Below the Ice — A Quintillion a Second: What Exascale Really Means for Science

This is the print twin of tonight's Below the Ice — our evening deep-dive, one topic told properly. Prefer it in your ears while you wind down? Listen to today's episode.

The morning wire ran a clean, tidy headline this week: Europe switches on its first exascale supercomputer; a stack of new AI-for-science tools lands at ISC in Hamburg. It scrolls past in three seconds. But under that one word — exascale — is a number so large it stops meaning anything, a decade of engineering spent fighting heat and physics, and a quiet shift in what a "supercomputer" is even for. So tonight we sit with it. We start from a single calculation, build up to a quintillion of them a second, and then ask the only question that matters for the rest of us: now that these machines exist, what actually changes?

What it is

Start with the smallest piece. A floating-point operation — a FLOP — is one arithmetic step on a number with a decimal point: multiply 3.14 by 2.0, add the result to something else. That's it. Every weather forecast, every protein fold, every climate model is, underneath, an unimaginable pile of those tiny steps.

Exascale is a threshold for how fast you can do them. An exaFLOP is one quintillion floating-point operations per second — a 1 with eighteen zeros after it, 1,000,000,000,000,000,000, every second. A machine that sustains at least that is an exascale computer. As of this month Europe has its first one: JUPITER, at Germany's Forschungszentrum Jülich, which NVIDIA describes in its exascale-science writeup as running on Grace Hopper Superchips stitched together with Quantum-X800 InfiniBand networking. It's only the handful-th machine on Earth to cross that line — you can watch the leaderboard shift on the twice-yearly TOP500 list.

So that's the definition, plainly: exascale is not a new kind of computer. It's an old kind that finally got fast enough to count past a number we have no intuition for.

How it actually works

Let's make the number real, because "a quintillion" is just noise until you anchor it.

Imagine every human being alive — about eight billion of us — each doing one calculation per second, by hand, without sleeping. To match what JUPITER does in a single second, all eight billion of us would have to keep going for nearly four years. That's the gap between a person and an exascale machine: roughly the whole human race, working for four years, versus one tick of the clock.

You don't get there with one heroic chip. You get there with tens of thousands of them, and here's the part the headline hides: the chips were never the hard problem. The plumbing was. Three walls stand between you and exascale, and all three are about moving things, not computing them.

1. Data has to travel. A calculation is useless if the number it needs is sitting on a chip across the room. At exascale you're shuttling oceans of data between thousands of processors thousands of times a second, so the network between chips — that Quantum-X800 InfiniBand fabric — matters as much as the chips themselves. A supercomputer is mostly a very, very good answer to the question "how do I get this number over there in time?"
2. Heat has to leave. Pack that many chips together at full tilt and you've built a furnace. NVIDIA's own cooling writeup describes running coolant hotter — up to 45°C, warmer than a hot tub — precisely because a higher coolant temperature is easier and cheaper to shed to the outside air. The counterintuitive trick of the whole field is that hotter cooling is more efficient cooling.
3. Power has to come from somewhere. These machines draw the electricity of a small town. The binding constraint on the next decade of computing isn't transistors — it's the grid. That's why "energy" keeps showing up next to "exascale" in every one of these announcements.

So the right mental model isn't a brain. It's a city: tens of thousands of workers (the chips), a road system that decides whether the city functions or gridlocks (the network), a cooling system keeping it from melting, and a power plant that caps how big it can ever grow.

Why it matters now

For most of their history, supercomputers ran simulations — set up the physics equations of a galaxy or a hurricane or a fusion plasma and grind them forward, step by step. Slow, exact, and enormously expensive.

The shift worth staying up for is this: that raw simulation is now blending with AI. The new pattern is the surrogate model — you run the slow, exact simulation (or gather real experimental data) once, train a neural network on the results, and from then on the network predicts in seconds what the simulation took weeks to compute. The canonical example is protein folding: NVIDIA's data-acquisition writeup notes that AlphaFold2's 2020 breakthrough leaned entirely on the ~170,000 protein structures scientists had painstakingly measured over decades. The simulation and the experiment generate the truth; the AI learns to shortcut to it.

