Ultra, and had just started
- 3:19shipping a third, Vera Rubin. And to see
- 3:22why the buildings can't just move
- 3:24faster, you have to look inside a single
- 3:26steel cabinet. That cabinet [music] is
- 3:29called a rack, the tall frame the
- 3:31servers slide into, and a classic cloud
- 3:33data center is basically [music]
- 3:35a warehouse full of them. Each rack runs
- 3:38its own small jobs, a website here, an
- 3:41email inbox there, and none of them care
- 3:43much what the neighbors are doing. An AI
- 3:45training hall works the opposite way.
- 3:48The whole floor is wired together to act
- 3:50like one enormous computer with tens of
- 3:53thousands of GPUs working on the same
- 3:55model and constantly passing results
- 3:58back and forth. Inside each Blackwell
- 4:00rack, 72 GPUs are tied together over
- 4:03thousands of copper cables, so the
- 4:06cabinet behaves almost like one giant
- 4:08chip. That changes what every rack has
- 4:11to carry. Across the industry, the
- 4:13average rack in a typical data center
- 4:15still draws under 8 kilowatts. Nvidia's
- 4:18Blackwell racks, the systems going into
- 4:20AI halls right now, each draw around 120
- 4:24kilowatts or more. Same cabinet, about
- 4:2815 times the power. That jump explains
- 4:31Temple better than anything else. Every
- 4:34cable, breaker, pipe, and air handler in
- 4:36a data center is sized for a certain
- 4:38load per rack. When that load jumps
- 4:4115-fold, none of that sizing holds
- 4:43anymore, and no simple upgrade fixes it.
- 4:47That's why Meta started over in Temple.
- 4:50600 kilowatts is the figure Nvidia has
- 4:52put on Kyber, the rack design it plans
- 4:55to ship in 2027. Every generation packs
- 4:58more GPUs and more power into the
- 5:00cabinet. Kyber is about five times the
- 5:03power of a Blackwell rack squeezed into
- 5:05the same kind of cabinet. And Nvidia has
- 5:08already talked about racks that need a
- 5:09full megawatt after that. Those are the
- 5:12racks the halls going up today will have
- 5:14to be ready for. That leaves every
- 5:17builder in a trap. A large data center
- 5:19takes years [music] to design, permit,
- 5:22and build. But Nvidia now brings out a
- 5:24new rack generation [music] roughly
- 5:26every year. So the concrete going in
- 5:28today is for chips that don't exist yet.
- 5:31Guess too low and you get Temple again,
- 5:33a building the chips outgrow before it
- 5:35opens. So builders now lean toward
- 5:38oversizing [music], even though it means
- 5:40paying for power and cooling that could
- 5:42sit unused for years. Many operators now
- 5:45hedge. They put in the power and cooling
- 5:48first and fill the halls with dense
- 5:50racks later, [music] adding them as the
- 5:52chips arrive.
- 5:54And filling those halls means breaking
- 5:56the oldest rule in the building, the one
- 5:58that says air can keep everything cool.
- 6:01Every kilowatt that goes into a rack
- 6:03comes back out as heat, and for decades
- 6:06air carried [music] it away. Cold air
- 6:09goes in the front, fans push it across
- 6:11the chips, and hot air comes out the
- 6:13back. The trouble is that air carries
- 6:15heat poorly. Somewhere around 30 to 40
- 6:18kilowatts per rack, you physically can't
- 6:21push enough of it through a cabinet
- 6:22[music] without the fans themselves
- 6:24eating a big slice of the power. Every
- 6:27Blackwell rack is far past that line.
