Pod Insights Summation with Auren Hoffman The LM Brief: The Energy Bottleneck Behind AI’s Growth About this episode This week’s World of DaaS LM Brief explores Big Data Wire's article on how electricity, not data or compute, is emerging as the ultimate constraint on AI progress. As global data center demand surges, performance per watt efficiency and access to reliable power are becoming the new competitive edge. Some nations and companies are pursuing nuclear-powered computation hubs to secure a strategic energy advantage. Listen to this short podcast summary, powered by NotebookLM.
Transcript Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.
0:01 Auren Hoffman? Hello, fellow data nerds. Welcome to the LM Brief. I'm your host, Oren Hoffman, CEO of Incubate and GP of Flex Capital. In the LM Brief, we'll bring you quick, high-value summaries of some of the most interesting DAS-themed articles, podcasts, and ideas, which is powered by Google's Notebook LM. Discover more episodes, get weekly data-as-a-service news, original content, and articles on data and more at worldofdas.com. That's worldofdas.com. Doc is in.
0:33 Speaker 01 So if you've been following the exclusive growth of AI, you probably know the basic recipe, right? Chips, data, and increasingly, electricity.
0:43 Speaker 02 That's right.
0:45 Speaker 01 We've all been watching the chip raise for years, seems like forever, but digging into the sources you sent over, there's this huge blinking warning light. Electricity, it's fast becoming the real bottleneck, the fundamental ingredient that's actually hardest to scale up.
1:00 Speaker 02 That really is the core tension, isn't it? We're kind of moving past the hype now and looking at these frankly staggering projections, especially from groups like the International Energy Agency, the IEA.
1:11 Speaker 01 Yeah. Okay, so let's get into it. Let's unpack the numbers that are ringing these alarm bells.
1:16 Speaker 02 Yes.
1:16 Speaker 01 Right, when we talk scale, we need the actual figures. What are the hard numbers the IEA is putting out for data centers right now?
1:23 Speaker 02 Okay, so they're starting with a global baseline for 2024, and it's approximately 415 terawatt-hours, TWH of electricity.
1:31 Speaker 01 The 415 TWH, okay, that sounds like a lot already.
1:34 Speaker 02 It is a lot. And get this, that figure already reflects consumption growing about 12% per year for the last five years, but it's the projection that's really jarring.
1:44 Speaker 01 Right, the forward-looking part. What do they see coming?
1:46 Speaker 02 That 415 TWH, they expect it to nearly double in just six years.
1:50 Speaker 01 Double?
1:50 Speaker 02 Yeah, rocketing up to around 945 TWH by 2030.
1:54 Speaker 01 Wow, okay, 945 TWH. To give people a sense of that scale, what does that compare to?
2:00 Speaker 02 Well, think about it like this. 945 TWH.
2:05 Speaker 02 That's not actually far off the current annual electricity use of a whole industrialized country, like, say, Germany.
2:12 Speaker 01 Germany uses around 500 TWH a year, right?
2:15 Speaker 02 Roughly, yeah. So we're talking about a brand-new economic sector demanding energy on par with a major European economy, almost double Germany's current use, just for data centers globally.
2:26 Speaker 01 Okay, that puts it in perspective. Beryl's comment makes perfect sense, then. Data center demand could genuinely rival entire country's usage.
2:33 Speaker 02 And you can see where the pressure points are emerging. The sources break it down regionally. The US, for example, is projected to account for nearly half of that global growth in electricity demand from data centers.
2:43 Speaker 01 Half, just in the US?
2:44 Speaker 02 Yeah, and in Japan. Data centers could account for over half of their total national demand growth. Even places like Malaysia, rapidly growing, data centers might be a fifth of all new electricity demand.
2:55 Speaker 01 That just sounds unsustainable without some massive changes. Huge investments needed, I guess.
3:00 Speaker 02 Exactly. Huge, immediate investment in generation, and, just as importantly, grid upgrades.
3:06 Speaker 01 Let's zoom in on those US numbers, then, because half the growth is, well, it's huge. What was the starting point?
3:12 Speaker 02 The 2023 baseline in the US was around 176 TWH. That was already about 4.4% of total national electricity demand.
3:21 Speaker 01 Okay, 4.4%. And by 2030?
3:23 Speaker 02 The projection suggests US consumption could approach 800 TWH.
