Transcriptly

The Story You’re Not Hearing About AI Data Centers | Ayșe Coskun | TED

11:57Added Aug 30, 2026

The race to build smarter AI is crashing into a physical limitation: the power grid simply can't keep up with the energy demands of data centers. Computer sc...

Watch on YouTube →
Contributed by 刘嘉琪

Transcript

Transcript format
  1. 00:04Right now, the world is in an AI race.
  2. 00:08Companies, governments, universities
  3. 00:11are all racing to build bigger models, smarter systems.
  4. 00:16And behind the scenes,
  5. 00:17they are racing to build more data centers to power AI.
  6. 00:22But there's a problem.
  7. 00:24We are running head first into the limits of our infrastructure.
  8. 00:28The power grid includes all the infrastructure,
  9. 00:31power plants, transmission lines and all
  10. 00:33to generate and deliver power to our homes, our businesses,
  11. 00:37and now to AI data centers.
  12. 00:40In the United States,
  13. 00:41the grid operators are reporting that new AI data center projects
  14. 00:46are requesting power loads equal to entire cities.
  15. 00:50In some regions, utilities simply can't keep up.
  16. 00:55So when you hear “AI data center,” what comes to mind?
  17. 01:00For many, it's one thing:
  18. 01:03energy hogs.
  19. 01:05And they are not wrong.
  20. 01:07AI is dramatically accelerating the electricity demand of data centers.
  21. 01:12Just training GPT-4
  22. 01:14is estimated to have consumed around the annual electricity use
  23. 01:19of thousands of US homes.
  24. 01:21In another striking example, in Ireland,
  25. 01:24nearly 20 percent of the nation's electricity
  26. 01:28is drawn by data centers today.
  27. 01:32And these are not just statistics.
  28. 01:35They are also community stories.
  29. 01:37In the data center alley in Virginia,
  30. 01:40residents recently saw higher electricity bills,
  31. 01:4420 percent higher already compared to just a few years ago,
  32. 01:48as utilities scramble to serve massive new AI facilities.
  33. 01:53So energy-hog label seems well deserved.
  34. 01:59But that's only half the story.
  35. 02:01Here is the new view.
  36. 02:04These facilities are not just energy-hungry brains.
  37. 02:08They can also be the muscles of the grid, flexing on demand.
  38. 02:14Unlike our homes or hospitals,
  39. 02:16AI data centers run jobs that are predictable,
  40. 02:21controllable and often delayable.
  41. 02:24That makes them ideal to help balance supply and demand on the grid.
  42. 02:30By making AI data centers power-flexible,
  43. 02:34we can connect them much more rapidly to the grid,
  44. 02:37while at the same time making electricity more affordable and resilient.
  45. 02:44What's more, the AI boom is arriving
  46. 02:48just as the renewable boom is also taking off.
  47. 02:52Wind and solar don't follow our schedules,
  48. 02:56but data centers can.
  49. 02:58Which means we can align the rise of AI
  50. 03:02with the rise of clean energy,
  51. 03:04if we are bold enough to rethink their role.
  52. 03:08All this transformation to power flexibility
  53. 03:12didn't just come out of thin air.
  54. 03:15It builds on decades of research
  55. 03:18on energy-efficient computing,
  56. 03:20scheduling, optimization and many others.
  57. 03:25I've lived this journey myself.
  58. 03:27Early in my career,
  59. 03:29I asked a question that many found unrealistic.
  60. 03:34Could computer systems adapt their behavior
  61. 03:40depending on power grid needs,
  62. 03:42but without breaking their performance promise
  63. 03:46to their users?
  64. 03:49At the time, this sounded radical
  65. 03:51because why would we ever design a system that would slow itself down
  66. 03:57on purpose?
  67. 03:59But then came the breakthroughs.
  68. 04:02First, we discovered
  69. 04:04not all computing tasks are urgent.
  70. 04:07Some can wait for minutes or hours,
  71. 04:10and some can be slowed down without anyone really noticing it.
  72. 04:15For example,
  73. 04:17a researcher analyzing hundreds of medical images with AI
  74. 04:22may be OK with waiting just a little longer.
  75. 04:25Or, if you are fine-tuning your AI model
  76. 04:28over the course of the next few days,
  77. 04:30you may be OK with slowing it down for just a few hours.
  78. 04:35This inherent flexibility in computing
  79. 04:38gives us the flexibility we need to manage power.
  80. 04:40Second,
  81. 04:42we reframed the problem.
  82. 04:45Instead of asking
  83. 04:47how do we compute as fast as possible,
  84. 04:50we asked,
  85. 04:52how do we make computer systems meet the constraints of the power grid,
  86. 04:57while at the same time still delivering on user performance agreements?
  87. 05:02This shift led to new strategies:
  88. 05:04capping power,
  89. 05:06shifting workloads
  90. 05:08and provisioning the data center as a flexible reserve to the grid.
  91. 05:13A key aspect here is that we do keep the performance promise to users,
  92. 05:18so it's not arbitrary.
  93. 05:20User experience remains as a key target.
  94. 05:24And better yet, it becomes more predictable.
  95. 05:28So we built prototypes on real data-center servers,
  96. 05:33and they worked.
  97. 05:34Systems that could follow a power target
  98. 05:37while still delivering results.
  99. 05:40But all this journey wasn't smooth.
  100. 05:42There were paper rejections, funding rejections,
  101. 05:47colleagues telling me this would never work.
  102. 05:51Well, since I was a kid, I was told I'm a persistent person.
  103. 05:55Perhaps stubborn at times.
