Latest / DX Today | No-Hype Podcast & News About AI & DX / China Plugs a 2 Gigawatt AI Campus Into Wind and Solar

Transcript

China Plugs a 2 Gigawatt AI Campus Into Wind and Solar · DX Today | No-Hype Podcast & News About AI & DX

• 0:00
  SPEAKER_01: Welcome to the DX Today Podcast, your daily deep dive into the AI ecosystem.
• 0:05
  SPEAKER_01: I'm Brian, and joining me as always is Jessica.
• 0:08
  SPEAKER_00: Thanks, Brian.
• 0:09
  SPEAKER_00: Today we are talking about electricity, which sounds boring until you realize it may decide who wins the AI race.
• 0:16
  SPEAKER_00: China is quietly wiring its data centers straight into wind farms and solar fields, and the contrast with America is striking.
• 0:23
  SPEAKER_01: I like that we are starting with power, because in every system I have ever run, the fancy software was never the bottleneck.
• 0:31
  SPEAKER_01: It was always the boring stuff underneath, like cooling, cables, and whether the utility would actually deliver the megawatts it promised.
• 0:38
  SPEAKER_00: Exactly.
• 0:39
  SPEAKER_00: And the boring stuff just got very expensive.
• 0:42
  SPEAKER_00: A new report from Inside Climate News, published this morning, lays out how China is using its enormous wind, solar, and battery industries to feed AI computing, while the United States leans harder on natural gas.
• 0:55
  SPEAKER_01: Let's make it concrete before we go big picture.
• 0:58
  SPEAKER_01: What is actually running out there right now, today, rather than sitting in a government planning document or a glossy press release about the future?
• 1:05
  SPEAKER_00: The flagship is a campus run by Envision Energy, which is the second largest wind turbine supplier in the world.
• 1:12
  SPEAKER_00: In August, it switched on a data center campus in Inner Mongolia that gets its electricity directly from wind and solar, backed by energy storage.
• 1:20
  SPEAKER_01: And how big are we talking?
• 1:22
  SPEAKER_01: Because the phrase AI data center now covers everything from a converted warehouse with a few racks to something that needs its own power plant and its own zip code.
• 1:31
  SPEAKER_00: This one is the second kind.
• 1:33
  SPEAKER_00: Envision calls it the Galaxy Campus.
• 1:35
  SPEAKER_00: It is designed for two gigawatts.
• 1:37
  SPEAKER_00: And the company says it could support around 1 million AI accelerators when fully built, which would make it a contender for the largest standalone AI data center anywhere.
• 1:46
  SPEAKER_01: Two gigawatts is a number people throw around casually now, but for context, that is roughly the output of two large nuclear reactors dedicated to running chips.
• 1:56
  SPEAKER_01: That is not a server room, that is a small industrial city that happens to think for a living.
• 2:01
  SPEAKER_00: And here is the part that gets the attention of every chief financial officer.
• 2:05
  SPEAKER_00: Envision says customers can lock in electricity at roughly 20% below grid prices, and they can hold that rate for a full decade, which is an eternity in energy markets.
• 2:14
  SPEAKER_01: That is the line that would make me lean forward in a budget meeting.
• 2:18
  SPEAKER_01: A 10-year fixed price that undercuts the grid is not a sustainability story, it is a cost of goods story.
• 2:24
  SPEAKER_01: Power is one of the biggest line items for anyone serving AI models at scale.
• 2:29
  SPEAKER_00: Right.
• 2:30
  SPEAKER_00: And the general manager of the data center unit, Ricky Jang, put it bluntly when they commissioned it.
• 2:35
  SPEAKER_00: He said the next frontier of AI is infrastructure.
• 2:39
  SPEAKER_00: Envision also has a wider plan called Mission Gobi, targeting five gigawatts of green AI computing in deserts worldwide by 2030.
• 2:47
  SPEAKER_01: Okay, so that is one company.
• 2:50
  SPEAKER_01: My natural skepticism says a single showcase campus does not make a national strategy.
• 2:55
  SPEAKER_01: What is Beijing actually committing to?
• 2:58
  SPEAKER_01: And are the numbers big enough to matter beyond a nice photo opportunity?
