Energy & AI

“A single Bitcoin transaction takes 77 kilowatt[-hours], which is what it takes to power one average American house for a week… it’s an insane amount of energy. So that’s where we are with Bitcoin.” — 00:08:45

“I’m honestly bullish on proof of work. I can imagine the future when there’s solar panels that are powering the heating of my water heater, and all those things are doing is hashing. That’s the future we could almost imagine.” — 00:16:42

“The biggest cost in a data center right now is not the cloud computing itself, it’s cooling of the systems… you have concentration of heat… in order to cool they really need to use enormous amounts of commercial power… The thing [is], when you decentralize systems, you effectively decentralize energy consumption itself.” — 00:39:06

“Texas has an unregulated grid and there’s a lot of overage compute power that comes from renewables in Texas. Folks that have infrastructure that can utilize this overage from renewables can effectively run data centers for profit, because most of the cost in a data center is cost for cooling — which is electricity, not the actual compute.” — 00:56:05

“This whole promise of decentralized systems is: it looks inward and it looks to be more efficient with usage, instead of outward and expansionary. I think that’s important.” — 00:20:48

“My mission in life is to create a more equitable computing platform, my mission… to create a more greener planet. That’s the kind of stuff that I want to focus on: infrastructure.” — 01:06:42

“Anyone that has capacity that’s not being used — it’s already burning carbon, so why not sell that capacity? It’s about improving efficiency of underutilized resources, so it’s inherently better for the planet. Whereas most commercial cloud providers are external, they’re expansionary — they look to acquire more compute, more chips.” — 00:04:22

“A lot of times when you have a server that’s sitting out there that’s turned on, the carbon emissions is not just the server but the cooling systems and the whole ecosystem around keeping those lights on. Instead of trying to be expansionary to get more power, we’re looking to be more reusable — we’re looking inwards to open up what we have already deployed and make that more usable.” — 00:02:55

“I look at decentralization as a way to improve global energy efficiency for computational resources, just pure and simple… hyperscalers are building new data centers to expand… whereas web3 and decentralization is inward-looking at what we have already deployed… web3 is greener for me. Like, web3 is about efficiency and improving the global energy usage efficiency.” — 00:30:16

“Web3 is greener for me — like, web3 is about efficiency and improving the global energy usage efficiency.” — 00:00:00

“The way I would describe compute is essentially a logical abstraction of energy. It is transforming energy, electricity, to logic, and there are different layers of that abstraction as we get higher towards this energy being consumed by a user.” — 00:03:43

“I read a stat somewhere that said about 5% of global GDP will be spent on AI compute… I think in 2030 or something, and it’s believable the way AI is going. It’s going to need a lot more chips and we don’t have capacity as humanity to produce [that] level of chips — the only way we can achieve that level of usage is by looking inward, [at] what we have right now that’s not being utilized.” — 01:00:34

“The real bottleneck for scalability is not compute at a scale, it’s really the power… and power is heavily regulated in the United States… decentralization of power is very needed.” — 00:07:57

“We’re not stopping at compute. I’m touring power facilities… it’s way more challenging to decentralize power than compute, but it’s not out of the reach… There was an attempt at creating a power market by a company called Enron — they messed up — but I think there’s another opportunity to create a truly decentralized power market.” — 00:48:01

“The energy needs to catch up… I think the prediction now is by 2030, 20% of US electricity production will be consumed by AI. So we don’t have infrastructure to even power that power requirement — we need to build nuclear reactors.” — 00:15:59

“GPT-4 was trained I believe on a 100 megawatt capacity data center, and that grows to about [2]50 megawatt capacity for the newer model… and the next tier is 1.2 gigawatt. When you get to 1.2 gigawatt, that means it’s a nuclear reactor — nuclear reactor produces 1 gigawatt. We can’t do solar… we just don’t have enough storage capacity for solar. So it is going to be nuclear, and Microsoft just bought a nuclear reactor in Pennsylvania.” — 00:17:23

“Grok 3… in Memphis data center takes about 50 megawatts, and if you ask me where they got 50 megawatts, they’re actually burning diesel… The next model is going to be 250 megawatt and a model after that is going to be 1.2 gigawatt. The only place to get a 1.2 gigawatt capacity is a nuclear reactor.” — 00:11:37

