YouTube2Text

Why AI Data Centers Are Becoming Impossible to Build — Transcript

by Tomorrow Engineered · 3,909 words · 622 segments · language en · Watch on YouTube

Full transcript

  1. 0:00In early 2023, Meta tore down a data
  2. 0:03center it had half-built in Temple
  3. 0:05[music], Texas. You'd assume something
  4. 0:07went wrong on the job site, but the
  5. 0:09building was out of date before it was
  6. 0:11even finished because the AI chips it
  7. 0:14was meant to hold had already outgrown
  8. 0:16it. It was one of the first signs of the
  9. 0:18strangest problem in the AI race. The
  10. 0:21richest companies on the planet can
  11. 0:23afford all the chips they want, and they
  12. 0:25still can't get the buildings, the
  13. 0:27cooling, and the power ready fast enough
  14. 0:30to switch them on.
  15. 0:31Meta is the company to follow here
  16. 0:33[music] because it has hit every wall in
  17. 0:35this story first.
  18. 0:37Its goal sounds simple. Build AI
  19. 0:40supercomputers measured in gigawatts
  20. 0:41[music], each one drawing as much
  21. 0:43electricity as a city, and build them
  22. 0:46fast enough to keep up with Nvidia's
  23. 0:48chips [music]. The obstacle is
  24. 0:50everything those chips have to plug
  25. 0:51into, from the cooling in the walls to
  26. 0:54the power lines outside.
  27. 0:56Temple is where that obstacle first
  28. 0:58showed up in steel and concrete.
  29. 1:01Satellite images of the site, out on the
  30. 1:03open land between Austin and Dallas,
  31. 1:05tell the story frame by frame.
  32. 1:08Early that year, a long H-shaped Meta
  33. 1:11data center stands there, half-finished,
  34. 1:14built the same way Meta had built its
  35. 1:15server farms for years.
  36. 1:17A couple of months later, the building
  37. 1:19is gone, and the ground [music] has been
  38. 1:21scraped clear.
  39. 1:23Publicly, Meta talked about a pause and
  40. 1:26a redesign.
  41. 1:27Shortly before the teardown [music], it
  42. 1:29had stopped work on the Temple site, put
  43. 1:31building projects in Idaho and Alabama
  44. 1:34on hold, and canceled an expansion in
  45. 1:36Odense, Denmark [music] outright.
  46. 1:38Its explanation was that running AI at
  47. 1:41scale needs a different kind of data
  48. 1:43center from the ones it had been
  49. 1:44building to serve Facebook and
  50. 1:45Instagram.
  51. 1:47The satellite pictures show how far that
  52. 1:49redesign went. The problem was heat.
  53. 1:53The old design was planned around
  54. 1:54ordinary server chips, and GPU chips
  55. 1:57pull far more power and throw off far
  56. 2:00more heat.
  57. 2:01Meta's engineers said that forced a full
  58. 2:03rethink with liquid cooling brought
  59. 2:06right up to the chips.
  60. 2:07The networking had to change, too,
  61. 2:09because the GPU chips in an AI cluster
  62. 2:12trade data with each other constantly.
  63. 2:15Meta's own planning documents, reported
  64. 2:16[music] at the time, called for 24 to 32
  65. 2:20times the networking capacity of its old
  66. 2:22halls.
  67. 2:23Finishing the old structure would have
  68. 2:25meant opening a data center that was
  69. 2:26obsolete on day one. So, it came down.
  70. 2:30Work restarted on the new design later
  71. 2:32that year.
  72. 2:34Meta claimed the new buildings would
  73. 2:35cost 31%less than the old ones and go up
  74. 2:39much faster.
  75. 2:40That's Meta's own number, so treat it as
  76. 2:42a target until someone outside the
  77. 2:44company checks it.
  78. 2:46The rebuild took a while.
  79. 2:48This past July, the Temple data center
  80. 2:51finally started serving traffic more
  81. 2:53than 4 years after Meta first shared
  82. 2:55plans for it. Meta says it's the first
  83. 2:58building to go live on its new AI
  84. 3:00design, cooled by a closed liquid loop
  85. 3:03that the company says [music] needs no
  86. 3:04water for most of the year.
  87. 3:064 years is a long time in this business.
