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

0:04 Right now, the world is in an AI race.

0:08 Companies, governments, universities

0:11 are all racing to build bigger models, smarter systems.

0:16 And behind the scenes,

0:17 they are racing to build more data centers to power AI.

0:22 But there's a problem.

0:24 We are running head first into the limits of our infrastructure.

0:28 The power grid includes all the infrastructure,

0:31 power plants, transmission lines and all

0:33 to generate and deliver power to our homes, our businesses,

0:37 and now to AI data centers.

0:40 In the United States,

0:41 the grid operators are reporting that new AI data center projects

0:46 are requesting power loads equal to entire cities.

0:50 In some regions, utilities simply can't keep up.

0:55 So when you hear “AI data center,” what comes to mind?

1:00 For many, it's one thing:

1:03 energy hogs.

1:05 And they are not wrong.

1:07 AI is dramatically accelerating the electricity demand of data centers.

1:12 Just training GPT-4

1:14 is estimated to have consumed around the annual electricity use

1:19 of thousands of US homes.

1:21 In another striking example, in Ireland,

1:24 nearly 20 percent of the nation's electricity

1:28 is drawn by data centers today.

1:32 And these are not just statistics.

1:35 They are also community stories.

1:37 In the data center alley in Virginia,

1:40 residents recently saw higher electricity bills,

1:44 20 percent higher already compared to just a few years ago,

1:48 as utilities scramble to serve massive new AI facilities.

1:53 So energy-hog label seems well deserved.

1:59 But that's only half the story.

2:01 Here is the new view.

2:04 These facilities are not just energy-hungry brains.

2:08 They can also be the muscles of the grid, flexing on demand.

2:14 Unlike our homes or hospitals,

2:16 AI data centers run jobs that are predictable,

2:21 controllable and often delayable.

2:24 That makes them ideal to help balance supply and demand on the grid.

2:30 By making AI data centers power-flexible,

2:34 we can connect them much more rapidly to the grid,

2:37 while at the same time making electricity more affordable and resilient.

2:44 What's more, the AI boom is arriving

2:48 just as the renewable boom is also taking off.

2:52 Wind and solar don't follow our schedules,

2:56 but data centers can.

2:58 Which means we can align the rise of AI

3:02 with the rise of clean energy,

3:04 if we are bold enough to rethink their role.

3:08 All this transformation to power flexibility

3:12 didn't just come out of thin air.

3:15 It builds on decades of research

3:18 on energy-efficient computing,

3:20 scheduling, optimization and many others.

3:25 I've lived this journey myself.

3:27 Early in my career,

3:29 I asked a question that many found unrealistic.

3:34 Could computer systems adapt their behavior

3:40 depending on power grid needs,

3:42 but without breaking their performance promise

3:46 to their users?

3:49 At the time, this sounded radical

3:51 because why would we ever design a system that would slow itself down

3:57 on purpose?

3:59 But then came the breakthroughs.

4:02 First, we discovered

4:04 not all computing tasks are urgent.

4:07 Some can wait for minutes or hours,

4:10 and some can be slowed down without anyone really noticing it.

4:15 For example,

4:17 a researcher analyzing hundreds of medical images with AI

4:22 may be OK with waiting just a little longer.

4:25 Or, if you are fine-tuning your AI model

4:28 over the course of the next few days,

4:30 you may be OK with slowing it down for just a few hours.

4:35 This inherent flexibility in computing

4:38 gives us the flexibility we need to manage power.

4:40 Second,

4:42 we reframed the problem.

4:45 Instead of asking

4:47 how do we compute as fast as possible,

4:50 we asked,

4:52 how do we make computer systems meet the constraints of the power grid,

4:57 while at the same time still delivering on user performance agreements?

5:02 This shift led to new strategies:

5:04 capping power,

5:06 shifting workloads

5:08 and provisioning the data center as a flexible reserve to the grid.

5:13 A key aspect here is that we do keep the performance promise to users,

5:18 so it's not arbitrary.

5:20 User experience remains as a key target.

5:24 And better yet, it becomes more predictable.

5:28 So we built prototypes on real data-center servers,

5:33 and they worked.

5:34 Systems that could follow a power target

5:37 while still delivering results.

5:40 But all this journey wasn't smooth.

5:42 There were paper rejections, funding rejections,

5:47 colleagues telling me this would never work.

5:51 Well, since I was a kid, I was told I'm a persistent person.

5:55 Perhaps stubborn at times.

5:58 And bold ideas require persistence

6:03 because change almost always looks impossible

6:07 before it looks obvious.