• 3:02
  SPEAKER_00: They are big.
• 3:03
  SPEAKER_00: China wants to expand its AI computing capacity more than sixfold by 2030.
• 3:08
  SPEAKER_00: Its computing centers used about 170 terawatt hours of electricity in 2025, which was around 1.6% of national use.
• 3:17
  SPEAKER_01: And let me guess where that goes.
• 3:19
  SPEAKER_01: Every one of these projections has a line that goes up and to the right.
• 3:22
  SPEAKER_01: And the only real question is how steep the line gets and who ends up paying for the slope.
• 3:27
  SPEAKER_00: Steep.
• 3:28
  SPEAKER_00: Beijing estimates roughly 800 terawatt hours by 2030, which would be about 6% of the entire country's electricity.
• 3:36
  SPEAKER_00: And the goal is for more than 80% of the power for new data centers in the major computing hubs to come from renewables.
• 3:42
  SPEAKER_01: 80% is an aggressive target.
• 3:45
  SPEAKER_01: In my experience, aggressive targets are where projects go to die quietly.
• 3:50
  SPEAKER_01: Is there any track record suggesting China can actually build renewable capacity fast enough to hit something like that?
• 3:56
  SPEAKER_00: This is where your skepticism runs into the data, Brian.
• 3:59
  SPEAKER_00: From 2021 through 2025, China's annual renewable generation ran about 20% above its own targets.
• 4:07
  SPEAKER_00: They are now planning to grow from about 4,000 terawatt hours of renewables to roughly 6,000.
• 4:13
  SPEAKER_01: Fair enough.
• 4:14
  SPEAKER_01: I will take that one on the chin.
• 4:15
  SPEAKER_01: Overshooting your own targets for five straight years is not a press release.
• 4:19
  SPEAKER_01: That is a pattern.
• 4:21
  SPEAKER_01: Give me the comparison number that really shows the gap between the two countries.
• 4:25
  SPEAKER_00: Here it is.
• 4:26
  SPEAKER_00: In 2025, China added an estimated 540 gigawatts of new generating capacity, and more than 80% of it was renewable.
• 4:35
  SPEAKER_00: The United States added 54 gigawatts, with about 60% of that from renewables.
• 4:40
  SPEAKER_01: So China added roughly 10 times as much new capacity in a single year.
• 4:45
  SPEAKER_01: That is the number I would put on the first slide, because it reframes the whole AI competition.
• 4:50
  SPEAKER_01: It is not only a chip race and a model race, it is a power plant race.
• 4:55
  SPEAKER_00: And that is precisely the argument Kyle Chan made.
• 4:58
  SPEAKER_00: He is a fellow at the Brookings Institution's China Center, and he testified to the House Select Committee on China in April.
• 5:05
  SPEAKER_00: His view is that America leads in large AI models and advanced chips, while China leads in building power plants quickly and at scale.
• 5:13
  SPEAKER_01: I appreciate that framing because it is honest about both sides.
• 5:17
  SPEAKER_01: Nobody gets to declare total victory.
• 5:19
  SPEAKER_01: America has the better chips and the frontier labs, and China has the shovels, the turbines, and the factories that make the batteries.
• 5:26
  SPEAKER_00: Chan also made a point I found sharp.
• 5:29
  SPEAKER_00: He said China does not want to choose between a faster data center build-out and its energy transition.
• 5:35
  SPEAKER_00: In other words, Beijing is trying to get both at once, speed and clean power, instead of trading one away for the other.
• 5:41
  SPEAKER_01: Now flip it around.
• 5:43
  SPEAKER_01: What is the American approach right now?
• 5:45
  SPEAKER_01: Because from where I sit, the answer seems to be that we will take any electron we can find as fast as we can find it and worry about the rest later.
• 5:53
  SPEAKER_00: That is a fair summary.
• 5:55
  SPEAKER_00: The Trump administration dropped a requirement that AI data centers on federal land be paired with clean energy.
• 6:02
  SPEAKER_00: And the consultancy Wood Mackenzie projects the United States will add between 13 and 16 gigawatts of new gas-fired capacity every year from 2026 through 2035.
• 6:13
  SPEAKER_01: And how does that compare to what we were doing before the AI boom?