  88. 3:10In that window [music], Nvidia brought
  89. 3:12out two whole generations of data center
  90. 3:14GPUs, [music] Blackwell and then
  91. 3:17Blackwell Ultra, and had just started
  92. 3:19shipping a third, Vera Rubin. And to see
  93. 3:22why the buildings can't just move
  94. 3:24faster, you have to look inside a single
  95. 3:26steel cabinet. That cabinet [music] is
  96. 3:29called a rack, the tall frame the
  97. 3:31servers slide into, and a classic cloud
  98. 3:33data center is basically [music]
  99. 3:35a warehouse full of them. Each rack runs
  100. 3:38its own small jobs, a website here, an
  101. 3:41email inbox there, and none of them care
  102. 3:43much what the neighbors are doing. An AI
  103. 3:45training hall works the opposite way.
  104. 3:48The whole floor is wired together to act
  105. 3:50like one enormous computer with tens of
  106. 3:53thousands of GPUs working on the same
  107. 3:55model and constantly passing results
  108. 3:58back and forth. Inside each Blackwell
  109. 4:00rack, 72 GPUs are tied together over
  110. 4:03thousands of copper cables, so the
  111. 4:06cabinet behaves almost like one giant
  112. 4:08chip. That changes what every rack has
  113. 4:11to carry. Across the industry, the
  114. 4:13average rack in a typical data center
  115. 4:15still draws under 8 kilowatts. Nvidia's
  116. 4:18Blackwell racks, the systems going into
  117. 4:20AI halls right now, each draw around 120
  118. 4:24kilowatts or more. Same cabinet, about
  119. 4:2815 times the power. That jump explains
  120. 4:31Temple better than anything else. Every
  121. 4:34cable, breaker, pipe, and air handler in
  122. 4:36a data center is sized for a certain
  123. 4:38load per rack. When that load jumps
  124. 4:4115-fold, none of that sizing holds
  125. 4:43anymore, and no simple upgrade fixes it.
  126. 4:47That's why Meta started over in Temple.
  127. 4:50600 kilowatts is the figure Nvidia has
  128. 4:52put on Kyber, the rack design it plans
  129. 4:55to ship in 2027. Every generation packs
  130. 4:58more GPUs and more power into the
  131. 5:00cabinet. Kyber is about five times the
  132. 5:03power of a Blackwell rack squeezed into
  133. 5:05the same kind of cabinet. And Nvidia has
  134. 5:08already talked about racks that need a
  135. 5:09full megawatt after that. Those are the
  136. 5:12racks the halls going up today will have
  137. 5:14to be ready for. That leaves every
  138. 5:17builder in a trap. A large data center
  139. 5:19takes years [music] to design, permit,
  140. 5:22and build. But Nvidia now brings out a
  141. 5:24new rack generation [music] roughly
  142. 5:26every year. So the concrete going in
  143. 5:28today is for chips that don't exist yet.
  144. 5:31Guess too low and you get Temple again,
  145. 5:33a building the chips outgrow before it
  146. 5:35opens. So builders now lean toward
  147. 5:38oversizing [music], even though it means
  148. 5:40paying for power and cooling that could
  149. 5:42sit unused for years. Many operators now
  150. 5:45hedge. They put in the power and cooling
  151. 5:48first and fill the halls with dense
  152. 5:50racks later, [music] adding them as the
  153. 5:52chips arrive.
  154. 5:54And filling those halls means breaking
  155. 5:56the oldest rule in the building, the one
  156. 5:58that says air can keep everything cool.
  157. 6:01Every kilowatt that goes into a rack
  158. 6:03comes back out as heat, and for decades
  159. 6:06air carried [music] it away. Cold air
  160. 6:09goes in the front, fans push it across
  161. 6:11the chips, and hot air comes out the
  162. 6:13back. The trouble is that air carries
  163. 6:15heat poorly. Somewhere around 30 to 40
  164. 6:18kilowatts per rack, you physically can't
  165. 6:21push enough of it through a cabinet
  166. 6:22[music] without the fans themselves
  167. 6:24eating a big slice of the power. Every
  168. 6:27Blackwell rack is far past that line.