- 6:30So, the industry turned to liquid. In a
- 6:32liquid-cooled rack, a metal cold plate
- 6:35sits right on top of each GPU with
- 6:37coolant flowing through tiny channels
- 6:39inside it, carrying the heat out of the
- 6:41hall to big coolers outside. Water can
- 6:45carry far more heat than air, which is
- 6:47why this one change makes the dense
- 6:49racks possible at all. It's the system
- 6:52Meta built into Temple. Liquid brings
- 6:54its own risks, though. If the pumps stop
- 6:57while the racks are at full load, the
- 6:58chips can overheat within seconds. So,
- 7:01the cooling needs its own backup pumps
- 7:03and backup power, just like the servers
- 7:05do. Then, Nvidia went further. A single
- 7:09Vera Rubin rack can draw around 200
- 7:12kilowatts. A household space heater on
- 7:14full blast draws about 1 and 1/2
- 7:16kilowatts. So, that single cabinet burns
- 7:19as much power as well over 100 of them.
- 7:22Nvidia cools every compute tray in that
- 7:24rack with liquid, and the trays have no
- 7:27hoses hanging off them and no fans at
- 7:29all. Think about what that does to a
- 7:31building designed around air. The giant
- 7:34air handlers, the carefully planned hot
- 7:36and cold aisles, the whole layout a
- 7:38classic data center was built around now
- 7:41has far less to do. In its place, you
- 7:44need pipes, pumps, heat exchangers, and
- 7:46leak [music] sensors running under every
- 7:48row. And how do you even get that much
- 7:50power into one cabinet? Inside today's
- 7:53racks, power travels at low voltage, and
- 7:56pushing hundreds of kilowatts at low
- 7:58voltage takes thick, heavy bars of
- 8:00copper. So, Nvidia is moving the
- 8:03industry to 800-V direct current, which
- 8:05is nearly seven times the voltage at a
- 8:07North American wall socket. The power
- 8:10gets converted from the grid once, then
- 8:12travels at that higher voltage, so far
- 8:14less copper can carry the same load.
- 8:17Nvidia says that system reaches full
- 8:19production alongside the Kyber racks.
- 8:22That's two redesigns at once, the
- 8:24cooling and the wiring, every time
- 8:26Nvidia moves up a generation. They're
- 8:28expensive, and every new AI hall now
- 8:31needs them. The equipment they call for
- 8:34is [music] where the real waiting
- 8:35starts.
- 8:36Start at the edge of the property, where
- 8:38the grid comes in.
- 8:40Electricity arrives over high-voltage
- 8:42lines, and before a [music] data center
- 8:44can use it, huge transformers have to
- 8:46step that voltage down.
- 8:48A large power transformer is
- 8:50custom-built, weighs hundreds of tons,
- 8:53and is wound from specialized steel and
- 8:55copper.
- 8:56Each one is built for a specific site,
- 8:59tested at the factory, and then hauled
- 9:01out on special rail cars because it's
- 9:03too heavy for most roads.
- 9:05Today, the average wait for a standard
- 9:07one is about 128 [music] weeks. That's
- 9:11roughly 2 and 1/2 years, nearly as long
- 9:13as it took Meta to finish Temple after
- 9:16work restarted. The biggest units can
- 9:18take 4 years or more, and prices have
- 9:21jumped by roughly 80 [music]%in just a
- 9:23few years. It isn't only transformers,
- 9:26either. Switch gear, the heavy switching
- 9:29equipment that routes and cuts power
- 9:31inside the building, is stuck in the
- 9:33same queue, along with circuit breakers,
- 9:35because utilities, factories, and AI
- 9:38builders are all ordering from the same
- 9:40small group of suppliers at once. The
- 9:43suppliers are responding. This June,
- 9:46Hitachi Energy broke ground in South
- 9:48Boston, Virginia, on a big new factory
- 9:50for these giant transformers.
- 9:52But a plant that's only now breaking
- 9:54ground can't ship anything to the halls
- 9:56going up today.
- 9:58This is where the problem stops being
- 10:00theoretical, because the chips are now
- 10:02arriving faster than the rooms to put
- 10:04them in.
- 10:06Late last year, Microsoft's chief
- 10:08executive, Satya Nadella, made an
- 10:10unusual admission on an investing
- 10:12podcast.
