261024 – Molten Salt Thorium Reactors – A Technical Assessment

by | Oct 24, 2026 | Physics | 0 comments

Molten Salt Thorium Reactors: A Technical Assessment

And whether they can power waterless data centers in remote places

Dr. Thomas Lee Abshier, ND — 23 October 2026

1. Why this analysis

Thorium molten salt reactors have become a popular answer to two problems at once: the need for abundant, clean, firm electricity, and the damage that AI data centers are doing to the communities they are built beside. A widely read September 2026 essay (“How the Law of Displacement Can Solve the Data Center Crisis,” The Forgotten Side of Medicine) proposes moving the gigawatt-scale data centers to remote, cold, sparsely populated places such as Alaska’s North Slope, powering them behind the meter, first with natural gas and eventually with thorium molten salt reactors.

This site has not yet published a technical assessment of the molten salt thorium reactor itself. This article provides one. It asks three questions:

  1. What does the thorium fuel cycle’s physics promise, and what does it demand?
  2. What engineering problems stand between the 1960s Oak Ridge demonstration and a commercial plant?
  3. Is it physically sound to place data centers in remote areas, power them with such reactors, and cool both the computers and the reactor without consuming fresh water?

My conclusion, stated at the outset: the physics is sound, and the remote, waterless, off-grid configuration is not only possible but is the natural fit for this kind of reactor. The obstacles are chemical, metallurgical, industrial, and regulatory, not physical. They are solvable, but they take years, and the first reactors to serve remote data centers will almost certainly be salt-cooled uranium designs, with thorium following as its fuel cycle is built.

2. The physics of the thorium fuel cycle

Fertile, not fissile

Natural thorium is almost entirely one isotope, Th-232. It does not sustain a chain reaction. It is fertile: when it absorbs a neutron it becomes fuel through a two-step decay chain:

Th-232 + n → Th-233 → (β⁻, 22 minutes) → Pa-233 → (β⁻, 27 days) → U-233

U-233 is fissile, and it is an excellent fissile isotope for a slow-neutron (thermal) reactor. The key figure of merit is η (eta), the average number of neutrons released per neutron absorbed in the fuel. For thermal neutrons:

Fissile isotope η (thermal)
U-233 ≈ 2.29
U-235 ≈ 2.07
Pu-239 ≈ 2.11

To breed (make at least as much fuel as you burn), a reactor needs one neutron per absorption to continue the chain and one more to convert a fertile atom into a new fissile atom. That requires η > 2, with the surplus covering every leakage and parasitic capture. U-233 is the only fissile isotope that clears this bar in a thermal spectrum. That is why thorium is interesting: it enables a breeder that runs on slow neutrons in a moderated, low-pressure, relatively gentle reactor, rather than requiring the fast-neutron, liquid-sodium designs the uranium-plutonium breeder program pursued.

But the margin is thin. Of the 2.29 neutrons, roughly 2.0 are committed. The remaining ~0.29 must cover every neutron lost to leakage, to structural materials, to the salt itself, to fission products, and to protactinium. Every engineering choice in a thorium breeder is ultimately a decision about how to spend that 0.29.

How much fuel

Each fission releases about 200 MeV, which works out to roughly 0.95 megawatt-days of heat per gram of heavy metal fissioned. A reactor delivering one gigawatt of electricity for a year at 45 percent thermal efficiency fissions about 0.85 tonnes of material. In a true thorium breeder at equilibrium, that is very nearly the thorium consumed: a little under one tonne per gigawatt-year.

By comparison, a conventional light-water reactor at 33 percent efficiency fissions about 1.2 tonnes per gigawatt-year, but it must load roughly 20 tonnes of enriched fuel, mined from about 150 to 200 tonnes of natural uranium, because it uses only a small fraction of the uranium it handles. Thorium is also about three times as abundant in the Earth’s crust as uranium and is already extracted as an unwanted byproduct of rare-earth mining. On fuel supply alone, the thorium breeder is effectively inexhaustible.

