Cyrus Reed: Hey, been a weird week — you catch that Jensen Huang thing from May?
Iris Holm: The Stanford class. Yeah.
Cyrus Reed: He's standing in front of CS153 at Stanford and he just — he says AI is going to need probably a thousand times more energy than we currently have. A thousand times. And I keep turning that number over and thinking... wait, is that even a claim? Like what's the methodology behind that?
Iris Holm: It's a narrative. The question is what work that number is doing for Nvidia.
Cyrus Reed: Okay but — wait, so that's actually where I want to start, because I went looking for what's actually breaking right now, not in 2030, and Bloomberg ran a piece on August sixth, this year, saying AI's volatile power demand is causing physical damage to data center assets. Not a projection. Not a model. Equipment. Breaking. Now.
Iris Holm: From volatility. Not from volume.
Cyrus Reed: Which completely reframes the whole — I mean, everyone's asking 'can we generate enough electricity' and the Bloomberg thing is saying the grid might not fail from too much draw, it might fail from how erratically AI systems pull what they're already getting.
Iris Holm: So the real question isn't volume. It's whether the infrastructure can handle the shape of the demand.
Cyrus Reed: And that shape question is actually what makes Huang's thousand-times number so slippery, because — wait, okay, think about it like a contractor telling a city it needs a thousand new roads. Maybe true! But the contractor also sells roads. That's the whole — that's the thing you have to hold onto.
Iris Holm: Nvidia sells the chips. Huang is the CEO of that company. He stands up at Stanford's CS153 class and says a thousand times more power — not in a technical paper, not in an engineering white paper, not in a regulatory filing. A university class.
Cyrus Reed: No published methodology. Like, nowhere.
Iris Holm: Nowhere. And the IEA — which does publish its methodology — projects data centers at roughly three percent of global electricity demand by 2030. Not dominant. Smaller than EVs, smaller than industrial motors, smaller than air conditioning.
Cyrus Reed: So how does three percent become a thousand times? I mean — wait, is Huang describing something the IEA number just isn't capturing, or is he narrating a capital story that happens to require Nvidia at the center of it?
Iris Holm: That's exactly the gap. And Lawrence Berkeley National Laboratory admitted — in their own DOE report — that their 2016 forecasting model completely missed 2018 AI server growth. So the honest answer is forecasting here is genuinely broken.
Cyrus Reed: Which means Huang could be right, but for reasons no one can actually verify yet.
Iris Holm: Right. And his specific framing — he called future AI 'generative and continuous,' always-on agentic systems, not periodic training runs — that's a real architectural claim. The question is whether the thousand-times number follows from it or just sounds like it does.
Cyrus Reed: So he's describing a physics reality, maybe, but the number does an enormous amount of work positioning Nvidia inside a multi-trillion-dollar infrastructure cycle — and we genuinely cannot tell which one is doing the heavy lifting.
Iris Holm: And that positioning question only gets sharper when you look at what's actually breaking. The Bloomberg piece isn't about volume. It's about load character.
Cyrus Reed: Right — and this is the part I keep getting stuck on. Because an AI data center can't just... dial down. Like a factory can pause a production line when the grid is stressed. A steel mill, a semiconductor fab — curtailable loads. But an AI facility running a training job? It's always-on, non-interruptible, and the draw doesn't stay flat — it spikes. Hard. Fast.
Iris Holm: Two hundred megawatts to eight hundred megawatts. Under a minute.
Cyrus Reed: Wait — that's a real scenario?
Iris Holm: That's the scenario. Grid operator watches a facility spin up a new training run — transformer serving that facility wasn't designed for that swing. Overheats. Fails eventually. She orders a replacement. Lead time now: five years. That facility runs degraded for sixty months because the pre-2020 assumption was twenty-four to thirty months and that world is gone.