“The notion is that every year the amount of power we need is growing by half an order [of] magnitude… Grok was [a] 50 megawatt [data] center — you don’t have 50 megawatt supply in a single source, so they literally are burning diesel… the next model is going to be 250 megawatt, and the model after that, the next year, is going to be 1.2 gigawatt… to get anything beyond a gigawatt you need a nuclear reactor.” — 00:39:56

“The energy needs are just going to be so high — by 2030, AI is going to consume 20% of [the] US electric grid, and that’s a given.” — 00:43:31

“We’re going to really reduce the price of electricity to a point that [it’s] going to be too cheap to meter — I think that’s going to be a reality. I’m really excited about that because I want to see human civilization reach Type 1 status in my lifetime.” — 00:45:41

“Especially with AI, the need for electricity and compute — I don’t think we’ll be able to even fulfill with solar or any of the sustainable methods. I think we’re going to have to go nuclear for powering AI.” — 00:18:10

“As LLMs get larger and larger and larger, the amount of energy they consume also goes in proportion to the size of the LLM that’s being trained… the growth of energy needed every year is about half an order magnitude — that means every year we are [5]x-ing the amount of energy we need to develop the state-of-the-art models for that year.” — 00:21:01

“The next big LLM will need [a] 250 megawatt data center, and a year after that is going to be 1.2 gigawatt data center, and the only place to get 1.2 gigawatts of capacity is a nuclear reactor. So in two years we’re going to need a new nuclear reactor to train the most powerful state-of-the-art model, and I think that is a huge bottleneck we don’t talk much about.” — 00:22:29

“We don’t pay enough attention to the energy bottlenecks. We just keep assuming these models are going to get larger and larger, but where do you get the power from? The only place you can get the power is China… only China can build a nuclear reactor in three years — they’ve done it — and the US is not that competitive anymore, sadly.” — 00:28:55

“Elon Musk is literally burning diesel because you cannot get energy anywhere to power 50 megawatt data center. And next year, 26, we’re going to need 250 megawatt data center, right? To make the better GPT, like, whatever, 5 or 6. And a year after that, we’re going to need 1.2 gigawatt data center.” — 00:29:23

“I think, like, AI is hitting a wall with energy. Some say it’s data, but I don’t believe it’s data. I think there’s a lot of data out there with the right incentives we can unlock the data, but we cannot get more power with the right incentives because the infrastructure doesn’t exist in the United States… the energy needs are growing at half an [order of] magnitude, about 5x every year.” — 00:28:01

“I’m building a 60 kilowatt capacity in my house, they can power, like, an H100 cluster, I mean, I think all of us are going to just be producing energy locally and having a lot of compute locally that will finally fulfill our dream of sovereignty because you need energy and you need compute.” — 00:36:54

“You look at what technologies can enable model training advancement, and if you look around, only decentralization — distributed training — is the only way out. That’s why, in two years I think we’re going to cap out as to how much you can do in a single data center.” — 00:19:12

“Using synthetic data and using mixture of experts mechanism you can actually solve the data problem, but what we cannot solve is the energy problem. That’s why it’s very, very important, if you’re doing training, to focus on distributing your training runs versus trying to go with the traditional mechanism of centralizing your training runs… because we’re going to hit a cap in two years and we have no solutions.” — 00:55:53

“I’m extremely skeptical if we are able to move as fast as we want in the hyperscaler market, but I think we have a bigger chance to move in the second-tier market, which is the one megawatt to 10 megawatt.” — 00:20:24

“The rate at which the energy needs are evolving for training state-of-the-art models is about doubling every two years right now… Grok 3 is reportedly, or believed to be, trained on a 35 megawatt capacity, and this year we’ll need about 70 megawatt capacity to train the next generation, I believe Grok 4.” — 00:10:07

“There are 600 of them in the US — we know exactly where they are… if there’s an adversary in the future, all they’re going to do is send EMP bombs to these data centers in drones… you’ve got to decentralize the grid to have a more resilient world. Instead of having a single place that can be attacked, I would have 10,000 if not a million of these places that generate energy, sharing energy in a peer-to-peer manner.” — 00:34:22

“In a couple years you’re going to need nuclear reactors, and there’s no way in hell you can power through coal, through onshore windmills, offshore windmills, hydro, geothermal… they’re not going to cut it. By a pessimistic estimate we need about 8% of global energy spent on AI; optimistic estimates, about 2%. It doesn’t matter what estimate you take, it’s still an enormous energy draw from the grid.” — 00:15:09