Waste

Because thorium (mass 232) starts lower on the ladder of neutron captures than U-238, far fewer neutron captures are needed to make the heavy transuranic elements (plutonium, americium, curium) that dominate the long-term radiotoxicity of conventional spent fuel. A thorium cycle therefore produces much less long-lived transuranic waste. It does not escape fission products: per unit of energy, every fission reactor produces about the same mass of them. Most decay to low levels within roughly three centuries; a few (technetium-99, iodine-129) are long-lived and must still be isolated.

3. Why liquid fuel

Thorium can be used in solid fuel, but its natural partner is the molten salt reactor (MSR), where the fuel is dissolved in a fluoride salt that also carries heat. The canonical salt is “FLiBe” (lithium fluoride and beryllium fluoride), with thorium and uranium fluorides dissolved in it.

The physical advantages follow directly from the fluid:

  • Low pressure. The salt boils at well over 1,400 °C, so a reactor running at 600–700 °C operates near atmospheric pressure. A light-water reactor runs at about 150 atmospheres; much of its cost and accident potential come from containing that pressure. A salt reactor has no stored pressure to drive radioactivity outward.
  • Negative temperature feedback. When the salt heats, it expands, pushing fuel out of the core and lowering reactivity. The reactor throttles itself.
  • Passive shutdown. Active cooling keeps a plug of frozen salt in a drain line frozen. If power is lost or the salt overheats, the plug melts, and gravity drains the fuel into subcritical tanks designed to shed decay heat passively.
  • No fuel meltdown in the usual sense. The fuel is already liquid; there is no fuel cladding to fail.
  • Gaseous fission products leave on their own. Xenon-135, the strongest neutron poison in any reactor, bubbles out of the salt and an off-gas system removes it. This improves the neutron economy, which matters greatly for a thorium breeder, and removes the xenon transients that complicate solid-fuel reactor operation.
  • Online refueling. Fuel can be added and the salt adjusted while the reactor runs.

These are genuine advantages, not marketing. The Molten Salt Reactor Experiment (MSRE) at Oak Ridge National Laboratory, an approximately 8 MW-thermal graphite-moderated reactor, demonstrated these in principle and operated from 1965 to 1969. In 1968, it became the world’s first reactor to run on U-233 (though that U-233 was produced elsewhere, not bred in the reactor). It didn’t breed; it had no thorium blanket, and it didn’t demonstrate the online chemical processing a breeder needs.

4. The hard problems

Here is where honesty matters most. The Oak Ridge program was canceled in the early 1970s, and the most commonly told story is that thorium lost for political reasons: it didn’t make weapons-grade plutonium, and the Atomic Energy Commission instead committed to the liquid-metal fast breeder. That account is substantially true. But the program also failed to solve several real problems, and those problems remain.

4.1 Protactinium-233

The intermediate isotope Pa-233 has a 27-day half-life and a substantial appetite for thermal neutrons (a capture cross-section of roughly 40 barns). Every Pa-233 nucleus that captures a neutron before decaying is a double loss: the neutron is gone, and the atom that would have become U-233 becomes Pa-234 instead, which leads to U-234, a non-fissile isotope. In a high-flux thermal breeder, this loss alone can consume much of the 0.29-neutron margin.

The Oak Ridge Molten Salt Breeder Reactor design (1 GWe, breeding ratio about 1.06) solved this on paper by continuously drawing salt through a chemical processing plant that extracted protactinium on a few-day cycle, held it outside the core while it decayed, and returned the resulting U-233. The extraction step uses reductive extraction into molten bismuth, a technically demanding, high-temperature chemical process demonstrated at laboratory scale but never at plant scale. Alternatives include a two-fluid design (thorium in a separate blanket salt, outside the high-flux region), lower power density, or accepting a conversion ratio below one and periodically adding fissile material. Each trades breeding performance for simplicity.

A second, less discussed consequence follows. Protactinium separated from the reactor and allowed to decay yields nearly pure U-233, without the U-232 contamination that otherwise makes reactor-bred U-233 intensely gamma-radioactive and hard to steal or weaponize. Protactinium separation is therefore the weakest point in the thorium cycle for proliferation resistance. Any design that separates protactinium must account for this.