• 6:17
  SPEAKER_01: Because gas plants are not new, and utilities have always built some of them.
• 6:21
  SPEAKER_01: The question is whether this is business as usual or a real shift.
• 6:25
  SPEAKER_00: It is a real shift.
• 6:27
  SPEAKER_00: That range is more than triple the average annual gas additions from 2023 through 2025.
• 6:33
  SPEAKER_00: And a Sierra Club analysis, reported by Bloomberg, projects that emissions from new off-grid data centers could reach the level of Mexico's total annual emissions by 2030.
• 6:43
  SPEAKER_01: I want to be fair to the American side, though.
• 6:45
  SPEAKER_01: Gas is dispatchable.
• 6:47
  SPEAKER_01: It runs when the wind stops and the sun goes down.
• 6:50
  SPEAKER_01: If you are an operator with a customer paying for guaranteed uptime, gas is the safe answer, and I have been that operator.
• 6:58
  SPEAKER_00: Totally fair.
• 6:59
  SPEAKER_00: And that is exactly why the storage piece matters so much in the Chinese model.
• 7:03
  SPEAKER_00: China is the dominant manufacturer of batteries, so it can pair solar and wind with storage cheaply.
• 7:09
  SPEAKER_00: Kyle Chan pointed directly to solar construction and battery manufacturing as the reasons China is building faster.
• 7:15
  SPEAKER_01: So the real advantage is not just sunshine in the desert, it is owning the entire supply chain for the hardware that makes renewable power reliable enough for a data center.
• 7:25
  SPEAKER_01: That is a manufacturing moat, not merely a geography moat.
• 7:28
  SPEAKER_00: Right.
• 7:29
  SPEAKER_00: But geography helps too.
• 7:32
  SPEAKER_00: China launched an initiative in 2022 called East Data, West Computing, which pushes computing infrastructure to the less crowded western regions with abundant wind and sun.
• 7:43
  SPEAKER_00: Inner Mongolia is reported to hold more than half of China's wind resources.
• 7:47
  SPEAKER_01: And it is huge.
• 7:48
  SPEAKER_01: If I remember the article correctly, Inner Mongolia is about 1.7 times the size of Texas, with one of the lowest population densities in the country.
• 7:58
  SPEAKER_01: That means almost nobody lives next door to a two-gigawatt campus.
• 8:02
  SPEAKER_00: Which brings us to the human side.
• 8:04
  SPEAKER_00: And to me, this is the most underrated part of the whole story.
• 8:07
  SPEAKER_00: A Washington Post, ABC News, and Ipsos poll found that 72% of Americans oppose building data centers in their own communities.
• 8:16
  SPEAKER_01: 72% is not a split electorate.
• 8:19
  SPEAKER_01: That is close to a consensus, and in American politics, consensus on anything is rare.
• 8:24
  SPEAKER_01: Where is that opposition showing up in real life beyond a number in a national poll?
• 8:29
  SPEAKER_00: Just this weekend in Louisville, Kentucky, a group of local residents marched to the home of the mayor.
• 8:34
  SPEAKER_00: They were protesting a proposed hyperscale data center campus near their neighborhood.
• 8:39
  SPEAKER_00: And rising electricity bills are now becoming a midterm election issue around the country.
• 8:43
  SPEAKER_01: That is the part engineers tend to underestimate.
• 8:47
  SPEAKER_01: You can have the land, the chips, and the financing all lined up and still lose a year or two to zoning hearings and lawsuits.
• 8:54
  SPEAKER_01: Permitting delays are a real constraint, every bit as real as a transformer shortage.
• 8:59
  SPEAKER_00: And David Fishman, a principal at the Lantau Group, which advises on Asian power markets, explained why China does not see the same backlash.
• 9:07
  SPEAKER_00: His point was that data center growth there has not pushed up retail power prices, and many sites are far from big cities.
• 9:14
  SPEAKER_01: He had a line in there that I thought was refreshingly candid.
• 9:17
  SPEAKER_01: He said if it does not affect your costs and it does not affect your power prices, what is there to worry about?
• 9:23
  SPEAKER_01: That is basically the entire politics of infrastructure in one sentence.