6:09 So you take that feedback, you reframe it again and again,

6:13 and you keep building.

6:15 You keep proving.

6:16 So what began as scribbles on a whiteboard 12 years ago,

6:21 is now running on real AI data centers.

6:25 Why does this matter now?

6:26 Because the power grids challenge

6:29 isn't just to generate more power.

6:32 It's about timing.

6:34 Solar gives us a glut of electricity at noon,

6:39 but demand might peak in the evening.

6:41 Wind might be abundant one day and scarce the next.

6:45 Nuclear takes decades and billions of dollars to build

6:51 and is often hard to locate in urban areas.

6:55 Batteries are critical,

6:57 but scaling them is costly, slow,

7:01 and often not environmentally clean.

7:03 Meanwhile, AI data centers themselves face five to seven-year wait times

7:10 just to connect to the grid

7:12 in places like Virginia.

7:14 In AI time,

7:15 where technologies shift in a major way every six months,

7:18 five to seven years is an eternity.

7:21 So here's the opportunity.

7:23 With the right orchestration,

7:25 AI data centers can be flexible today.

7:28 No waiting, no new massive power infrastructure construction.

7:33 They can soak up excess solar in the afternoon,

7:38 scale down at peak times

7:40 and act as virtual batteries today.

7:43 And the stakes are real.

7:44 Take Texas, August 23.

7:47 During a brutal heat wave,

7:50 the rising electricity demand pushed the grid to its limits.

7:55 Wholesale electricity prices spiked over 800 percent

8:00 in a single afternoon.

8:02 So flexible loads, if they were widely available,

8:06 could have reduced the costs

8:08 and could have prevented the emergency alerts that went to the consumers.

8:12 So we have two opportunities here.

8:14 One, we can make current data centers flexible

8:19 and help prevent blackouts

8:20 and reduce electricity costs.

8:23 Two, and perhaps the more significant,

8:26 by making future data centers power-flexible,

8:31 we can connect them much earlier

8:33 without waiting for major power grid upgrades.

8:36 If we ignore this opportunity,

8:40 we are not just wasting renewable energy

8:43 and we are not just raising our electricity bills.

8:46 We are also slowing AI adoption,

8:49 making it delayed,

8:51 more expensive and less accessible to society.

8:55 But there's a catch.

8:58 Orchestrating this flexibility is not easy.

9:02 Prices change hourly.

9:05 Workloads may arrive unpredictably.

9:08 Grid rules change across states, across countries.

9:12 So no human operator

9:14 and no single fixed data center management policy can keep up.

9:18 This is where AI itself comes back into the story.

9:23 The very technology driving this unforeseen demand

9:27 is also probably the only thing smart enough to tame it.

9:31 AI can learn patterns, anticipate grid needs

9:36 and coordinate across data centers, across utilities,

9:40 even nations in real time.

9:43 Imagine a data center

9:44 or a whole network of them,

9:46 as an orchestra,

9:48 with hundreds of instruments, all playing at once.

9:52 Left on their own, it can sound like chaos.

9:57 But bring in a conductor,

9:59 suddenly all that noise turns into music.

10:02 The conductor in this case is AI.

10:06 AI can direct data center operation

10:10 so that the data center can precisely match power constraints,

10:15 depending on what the grid needs, what power is available

10:19 and what users demand.

10:21 The result is harmony.

10:24 Reliable electricity, efficient computing

10:27 and a system that works beautifully together.

10:30 And that's exactly what we've built.

10:33 We built software that slows down, speeds up,

10:37 or pauses workloads in a data center,

10:40 or shifts workload among data centers.

10:43 Our conductor platform tunes performance and power at real time,

10:49 all the while respecting user and cloud-provider performance needs.

10:54 In this way, by flexing when needed,

10:57 we can connect AI data centers much faster to the grid.

11:03 Make better use of the available power in the power grid

11:07 and enable faster AI adoption.

11:11 I've been inside this story

11:12 from an idea that once seemed impossible

11:15 to prototypes in a lab,

11:17 to systems now running in the field,

11:19 and I believe this is just the beginning.

11:21 AI is already reshaping how we compute,

11:24 but it could also reshape how we power the world.

11:28 So the question isn't how much energy AI consumes.

11:33 The real question is how much flexibility, resilience

11:38 and clean power can AI unlock?

11:41 If we are bold enough to rethink AI data centers,

11:44 the very machines that now seem like a burden

11:48 could be our greatest assets

11:50 in building a sustainable AI future.

11:54 Thanks.

11:55 (Applause)