4.2 Online chemical processing

Beyond protactinium, a true breeder must continuously remove fission products, especially neutron-absorbing lanthanides, and keep the salt chemistry within tight limits. The reactor is, in effect, joined to a small, radioactive, high-temperature chemical plant. No one has built that plant at scale. This is the single largest technical gap between the thorium breeder and a working commercial machine. Most current designs sidestep it by not breeding: they run as converters (making some but not all of their own fuel) or burn low-enriched uranium in salt and leave thorium for later.

4.3 Materials

The fluoride salts are chemically stable but corrosive to most metals at 650–700 °C, especially with impurities or oxidizing conditions. Oak Ridge developed Hastelloy-N, a nickel-molybdenum alloy, which performed well, but the MSRE revealed shallow intergranular cracking caused by fission-product tellurium. Later work showed that controlling the salt’s redox state (the ratio of UF₄ to UF₃) and modifying the alloy with niobium or titanium could suppress the cracking. These fixes are promising but lack decades of service data.

Graphite, the moderator, swells and then shrinks under fast-neutron bombardment and must be replaced every several years in a high-power core. That is a design and maintenance requirement, not an obstacle, but it shapes reactor size and cost.

4.4 Lithium-7

Natural lithium contains about 7.5 percent lithium-6, which absorbs neutrons voraciously and produces tritium. The salt must therefore use lithium enriched to better than 99.99 percent lithium-7. The United States stopped producing enriched lithium in 1963, and today China and Russia supply it. A single large molten salt reactor needs tens of tonnes. Domestic lithium-7 enrichment capacity is a prerequisite for an American thorium program, and it should be treated as such.

4.5 Tritium and beryllium

Even with highly enriched lithium-7, neutron reactions with lithium and beryllium produce tritium, a radioactive hydrogen isotope that readily diffuses through hot metal, including heat exchanger walls. Tritium management (capture, barriers, permeation control) is an active engineering problem for every FLiBe system. Beryllium is also a chemical toxin; industry has well-understood handling, but it adds cost and procedure.

4.6 A liquid that is highly radioactive

The same fluidity that eliminates meltdown means the primary system carries highly radioactive liquid through pumps, pipes, and heat exchangers. A leak does not explode or flash to steam, and the salt freezes on contact with cooler surfaces, but maintenance of a radioactive liquid loop must be done remotely. This is manageable and was managed at Oak Ridge, but it is a real cost.

4.7 Licensing

No licensing framework anywhere in the world exists for a liquid-fueled thorium reactor with online processing. American regulation grew up around solid-fuel light-water reactors. New pathways (Department of Energy authorization of test reactors and statutory deadlines for NRC decisions) have shortened the road for demonstration machines, but a commercial liquid-fuel plant will require rules that don’t yet exist.

5. What has actually been demonstrated

It is important to distinguish three different things that are often blurred together.

A liquid-fueled thorium reactor. China’s TMSR-LF1, a 2 MW-thermal experimental molten salt reactor in the Gobi Desert near Wuwei, Gansu Province, reached criticality in October 2023 and full power in June 2024. In November 2025, the Shanghai Institute of Applied Physics announced the first thorium-to-uranium conversion in an operating molten salt reactor. The reactor runs on uranium enriched below 20 percent, with about 50 kg of thorium and a conversion ratio of about 0.1. In plain terms, it proves the chemistry and the conversion in a working machine; it does not yet breed, and it is very small. China has announced a ladder of larger demonstration units: roughly 10 MW-electric (60 MW-thermal) around 2030 and a 100 MW-thermal unit targeted for 2035. Notably, China sited the reactor in the desert because a salt reactor doesn’t need a river or coastline for cooling.

A salt-cooled reactor with solid fuel. America’s leading salt program, Kairos Power, uses FLiBe as a coolant but keeps its uranium fuel in solid TRISO pebbles. Its Hermes test reactor (35 MW-thermal, no power production) is under construction in Oak Ridge, with completion now expected in 2028. Its Hermes 2 plant, which will produce about 50 MW of electricity for the TVA grid serving Google’s data centers, broke ground in April 2026 and is expected to operate by 2030. This is not a thorium reactor or a liquid-fuel reactor, but it is building the industrial capabilities a thorium program will need: purified fluoride salt production, salt handling, high-temperature heat exchange, and a licensing precedent for a non-water reactor.