3:28 Speaker 01 800? From 176, that's more than quadrupling in, what, seven years?
3:33 Speaker 02 Yeah, and think about what that means for the grid. To support that kind of jump, the US would need over 80 gigawatts of extra capacity.
3:40 Speaker 01 80 gigawatts? That's like building dozens of new, large power plants, just for data centers, in just a few years.
3:47 Speaker 02 That's the scale of the challenge. And if you narrow it down even further, just looking at the AI part, the training, the inference for these big models. Deloitte's projections are even more pointed. Power demand just from AI data centers could climb from about 4 gigawatt this year, 2024, something like 123 gigawatt by 2035.
4:04 Speaker 01 Whoa, 123 gigawatts, just for AI?
4:07 Speaker 02 It's an immense number. You can see why utility companies and energy regulators are suddenly, you know, having sleepless nights about generative AI. It's crossed over from being a tech problem to a physical infrastructure problem. A big one.
4:21 Speaker 01 It's easy to get stuck on the terawatts and gigawatts, these huge abstract numbers. But maybe we should break down why. Why are these AI systems just so incredibly power hungry? What's actually happening?
4:32 Speaker 02 Yeah, good question. We can sort of break down the energy draw into three main pieces, and they're all linked. First, the actual computation. Second, getting rid of the heat the computation generates, so cooling. And third, the sheer intensity, the sustained nature of the workload.
4:48 Speaker 01 Okay, let's start with computation. That's the obvious one, the chips doing the work.
4:52 Speaker 02 Right. Specialized chips, GPUs, TPUs, all those accelerators, they're the main engine. They're performing just billions, trillions of calculations a second. They need a lot of power, often running near their peak.
5:02 Speaker 01 Is there like a simple way to picture that?
5:05 Speaker 02 Well, Google actually gave an example, trying to make it tangible. They estimated that an average text prompt sent to their Gemini apps uses about 0.24 watt hours of electricity.
5:16 Speaker 01 0.24 watt hours, okay, that sounds tiny, like completely negligible.
5:20 Speaker 02 Exactly, and that's the catch. It is tiny for one prompt.
5:24 Speaker 01 Ah.
5:24 Speaker 02 But then you realize, Google handles, what, billions of searches daily?
5:29 Speaker 01 Yeah.
5:29 Speaker 02 And you scale that up, millions of users, making millions, maybe billions of prompts. That tiny fraction of a watt hour, multiplied globally, running 247, that's how you get to terawatt hours.
5:41 Speaker 01 It's the sheer volume, the scale, makes the small unit cost massive in aggregate.
5:46 Speaker 02 Precisely, it's consumption by volume, absolutely.
5:48 Speaker 01 Okay, so that's the compute cost. But then physics kicks in. All that energy use creates heat. This feels like maybe the biggest physical limit, right? Thermodynamics.
5:58 Speaker 02 It really is.
5:59 Speaker 01 Yeah.
5:59 Speaker 02 The forces are very clear on this. These GPUs, they basically turn electricity directly into heat. It's incredibly inefficient from that perspective.
6:06 Speaker 01 So you spend energy doing the calculation, and then you immediately have to spend more energy just to get rid of the waste heat.
6:11 Speaker 02 That's the cycle. You need massive cooling systems, pumps, chillers, huge air handlers running nonstop. And they consume significant power, too, just managing the thermal load.
6:22 Speaker 01 How much power are we talking about, like per rack of servers?
6:25 Speaker 02 The comparison in the sources is stark. An older server rack might draw a few kilowatts. A single rack today, filled with modern AI accelerators, that can easily consume 30 to 50 kilowatts.
6:37 Speaker 01 30 to 50. Per rack.
6:38 Speaker 02 Yes, and that's kind of the baseline now. High-density AI racks are already pushing 80 kilowatts. Some designs are even targeting 120 kilowatts or more per rack.
6:48 Speaker 01 120 kilowatts for one cabinet of gear. You need serious power infrastructure just for that, let alone the cooling.
6:55 Speaker 02 Exactly. You might have, say, tens of thousands of dollars of chips in that rack, but you need a comparable investment in the power and cooling infrastructure just to keep it running.
7:03 Speaker 01 And the third factor was intensity. How does that differ from older data centers?
