Fusion's 3.15 MJ milestone, the number that wasn't on the slide, and the chasm between every “it works!” demo and a thing you can ship.
On December 13, 2022, the U.S. Secretary of Energy stood at a podium in Washington and announced a milestone physicists had chased since the 1950s. Eight days earlier, at Lawrence Livermore's National Ignition Facility, 192 lasers had fired a pulse a few billionths of a second long — 2.05 megajoules of light — into a peppercorn-sized capsule of deuterium-tritium fuel, and the fuel had fired back: 3.15 megajoules of fusion energy. More out than in. Gain greater than one. For the first time in history, a controlled fusion reaction had released more energy than was delivered to it. The headlines wrote themselves: limitless clean power, a star in a bottle, the energy crisis solved.
Here is the number that wasn't on the slide. To make those 2.05 megajoules of laser light, the facility drew roughly 300 megajoules from the wall. The reaction returned 3.15. So the honest energy ledger for the most celebrated fusion experiment in history reads: about 300 in, about 3 out — a net loss north of 99%. The physics achieved gain. The facility ran at roughly one percent efficiency. Both of those sentences are completely true at the same time, and the distance between them is the most important thing the press conference didn't explain.
It is also the exact distance between every “we got it working!” demo and a thing you can actually ship.
The 3.15-megajoule result was real science — the validation of a triple-product physics problem that took the better part of seventy years to solve. What it was not was a power plant, and the reasons map cleanly onto any production system. There are three gaps between NIF and commercial inertial fusion energy, and the instructive part is that they are wildly different in size.
Efficiency. NIF converts wall power to fusion output at about 1%. A commercial plant needs better than 10% just to break even on the grid connection, never mind sell electricity. Call that a 10× gap — one order of magnitude. It's hard. It is, relatively speaking, the easy one.
Repetition rate. NIF fires roughly once per day; between shots, the optics are inspected and reset and the next hand-built target is aligned. A power plant needs to do this about ten times a second, continuously, forever. Ten shots a second against one a day is a gap of nearly a million-fold — close to six orders of magnitude. This is not “the same problem but harder.” It is a different category of problem.
Manufacturability. Each NIF target is a precision artifact — a machined capsule held to nanometer tolerances inside a gold can, hand-assembled, hand-inspected, hand-positioned, at a cost that runs into the tens or hundreds of thousands of dollars per shot. A power plant firing ten times a second needs roughly 864,000 targets a day, at pennies each, stamped out and fed in like cartridges. That, too, is a six-orders-of-magnitude leap, on an axis (mass manufacturing) that the science facility never had to think about at all.
The physics milestone was one achievement on one axis. Commercial fusion needs many orders of magnitude of engineering across several axes at once — and the new private inertial-fusion companies (Focused Energy, Xcimer, Marvel Fusion, and Longview Fusion Systems, the last founded by veterans of the NIF team itself) exist precisely to grind through that engineering. The DOE's own 2025 fusion roadmap treats commercial electricity as a multi-decade program. The physics took seventy years. The engineering is widely expected to take another twenty or thirty.
Here is the detail that turns this from a one-time caveat into a law of nature, and it's the part most coverage missed because it happened after the cameras left.
NIF did not ignite once. Since December 2022 it has ignited again and again. By February 2024 a shot produced 5.2 megajoules from 2.2 in — a gain near 2.4, comfortably better than the famous result. Later shots pushed the yield higher still; by late 2025 the facility had logged on the order of ten ignitions, with one shot reaching about 8.6 megajoules (figures from LLNL's reporting and a 2025 GAO oversight report). The science number got dramatically better over three years.
And it is still not a power plant. The repetition rate is still about one shot a day. The wall-plug efficiency is still about 1%. The targets are still hand-built. Tripling the gain moved the production metrics exactly zero, because improving the thing the demo measures and building the thing production needs are different problems solved by different disciplines. You can polish the breakthrough indefinitely and never accidentally back into a product. That is the whole lesson, and fusion just spent three years proving it in public.
If you build software, you have lived this, probably this quarter:
Every one of these is gain greater than one at one shot a day. It worked — genuinely, the science is sound, the thing is possible. And it is production-irrelevant in its current form. Both true at once. The demo is not a lie; it's an ignition. It just isn't electricity.
Notice what everyone reported about NIF: 3.15 out, 2.05 in, gain 1.5. Notice what almost nobody reported: 3.15 out, 300 in, efficiency 1%. The number that made the headline was the gain — the demo metric, computed over the most favorable denominator available (energy that actually reached the fuel). The number that determines whether you have a power plant was the wall-plug efficiency — the production metric, computed over the denominator that actually bills you. The first was on every slide. The second you had to go digging for.
This is the single most portable habit in this essay. When someone shows you a breakthrough, find the denominator they left off the slide. The model is 95% accurate — over what input distribution, and what's the p99 latency and the cost per call at production throughput, not demo throughput? The pipeline answers the question — at what hallucination rate across adversarial real queries, and at what unit economics when it's ten thousand queries a second instead of the one you just watched? The gain is the number that gets the press conference. The wall-plug number is the one that decides whether you have a business. Always ask: what's the 300 megajoules?