“Each prompt, each call takes about 2.9 watt hours… 2025 we’re going to need about half a gigawatt capacity to serve AI… and every year it’s growing at 37% — the energy needs for AI, that includes inference and training.” — 00:14:28

“In my lifetime I want to see humanity reach Kardashev scale type one civilization… a civilization that harnesses all the energy of its home planet. There’s a lot of energy that comes to our planet — we just don’t harness that because we don’t have incentives to do [it], and the economics don’t play out really well… but with this it does.” — 00:29:29

“The Colossus xAI data center in Memphis has about 150 megawatt capacity, has about 200,000 H100 chips… it’s drawing about 7 megawatts from the grid… They’re burning gas… The problem is so bad in Memphis they have a huge pollution problem — it’s causing elevated issues of asthma, respiratory diseases… there’s a lawsuit that’s looming.” — 00:10:51

“It’s reasonable to expect AI will grow at least as fast as mobile and internet… in the next 15 years we’re going to see an increase from five to 150 queries a day [per user]… we are going to literally grow by 37% year-over-year… the next 15 years, the world will consume 2% — this is a conservative estimate — of global energy supply on AI… Gartner or any of the bigger firms say it’s 8%… Even 2% is an additional 100 nuclear reactors we need.” — 00:22:44

“China is the only country that solved their energy problem, because they overbuilt for the last decade and they actually have a terawatt installed capacity — that’s four times more than the US… They can build a nuclear reactor in 3 years… unless you’re in China, you cannot scale AI. So that’s a big challenge for America, national security. China is going to out-accelerate us because they have the energy. So people will start talking about this in 2 years, and you heard here first.” — 00:17:00

“In the US we have Gen 1 and Gen 2 combined, about 96 reactors, and they’re utilized fully 93% of the time. The last one available was Three Mile Island — that got swept up by Microsoft for 20 years. So there are no nuclear reactors that are free anymore… you can’t build new ones, because the last one we built took about 14 years and 34 billion dollars.” — 00:09:07

“If you look at the amount of installed capacity needed to train state-of-the-art models… it’s doubling every 2 years. So Grok-3 roughly took about 25, 35 megawatt capacity to train… right now the state of the art is 35 MW; in [2026] it will be 70 MW, and [in] ‘28 we’re looking at 900 MW… It’s nearly impossible to get over 10 MW capacity in America.” — 00:07:39

“The xAI data center in Memphis, owned by Elon Musk, is only able to get 7 MW from the grid, but it’s 150 MW data center… they’re literally burning [liquefied natural] gas to power the data centers, and that’s causing environmental concerns in South Memphis… we’re at a point that we have to burn fossil to make AI. And that’s not sustainable… it’s doubling every 2 years.” — 00:14:52

“If we can finance small modular data centers — we’re talking about 35 kilowatt to 50 kilowatt data centers, cost you anywhere from $2 to $3 million… if you can power them using the solar panels, you can effectively create an alternative to these hyperscale data centers. And… by 2040, with about 2 million of these, we can solve global energy problems.” — 00:31:16

“Today we are looking at… a billion users… doing about five queries a day, five prompts a day. And that’s projected to grow… to about 150 prompts a day. And that’s very conservative… The AI energy need alone is, I believe, growing at around 37% year over year.” — 00:19:48

“We cannot build nuclear reactors fast enough to demand. So by 2030, or even sooner, maybe, we will run out of energy to train new models. So AI will not evolve in terms of how big it can get in terms of models by 2030 or sooner.” — 00:16:41

“If you can actually train using distributed clusters, that means we can have distributed clusters in places that can have about a hundred kilowatt capacity. Now, how can we get a hundred kilowatt capacity? Solar… a 4,000 square foot roof can get you a hundred kilowatt capacity.” — 00:27:55

“The data center [that trained] Grok is based in Memphis, Tennessee. And that is currently drawing about 150 megawatts… and it’s only drawing about seven megawatts from the grid. So the rest of the energy, they’re actually burning… LNG to power, because in America, it’s nearly impossible now to get anything over 10 megawatts.” — 00:12:57

“We’ll consume about 8% of global energy just to power — and that’s a big chunk, 8% is a lot — and we’re talking about the West, right? Most of the energy today is produced in the East, in China really. China has about 8 terawatt install capacity, US has about 2 terawatt install capacity.” — 00:13:52