Non-nuclear prototypes. Companies such as Copenhagen Atomics (Denmark) and Flibe Energy (United States) are developing liquid-fuel thorium designs and testing full-scale non-nuclear salt loops, with nuclear testing no earlier than the late 2020s.

The honest summary: as of late 2026, the world’s only operating thorium molten salt reactor is Chinese and produces 2 MW of heat. No one has operated a thorium breeder with online processing.

6. Cost

Published techno-economic models place mature molten salt thorium power among the cheapest sources of electricity, roughly $20–55 per megawatt-hour. Those numbers project mature, serially manufactured plants. No commercial plant exists. Every first-of-a-kind nuclear plant in history has cost more than projected, and these will too.

The physical reasons to expect low eventual costs are nevertheless real:

  • No pressure vessel rated for 150 atmospheres, and much smaller containment.
  • High operating temperature, giving higher thermal efficiency (about 45 percent versus about 33 percent for light-water reactors), so less reactor per kilowatt delivered.
  • Negligible fuel cost: under a tonne of thorium per gigawatt-year.
  • Factory-built modules, because the low-pressure components are lighter and simpler.
  • Freedom from cooling-water siting constraints, which opens cheap, remote land (see §8).

The cost will depend on how quickly chemical processing, materials qualification, lithium-7 supply, and licensing are resolved, and on whether plants can be built in series rather than one at a time.

7. Cooling the data center without water

Now to the question the reader asked: can remote data centers use power to run cooling systems rather than consuming fresh water? The answer is yes, and the physics is straightforward.

The heat budget

Essentially all of the electricity a data center consumes becomes heat inside the building. A one-gigawatt computing load is a one-gigawatt heater that must be cooled continuously. There are two ways to get rid of that heat.

Evaporation. Evaporating water carries away about 2.43 megajoules per kilogram. If all of a gigawatt were rejected by evaporation, it would consume about 410 kilograms of water per second, or about 9.4 million gallons per day, before counting the “blowdown” water discharged to keep minerals from building up. Evaporation is energy-cheap, which is why it became the industry default, but it is water-expensive, and the source of the water conflicts communities are experiencing.

Dry (sensible) cooling. The alternative is a closed loop: a coolant circulates through the computers, then through outdoor heat exchangers (essentially large radiators) where fans blow air across them. No water is consumed after the initial fill. To carry away one gigawatt while warming the air by 15 °C requires about 66,000 kilograms of air per second (about 53,000 cubic meters per second). With realistic fan pressures and efficiencies, the fans draw roughly 13 megawatts, about 1.3 percent of the load. That is a modest price to eliminate water consumption entirely.

When is refrigeration needed?

The reader’s phrasing, “using power to run refrigerant systems,” describes the hot-climate version of dry cooling. A refrigerant chiller (a heat pump) is needed only when the outdoor air is too warm to absorb the heat directly. In a hot summer climate, air-cooled chillers might consume 20 to 35 percent of the computing load, which is a serious energy penalty; industry analyses commonly cite 10 to 20 percent penalties for dry cooling across typical U.S. climates.

In a cold, remote site, this penalty nearly disappears. Modern AI accelerators are increasingly liquid-cooled at the chip, with coolant supplied at around 30–45 °C. Any time the outdoor air is cooler than that, which on Alaska’s North Slope is essentially always (the proposed site there averages about 12 °F), heat flows to the air through radiators with no compressor at all. The cold site does not need refrigeration; it needs fans, pumps, and freeze protection (glycol in the outdoor loop). The engineering problem in the Arctic is the opposite of the one in Texas: keeping the outdoor coolers from freezing and managing very dry air indoors.

So the reader is right, and in a cold place the idea is even better than stated: the remote data center can run water-free on radiators, spending perhaps 1–3 percent of its power on fans and pumps. In a hot remote place (a desert), it can still run water-free using refrigeration, at a larger energy cost.