Cyrus Reed: So the August sixth Bloomberg report — the physical damage — that's not a forecast, that's the transformer scenario playing out now, and the replacement queue is five years long. That's — okay, that actually reframes the hundred-gigawatt gap Bank of America found. Because it's not just that we need more generation capacity, it's that the hardware to route whatever we build is already backlogged past the point where demand projections will have shifted again.
Iris Holm: Bhargavi Vepuri at Prudential Financial put it as a stack — compute, energy, governance, cost, all simultaneously constrained. The load inflexibility argument fits that framing exactly. It's not one bottleneck. Each layer compounds the one below it.
Cyrus Reed: Which means Huang's thousand-times number — even if the physics are real — is almost solving the wrong problem? Like, building more generation partially misses it if the actual failure mode is the shape of demand, not the size.
Iris Holm: And there's a layer under that we haven't touched — whether efficiency gains actually reduce the load or just make room for more of it. That one gets uncomfortable fast.
Cyrus Reed: But wait — that's where the efficiency story gets weird, because everyone's citing tokens-per-watt as the thing that saves us. More output per joule, problem solved, right? Except — okay, Jevons Paradox. We made cars more fuel-efficient and people drove more miles. Not fewer. So if AI chips get more efficient, do we just... spin up more agents?
Iris Holm: That's exactly what the IEA data suggests is already happening. Global data-center consumption doubles between 2025 and 2030 — even as tokens-per-watt improves. Efficiency absorbed by volume.
Cyrus Reed: Doubles. Despite the efficiency gains.
Iris Holm: The DOE number makes it concrete. Four point four percent of U.S. power in 2023 — Lawrence Berkeley ran that analysis — projected to hit twelve percent by 2028. That's a nearly three-times jump in five years. Efficiency didn't bend the curve. It just made room for more workload.
Cyrus Reed: So cheaper tokens-per-watt makes it economical to run always-on agents continuously that you — wait, no, that you couldn't justify running before. Like, the efficiency gain doesn't retire demand, it unlocks demand that was waiting.
Iris Holm: And Goldman Sachs already has us at thirty-one gigawatts in 2025, sixty-six by 2027. Bank of America says we need over two hundred and thirty gigawatts of new generation through 2031 — regulated utilities expected to add ninety-three. That's a hundred-gigawatt gap. Goldman's grid investment figure to close it: seven hundred and twenty billion dollars.
Cyrus Reed: And those numbers are using different denominators, different time horizons — I mean, you can't just stack them like they're measuring the same thing.
Iris Holm: Frankly, no. But directionally? They all point the same way. And the punchline isn't the generation gap — it's who wins. If efficiency gains don't reduce total consumption, the competitive edge shifts. Not to whoever engineered the best chip. To whoever locked in transformer supply chains and permitting relationships three years ago.
Cyrus Reed: So the race isn't silicon anymore. It's — huh — it's who signed infrastructure deals before the lead times stretched to five years.
Iris Holm: And that's the part that doesn't resolve cleanly. Because even if Huang's thousand-times number is strategically inflated — and I think it probably is — the transformer bottleneck is real. The Bloomberg damage report is real. Five-year lead times are real. The narrative might be doing work for Nvidia and the underlying crisis might still be genuine. Both at once.
Cyrus Reed: Which means — wait, so you can't actually fact-check your way out of it. Like, you debunk the thousand-times methodology and you've still got equipment breaking in August 2026 and a hundred-gigawatt gap that regulated utilities won't close. The number might be wrong and the problem it's pointing at is still sitting there.
Iris Holm: That's the uncomfortable part. The winners here probably aren't whoever built the most efficient chip. They're whoever secured transformer supply and permitting before 2021, before the lead times doubled. That's — that's not a technology story at all.
Cyrus Reed: We started with Huang in a Stanford classroom saying a thousand times. And I think I land somewhere like — yeah, maybe, but the actual story is a grid operator watching a transformer overheat because a training run spiked from 200 to 800 megawatts too fast, and the replacement is five years out. That's the version of the thousand-times problem that's already here.
Iris Holm: And nobody needs a white paper to be nervous about that. Thanks for working through it.