“If you’re making about 10 cents a kilowatt hour, using Akash you’re literally selling energy for… a dollar fifty, because you’re adding GPUs on top of it. Think of GPUs as a wrapper on energy in terms of income.” — 00:32:40

“The bottom line is we cannot meet the energy demand in next six years. The only way to solve the problem is to distribute it.” — 00:15:18

“That’s going to double in two years, 300 megawatt, double in two more years, you know, 300, 600 and 1.2. So you can see, in the next six years the only way you’re going to power a training run is using a nuclear reactor, because a nuclear reactor is the only one that can produce over a gigawatt capacity.” — 00:12:24

“The xAI data center in Memphis is only deriving about 7 megawatts from the grid. The rest of, like, the 143 megawatt comes from burning the gas. We’re talking about Elon Musk burning gas.” — 00:16:00

“Department of Energy report recently said 12%, and DOE is known to take a very conservative approach and they always miss their estimation. So when DOE says 12% of US energy will be consumed on AI by 2028, I think in reality it’ll be 30% by 2030.” — 01:51:13

“In 2023, US data centers consumed 176 terawatt hours of electricity. That’s about 4.4% of the national grid — more than double what they used just 5 years ago. And the curve isn’t slowing down. By 2028 the number could reach 12%, more than every household in California combined. The problem isn’t silicon, it’s supply.” — 04:27:34

“Considering the energy need, particularly concentrated energy need for training, is doubling every two years, the expectation by 2030 the only way to train the state-of-the-art model is with a gigawatt supply. So nuclear will be the only way to do so. Unfortunately, the last nuclear reactor we built in America took about 14 years.” — 01:39:34

“Workloads can migrate to where the power is greenest, cheapest, and most available. When it’s sunny in Arizona, we train there. When wind peaks in Kansas, we infer there. This isn’t a small optimization. It is a fundamental shift. It turns compute from static burden into a dynamic asset. Instead of straining the grid, it balances it. This is the beginning of a new symbiosis between AI and clean energy.” — 04:40:10

“Capturing energy is one thing, but transmitting the energy is very hard… a lot of times you may see a source but it may not be big enough to power a data center… but trust me, right now with decentralization of energy and compute along with it, a lot of these untapped sources like geothermal or any of the sources that maybe have small production capability will be used a lot more, and that’s the vision that we’re going after when it comes to decentralized compute and energy.” — 02:03:40

“Every leap forward is sparked by a new kind of fire… Now AI is igniting a new era. To reach Kardashev level one, we need 10 quadrillion watts. That’s about 500 times our current global energy use… We’re not just building better, smarter software. We’re building the backbone of a new civilization, one that’s built not just on data, but on decentralized energy and compute.” — 04:23:53

“If you look at the math, by 2028, 12% of US electricity will be used for AI — that’s from a government DOE report — and the reality is about 30%. So every state now has an opportunity to become an energy powerhouse. And I think if you’re not innovating in your policy, you’re going to be left behind.” — 04:52:28

“Gartner predicts by 2026, 40% of all new AI data centers may not be able to get enough power to run. That’s not a future risk. It’s a near-term constraint. In this new era, watts are the new silicon. Power isn’t just an input. It’s the bottleneck and the battleground.” — 04:28:19

“Workloads can migrate in real time to where the power is the cheapest, cleanest, or most available. When it’s sunny in Arizona, train there. When the wind is up in Kansas, infer there. This turns compute from a burden into a flexible grid-balancing asset. It’s [the] beginning of a new symbiosis between AI and energy.” — 00:11:16

“According to Department of Energy, by 2028, data centers could use as much as 12% of US electricity. That’s more than every home in California combined. We’re not just seeing growth, we’re seeing a surge. Data centers now represent 44% of all projected new load on the grid.” — 00:03:01

“Every great leap is lit by a new kind of fire. Steam ushered in the industrial age. Electricity powered cities. The internet connected the minds… Now AI is lighting a new kind of fire. To reach [Kardashev] level one, we need 10 quadrillion watts. That’s about 500 times our global current energy use.” — 00:00:04

“By 2026, Gartner warns that 40% of all new AI data centers may be unable to secure sufficient power to operate. Some utilities are already pushing back — in parts of Texas and Virginia, projects are being paused or denied outright. This isn’t because the chips aren’t ready. It’s because the circuits aren’t. We’re facing a hard truth: in the age of AI, watts are the new silicon.” — 00:04:34