- 10:14He said Microsoft's biggest issue had
- 10:16become power [music] and getting
- 10:18buildings finished close enough to that
- 10:20power.
- 10:21His warning was that Microsoft might be
- 10:23left holding Nvidia chips it had no
- 10:25place to switch on because it was short
- 10:27of what he called warm shells, finished
- 10:30buildings with the power and cooling
- 10:32ready to go.
- 10:33Think about who was saying that.
- 10:36Microsoft is one of Nvidia's biggest
- 10:38customers with every reason to talk up
- 10:40its AI buildout, and its boss was
- 10:42telling the world the limit was
- 10:44buildings and electricity.
- 10:46He even said he'd rather not buy too
- 10:48much of any one generation of Nvidia
- 10:50chips because the next one shows up so
- 10:53quickly.
- 10:54For years, the big worry in AI was
- 10:56whether TSMC could [music] make enough
- 10:58chips. Nadella was adding a new worry to
- 11:01that list, one that sits at the
- 11:03substation.
- 11:05Microsoft's own answer is Fairwater
- 11:06[music], a campus in Mount Pleasant,
- 11:08Wisconsin, built to hold hundreds of
- 11:11thousands of Nvidia GPUs.
- 11:13Microsoft stacked its racks on two
- 11:15floors so the GPUs could sit closer
- 11:18together and the cables between them
- 11:20could stay short. Beneath the site run
- 11:22about 190 km [music] of underground
- 11:25cable, more than 4 and 1/2 marathons
- 11:28laid end to end.
- 11:29It's cooled by a closed liquid loop
- 11:31[music] that Microsoft says is backed by
- 11:33the second largest water-cooled chiller
- 11:35plant on the planet. And every bit of it
- 11:38still depends on power from outside the
- 11:40fence, which is where [music] this
- 11:42bottleneck gets a lot bigger.
- 11:45A grid connection is slower still
- 11:47because it depends on power lines
- 11:49[music], substations, and new power
- 11:51plants, and those take years of planning
- 11:54and permits.
- 11:55In Northern Virginia, the biggest data
- 11:57center market in the world, the local
- 11:59utility, Dominion Energy, now has
- 12:01requests from large customers adding up
- 12:04to nearly [music] three times the most
- 12:06power its whole system has ever
- 12:08delivered at once.
- 12:09A big new data center there can wait as
- 12:11long as 7 [music] years for its
- 12:13connection.
- 12:14Virginia's regulators have already
- 12:15created a separate electricity rate for
- 12:17the biggest users [music], aimed at
- 12:19keeping the cost of new lines off
- 12:21everyone else's bill. So, where do
- 12:24builders go when Northern Virginia is
- 12:26full?
- 12:27Texas was supposed to be the escape
- 12:29route. It has open land, plenty [music]
- 12:31of gas and wind, and a grid of its own.
- 12:34So, the industry piled in, Meta
- 12:36included. By this summer, the line of
- 12:38requests to plug into the Texas grid had
- 12:41reached around 474 gigawatts, nearly all
- 12:44of it from data centers. That's more
- 12:47than five times the most electricity the
- 12:49entire Texas grid has [music] ever
- 12:51delivered at one moment. Plenty of those
- 12:53requests are speculative, the same
- 12:55developer reserving power at several
- 12:57sites and planning to build at only one.
- 13:00Texas had already passed a law making
- 13:02big new users pay $100,000 just to apply
- 13:06and own up to any duplicate requests.
- 13:08[music] Even so, a small slice of that
- 13:11line would swamp the grid. And then
- 13:13Texas did something that would have
- 13:15sounded unthinkable when Meta first
- 13:17broke ground in Temple. In early August,
- 13:19Governor Greg Abbott ordered a full
- 13:21audit of every data center waiting to
- 13:23connect to the Texas grid [music], and
- 13:26no new data center could move forward
- 13:27until it was done.
