The Bill Is Coming

The Bill Is Coming

AI needs more power than the grid can provide.

Fusion was supposed to be the answer. Here’s why it keeps arriving late, and what that actually means.

Somewhere right now, a data center is being built. Then another. Then another after that. The race to train larger AI models, run more inference, and expand cloud infrastructure is creating an electricity demand that is genuinely difficult to overstate. Microsoft, Google, Meta, and Amazon are collectively spending hundreds of billions of dollars on AI infrastructure and all of it needs power. Constant, reliable, round-the-clock power. Not power that works when the sun shines or the wind blows. Power that never stops.

This is not a future problem. It is a present one. Tech companies are buying stakes in nuclear plants, signing deals with energy developers, and in some cases trying to restart reactors that were previously shut down. The grid, as it currently exists, was not designed for what AI is about to ask of it.

So when the world’s largest fusion reactor ITER, a $28 billion machine being assembled in southern France — announced that its full operational test would not happen until 2039, the story was filed under science news. It probably should have been filed under infrastructure crisis.

Fusion is the energy technology that could change everything. The question is whether it will arrive before the bill does.

Two Ways to Make Power Without Burning Things

Before we talk about what’s gone wrong, it helps to understand what we’re actually building toward. So let’s start with the basics.

Every nuclear power plant you have ever seen — the ones with the cooling towers, the ones that generate reliable baseload electricity, the ones that have been running since the 1970s — those are fission reactors. Fission means splitting. You take a heavy atom, like uranium, and you break it apart. When it splits, it releases energy. That energy heats water into steam. The steam spins a turbine. The turbine generates electricity.

Fission works. It is proven, commercial, and operating at grid scale today. The debate around fission is not whether it works — it does — but whether the trade-offs are acceptable. The waste it produces stays radioactive for a very long time. The plants are expensive to build and politically complicated to site. But when a fission plant is running, it runs reliably for decades and produces enormous amounts of carbon-free electricity. For AI data centers that need constant power, that is not a small thing.

Fusion is different. Fusion means joining. Instead of breaking heavy atoms apart, you push light atoms together. Specifically, isotopes of hydrogen. When you force them to fuse, a small amount of their mass converts directly into energy — enormous amounts of energy, cleanly, with far less long-lived radioactive waste than fission.

This is what powers the sun. The sun is essentially a giant fusion reactor running on hydrogen, held together by its own gravity, producing light and heat that has sustained life on Earth for four billion years. What physicists have been trying to do for the past 70 years is reproduce that process in a machine small enough to build on the ground.

Fission is breaking a log to make fire. Fusion is building a small star in a bottle. The first one we know how to do. The second one we are still working on.

The reason fusion is so attractive is the combination of qualities it promises: abundant fuel, carbon-free operation, and a waste profile that is far more manageable than conventional fission. Deuterium — one of the hydrogen isotopes used as fusion fuel — can be extracted from seawater. The energy potential is effectively limitless. If fusion works commercially, it is not an incremental improvement on existing energy. It is a different category of energy entirely.

The reason fusion is so difficult is that reproducing stellar conditions on Earth requires heating plasma to temperatures far hotter than the surface of the sun — and then holding that superheated plasma stable inside a machine long enough for the fusion reaction to happen and produce more energy than you put in. Every engineering problem in fusion is, at its core, a variation of the same challenge: how do you contain something that hot without it touching the walls.

The Biggest Bet Ever Made

ITER — which stands for International Thermonuclear Experimental Reactor, and is also Latin for ‘the way’ — is the world’s attempt to answer that question definitively. Thirty-five nations are contributing to it. It contains the world’s most powerful magnet. It will weigh 23,000 tonnes when complete. The superconducting magnets alone weigh thousands of tonnes and are connected by 200 kilometers of superconducting cable, all kept at minus 269 degrees Celsius.

It is, by any reasonable measure, the most complex engineering project in human history.