7:08 Speaker 02 Think of old workloads as having peaks and valleys. You know, processing web requests, database queries, usage spikes, then drops.
7:15 Speaker 01 Right.
7:15 Speaker 02 Modern AI workloads, especially training these huge models, they're like marathon runners, not sprinters. They operate close to full capacity, flat out, for days, sometimes even weeks at a time.
7:26 Speaker 01 Constant sustained pressure.
7:28 Speaker 02 Exactly. Constant pressure on power delivery, constant pressure on cooling systems. It means energy efficiency isn't just about saving money on the electricity bill anymore. It's fundamental. Can the system even operate reliably at this scale without literally melting down?
7:43 Speaker 01 Okay, that constant pressure, that intensity, leads us right to the next point. The sources mention this performance per watt paradox, which sounds counterintuitive. If the new chips, like Nvidia's H200, are more efficient, shouldn't consumption be slowing down?
7:59 Speaker 02 You'd think so, right?
8:00 Speaker 01 Yeah.
8:00 Speaker 02 But the problem is, the efficiency gains per chip are being completely swamped by the sheer growth in ambition.
8:06 Speaker 01 How so?
8:07 Speaker 02 Well, you give developers a more efficient chip, and what do they do? They don't run the same old models faster, they build bigger models, they run training for longer, they move exponentially more data around between the chips, the memory, the network.
8:19 Speaker 01 Ah, so the efficiency gain at the chip level gets eaten up by doing much, much more at the system level.
8:24 Speaker 02 Exactly, and chip efficiency doesn't account for all the other power costs, the interconnects, the memory access, and especially that massive cooling overhead we just talked about. System-level efficiency is what matters.
8:36 Speaker 01 So the old race, just pure speed, fastest clock speed. That's kind of over.
8:42 Speaker 02 It's been replaced. The new metric, the new king, according to these sources, is speed to power.
8:48 Speaker 01 Speed to power, define that for us.
8:50 Speaker 02 It's basically the critical balance. How much actual useful computation can you get done for every watt of electricity you consume? That ratio is now driving almost every design decision.
9:00 Speaker 01 From chip architecture to building layout.
9:03 Speaker 02 Absolutely, everything. Choice of cooling air versus liquid, the physical location of the data center itself. It's even changing how cloud providers charge customers.
9:12 Speaker 01 Right, I saw that. It's not just about how long your code runs anymore. It's about how much power your job draws.
9:17 Speaker 02 Yeah, if your code is inefficient, power hungry, you're going to pay more. It directly incentivizes optimizing for that speed to power metric.
9:26 Speaker 01 And physically, inside the data centers.
9:28 Speaker 02 The whole design philosophy is flipped. It used to be about maximizing square footage. How many racks can we cram in? Now it's designed around megawatt budgets.
9:37 Speaker 01 Nigawatt budgets, okay.
9:38 Speaker 02 Architects are talking to utility companies years ahead. Can your substation handle another 50 megawatts? 100 megawatts? They're shifting heavily to liquid cooling because it's more power efficient for dense heat loads, even though it requires totally different plumbing and infrastructure.
9:55 Speaker 01 It's a complete reversal. It's not how many chips can we afford? It's how much power can we reliably get?
10:01 Speaker 02 That's the core question. How much power can we physically deliver and then effectively cool? The infrastructure built for yesterday's internet, the wires, the transformers, the cooling towers, it's straining to keep up with today's AI demands.
10:15 Speaker 01 Which brings us to the really big strategic shift, the geographical one. If power is the main constraint, the sources argue the whole calculus for where you build these massive tech hubs has changed.
10:26 Speaker 02 Completely changed. Forget just chasing cheap land or low latency network routes or even tax breaks, though those still matter a bit.
10:32 Speaker 01 So what are they chasing now?
10:34 Speaker 02 They're chasing stable grids, first and foremost. Access to reliable, large-scale, preferably low-cost electricity. And significantly, they need space, often nearby, for dedicated renewable generation or other power sources.
10:48 Speaker 01 So the source material flat out states, "Power availability now dictates where the next big cluster gets billed, not fiber optic routes."
10:55 Speaker 02 Yes, that's the new reality. Power dictates location.
10:59 Speaker 01 This creates what the sources call a strategic electricity advantage. That's a fascinating idea. Energy isn't just a cost, it's a competitive weapon.