“A type one civilization doesn’t emerge from one cloud provider. It emerges from a network of millions — coordinated, open, and alive… Let’s build a cloud that reflects our values. A cloud that can’t be shut down. A cloud we own together. Let’s meet at Kardashev 1.” — 00:19:24

“In 2023, US data centers consumed 176 terawatt hours — about 4.4% of the entire national electricity output. And the Department of Energy now warns that by 2028 that could be more than triple, to as much as 12% of the grid… AI isn’t just a user of compute, it’s a devourer of energy.” — 00:01:28

“Workloads can migrate to where the power is greenest, cheapest, and most available. When it’s sunny in Arizona, we train there. When wind peaks in Kansas, we infer there. This isn’t a small optimization. It is a fundamental shift. It turns compute from static burden into a dynamic asset. Instead of straining the grid, it balances it.” — 00:16:05

“When you decentralize compute, it becomes mobile. It can move with energy. Imagine AI training jobs chasing solar peaks across Arizona rooftops, inference pipelines migrating dynamically with Kansas wind, GPUs spinning up where renewables are cheap and idling when they’re not. Instead of compute demanding power, power attracts compute. This flips the entire model.” — 00:14:15

“To reach Kardashev level one, we need 10 quadrillion watts. That’s about 500 times our current global energy use. That scale of ambition we are stepping into.” — 00:00:03

“One GB200 NVL72 rack draws around 120 kW and contains 72 of the world’s most advanced AI chips. To build one gigawatt AI cloud, we need around 8,300 of these racks. But here’s the key: they don’t have to live in one facility. They can live everywhere — distributed across regions, climates, and use cases.” — 00:01:26

“In 2023, US data centers consumed 176 terawatt hours of electricity. That’s about 4.4 [percent] of [the] national grid… By 2028, the number could reach 12% — more than every household in California combined… Gartner predicts by 2026, 40% of all new AI data centers may not be able to get enough power to run. That’s not a future risk. It’s a near-term constraint.” — 00:03:42

“In this new era, watts are the new silicon. Power isn’t just an input. It’s the bottleneck and the battleground.” — 00:04:25

“Some data came out last week actually that most data centers now have a 10-year interconnection queue… which means I can start constructing it [in] 2035. That’s not going to work for where we’re headed. And so this intersection of distributed energy, distributed compute is I think the next generation of how we build infrastructure.” — 00:11:27

“I’m aligned with you there. I mean, Department of Energy report recently said 12%, and DOE is known to take a very conservative approach and they always miss their estimation. So when DOE says 12% of US energy will be consumed on AI by 2028, I think in reality it’ll be 30% by 2030.” — 00:19:11

“Considering the energy need — particularly concentrated energy need for training — is doubling every two years, so the expectation by 2030 the only way to train the state-of-the-art model is with a gigawatt supply. So nuclear will be the only way to do so. Unfortunately, the last nuclear reactor we built in America took about 14 years, and the interconnect requests now average anywhere from 10 years here for new data center builds.” — 00:07:25

“But also there is this Jevons paradox, right — the more efficient something gets, the more usage it gets. I mean, I always want my ChatGPT to go faster… but I’m able to do a lot more things that I couldn’t.” — 00:14:50

“What are your thoughts on decentralization in terms of energy — so small micro grids producing or capturing energy, maybe even consuming at the edge, and interacting in a peer-to-peer manner, maybe contributing back to a larger hub?” — 00:16:59

“What are your general thoughts on renewables through solar versus small modular reactors that can produce 24-hour continuous power? As I know solar is great for residential, but when you’re doing utility scale — I mean 1.2 gigawatt — you’re looking at essentially a nuclear reactor.” — 00:04:24

“In 2024, US data centers use 200 terawatt hours of electricity, enough to power Thailand for a year. The same estimate holds that by 2028, AI power usage is predicted to reach between 165 and 326 terawatt hours annually, enough to power 22% of US households.” — 00:00:46

“Chinese are building a gigawatt of solar every two days. I mean, that’s equivalent to one nuclear reactor every two days right now. And they’re very, very good at it. We’re not doing that. And we can’t do that. We’re not Chinese, right? What we can do is decentralization. That we are very good at.” — 00:44:00

“They’ve taken a very conservative estimate and said, hey, at the minimum, I think we’re going to do, like, 12%. Reality, I think people project anywhere from 30% to 35%, and it’s actually resonating.” — 00:06:28