What dry cooling does not eliminate

Two caveats: First, if the extra electricity for dry cooling comes from a water-cooled power plant elsewhere, some water use is simply moved off-site. That is why the reactor’s own cooling matters (§8). Second, the large fans and pumps are themselves sources of low-frequency noise. In a remote site with no neighbors, this matters little; near homes, it does.

8. Cooling the reactor without water

Every heat engine must reject heat to a cold reservoir. This is the Second Law, and it applies to the reactor as much as to the computers.

A conventional light-water reactor delivers steam at about 300 °C. Its efficiency is about 33 percent, so for every gigawatt of electricity it rejects two gigawatts of heat, usually to a river, lake, ocean, or evaporative cooling tower. Dry-cooling a light-water reactor is possible but costly: the lower steam temperature drops efficiency by several points when the cold side is air instead of water, and the radiators must be enormous.

The molten salt reactor changes this because of its temperature. With a reactor outlet of 650–700 °C, the heat can drive a gas Brayton cycle (the same thermodynamic cycle as a jet engine or gas turbine), either an open air cycle or a closed supercritical-CO₂ cycle. Brayton cycles naturally reject heat to air. High-temperature salt reactors have been studied for decades precisely because they make dry cooling affordable and free nuclear plants from riverbanks and coastlines. Developers advertise efficiencies above 45 percent with either dry or wet cooling.

The cold site helps here too. The ideal (Carnot) efficiency of an engine running from 700 °C (973 K) to a 35 °C summer heat sink is 68 percent; to a −10 °C Arctic sink it is 73 percent. Real cycles capture only part of this, but the direction is unambiguous: the colder the sink, the more electricity per unit of heat. For supercritical-CO₂ cycles in particular, cold ambient air lets the compressor inlet approach the CO₂ critical point (31 °C), which is where those cycles are most efficient, something difficult to achieve with air cooling in a warm climate.

In a cold region, the rejected heat is also a resource rather than a waste: it can warm buildings, greenhouses, and nearby communities.

9. A worked example: one gigawatt of computing in the Arctic

Assume a 1 GW computing load on Alaska’s North Slope.

Quantity Value
Computing load 1.00 GW
Cooling overhead (fans, pumps; PUE ≈ 1.1) 0.10 GW
Electricity required 1.10 GW
Reactor thermal output at 45% efficiency 2.44 GW-thermal
Heat rejected by the reactor (to air) 1.34 GW
Heat rejected by the data center (to air) 1.10 GW
Fresh water consumed, steady state near zero (makeup only)
Thorium consumed (equilibrium breeder) about 0.95 tonnes per year
Grid connection to residential customers none

Every watt of that energy ends as warm air over empty tundra. No river is drawn down, no town’s water pressure falls, no residential ratepayer funds the generation or transmission, and no household hears the fans.

A note on electrical quality belongs here as well. Large data centers draw power through tens of thousands of switching power supplies, which inject harmonic distortion onto the wiring that serves them. On a shared grid, that distortion can spread to neighboring customers. On an isolated, behind-the-meter system, the generator and the load share their own network; the distortion is an internal engineering problem with well-known solutions (active harmonic filters, multi-pulse rectifiers, properly sized transformers), and no neighbor is connected to it at all. Whatever one concludes about the health questions raised about “dirty electricity,” the off-grid configuration removes the exposure.

10. Timeline, and what to build in the meantime

The order of events, if pursued seriously, looks like this:

  1. Now to about 2030: gas and salt-cooled uranium. Remote campuses start on local natural gas (on the North Slope, gas currently reinjected because there’s no pipeline) or on the first small reactors. A 1–3 GW gas-powered campus has been proposed near the Dalton Highway south of Deadhorse; the state issued a preliminary decision in favor of a 50-year lease in May 2026, and public comment ran heavily against it. The lesson is that remoteness alone does not secure consent: the lease terms, the gravel, the gas, and who benefits all matter.
  2. Build the thorium supply chain in parallel, not afterward. Develop lithium-7 enrichment, fuel-salt production, thorium separation from rare-earth processing, and, above all, online chemical processing and protactinium handling while the first reactors run, because they set the pace.
  3. Late 2020s to early 2030s: an American thorium test reactor. With Department of Energy authorization, a liquid-fueled thorium test reactor could operate within a few years of a funded decision. It must then accumulate years of data on salt chemistry, corrosion, tritium, and processing before a commercial unit follows.
  4. Mid-2030s onward: commercial thorium power, sited first where it competes best: remote, cold, off-grid sites with large, steady loads and customers able to pay for first-of-a-kind plants. Data centers are such customers.