11:08 Speaker 02 Absolutely.
11:09 Speaker 01 Yeah.
11:09 Speaker 02 Countries, even specific regions or companies, they can generate reliable power at the immense scale needed. They gain a huge tangible advantage in the global AI race and frankly, in the whole digital economy.
11:19 Speaker 01 So we're moving from talking about cloud regions to talking about power regions.
11:23 Speaker 02 That's the emerging concept. Where's the power reliable and scalable? That's where the next generation of AI infrastructure will likely concentrate.
11:31 Speaker 01 Okay, so practically speaking, how are companies and even governments planning for this? If the existing grid is strained.
11:39 Speaker 02 They're pursuing future solutions right now. They know they can't just wait for the traditional grid upgrades, which can take decades. One major area discussed is investment in what are being called nuclear computation hubs.
11:50 Speaker 01 Nuclear computation hubs. Okay, that sounds futuristic. Break that down.
11:54 Speaker 02 It's actually pretty straightforward, conceptually. These are planned zones that co-locate new small modular reactors, or SMRs.
12:03 Speaker 01 Smaller, more flexible nuclear reactors.
12:06 Speaker 02 Exactly. Co-locating those directly with hyperscale data centers. The idea is the data center gets its own dedicated, reliable 24/7 baseload power source right next door. It bypasses the constraints of the public grid.
12:18 Speaker 01 That completely changes the game in terms of capacity and reliability for that specific facility.
12:23 Speaker 02 It does. And we're also seeing plans for hybrid projects, maybe combining dedicated nuclear with gas for backup or large-scale solar and batteries. The goal is maximizing uptime and power security.
12:35 Speaker 01 And the alternative, if these investments don't happen fast enough.
12:38 Speaker 02 The sources warn pretty clearly. Without much faster investment in grid expansion and these kinds of new dedicated power sources, the US, for example, could face real electricity reliability risks by the early 2030s, simply because data center demand starts crowding out everyone else.
12:56 Speaker 01 So the pace of AI development, the digital revolution, it's now physically tethered to the speed of building concrete, steel, wires, and power plants in the real world.
13:07 Speaker 02 The limits of physics and civil engineering are setting the pace now, not just semiconductor physics.
13:11 Speaker 01 All right, let's try to wrap this up for you, the listener. The core transformation we've seen today is that AI energy demand is just exploding, doubling globally in maybe six years. And that's forcing this huge pivot.
13:23 Speaker 02 Yeah, the focus isn't just maximizing raw compute power anymore. It's shifted decisively to maximizing power availability and power efficiency. It's the dawn of the speed to power era.
13:33 Speaker 01 Right. The simple question used to be, can we make the chip faster? Now it's this incredibly complex engineering puzzle. Can the physical world, the power grids, the cooling systems, the substations, actually keep up with the machines we're building?
13:46 Speaker 02 That's the heart of it. The big takeaway here is that computational science and energy infrastructure planning are now completely inseparable. You can't advance one without hitting limits on the other. This isn't a problem just chip designers can solve alone anymore.
14:00 Speaker 01 So a final thought for people to chew on.
14:02 Speaker 02 Well, given the sheer scale we've talked about and AI's reliance on constant sustained power, the real question maybe isn't just can we power AI? It's more fundamental.
14:12 Speaker 01 Yeah.
14:12 Speaker 02 Can our existing energy infrastructure, which is built for a different world and changes slowly, actually evolve fast enough to keep pace? And building on that, what role should governments play in managing access to this emerging strategic electricity advantage? Who gets the power and how do we ensure it's used responsibly?
14:32 Auren Hoffman? Thanks for listening. If you haven't already done so, please subscribe to World of DAS wherever you consume your podcasts, YouTube, Apple, Spotify and more. And please help us get discovered by leaving a review. And check out worldofdas.com. That's worldofdas, D-A-A-S, dot com. And of course, connect with me on Twitter, at Auren. That's A-U-R-E-N. Would absolutely love to hear from you.
Previous Clay COO Varun Anand - why cutting-edge GTM teams are moving away from traditional CRMs Next Astroten Founder Richard Shotton - why your A/B tests fail and how behavioral science can fix them Page updated November 1, 2026 · transcript and summary are machine-generated