- 13:29The grid operator paused approvals to
- 13:31switch on new data centers and crypto
- 13:33mines while it worked through the list.
- 13:35Just days earlier, when the grid
- 13:37operator sorted that line, only about 4
- 13:40gigawatts and every 10 [music] came from
- 13:42projects that had cleared its first
- 13:43eligibility check. Now look at the
- 13:46timing.
- 13:47Meta's Temple building, the one torn
- 13:49down and rebuilt for the AI era, started
- 13:51serving traffic in late [music] July, on
- 13:54the same day the Texas grid set its
- 13:56all-time record for demand. Less than 2
- 13:59weeks later, the state it sits in froze
- 14:01the line of data centers waiting behind
- 14:03it. One of the most building-friendly
- 14:05places in America [music]
- 14:06had decided the power question came
- 14:08first. And plenty of builders were fine
- 14:11with that. At least 43 companies
- 14:13building data centers [music] publicly
- 14:15backed the audit, and clearing the
- 14:17speculative requests out could move real
- 14:20projects up the line. If you're getting
- 14:22something out of this, subscribe.
- 14:24Tomorrow Engineered covers the machines,
- 14:26the factories, and the companies racing
- 14:29to build the world's most advanced
- 14:31chips, and who's really ahead.
- 14:331.2 gigawatts is how much power Elon
- 14:36Musk's AI company, XAI, now has
- 14:39permission to generate from gas turbines
- 14:41on a single site in Mississippi.
- 14:44All of it for its own supercomputers.
- 14:47That's enough electricity for roughly
- 14:49900,000 American homes.
- 14:52While the grid queues kept growing, XAI
- 14:54decided not to stand in them at all.
- 14:57For its Colossus 2 supercomputer in
- 14:59Memphis, Tennessee, XAI bought the site
- 15:02of an old power plant just across the
- 15:04state line in Southaven, Mississippi,
- 15:06and started filling it with gas
- 15:08turbines, [music] beginning with
- 15:09portable units it could truck in.
- 15:11Mississippi regulators then approved a
- 15:13permit for 41 turbines there.
- 15:16At its first Memphis site, [music] XAI
- 15:18ran temporary turbines while it waited
- 15:20for the grid, and now that its
- 15:22substations are online, it has been
- 15:24taking some of them away.
- 15:26It hasn't been smooth. Local
- 15:28environmental and civil rights groups
- 15:30accused XAI of running turbines before
- 15:33it had the right permits, and the fight
- 15:35over air quality in South Memphis is
- 15:37still going.
- 15:38Shelby County has since granted a permit
- 15:40for turbines at the Memphis site, and
- 15:42that permit [music] is under appeal.
- 15:45But the logic is easy to follow. A
- 15:47turbine you can buy and truck in can be
- 15:49running long before a grid connection
- 15:51that might take years.
- 15:53Except [music] the turbines are running
- 15:55out, too.
- 15:56GE Vernova, one of the world's biggest
- 15:59turbine makers, is sold out of gas
- 16:01[music] turbines through 2028, and it
- 16:03expects data centers to drive about a
- 16:05third of its American [music] gas power
- 16:07deals.
- 16:08A builder who decides today that it
- 16:10wants its own gas plant joins a line
- 16:12stretching toward [music] the end of the
- 16:14decade. So, even going off the grid now
- 16:17means waiting years for equipment.
- 16:20So the biggest players are locking
- 16:21[music] up power years ahead anyway they
- 16:24can. Microsoft went back to a name
- 16:26almost everyone [music] knows. It signed
- 16:29a long-term deal that's bringing a
- 16:31shutdown reactor at Three Mile Island in
- 16:33Pennsylvania back to life. Now renamed
- 16:36the Crane Clean Energy Center with power
- 16:39aimed [music] for next year.