  169. 6:30So, the industry turned to liquid. In a
  170. 6:32liquid-cooled rack, a metal cold plate
  171. 6:35sits right on top of each GPU with
  172. 6:37coolant flowing through tiny channels
  173. 6:39inside it, carrying the heat out of the
  174. 6:41hall to big coolers outside. Water can
  175. 6:45carry far more heat than air, which is
  176. 6:47why this one change makes the dense
  177. 6:49racks possible at all. It's the system
  178. 6:52Meta built into Temple. Liquid brings
  179. 6:54its own risks, though. If the pumps stop
  180. 6:57while the racks are at full load, the
  181. 6:58chips can overheat within seconds. So,
  182. 7:01the cooling needs its own backup pumps
  183. 7:03and backup power, just like the servers
  184. 7:05do. Then, Nvidia went further. A single
  185. 7:09Vera Rubin rack can draw around 200
  186. 7:12kilowatts. A household space heater on
  187. 7:14full blast draws about 1 and 1/2
  188. 7:16kilowatts. So, that single cabinet burns
  189. 7:19as much power as well over 100 of them.
  190. 7:22Nvidia cools every compute tray in that
  191. 7:24rack with liquid, and the trays have no
  192. 7:27hoses hanging off them and no fans at
  193. 7:29all. Think about what that does to a
  194. 7:31building designed around air. The giant
  195. 7:34air handlers, the carefully planned hot
  196. 7:36and cold aisles, the whole layout a
  197. 7:38classic data center was built around now
  198. 7:41has far less to do. In its place, you
  199. 7:44need pipes, pumps, heat exchangers, and
  200. 7:46leak [music] sensors running under every
  201. 7:48row. And how do you even get that much
  202. 7:50power into one cabinet? Inside today's
  203. 7:53racks, power travels at low voltage, and
  204. 7:56pushing hundreds of kilowatts at low
  205. 7:58voltage takes thick, heavy bars of
  206. 8:00copper. So, Nvidia is moving the
  207. 8:03industry to 800-V direct current, which
  208. 8:05is nearly seven times the voltage at a
  209. 8:07North American wall socket. The power
  210. 8:10gets converted from the grid once, then
  211. 8:12travels at that higher voltage, so far
  212. 8:14less copper can carry the same load.
  213. 8:17Nvidia says that system reaches full
  214. 8:19production alongside the Kyber racks.
  215. 8:22That's two redesigns at once, the
  216. 8:24cooling and the wiring, every time
  217. 8:26Nvidia moves up a generation. They're
  218. 8:28expensive, and every new AI hall now
  219. 8:31needs them. The equipment they call for
  220. 8:34is [music] where the real waiting
  221. 8:35starts.
  222. 8:36Start at the edge of the property, where
  223. 8:38the grid comes in.
  224. 8:40Electricity arrives over high-voltage
  225. 8:42lines, and before a [music] data center
  226. 8:44can use it, huge transformers have to
  227. 8:46step that voltage down.
  228. 8:48A large power transformer is
  229. 8:50custom-built, weighs hundreds of tons,
  230. 8:53and is wound from specialized steel and
  231. 8:55copper.
  232. 8:56Each one is built for a specific site,
  233. 8:59tested at the factory, and then hauled
  234. 9:01out on special rail cars because it's
  235. 9:03too heavy for most roads.
  236. 9:05Today, the average wait for a standard
  237. 9:07one is about 128 [music] weeks. That's
  238. 9:11roughly 2 and 1/2 years, nearly as long
  239. 9:13as it took Meta to finish Temple after
  240. 9:16work restarted. The biggest units can
  241. 9:18take 4 years or more, and prices have
  242. 9:21jumped by roughly 80 [music]%in just a
  243. 9:23few years. It isn't only transformers,
  244. 9:26either. Switch gear, the heavy switching
  245. 9:29equipment that routes and cuts power
  246. 9:31inside the building, is stuck in the
  247. 9:33same queue, along with circuit breakers,
  248. 9:35because utilities, factories, and AI
  249. 9:38builders are all ordering from the same
  250. 9:40small group of suppliers at once. The
  251. 9:43suppliers are responding. This June,
  252. 9:46Hitachi Energy broke ground in South
  253. 9:48Boston, Virginia, on a big new factory
  254. 9:50for these giant transformers.
  255. 9:52But a plant that's only now breaking
  256. 9:54ground can't ship anything to the halls
  257. 9:56going up today.
  258. 9:58This is where the problem stops being
  259. 10:00theoretical, because the chips are now
  260. 10:02arriving faster than the rooms to put
  261. 10:04them in.