  104. 05:58And bold ideas require persistence
  105. 06:03because change almost always looks impossible
  106. 06:07before it looks obvious.
  107. 06:09So you take that feedback, you reframe it again and again,
  108. 06:13and you keep building.
  109. 06:15You keep proving.
  110. 06:16So what began as scribbles on a whiteboard 12 years ago,
  111. 06:21is now running on real AI data centers.
  112. 06:25Why does this matter now?
  113. 06:26Because the power grids challenge
  114. 06:29isn't just to generate more power.
  115. 06:32It's about timing.
  116. 06:34Solar gives us a glut of electricity at noon,
  117. 06:39but demand might peak in the evening.
  118. 06:41Wind might be abundant one day and scarce the next.
  119. 06:45Nuclear takes decades and billions of dollars to build
  120. 06:51and is often hard to locate in urban areas.
  121. 06:55Batteries are critical,
  122. 06:57but scaling them is costly, slow,
  123. 07:01and often not environmentally clean.
  124. 07:03Meanwhile, AI data centers themselves face five to seven-year wait times
  125. 07:10just to connect to the grid
  126. 07:12in places like Virginia.
  127. 07:14In AI time,
  128. 07:15where technologies shift in a major way every six months,
  129. 07:18five to seven years is an eternity.
  130. 07:21So here's the opportunity.
  131. 07:23With the right orchestration,
  132. 07:25AI data centers can be flexible today.
  133. 07:28No waiting, no new massive power infrastructure construction.
  134. 07:33They can soak up excess solar in the afternoon,
  135. 07:38scale down at peak times
  136. 07:40and act as virtual batteries today.
  137. 07:43And the stakes are real.
  138. 07:44Take Texas, August 23.
  139. 07:47During a brutal heat wave,
  140. 07:50the rising electricity demand pushed the grid to its limits.
  141. 07:55Wholesale electricity prices spiked over 800 percent
  142. 08:00in a single afternoon.
  143. 08:02So flexible loads, if they were widely available,
  144. 08:06could have reduced the costs
  145. 08:08and could have prevented the emergency alerts that went to the consumers.
  146. 08:12So we have two opportunities here.
  147. 08:14One, we can make current data centers flexible
  148. 08:19and help prevent blackouts
  149. 08:20and reduce electricity costs.
  150. 08:23Two, and perhaps the more significant,
  151. 08:26by making future data centers power-flexible,
  152. 08:31we can connect them much earlier
  153. 08:33without waiting for major power grid upgrades.
  154. 08:36If we ignore this opportunity,
  155. 08:40we are not just wasting renewable energy
  156. 08:43and we are not just raising our electricity bills.
  157. 08:46We are also slowing AI adoption,
  158. 08:49making it delayed,
  159. 08:51more expensive and less accessible to society.
  160. 08:55But there's a catch.
  161. 08:58Orchestrating this flexibility is not easy.
  162. 09:02Prices change hourly.
  163. 09:05Workloads may arrive unpredictably.
  164. 09:08Grid rules change across states, across countries.
  165. 09:12So no human operator
  166. 09:14and no single fixed data center management policy can keep up.
  167. 09:18This is where AI itself comes back into the story.
  168. 09:23The very technology driving this unforeseen demand
  169. 09:27is also probably the only thing smart enough to tame it.
  170. 09:31AI can learn patterns, anticipate grid needs
  171. 09:36and coordinate across data centers, across utilities,
  172. 09:40even nations in real time.
  173. 09:43Imagine a data center
  174. 09:44or a whole network of them,
  175. 09:46as an orchestra,
  176. 09:48with hundreds of instruments, all playing at once.
  177. 09:52Left on their own, it can sound like chaos.
  178. 09:57But bring in a conductor,
  179. 09:59suddenly all that noise turns into music.
  180. 10:02The conductor in this case is AI.
  181. 10:06AI can direct data center operation
  182. 10:10so that the data center can precisely match power constraints,
  183. 10:15depending on what the grid needs, what power is available
  184. 10:19and what users demand.
  185. 10:21The result is harmony.
  186. 10:24Reliable electricity, efficient computing
  187. 10:27and a system that works beautifully together.
  188. 10:30And that's exactly what we've built.
  189. 10:33We built software that slows down, speeds up,
  190. 10:37or pauses workloads in a data center,
  191. 10:40or shifts workload among data centers.
  192. 10:43Our conductor platform tunes performance and power at real time,
  193. 10:49all the while respecting user and cloud-provider performance needs.
  194. 10:54In this way, by flexing when needed,
  195. 10:57we can connect AI data centers much faster to the grid.
  196. 11:03Make better use of the available power in the power grid
  197. 11:07and enable faster AI adoption.
  198. 11:11I've been inside this story
  199. 11:12from an idea that once seemed impossible
  200. 11:15to prototypes in a lab,
  201. 11:17to systems now running in the field,
  202. 11:19and I believe this is just the beginning.
  203. 11:21AI is already reshaping how we compute,
  204. 11:24but it could also reshape how we power the world.
  205. 11:28So the question isn't how much energy AI consumes.
  206. 11:33The real question is how much flexibility, resilience
  207. 11:38and clean power can AI unlock?
  208. 11:41If we are bold enough to rethink AI data centers,
  209. 11:44the very machines that now seem like a burden
  210. 11:48could be our greatest assets
  211. 11:50in building a sustainable AI future.
  212. 11:54Thanks.
  213. 11:55(Applause)