That's the turn the ISC announcements are all circling. NVIDIA used the conference to introduce a clutch of AI-for-science software — the DAQIRI library and ALCHEMI microservices, aimed at everything from chemistry and materials discovery to the search for dark matter — plus new agentic-AI systems at Los Alamos built on its Vera CPUs. And it's not only the national labs: the U.S. National Science Foundation's NAIRR pilot has already put serious AI infrastructure behind 700+ research projects, from protein prediction to tracking infectious-disease outbreaks.

Here's why a builder winding down on the couch should care, even if you'll never touch one of these machines. The pattern — run the expensive ground-truth process once, train a fast model to imitate it, then call the model a thousand times — is not exclusive to dark-matter labs. It's the same move you make when you cache an expensive computation, fine-tune a small model on a big model's outputs, or build a cheap classifier from labeled examples. Exascale science is the most dramatic version of a technique that scales all the way down to your own projects.

What is overhyped

Now the honest part, because the word "exascale" is doing a lot of marketing work.

First, the number itself is a peak, not a promise. That quintillion-a-second figure is a benchmark score — the machine flat-out, running an idealized test. Real scientific code, with all its messy data movement and waiting, rarely sustains anything close to the peak. "Exascale" on the brochure and exascale delivered to your experiment are different quantities, and the gap is where a lot of the honesty lives.

Second, AI-for-science is acceleration, not oracle. A surrogate model is only as good as the simulations and measurements it learned from, and it can be confidently, fluently wrong the moment you ask it about something outside that training distribution. These tools speed up the search for an answer; they don't remove the obligation to check the answer against reality. Treating a fast prediction as a verified result is exactly the failure mode to avoid.

Third — and we say this plainly — almost every source above is one vendor, narrating from its own conference booth. The hardware is real, the science running on it is real, and the people at Jülich and Los Alamos are doing genuinely hard work. But "our chips power the frontier of science" is also a product pitch, and the framing deserves the same skepticism you'd give any keynote. The discovery is the story; the logo on the cooling rack is the advertisement.

What to watch

Three concrete things, the way we close every dive.

1. Whether AI surrogates earn trust in high-stakes science. Watch for validation standards — how a drug-discovery or climate group proves a model's fast prediction holds up against ground truth before anyone acts on it. The science-for-real test isn't speed, it's trusted speed, and that's still being negotiated.
2. Who actually gets access. Exascale could concentrate in a few national labs, or open up. Watch programs like NAIRR in the U.S. and EuroHPC in Europe — whether a small university lab can get real time on these machines decides whether "AI for science" means all science or just the well-funded kind.
3. The energy and cooling ceiling. Watch the power story as closely as the compute story. The 45°C cooling threshold and the scramble for efficient, even renewable, power aren't side notes — they're the wall the whole field is accelerating toward. The next constraint on science isn't ideas or chips. It's watts.

The reassuring version of tonight's story is the headline: biggest machine ever, science unleashed. The truer version is quieter and more demanding. We built something that can do in one second what humanity could not do in four years by hand — and the moment we have it, the real work becomes restraint: checking the fast answers, sharing the scarce access, and paying the power bill the planet can actually afford. The quintillion is the easy part. What we choose to point it at is the rest.

---

That's tonight's Below the Ice. The full episode — same topic, slower and out loud — is up now: listen to today's episode. More deep-dives at penguinalley.com.

Sources: At ISC, JUPITER Shows What Exascale Science Looks Like — NVIDIA · New NVIDIA AI Software Unlocks Scientific Discoveries (ISC, Hamburg) · NVIDIA Vera CPU Opens Agentic Scientific AI at Los Alamos · NAIRR Reshapes Scientific Research · The 45°C Breakthrough to Cool AI's Biggest Machines · Real-Time AI for High-Speed Data Acquisition with DAQIRI · TOP500 supercomputer list