The goal of ITER is not to produce electricity. This is important to understand. ITER is an experiment. Its purpose is to prove that a fusion reaction can produce more energy than it consumes — what physicists call Q greater than 1 — and to demonstrate that the engineering required to do this at scale actually works. It is the proof of concept that is supposed to unlock the commercial reactors that come after it.

Originally, ITER was supposed to cost around $5 billion and achieve first plasma in 2020. It now costs over $28 billion and full deuterium-tritium operation — the actual fusion reaction — is scheduled for 2039 at the earliest. That is a budget that more than quintupled and a timeline that slipped nearly two decades.

The most charitable reading of ITER is that it is an unprecedented engineering challenge and delays were inevitable. The least charitable reading is that it is a cautionary tale about what happens when you optimize a project for political survival rather than engineering outcomes. Both readings contain truth.

How a $28 Billion Project Loses 15 Years

Here is a thought experiment. Imagine you are building a house. But the architect is from one country. The electrician is from another. The plumber is from a third. Each of them is accountable to their own government, working to their own national standards, and shipping their components from their home country to the build site. Every time something doesn’t fit, the negotiation to fix it goes back through three layers of international governance before anyone picks up a wrench.

That is approximately how ITER is built. Thirty-five nations contribute components manufactured domestically, shipped to France, and assembled on site. The structure distributes cost and political buy-in across a wide coalition — which is how you get 35 governments to fund something this expensive. But it also distributes responsibility in a way that makes fast correction nearly impossible.

When something goes wrong in a company, a single determined builder can make a decision and execute it by next week. When something goes wrong in ITER, the correction has to move through an international governance structure where every member state has an interest in the outcome. The result is that each individual delay is explainable and each individual decision is defensible, but the accumulated effect is a project that has slipped its schedule in almost every revision it has ever published.

There is also a pattern that engineers and project managers call scope creep. ITER’s 2025 update described the new baseline as prioritizing a ‘more complete machine than initially planned’ before the first operational phase. Supporters can read that as prudent engineering. Critics can read it as a classic megaproject move: when the schedule breaks, redefine what the starting line means so the new timeline looks less catastrophic than it actually is.

There is a useful concept in project management: the difference between a project that is late and a project that has redefined what ‘on time’ means. ITER has done the second thing several times now.

None of this means the science is wrong. The physics of fusion is not in serious dispute among people who study it. What is in dispute is whether this particular institutional structure — sprawling, consensus-driven, diplomatically constrained — is capable of delivering a first-of-its-kind engineering achievement on a timeline that matters for the problems we actually need to solve.

The director general of ITER himself said it plainly at a 2024 press conference: ‘I am not going to be a person that will tell you fusion will solve all problems at the time that they should be solved, because the time to be solved is not the time that they should be solved.’ That is not a statement of scientific failure. It is an admission that the institutional vehicle may not match the urgency of the destination.

The Food Truck Across the Street

While ITER has been under construction, something interesting has happened. A different kind of fusion program has emerged, and it looks nothing like a 35-nation megaproject.

Commonwealth Fusion Systems, Helion Energy, TAE Technologies, and a growing list of private fusion companies are pursuing commercial fusion on timelines measured in years rather than decades. They are smaller, faster, and willing to make bets that an international treaty organization cannot make. Helion has a contract with Microsoft to deliver fusion power by 2028. That is either the most audacious promise in energy history or a genuine signal that the private model can move at a different speed than the institutional one.

Think of it this way. ITER is like a restaurant that has been under renovation for 30 years. The kitchen is going to be extraordinary. The team is brilliant. The ambition is real. But the sign in the window keeps changing the opening date, and at some point the sign stops being a promise and starts being a warning.

The private fusion companies are the food truck that opened across the street while the restaurant was still renovating. Smaller menu. Less ceremony. But they are actually serving food.