- 16:41Once it restarts, the reactor is
- 16:42expected to put about 835 [music]
- 16:45MW back on the grid. Google went a step
- 16:48further and bought an energy developer
- 16:50outright. It's parent company, Alphabet,
- 16:53paid $4.75 billion in cash for
- 16:56Intersect, a company that builds power
- 16:59plants, solar farms, and batteries right
- 17:01alongside data centers [music], starting
- 17:03with a shared site in Haskell County,
- 17:05Texas. That's about four times what Meta
- 17:08says it spent on the entire Temple
- 17:10[music] data center just to own the
- 17:12power side of the business. Meta took
- 17:14the idea further than anyone. In New
- 17:17Albany, Ohio, outside Columbus, it's
- 17:19bringing up Prometheus, its first
- 17:21gigawatt-class AI cluster with part of
- 17:24it already running. By outside
- 17:26estimates, the rebuilt Temple site runs
- 17:28on roughly 150 MW of computing power.
- 17:32Prometheus is designed for more than six
- 17:35times [music] that. To get it built
- 17:36faster, Meta put part of it inside giant
- 17:39weather-proof tents, aluminum frames
- 17:42wrapped in fabric that go up far quicker
- 17:44than a concrete hall.
- 17:46Mark Zuckerberg has said the tents are
- 17:47there for speed. Meta isn't relying on
- 17:50the grid alone, either. Filings show the
- 17:53site has its own power generation yard
- 17:55tied to the data halls by a wide utility
- 17:58corridor. And Ohio regulators approved
- 18:00two dedicated natural gas plants to feed
- 18:03the campus directly. Meta has also lined
- 18:06up nuclear power from Vistra,
- 18:08TerraPower, and Oaklo to back it. Some
- 18:10of it from reactors that haven't been
- 18:12built yet.
- 18:145 GW of computing power is what Meta
- 18:17wants from Hyperion, a campus it's
- 18:19building in Richland Parish, out in the
- 18:21rural northeast of Louisiana.
- 18:24Zuckerberg has said its footprint would
- 18:26cover a big [music] part of Manhattan.
- 18:28To pay for it, Meta set up a $27 billion
- 18:32joint venture [music] with the
- 18:33investment firm Blue Owl Capital, which
- 18:35keeps much of the debt off Meta's own
- 18:37books.
- 18:38To power it, Meta didn't just sign an
- 18:40electricity contract. Regulators had
- 18:43already approved new gas plants for the
- 18:45site.
- 18:46Then, this March, Meta and the utility
- 18:49Entergy announced [music] a much bigger
- 18:51deal.
- 18:52Hyperion is now set to be served by 10
- 18:55gas power plants, more than 7 gigawatts
- 18:58of capacity for a single customer.
- 19:01That's more than half of all the power
- 19:03Entergy currently generates across the
- 19:05entire state of Louisiana.
- 19:07It's a fleet of power plants big enough
- 19:09to reshape a state's grid. And the new
- 19:12[music] ones still need sign-off from
- 19:13Louisiana's regulators.
- 19:15That review is where [music] the real
- 19:17fight is, over who pays if plants built
- 19:20to run for decades outlive [music] the
- 19:22data center they were built for.
- 19:24Meta and Entergy say Meta will cover its
- 19:26own costs. Put it all together, and the
- 19:29biggest names in AI are turning into
- 19:31energy companies. They're paying for
- 19:33power plants, restarting reactors, and
- 19:36building their own supply, because the
- 19:38chips they've already paid for need
- 19:40somewhere to run.
- 19:42So, why not just make the chips use less
- 19:44power?
- 19:45If every chip did more work per watt,
- 19:48the same building could do more without
- 19:50a single new power plant. Nvidia says
- 19:53that's exactly what it's delivering.
- 19:55It claims Vera Rubin produces up to 10
- 19:58times more AI output per megawatt than
- 20:00Blackwell. In plain terms, a hall that
- 20:03used to answer one AI request could
- 20:05answer 10 on the same electricity.
- 20:08That's Nvidia's own number, measured on
- 20:10a workload Nvidia chose.