  262. 10:06Late last year, Microsoft's chief
  263. 10:08executive, Satya Nadella, made an
  264. 10:10unusual admission on an investing
  265. 10:12podcast.
  266. 10:14He said Microsoft's biggest issue had
  267. 10:16become power [music] and getting
  268. 10:18buildings finished close enough to that
  269. 10:20power.
  270. 10:21His warning was that Microsoft might be
  271. 10:23left holding Nvidia chips it had no
  272. 10:25place to switch on because it was short
  273. 10:27of what he called warm shells, finished
  274. 10:30buildings with the power and cooling
  275. 10:32ready to go.
  276. 10:33Think about who was saying that.
  277. 10:36Microsoft is one of Nvidia's biggest
  278. 10:38customers with every reason to talk up
  279. 10:40its AI buildout, and its boss was
  280. 10:42telling the world the limit was
  281. 10:44buildings and electricity.
  282. 10:46He even said he'd rather not buy too
  283. 10:48much of any one generation of Nvidia
  284. 10:50chips because the next one shows up so
  285. 10:53quickly.
  286. 10:54For years, the big worry in AI was
  287. 10:56whether TSMC could [music] make enough
  288. 10:58chips. Nadella was adding a new worry to
  289. 11:01that list, one that sits at the
  290. 11:03substation.
  291. 11:05Microsoft's own answer is Fairwater
  292. 11:06[music], a campus in Mount Pleasant,
  293. 11:08Wisconsin, built to hold hundreds of
  294. 11:11thousands of Nvidia GPUs.
  295. 11:13Microsoft stacked its racks on two
  296. 11:15floors so the GPUs could sit closer
  297. 11:18together and the cables between them
  298. 11:20could stay short. Beneath the site run
  299. 11:22about 190 km [music] of underground
  300. 11:25cable, more than 4 and 1/2 marathons
  301. 11:28laid end to end.
  302. 11:29It's cooled by a closed liquid loop
  303. 11:31[music] that Microsoft says is backed by
  304. 11:33the second largest water-cooled chiller
  305. 11:35plant on the planet. And every bit of it
  306. 11:38still depends on power from outside the
  307. 11:40fence, which is where [music] this
  308. 11:42bottleneck gets a lot bigger.
  309. 11:45A grid connection is slower still
  310. 11:47because it depends on power lines
  311. 11:49[music], substations, and new power
  312. 11:51plants, and those take years of planning
  313. 11:54and permits.
  314. 11:55In Northern Virginia, the biggest data
  315. 11:57center market in the world, the local
  316. 11:59utility, Dominion Energy, now has
  317. 12:01requests from large customers adding up
  318. 12:04to nearly [music] three times the most
  319. 12:06power its whole system has ever
  320. 12:08delivered at once.
  321. 12:09A big new data center there can wait as
  322. 12:11long as 7 [music] years for its
  323. 12:13connection.
  324. 12:14Virginia's regulators have already
  325. 12:15created a separate electricity rate for
  326. 12:17the biggest users [music], aimed at
  327. 12:19keeping the cost of new lines off
  328. 12:21everyone else's bill. So, where do
  329. 12:24builders go when Northern Virginia is
  330. 12:26full?
  331. 12:27Texas was supposed to be the escape
  332. 12:29route. It has open land, plenty [music]
  333. 12:31of gas and wind, and a grid of its own.
  334. 12:34So, the industry piled in, Meta
  335. 12:36included. By this summer, the line of
  336. 12:38requests to plug into the Texas grid had
  337. 12:41reached around 474 gigawatts, nearly all
  338. 12:44of it from data centers. That's more
  339. 12:47than five times the most electricity the
  340. 12:49entire Texas grid has [music] ever
  341. 12:51delivered at one moment. Plenty of those
  342. 12:53requests are speculative, the same
  343. 12:55developer reserving power at several
  344. 12:57sites and planning to build at only one.
  345. 13:00Texas had already passed a law making
  346. 13:02big new users pay $100,000 just to apply
  347. 13:06and own up to any duplicate requests.
  348. 13:08[music] Even so, a small slice of that
  349. 13:11line would swamp the grid. And then
  350. 13:13Texas did something that would have
  351. 13:15sounded unthinkable when Meta first
  352. 13:17broke ground in Temple. In early August,
  353. 13:19Governor Greg Abbott ordered a full
  354. 13:21audit of every data center waiting to
  355. 13:23connect to the Texas grid [music], and
  356. 13:26no new data center could move forward
  357. 13:27until it was done.