“If we can figure out how to decentralize or instead of bringing the power to the AI, let the AI go to the power source. So if you have wind in Texas during the evening, let’s go train an AI model there. Because now you have asynchronous. You can pretty much go anywhere you want.” — 00:24:57

“The 4.5 gigawatt facility in Texas, if you burn fossil, because there’s no other way, will generate about 2.5 million to 2.7 million tons of carbon every year. That’s more than two-thirds of all the CO2 emissions done by Vermont in 2023. It’s almost as much as a U.S. state emitting. We’re talking about one company.” — 00:10:19

“Elon said they want to build… They want to bring about 50 million GPUs by next five years. H100 equivalents… H100 takes about a kilowatt in energy. You can do the math… So the predictions from Department of Energy, I think, are very, very conservative.” — 00:08:45

“The cheapest is solar and a home and wind. Well, every house has a rooftop. We just don’t use a rooftop. Like in Texas, a typical ranch house has 4,000 square feet of rooftop. That generates about 80 kilowatt energy for six hours a day.” — 00:28:40

“The last nuclear reactor took about 14 years in the U.S…. We have about 96 reactors in America. They’re fully utilized 93% of the time. The last remaining reactor was the Three Mile Island in Pennsylvania that was scooped up by Microsoft and a 20-year lease. So there are no free reactors in America.” — 00:07:15

“The conservative estimates amount to about 500 plus, I think 508 [terawatt] hours by 2028. For context, that’s about 12% of US energy consumption for AI… it’s an incredible jump in terms of percentage consumption.” — 00:02:11

“The US grid grew about 1.2% in the last 20 years… Gartner predicts by 2026 — that’s next year — about 40% of AI data centers will be out of power, because we will be forced to a position where we have to make a decision as to who gets the energy: is it the homes or is it going to be data centers?” — 00:02:56

“The amount of energy we need to train state-of-the-art models… is doubling every two years. Right now the OpenAI data center that’s operational in Texas draws about 300 megawatts of capacity. So doubling that means by 2030 we’ll need a nuclear reactor at the minimum to train the state-of-the-art model. So ChatGPT 6 or 7 is going to be a nuclear-powered data center.” — 00:11:00

“Why burn energy but instead use that energy to train AI? Where it’s windy in Kansas, move the AI workload there. When it’s sunny in California, move the AI workload. The AI workload [has] to be elastic and has to be asynchronous… Akash, we are developing an energy-aware [scheduler] that will pick the lowest energy. So we’ll start advertising the cost per energy now.” — 00:46:49

“It’s not slowing down, the demand for GPUs and the demand for energy. Now if we don’t do something about it, we will start seeing rolling blackouts by next year. We will start seeing an energy crisis similar to the oil crisis in the 1970s, or even much larger scale.” — 00:12:27

“4.5 gigawatt is about five nuclear facilities equivalent. There’s no way in hell they’re getting that from nuclear… That much amount of LNG will produce anywhere between 2.5 to 2.7 million tons of net new carbon into the environment, and that’s equivalent to about 2/3 of all the CO2 emissions from Vermont.” — 00:08:04

“One of the worst things they did was to ban all fossil vehicles from 2030 without thinking about how they’re going to upgrade the grid. Along with the AI challenge, now the EV challenge is catching up. So states that have EV mandates are going to suffer the most with AI… I think it’s just going to get worse and worse, the energy problem.” — 00:32:33

“If you study the grid as much as I’ve studied, you would want to go full sovereign. Like I’m building a house in Texas that’s completely off-grid, because it’s so fragile… everything is in a little balance; a little extra thing will destroy the grid.” — 00:24:01

“The amount of energy we need is doubling every two years… the problem is so pervasive now that Congress invited me to testify on energy, and I did in May. The amount of energy is doubling to a point where we’ll need a nuclear reactor by 2030 to train the state-of-the-art AI.” — 00:18:26

“Three trillion dollars worth of investment went into AI, mostly buying chips and building data centers… Bloomberg literally wrote a report saying that in Baltimore, where you have data centers, the cost of energy went up by 280% in five years — that’s because of AI centralization… I don’t see any other way that this industry can come together without decentralization.” — 00:39:18

