Fusion Dreams: Tokamaks Meet Training Runs

"Fusion is thirty years away, and always has been."

This joke has circulated among physicists for decades. It captures the field's history of perpetual promise and perpetual delay. Since the 1950s, fusion—the process that powers the sun—has been "almost ready" for commercial deployment. And yet here we are, still waiting.

But something changed in the early 2020s. Private investment flooded into fusion startups. Billions of dollars. Dozens of companies. Google, Chevron, and Bill Gates backed competing approaches. Sam Altman, CEO of OpenAI, invested $375 million in Helion Energy.

The AI industry is betting on fusion. Not as charity or greenwashing, but as a solution to the compute ceiling. If AI needs more power than existing sources can provide, maybe the answer is to build an entirely new kind of power source.

The question is whether fusion will arrive in time—and whether it can scale fast enough to matter.

The Physics of Fusion

Fusion releases energy by combining light nuclei—typically isotopes of hydrogen—into heavier nuclei. The most common reaction uses deuterium (hydrogen with one neutron) and tritium (hydrogen with two neutrons) to create helium and a neutron.

The energy released per reaction is enormous. Fusing a kilogram of deuterium-tritium fuel releases about four times as much energy as fissioning a kilogram of uranium, and about ten million times as much energy as burning a kilogram of coal.

Fusion fuel is abundant. Deuterium can be extracted from seawater—there's enough in the oceans to power civilization for billions of years. Tritium is rarer but can be bred from lithium, which is also plentiful.

And fusion is clean. No carbon emissions. No long-lived radioactive waste. No risk of meltdown—if something goes wrong, the reaction simply stops.

On paper, fusion is the perfect energy source. The problem is the physics is brutally hard.

Why Fusion Is Hard

To fuse, nuclei must get close enough for the strong nuclear force to overcome their electrostatic repulsion. Since both nuclei are positively charged, they repel each other intensely at short distances.

Overcoming this repulsion requires extreme temperatures—on the order of 100 million degrees Celsius. At these temperatures, matter becomes plasma: a soup of electrons and bare nuclei, no longer bound into atoms.

Containing plasma at 100 million degrees is the central challenge. No solid material can withstand such temperatures. You can't just build a container and pour hot plasma into it.

Two main approaches have dominated fusion research:

Magnetic confinement uses powerful magnetic fields to contain the plasma, keeping it away from any physical walls. The most common device is the tokamak, a donut-shaped chamber with magnetic fields that spiral around the plasma, holding it in place.

Inertial confinement uses lasers or particle beams to compress tiny fuel pellets so quickly that fusion occurs before the plasma can disperse. The compression must be extraordinarily symmetric; any asymmetry causes the pellet to squirt sideways instead of imploding.

Both approaches have achieved fusion reactions in laboratory settings. The challenge is achieving net energy gain—getting more energy out of the fusion reactions than you put in to heat and confine the plasma.

For decades, fusion experiments got tantalizingly close to this threshold but never crossed it. The plasma instabilities were always worse than predicted. Energy leaked faster than expected. The goal retreated as researchers approached.

The Breakthrough: Ignition Achieved

On December 5, 2022, the National Ignition Facility (NIF) at Lawrence Livermore National Laboratory achieved ignition: a fusion reaction that produced more energy than the lasers delivered to the fuel.

The lasers delivered 2.05 megajoules. The fusion reaction released 3.15 megajoules. Energy gain of about 1.5.

This was a genuine scientific milestone. After seventy years of effort, fusion had crossed the threshold. The physics works. Net energy gain is possible.

But—and this is crucial—it wasn't net energy gain in any practical sense. The lasers are only about 1% efficient. The 2.05 megajoules they delivered required about 300 megajoules of electricity to generate. From a full-system perspective, the experiment consumed roughly 100 times more energy than it produced.

Furthermore, NIF isn't designed for power production. It was built for nuclear weapons research. The laser can fire only a few times per day. Turning this into a power plant would require firing continuously, which the current technology cannot do.

Still, ignition matters. It proves that the physics barrier has been crossed. The remaining challenges are engineering, not fundamental science. Engineering challenges are solvable, given enough money and time.

The Startups Racing Toward Reactors

The 2020s have seen an explosion of private fusion companies, each betting on different approaches to reach commercial viability faster than traditional research programs.

Commonwealth Fusion Systems (CFS), spun out of MIT, uses high-temperature superconducting magnets to build smaller, cheaper tokamaks. Their SPARC device, under construction, aims to demonstrate net energy gain from magnetic confinement in the late 2020s. They've raised over $2 billion and have agreements to supply power to several companies.

Helion Energy, backed by Sam Altman, uses a pulsed approach where two plasma masses are accelerated toward each other and compressed. They claim this could achieve net energy gain more cheaply than traditional tokamaks and have signed a power purchase agreement with Microsoft for electricity delivery in 2028.

TAE Technologies uses a particle accelerator approach, aiming to fuse hydrogen and boron rather than deuterium-tritium. This reaction produces no neutrons, simplifying reactor design, but requires even higher temperatures.

General Fusion uses a mechanical compression system—pistons compressing liquid metal around plasma. It's a hybrid of magnetic and inertial approaches.

Zap Energy, backed by Chevron, uses a z-pinch configuration where electrical current flowing through the plasma creates its own confining magnetic field.

These companies share a strategy: use modern technology—better magnets, better computing, better materials—to achieve what government programs couldn't. They're betting that advances in adjacent fields have lowered the barrier to fusion success.