- 20:13But the direction is real, and Nvidia
- 20:15says that across six generations of its
- 20:17chips, AI output per megawatt has
- 20:20improved about a million times over.
- 20:22Vera Rubin itself shows what happens to
- 20:25those gains. It's far more efficient
- 20:27than Blackwell, and each of its racks
- 20:29still draws more power than a Blackwell
- 20:32rack, because Nvidia spends the savings
- 20:34on packing more work into every cabinet.
- 20:37The whole industry does the same thing,
- 20:39which is why the power bill keeps
- 20:41climbing. When AI output gets cheaper to
- 20:44produce, companies don't bank the
- 20:46savings and switch off the lights in the
- 20:48hall. They run bigger models, longer
- 20:50conversations, and more AI agents
- 20:53working around the clock. Britain went
- 20:55through the same thing with coal in the
- 20:571800s, when better steam engines made
- 20:59coal power cheaper, and the country
- 21:01ended up burning more of it. The
- 21:03International Energy Agency has put a
- 21:05figure on where that leads. By 2030, it
- 21:09expects global data center electricity
- 21:11use to more than double, with AI as the
- 21:13biggest driver [music]. That would be
- 21:15around 945 terawatt hours a year,
- 21:19[music] more than all of Japan uses
- 21:21today.
- 21:22That's the hardest obstacle of all,
- 21:25because more money doesn't make any of
- 21:27it arrive sooner.
- 21:28Nvidia brings out a new rack generation
- 21:31roughly every year, while every fix in
- 21:33this story, the transformers [music],
- 21:35the turbines, the reactors, takes years.
- 21:38Hyperion's newest power [music] plants
- 21:40are still waiting on regulators, and the
- 21:42Texas audit is still working through its
- 21:44list.
- 21:45The neighbors are pushing back, too.
- 21:48[music] In Temple, a local group has
- 21:49formed to fight more data center
- 21:51construction, worried about the heat
- 21:53these sites throw off. In Virginia,
- 21:55Texas, and Louisiana, regulators are now
- 21:58arguing over who should pay for the new
- 22:00power lines and plants, the tech
- 22:02companies or the families living nearby.
- 22:05Every one of those fights adds months to
- 22:07projects that were already running late.
- 22:10Temple is running now, liquid-cooled and
- 22:13full of AI hardware, on a design that
- 22:15didn't exist [music] when its first
- 22:17walls went up. But Kyber is coming, and
- 22:20every hall built around today's racks
- 22:22[music] will face the same question
- 22:23Temple did. For now, the physical world
- 22:26is a harder limit on AI than the chips
- 22:28are because Nvidia can design a new rack
- 22:31faster than anyone can build the hall,
- 22:33the substation, and the power plant to
- 22:35plug it into.
- 22:37The leaders are the companies that bring
- 22:38their own [music] power. xAI moves
- 22:41fastest, Meta has made the biggest bets,
- 22:44and Microsoft has warned its chips could
- 22:46end up waiting on buildings. What could
- 22:48change that is more [music] transformer
- 22:50and turbine factories and chips that cut
- 22:53the power each job needs. The one thing
- 22:55to watch is 2027 when Nvidia plans to
- 22:59ship the first Kyber racks and we find
- 23:01[music] out who has a hall ready to
- 23:03switch them on.
- 23:05If the power coming in can't grow fast
- 23:07enough, the savings have to come from
- 23:09inside the rack where GPUs burn energy
- 23:12passing data over copper. Chip makers
- 23:14from TSMC to Nvidia are racing to swap
- 23:17[music] that copper for light using tiny
- 23:20lasers beside the chip.
- 23:22Watch this video [music] next. It's on
- 23:24your screen right now.
About this transcript
This page contains the full transcript of Why AI Data Centers Are Becoming Impossible to Build by Tomorrow Engineered, generated from the public captions YouTube serves with the video. The transcript has 3,909 words across 622 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.