“In my testimony I’ve given before Congress, I made one claim: AI is becoming the substrate of civilization, and AI is a foundational layer. But foundations only work if the power stays on. We are approaching an energy bottleneck that could limit who gets to build, ship, and even run AI. We can fix that by distributing where, when, and how we compute.” — 00:00:43

“The US Department of Energy report estimates US data centers consumed about 176 terawatt hours in 2023 — about 4.4% of US electricity. By 2028, the range is 325 to 580 terawatt hours, or 6.7% to 12% of US electricity. At the high end, that’s 1 in 8 watts on the grid going to data centers… Analysts are warning by 2026, 40% of AI data centers could be stuck waiting for power, even as some facilities spin up diesel at peak to stay online.” — 00:02:11

“Because the fabric is location flexible, schedulers can follow the sun and the wind — route PyTorch jobs to green hotspots at 11:00 a.m., shift to night wind regions at 2 a.m. Treated as flexible demand, AI becomes a grid asset, not a burden, absorbing surplus and backing off during peaks. That’s how we scale sustainably.” — 00:17:00

“Now imagine scaling this with small modular data centers — containerized racks on campuses, in offices, or community sites parked next to solar, wind or hydro. Compute rides on local clean energy, slashing transmission losses, soaking up daytime solar peaks and spinning down at night. Instead of one 300 megawatt site, deploy [thousands of small] pods. Puzzle pieces, not mega projects.” — 00:16:15

“Just yesterday, I believe, multiple data centers in San Jose are running out of power. They’re shut down. This was predicted by Gartner a few years ago — they said by 2026, 40% of data centers will go out of power, without energy… because we cannot build energy infrastructure as fast as we can build data centers.” — 00:03:37

“The last [US nuclear reactor] we built took about 14 years. Nuclear is great but we cannot build them fast enough… The SMRs, or small modular reactors, are still not market ready, and fusion is 10 years away. So we have the biggest nuclear reactor right above our roofs, right in the sky, called the sun. If we can leverage solar effectively we can solve a lot of problems — and the way we do that is homes.” — 00:10:08

“A grid, if you break it down, is just a bunch of copper wires… nobody wants new copper wires… digging the copper wires underground is extremely expensive and complicated. So it’s very hard to move energy… So the only really — I mean, real practical — way to solve the problem is to move the workload to where the energy is.” — 00:04:23

“It’s not like we don’t have enough energy. We have a lot of energy — it’s just we cannot transmit [it] well… What do I do with excess energy? I can train AI with excess energy. So AI becomes now a load balancer for energy, which is a great solution to the power problem, because the grid won’t want to take your energy… the grid is very fragile.” — 00:16:03

“For the most part, the data centers that AI demanded in ‘24 was satisfied by crypto companies, like it or not. At some point we’ll start seeing crypto and AI and all these technologies just [as] energy distributors with load balancers.” — 00:01:40

“GPUs were just the tip of the iceberg. Now we have memory, now we have storage, now you have energy, which is a big unknown. I mean, it’s clear to me, I think, how we can solve the energy problem, but it’s unknown for a lot of people.” — 00:12:30

“I don’t think [the IEA] has taken into account what Peter did or [OpenClaw] did, in terms of energy usage. I think that’s going to cause a lot more problems. I think that’s why I’m super bullish on distributed training.” — 03:20:49

“2024 took about 415 terawatt [hours] of capacity for energy… The International Energy Association predicts their base case predictions about 930 terawatt hours by 2030.” — 03:09:26

“That’s why [CoreWeave] crashed — because they couldn’t get more energy. It’s not that they can’t get chips. Chips are getting easier and easier… models are smaller and yada yada yada. Energy is not.” — 03:10:10

“A nuclear reactor from permit to power takes about 15 years — last one we built was 14 years… This energy doesn’t move as fast as data centers do, and Elon was very clear… we cannot build more energy on earth, we got to go to space.” — 03:06:35

“No one’s talking about one problem, and that’s energy… I think that’s a dark horse for AI. Why? Because we have about $2 trillion in commitments, sovereign commitments. Most of that money is to develop energy infrastructure. But AI moves in months. Energy doesn’t. I mean, to get a transformer today it’s about seven years.” — 03:05:50

“At this rate… they will be out of reliability standards by June 2027 — that means the grid is no longer reliable… The energy cost increased by 200% for Northern Virginia residents.” — 03:10:10

“They’re producing about a gigawatt of solar every 36 hours. So, one nuclear reactor equivalent of solar, just solar, every 36 hours. And that’s the scale China is operating at. US is not going to catch up. I think it’ll be impossible for us to even catch up.” — 00:28:37