The timelines are aggressive. Multiple companies claim they'll demonstrate net energy gain by the late 2020s and commercial power by the early 2030s. These claims should be viewed skeptically—fusion timelines have a history of slipping. But the concentration of talent and capital is unprecedented.

Why AI Investors Care

The AI industry's interest in fusion isn't abstract. It's about survival.

As we discussed earlier in this series, AI training costs are escalating exponentially. The energy demands are straining existing infrastructure. Nuclear provides near-term relief, but nuclear is politically constrained and slow to build. Renewables are intermittent. Natural gas is carbon-intensive and geopolitically fraught.

Fusion, if it works, solves multiple problems simultaneously:

Abundant fuel. Deuterium from seawater is essentially unlimited. No dependence on mines in specific countries.

High power density. Fusion plants could be compact relative to their output, fitting into existing power infrastructure.

No carbon emissions. AI companies face pressure to decarbonize, and fusion provides a path to doing so without compromising on power.

Continuous baseload. Unlike solar and wind, fusion runs whenever you want it to. No intermittency to manage.

Reduced geopolitical risk. No dependence on uranium enrichment, oil imports, or rare earth elements. Fusion is energy independence.

If you're planning a company around AI that might need ten times current power consumption within a decade, fusion looks like the ultimate solution. The bet isn't that fusion will work for sure—it's that if fusion works, early investors will capture enormous value, and if AI scaling continues, there's no alternative anyway.

Sam Altman has been explicit about this connection. His investments span both AI (OpenAI) and energy (Helion, nuclear). The strategy is to build the AI and build the power it requires in parallel.

The Timing Problem

Here's the tension: AI needs power now. Fusion is, optimistically, a decade away from commercial deployment.

Even if everything goes well for fusion startups—if they demonstrate net energy gain by 2028, build pilot plants by 2030, and begin commercial deployment by 2032—that's many years during which AI must rely on existing power sources.

And "everything going well" is a big if. Fusion has surprised researchers with unexpected instabilities and engineering challenges at every stage. The private sector may be more nimble than government labs, but it's not immune to physics.

The realistic scenario is probably:

- 2024-2030: AI relies on existing grid power, new nuclear (restarts and SMRs), and aggressive efficiency improvements. - 2030-2035: First commercial fusion plants come online, if the technology works. Initial deployment is limited—a few plants, learning curves, debugging. - 2035-2040: Fusion scales if it proves economical. AI data centers begin shifting to fusion power.

This timeline means fusion isn't a solution to the current compute ceiling. It's a potential solution to the ceiling a decade from now. The question is whether AI development can continue at current pace long enough for fusion to arrive.

The answer may be no. AI scaling might hit energy walls before fusion can help. If that happens, AI development slows, regardless of fusion progress. The timing has to match.

If Fusion Works

Assume the optimists are right. Fusion achieves commercial viability in the 2030s. Energy becomes abundant and cheap. What then?

The implications extend far beyond AI:

Abundant computation. If energy ceases to be the binding constraint, AI training costs collapse. Models can be larger. Experiments can be more numerous. The pace of AI development accelerates dramatically.

Industrial revolution 2.0. Cheap energy enables energy-intensive processes that are currently uneconomical: carbon capture, desalination, synthetic fuel production, space launch. Fusion is civilization-altering if it delivers on its promise.

Geopolitical realignment. Countries dependent on fossil fuel exports lose leverage. Energy-importing nations gain autonomy. The petrodollar system erodes.

Reduced climate pressure. Abundant clean energy provides a path to decarbonizing everything—electricity, transportation, industry, heating. Climate change becomes a solvable problem rather than an existential threat.

New constraints emerge. If energy is no longer the bottleneck, something else will be. Materials? Cooling? Skilled labor? Regulatory approval? Every solved constraint reveals the next constraint.

Fusion doesn't solve all problems. But it changes the game entirely. The whole framing of this series—the energy limits on intelligence—becomes obsolete if fusion works.

If Fusion Doesn't Work

The alternative scenario: fusion remains thirty years away. Private investment dries up after disappointing results. The physics is simply too hard to engineer economically.

In this case, AI development faces sustained energy constraints. Progress continues but more slowly. Efficiency improvements become paramount. The neuromorphic and biological approaches we discussed earlier become not just interesting but essential.

The world muddles through on a mix of nuclear, renewables, and diminishing fossil fuels. Climate change remains a struggle. Energy costs remain a constraint on computation and everything else.

This isn't catastrophic. Humanity has faced energy constraints before and found ways to work within them. But it's a slower, more constrained future than the fusion optimists imagine.

The Wager

The AI industry is making a wager: bet on fusion now, because if it works, the payoff is enormous, and if AI scaling continues, there's no alternative that fully solves the problem.

It's a calculated risk. The money invested in fusion—billions of dollars across the sector—is tiny compared to what AI companies spend on chips and data centers. It's insurance, a portfolio hedge against the future.

The compute ceiling may force AI to transcend its energy constraints or accept its limits. Fusion is the most ambitious transcendence available: a technology that would change everything if it works.

The dreamers are building tokamaks. The trainers are burning watts. Everyone's hoping the timelines align.

Further Reading

- Hurricane, O. A., et al. (2023). "Achieving Lawson Criterion in an Inertially Confined Fusion Implosion." Nature. - Ball, P. (2021). "The Chase for Fusion Energy." Physics World. - Commonwealth Fusion Systems. (2024). "SPARC Project Overview."

This is Part 8 of the Intelligence of Energy series. Next: "Reversible Computing: Escaping Landauer."