  358. 13:29The grid operator paused approvals to
  359. 13:31switch on new data centers and crypto
  360. 13:33mines while it worked through the list.
  361. 13:35Just days earlier, when the grid
  362. 13:37operator sorted that line, only about 4
  363. 13:40gigawatts and every 10 [music] came from
  364. 13:42projects that had cleared its first
  365. 13:43eligibility check. Now look at the
  366. 13:46timing.
  367. 13:47Meta's Temple building, the one torn
  368. 13:49down and rebuilt for the AI era, started
  369. 13:51serving traffic in late [music] July, on
  370. 13:54the same day the Texas grid set its
  371. 13:56all-time record for demand. Less than 2
  372. 13:59weeks later, the state it sits in froze
  373. 14:01the line of data centers waiting behind
  374. 14:03it. One of the most building-friendly
  375. 14:05places in America [music]
  376. 14:06had decided the power question came
  377. 14:08first. And plenty of builders were fine
  378. 14:11with that. At least 43 companies
  379. 14:13building data centers [music] publicly
  380. 14:15backed the audit, and clearing the
  381. 14:17speculative requests out could move real
  382. 14:20projects up the line. If you're getting
  383. 14:22something out of this, subscribe.
  384. 14:24Tomorrow Engineered covers the machines,
  385. 14:26the factories, and the companies racing
  386. 14:29to build the world's most advanced
  387. 14:31chips, and who's really ahead.
  388. 14:331.2 gigawatts is how much power Elon
  389. 14:36Musk's AI company, XAI, now has
  390. 14:39permission to generate from gas turbines
  391. 14:41on a single site in Mississippi.
  392. 14:44All of it for its own supercomputers.
  393. 14:47That's enough electricity for roughly
  394. 14:49900,000 American homes.
  395. 14:52While the grid queues kept growing, XAI
  396. 14:54decided not to stand in them at all.
  397. 14:57For its Colossus 2 supercomputer in
  398. 14:59Memphis, Tennessee, XAI bought the site
  399. 15:02of an old power plant just across the
  400. 15:04state line in Southaven, Mississippi,
  401. 15:06and started filling it with gas
  402. 15:08turbines, [music] beginning with
  403. 15:09portable units it could truck in.
  404. 15:11Mississippi regulators then approved a
  405. 15:13permit for 41 turbines there.
  406. 15:16At its first Memphis site, [music] XAI
  407. 15:18ran temporary turbines while it waited
  408. 15:20for the grid, and now that its
  409. 15:22substations are online, it has been
  410. 15:24taking some of them away.
  411. 15:26It hasn't been smooth. Local
  412. 15:28environmental and civil rights groups
  413. 15:30accused XAI of running turbines before
  414. 15:33it had the right permits, and the fight
  415. 15:35over air quality in South Memphis is
  416. 15:37still going.
  417. 15:38Shelby County has since granted a permit
  418. 15:40for turbines at the Memphis site, and
  419. 15:42that permit [music] is under appeal.
  420. 15:45But the logic is easy to follow. A
  421. 15:47turbine you can buy and truck in can be
  422. 15:49running long before a grid connection
  423. 15:51that might take years.
  424. 15:53Except [music] the turbines are running
  425. 15:55out, too.
  426. 15:56GE Vernova, one of the world's biggest
  427. 15:59turbine makers, is sold out of gas
  428. 16:01[music] turbines through 2028, and it
  429. 16:03expects data centers to drive about a
  430. 16:05third of its American [music] gas power
  431. 16:07deals.
  432. 16:08A builder who decides today that it
  433. 16:10wants its own gas plant joins a line
  434. 16:12stretching toward [music] the end of the
  435. 16:14decade. So, even going off the grid now
  436. 16:17means waiting years for equipment.
  437. 16:20So the biggest players are locking
  438. 16:21[music] up power years ahead anyway they
  439. 16:24can. Microsoft went back to a name
  440. 16:26almost everyone [music] knows. It signed
  441. 16:29a long-term deal that's bringing a
  442. 16:31shutdown reactor at Three Mile Island in
  443. 16:33Pennsylvania back to life. Now renamed
  444. 16:36the Crane Clean Energy Center with power
  445. 16:39aimed [music] for next year.