“Most of them are residential use for solar. But if you can extend the residential use to AI training, you win.” — 00:29:21

“The amount of energy required to train state-of-the-art models is doubling roughly every 2 years… by 2030 we’ll need about a gigawatt power install capacity to train the state-of-the-art AI if this trend continues, which looks like it’s continuing. Only thing that can provide reliably a gigawatt power is a nuclear reactor.” — 00:17:09

“Our copper wire grid is extremely fragile… you can’t produce energy and transmit it elsewhere even though you have solar at home. But if you put a GPU on top of it, now you abstracted energy — you abstracted the energy in a GPU that you can pass around.” — 02:22:21

“Look at AI data centers alone: first we saw a chip shortage, then we are seeing a memory shortage — memory has gone up four times the cost… then we’re starting to see energy shortage. We can’t even buy a transformer without waiting four years; we can’t buy a turbine without waiting 14 years. The amount of stress on resources — anything AI touches — is just unbelievable.” — 02:12:26

“One way to look at Bitcoin is a global energy load balancer… when you have excess energy you can use that energy to mine bitcoin, and when you want energy you can use that bitcoin to get more energy. Same concept extending to AI: when you have excess energy, you trade using AI.” — 02:23:04

“Yesterday OpenAI cancelled their plans for new data center buildouts. Why? Because they cannot get energy. It is happening right now… we’re going to see a lot of cancellations, and I’m afraid all the big leg up we got with GDP growth in the last quarter is going to be cancelled this quarter, because you’re going to find out building data centers is not as easy as it seemed to be.” — 02:13:50

“It’s very, very hard to get anything above 100 megawatts right now in terms of data centers… training scale is impossible because of energy… [Nvidia] literally quadrupled their supply chain. Energy doesn’t work that way — there is a four-year lead time for transformers right now, like four years, and about a 14-year lead time for the turbines and generators.” — 00:56:53

“Homes are essentially mini power plants if you really have solar… there are homes where there’s great wind, there are homes where there’s great sun, there’s homes where there’s great geothermal… and there are homes everywhere. So why not use these homes as power plants for AI?” — 00:35:41

“In space you have near-continuous power for solar — you can follow the sun, always be on the sunny side, you don’t need batteries. The economics don’t make sense yet, but they’re starting to make sense as space gets cheaper and cheaper. In like three to four years, maybe in five years, you can see space data centers becoming a real thing if energy is not solved on Earth.” — 00:59:05

“The challenge in solar is storage and transmission… our grid is not designed to be peer-to-peer — it’s designed to be hubbed. That means you generate energy in one place and you distribute the energy. It’s never designed to generate energy at the edge and distribute back to the grid.” — 00:53:22

“We have a gigantic nuclear reactor in the sky called the sun… it just doesn’t make sense to me that we don’t use most of it — most of it goes to waste — while we are building such infrastructure to compensate for our lack of motivation to use the sun’s energy. I’m a big proponent of renewables… my testimony before Congress was about leveraging renewables to train AI.” — 00:51:15

“When we said GPUs are going to [be] scarce, we also called out energy scarcity — energy is going to be scarce. And now we’re seeing Elon Musk going on stage and going on podcasts talking about energy. I testified before Congress about energy last year… I feel like every resource that AI touches is getting scarce right now.” — 00:15:11

“The lead time for a turbine today is 14 years. The lead time for a transformer is 4 years… They paused the build-out because they can’t get transformers.” — 00:27:49

“I testified before Congress about this last year talking about the energy crisis that’s coming and my testimony included an innovative approach as a solution where we want to do distributed data centers.” — 00:29:14

“I think by 2028 or 29, we’ll need a gigawatt data center to train the latest AI. A gigawatt data center is a nuclear reactor… And we know that’s not sustainable. We can’t build nuclear reactors fast enough.” — 00:31:25

“The reality today is most data center projects that were announced by OpenAI, as part of the Stargate program, by CoreWeave, several other players that announced building large data center projects are all stalled because of local community pushback, primarily because of energy. So we’re in a crisis beyond historical proportions.” — 00:03:16

“To remind folks, 2024, I think overall data center usage was around 415 terawatt hours. And that was expected to grow to around 980 terawatt hours by 2030. It’s about two and a half times growth.” — 00:02:23