• 9:27
  SPEAKER_00: He also said China can enjoy the luxury of insisting on green data centers precisely because it is so good at building conventional power assets, too.
• 9:36
  SPEAKER_00: That is the second-order effect people miss.
• 9:39
  SPEAKER_00: A strong grid lets you be picky.
• 9:41
  SPEAKER_00: A strained grid forces you to take whatever is available.
• 9:44
  SPEAKER_01: Let me play the skeptic one more time, because that is my job.
• 9:47
  SPEAKER_01: Isn't some of this just a story China likes to tell?
• 9:50
  SPEAKER_01: Remote desert data centers have real problems, like latency to users, water for cooling, and the challenge of getting skilled staff to live out there.
• 9:58
  SPEAKER_00: Those are real issues, and nobody serious says it is a solved problem.
• 10:02
  SPEAKER_00: But for training large models, latency matters much less than it does for serving consumers.
• 10:08
  SPEAKER_00: You can train in the desert and serve from the coast.
• 10:11
  SPEAKER_00: Training is where the gigawatts really go.
• 10:14
  SPEAKER_01: That is a good distinction, and it is one I would want every infrastructure planner to internalize.
• 10:20
  SPEAKER_01: Not all AI workloads need to live in the same place.
• 10:23
  SPEAKER_01: Training is a batch job that wants cheap, steady power, while inference wants to be close to the people asking questions.
• 10:30
  SPEAKER_00: And that split creates a strategic opportunity for American companies too.
• 10:35
  SPEAKER_00: Nothing stops United States firms from building training campuses near large wind and solar resources in places like West Texas or the Great Plains, where land is cheap and neighbors are few.
• 10:46
  SPEAKER_01: So what actually changes for people listening on Monday morning?
• 10:50
  SPEAKER_01: If I am running infrastructure at a mid-sized company or advising one, what is the practical takeaway from a story about wind farms in Inner Mongolia?
• 10:58
  SPEAKER_00: First, energy price is becoming a core part of your AI cost model, not a footnote.
• 11:04
  SPEAKER_00: When you compare cloud regions or providers, ask where the power comes from and whether the price is locked.
• 11:10
  SPEAKER_00: A 20% difference in electricity compounds fast at scale.
• 11:13
  SPEAKER_01: Second, I would add, take community relations seriously from day one.
• 11:18
  SPEAKER_01: If you are planning any facility, even a modest one, the old playbook of quietly buying land and announcing later is dying.
• 11:26
  SPEAKER_01: 72% opposition means you need neighbors on your side early.
• 11:30
  SPEAKER_00: And third, watch the policy signals closely.
• 11:33
  SPEAKER_00: The decision to drop clean energy pairing on federal land tells you the current approach in Washington is speed first.
• 11:40
  SPEAKER_00: That could change with the next election, and long-lived infrastructure has to survive several of those.
• 11:45
  SPEAKER_01: I think the bigger lesson is that the AI race is turning into an industrial race.
• 11:49
  SPEAKER_01: The winners will not just have the best models, they will have the cheapest, most reliable electrons, and they will have communities that are willing to host them.
• 11:58
  SPEAKER_00: Kyle Chan said it well.
• 12:00
  SPEAKER_00: Both countries are trying to play to their strengths and address their own weaknesses.
• 12:04
  SPEAKER_00: The open question for China is whether its renewable build-out can keep pace with its computing ambitions and for America, whether it can expand supply without a voter revolt.
• 12:13
  SPEAKER_01: That is a great place to land it.
• 12:16
  SPEAKER_01: Power, people, and policy are now as important as parameters.
• 12:20
  SPEAKER_01: Thanks, Jessica.
• 12:21
  SPEAKER_01: This one changed how I think about the competition, and I suspect it will for a lot of our listeners as well.
• 12:26
  SPEAKER_00: Thanks, Brian.
• 12:27
  SPEAKER_00: If you're building anything with AI this year, remember to follow the electricity.
• 12:32
  SPEAKER_00: It will tell you more about the future of this industry than almost any benchmark, product launch, or funding announcement you will read about this week.
• 12:39
  SPEAKER_01: That's all for today's episode of the DX Today podcast.
• 12:42
  SPEAKER_01: Thanks for listening, and we'll see you next time.