And the denominators stack, which is worse. NIF uses indirect drive: the lasers don't strike the fuel at all — they heat the inside of a tiny gold can, a hohlraum, which re-radiates X-rays that do the actual compressing. That conversion discards something like 90% of the laser energy before it ever reaches the fuel. So you can quote this one experiment at least three ways. Measure the fusion output against the energy actually coupled to the capsule and the “gain” balloons toward fifteen. Measure it against the 2.05 megajoules of laser light and you get the famous 1.5. Measure it against the 300 megajoules from the wall and you get a 99% loss. Same shot. Three numbers spanning more than an order of magnitude — and the one that made the headline was, of course, the most flattering. So the discipline isn't only finding the denominator they left off; it's noticing that a complex system has several denominators, and that a demo will always, honestly and without lying, quote the kindest one. Your model's “95% accuracy” is somebody's hohlraum number: true, and measured at the most favorable layer of the stack.
The mapping from fusion's three gaps to software's three production requirements is almost one to one. Efficiency becomes cost-per-unit-of-work: the demo runs at a dime a query, production needs a hundredth of a cent. Manufacturability becomes reproducibility: the demo works on your machine with your data, production has to work everywhere, every time, on inputs you've never seen. And repetition rate becomes throughput — and just as in fusion, this is the brutal one.
Improving a model's accuracy from 95% to 99% is hard, but it is the same kind of hard as the work that got you to 95: more data, better training, sharper evaluation — science, roughly. Scaling that same model from one query a second to ten thousand a second while holding the accuracy, the latency, and the cost is not more of that work. It is a different discipline — systems engineering, capacity planning, caching, batching, failure handling, the whole unglamorous machinery of running a thing continuously under load. Going from one shot a day to ten a second was never going to be solved by a better physics result, and going from one QPS to ten thousand is never going to be solved by a better model. The axis that looks like “just scale it up” is, in both fields, where the decade actually goes.
Throughput also changes the arithmetic on rare failures, and that is the part that bites hardest. A one-in-a-million target defect is a non-event when you fire once a day — you'd meet it about once every three thousand years. At ten shots a second you meet it twice a day. A failure mode that was statistically invisible during the demo becomes a daily operational fact at production rate, and you find out about it the way everyone finds out about these things: in production. The same arithmetic governs software exactly. The malformed input your demo never happened to feed it, the adversarial query, the pathological concurrency race — each is vanishingly rare per request and therefore certain at scale. Throughput doesn't merely multiply your successes; it multiplies your exposure to every tail you didn't test, which is why the system that purred through a thousand demo queries can fall over in its first production hour.
The move here is not cynicism. Ignition was a genuine, seventy-year achievement; the team earned every bit of the champagne, and “this has never been done” is exactly the kind of thing worth standing at a podium for. The failure mode is not celebrating the milestone — it's confusing the milestone for the finish line, and then being blindsided when the decade of engineering that was always going to follow turns out to be there.
It's worth being honest about why that confusion is so durable, because it is not stupidity. Demos are funded, celebrated, and photographed; the production engineering that follows is invisible, unglamorous, and consists mostly of fixing things that already appear to “work.” The gain number is press-conference-ready and the wall-plug number is a footnote, so the gain number is what gets reported, internalized, and — most dangerously — planned around. Roadmaps and valuations get built on the demo metric because the demo metric is the one everyone saw. The fix isn't to distrust breakthroughs; it's to drag the production denominator onto the same slide as the headline from day one — to quote the wall-plug efficiency next to the gain, the cost-at-real-throughput next to the latency, the accuracy-on-uncurated-traffic next to the benchmark — so that nobody in the room can quietly mistake the moment the science works for the moment the product does.
So separate the two questions, explicitly, every time, because they are answered by different people on different timescales. Does it work at all? is a science question — answered once, by a breakthrough, often after years of trying, and genuinely worth celebrating when the answer is yes. Will it work at throughput, at cost, repeatably, and manufacturably? is an engineering question — answered over years, by a different team, and it is where the time, the money, and the disappointments live.
When your model nails the demo, you have had your ignition: a real result, hard-won, more out than in, worth a moment of genuine pride. You also have one shot a day at one percent efficiency, with a hand-built target. Pop the cork — and then go find out what your 300 megajoules is, and budget for the decade, because that is the part that was always going to be the work.
Put the wall-plug number on the same slide as the gain.
The whole discipline here is refusing to let the demo metric stand in for the production one — quoting the accuracy-on-uncurated-traffic next to the benchmark, the cost-at-real-throughput next to the latency. That requires a record of what your system actually did in production, not what it scored in the demo. Chain of Consciousness anchors every agent action to a verifiable external record — an auditable trail of real behavior at real throughput, so the wall-plug number isn't a footnote you have to go digging for. It's the receipt that sits next to the headline.
See a verified provenance chain · Hosted Chain of Consciousness
pip install chain-of-consciousness · npm install chain-of-consciousness