ITER itself has started to notice. In late 2024, the organization launched a Private Sector Fusion Engagement effort to exchange information with fusion startups and lower barriers to collaboration. You can read that charitably as openness and institutional evolution. You can also read it as the most established fusion program in the world quietly acknowledging that it needs to learn something from the companies it once overshadowed.

When the institution starts learning from the startups, you are either witnessing a healthy ecosystem or a quiet admission that the original model has limits. Possibly both.

AI Helping Fusion. Fusion Eventually Helping AI.

Here is where the story gets genuinely interesting, and where the two frontiers — artificial intelligence and nuclear fusion — start to fold into each other.

The core engineering challenge of fusion is plasma stability. The superheated plasma inside a tokamak reactor is turbulent, unpredictable, and prone to disruptions that can damage the machine. Controlling it requires making thousands of real-time decisions about magnetic field configurations, based on sensor data that changes faster than any human operator can track.

This is exactly the kind of problem that machine learning was built for. Google DeepMind and Commonwealth Fusion Systems have both published work on using AI to improve plasma prediction and real-time tokamak control. In plain terms: AI can spot patterns in fusion data that humans miss, make control adjustments faster than humans can react, and model the complex physics of plasma behavior in ways that accelerate experimental cycles.

AI is already helping fusion happen faster. That is not a speculation. It is happening in laboratories now.

The second half of the relationship runs in the other direction, and it matters enormously for the energy problem we started with. If private fusion companies deliver commercial plants in the 2030s — and that is still a meaningful if — they could offer exactly what AI infrastructure needs: high-output, firm, carbon-free electricity that runs 24 hours a day regardless of weather. A fusion plant next to a hyperscale data center is not a science fiction scenario. It is a plausible design for how the most energy-hungry technology in human history gets powered cleanly.

The feedback loop is real: AI accelerates fusion development, and successful fusion eventually removes one of AI’s most serious constraints. The question is timing. The AI energy problem is urgent now. Fusion, even in the optimistic private-sector scenario, is a 2030s answer to a 2020s question.

For the next decade, the honest answer to AI’s energy problem is fission — the technology we already know how to build. Fusion is the answer to the decade after that, if we build the right institutions to get there.

What ITER Actually Proves

ITER will almost certainly teach the world something important about fusion physics. The science will advance. The data will be valuable. The engineering lessons, however painful, will inform every reactor that comes after it.

But ITER also proves something that has nothing to do with plasma physics. It proves that the hardest part of building the future is not the science. It is the governance. It is the question of who is accountable, who can say no, who can move fast when something breaks, and whether the institution that carries the technology is built for the problem or built for the politics.

Fusion has been 30 years away for 70 years. The joke survives not because the physics is impossible, but because the accumulated weight of institutional compromise keeps pushing the horizon forward. Every individual decision that caused a delay was defensible. The aggregate is the cautionary tale.

The private players moving fast with AI assistance are not guaranteed to succeed. Making a fusion reactor work in a laboratory is not the same as making one economically competitive with solar, wind, and advanced fission for 30 years of operation. The commercialization challenge is real and has not been solved by any team yet.

But the private model has one structural advantage that no amount of diplomatic engineering can replicate: someone is actually accountable for the outcome. There is a throat to choke. There is a founder whose name is on the door and whose capital is on the line. That is not sufficient for success, but it turns out to be remarkably useful for moving fast.

The lesson of ITER is not that big science is wrong. It is that big science needs a builder, not just a consortium. The difference between a project and a product is someone who cannot afford for it to be late.

The bill for AI’s energy appetite is coming. Fusion is one of the few technologies that could pay it cleanly and at scale. Whether it arrives in time depends less on the physics — which is beautiful and real — and more on whether the people building it are organized to win, or just organized to survive.

That is the question ITER cannot answer. But the food truck across the street might.

@currentlyted · Medium · Science & Technology

If this changed how you think about energy and AI — share it. If it didn’t — tell me why.