China’s published schedule puts its demonstration units in roughly this same window. A nation that begins only after China’s are running will buy the technology rather than build it.

11. Conclusion

The thorium molten salt reactor is not speculative. Its central physics is sound: U-233’s neutron yield permits thermal breeding; liquid fuel at atmospheric pressure removes the stored energy that drives light-water accidents; high temperature permits efficient, air-cooled power conversion; and a tonne of thorium a year can power a gigawatt. What remains unproven is the chemistry and metallurgy of running such a reactor for decades, the chemical plant a true breeder requires, and the industrial supply chain beneath it, starting with lithium-7.

The reader’s intuition is correct. Remote data centers powered by their own reactors and cooled by air instead of water are physically sound, and the molten salt reactor is especially well suited to that role because it, too, can be cooled by air. In a cold place, neither the computers nor the reactor needs refrigeration or fresh water; both need radiators and fans. The barrier is not physics. It is the decision to build what Oak Ridge demonstrated sixty years ago and then set aside.

Sources and further reading

  • A Midwestern Doctor, “How the Law of Displacement Can Solve the Data Center Crisis,” The Forgotten Side of Medicine (Substack), September 6, 2026.
  • World Nuclear Association, “Molten Salt Reactors” (information library): https://world-nuclear.org/information-library/current-and-future-generation/molten-salt-reactors
  • World Nuclear News, “Chinese molten salt reactor achieves conversion of thorium-uranium fuel”: https://www.world-nuclear-news.org/articles/chinese-msr-achieves-conversion-of-thorium-uranium-fuel
  • POWER Magazine, “China’s Molten Salt Reactor Reaches Thorium-Uranium Conversion Milestone”: https://www.powermag.com/chinas-molten-salt-reactor-reaches-thorium-uranium-conversion-milestone/
  • Rykhlevskii, Lindsay & Huff, “Online Reprocessing Simulation for Thorium-Fueled Molten Salt Breeder Reactor,” University of Illinois, 2017: https://arfcgroup.web.engr.illinois.edu/pres/2017-10-30-online.pdf
  • Forsberg et al., MIT/UC Berkeley, “High-Temperature Salt-Cooled Reactor for Power and Process Heat”: https://neup.inl.gov/content/uploads/14/2024/11/11-3272_Technical_AbstractMIT.pdf
  • Flibe Energy, LFTR specifications: https://flibe.com/lftr/
  • Nuclear Engineering International, NRC extends Hermes 1 construction deadline: https://www.neimagazine.com/?p=104127
  • American Nuclear Society, “Kairos Power breaks ground on first power-producing reactor in Oak Ridge,” April 2026: https://www.ans.org/news/tag-hermes%201/
  • Trane Technologies, data center water conservation white paper (closed-loop and dry cooling): https://elibrary.tranetechnologies.com/public/commercial-hvac/Literature/White%20Paper/DC-WPR004A-EN_04162026.pdf
  • GHD Nexus, “Cooling the hysteria: Reframing the narrative around data centre water usage,” March 2026: https://nexus.ghd.com/read—cooling-the-hysteria
  • Alaska Beacon, “A huge data center could rise on Alaska’s North Slope,” May 2026: https://alaskabeacon.com/2026/05/14/a-huge-data-center-could-rise-on-alaskas-north-slope/
  • Alaska Beacon, “Data center critics flood Alaska land managers with opposition to North Slope project,” July 2026: https://alaskabeacon.com/2026/07/06/data-center-critics-flood-alaska-land-managers-with-opposition-to-north-slope-project/

Submit a Comment