  446. 16:41Once it restarts, the reactor is
  447. 16:42expected to put about 835 [music]
  448. 16:45MW back on the grid. Google went a step
  449. 16:48further and bought an energy developer
  450. 16:50outright. It's parent company, Alphabet,
  451. 16:53paid $4.75 billion in cash for
  452. 16:56Intersect, a company that builds power
  453. 16:59plants, solar farms, and batteries right
  454. 17:01alongside data centers [music], starting
  455. 17:03with a shared site in Haskell County,
  456. 17:05Texas. That's about four times what Meta
  457. 17:08says it spent on the entire Temple
  458. 17:10[music] data center just to own the
  459. 17:12power side of the business. Meta took
  460. 17:14the idea further than anyone. In New
  461. 17:17Albany, Ohio, outside Columbus, it's
  462. 17:19bringing up Prometheus, its first
  463. 17:21gigawatt-class AI cluster with part of
  464. 17:24it already running. By outside
  465. 17:26estimates, the rebuilt Temple site runs
  466. 17:28on roughly 150 MW of computing power.
  467. 17:32Prometheus is designed for more than six
  468. 17:35times [music] that. To get it built
  469. 17:36faster, Meta put part of it inside giant
  470. 17:39weather-proof tents, aluminum frames
  471. 17:42wrapped in fabric that go up far quicker
  472. 17:44than a concrete hall.
  473. 17:46Mark Zuckerberg has said the tents are
  474. 17:47there for speed. Meta isn't relying on
  475. 17:50the grid alone, either. Filings show the
  476. 17:53site has its own power generation yard
  477. 17:55tied to the data halls by a wide utility
  478. 17:58corridor. And Ohio regulators approved
  479. 18:00two dedicated natural gas plants to feed
  480. 18:03the campus directly. Meta has also lined
  481. 18:06up nuclear power from Vistra,
  482. 18:08TerraPower, and Oaklo to back it. Some
  483. 18:10of it from reactors that haven't been
  484. 18:12built yet.
  485. 18:145 GW of computing power is what Meta
  486. 18:17wants from Hyperion, a campus it's
  487. 18:19building in Richland Parish, out in the
  488. 18:21rural northeast of Louisiana.
  489. 18:24Zuckerberg has said its footprint would
  490. 18:26cover a big [music] part of Manhattan.
  491. 18:28To pay for it, Meta set up a $27 billion
  492. 18:32joint venture [music] with the
  493. 18:33investment firm Blue Owl Capital, which
  494. 18:35keeps much of the debt off Meta's own
  495. 18:37books.
  496. 18:38To power it, Meta didn't just sign an
  497. 18:40electricity contract. Regulators had
  498. 18:43already approved new gas plants for the
  499. 18:45site.
  500. 18:46Then, this March, Meta and the utility
  501. 18:49Entergy announced [music] a much bigger
  502. 18:51deal.
  503. 18:52Hyperion is now set to be served by 10
  504. 18:55gas power plants, more than 7 gigawatts
  505. 18:58of capacity for a single customer.
  506. 19:01That's more than half of all the power
  507. 19:03Entergy currently generates across the
  508. 19:05entire state of Louisiana.
  509. 19:07It's a fleet of power plants big enough
  510. 19:09to reshape a state's grid. And the new
  511. 19:12[music] ones still need sign-off from
  512. 19:13Louisiana's regulators.
  513. 19:15That review is where [music] the real
  514. 19:17fight is, over who pays if plants built
  515. 19:20to run for decades outlive [music] the
  516. 19:22data center they were built for.
  517. 19:24Meta and Entergy say Meta will cover its
  518. 19:26own costs. Put it all together, and the
  519. 19:29biggest names in AI are turning into
  520. 19:31energy companies. They're paying for
  521. 19:33power plants, restarting reactors, and
  522. 19:36building their own supply, because the
  523. 19:38chips they've already paid for need
  524. 19:40somewhere to run.
  525. 19:42So, why not just make the chips use less
  526. 19:44power?
  527. 19:45If every chip did more work per watt,
  528. 19:48the same building could do more without
  529. 19:50a single new power plant. Nvidia says
  530. 19:53that's exactly what it's delivering.
  531. 19:55It claims Vera Rubin produces up to 10
  532. 19:58times more AI output per megawatt than
  533. 20:00Blackwell. In plain terms, a hall that
  534. 20:03used to answer one AI request could
  535. 20:05answer 10 on the same electricity.
  536. 20:08That's Nvidia's own number, measured on
  537. 20:10a workload Nvidia chose.
  538. 20:13But the direction is real, and Nvidia
  539. 20:15says that across six generations of its
  540. 20:17chips, AI output per megawatt has
  541. 20:20improved about a million times over.
  542. 20:22Vera Rubin itself shows what happens to
  543. 20:25those gains. It's far more efficient
  544. 20:27than Blackwell, and each of its racks
  545. 20:29still draws more power than a Blackwell
  546. 20:32rack, because Nvidia spends the savings
  547. 20:34on packing more work into every cabinet.
  548. 20:37The whole industry does the same thing,
  549. 20:39which is why the power bill keeps
  550. 20:41climbing. When AI output gets cheaper to
  551. 20:44produce, companies don't bank the
  552. 20:46savings and switch off the lights in the
  553. 20:48hall. They run bigger models, longer
  554. 20:50conversations, and more AI agents
  555. 20:53working around the clock. Britain went
  556. 20:55through the same thing with coal in the
  557. 20:571800s, when better steam engines made
  558. 20:59coal power cheaper, and the country
  559. 21:01ended up burning more of it. The
  560. 21:03International Energy Agency has put a
  561. 21:05figure on where that leads. By 2030, it
  562. 21:09expects global data center electricity
  563. 21:11use to more than double, with AI as the
  564. 21:13biggest driver [music]. That would be
  565. 21:15around 945 terawatt hours a year,
  566. 21:19[music] more than all of Japan uses
  567. 21:21today.
  568. 21:22That's the hardest obstacle of all,
  569. 21:25because more money doesn't make any of
  570. 21:27it arrive sooner.
  571. 21:28Nvidia brings out a new rack generation
  572. 21:31roughly every year, while every fix in
  573. 21:33this story, the transformers [music],
  574. 21:35the turbines, the reactors, takes years.
  575. 21:38Hyperion's newest power [music] plants
  576. 21:40are still waiting on regulators, and the
  577. 21:42Texas audit is still working through its
  578. 21:44list.
  579. 21:45The neighbors are pushing back, too.
  580. 21:48[music] In Temple, a local group has
  581. 21:49formed to fight more data center
  582. 21:51construction, worried about the heat
  583. 21:53these sites throw off. In Virginia,
  584. 21:55Texas, and Louisiana, regulators are now
  585. 21:58arguing over who should pay for the new
  586. 22:00power lines and plants, the tech
  587. 22:02companies or the families living nearby.
  588. 22:05Every one of those fights adds months to
  589. 22:07projects that were already running late.
  590. 22:10Temple is running now, liquid-cooled and
  591. 22:13full of AI hardware, on a design that
  592. 22:15didn't exist [music] when its first
  593. 22:17walls went up. But Kyber is coming, and
  594. 22:20every hall built around today's racks
  595. 22:22[music] will face the same question
  596. 22:23Temple did. For now, the physical world
  597. 22:26is a harder limit on AI than the chips
  598. 22:28are because Nvidia can design a new rack
  599. 22:31faster than anyone can build the hall,
  600. 22:33the substation, and the power plant to
  601. 22:35plug it into.
  602. 22:37The leaders are the companies that bring
  603. 22:38their own [music] power. xAI moves
  604. 22:41fastest, Meta has made the biggest bets,
  605. 22:44and Microsoft has warned its chips could
  606. 22:46end up waiting on buildings. What could
  607. 22:48change that is more [music] transformer
  608. 22:50and turbine factories and chips that cut
  609. 22:53the power each job needs. The one thing
  610. 22:55to watch is 2027 when Nvidia plans to
  611. 22:59ship the first Kyber racks and we find
  612. 23:01[music] out who has a hall ready to
  613. 23:03switch them on.
  614. 23:05If the power coming in can't grow fast
  615. 23:07enough, the savings have to come from
  616. 23:09inside the rack where GPUs burn energy
  617. 23:12passing data over copper. Chip makers
  618. 23:14from TSMC to Nvidia are racing to swap
  619. 23:17[music] that copper for light using tiny
  620. 23:20lasers beside the chip.
  621. 23:22Watch this video [music] next. It's on
  622. 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.