Ask ten people in fusion what a first-of-a-kind (FOAK) plant costs and you will get ten answers, most of them defended with more energy than the question deserves. The better question is which accounts in the plant will fall between FOAK and nth-of-a-kind (NOAK), which will not, and what you can do at the component level now to move an account from the second group to the first.
A plant is a stack of accounts, not a number
Plant-level costing only becomes useful when it is decomposed into a standard chart of accounts. The framework I consolidated from ARPA-E work between 2017 and 2024 maps every subsystem estimate onto an account structure in the IAEA, GEN-IV EMWG and EPRI lineage, so that magnets, power supplies, buildings and turbine plant can each be examined on their own [1]. In the worked 636.75 MWe (net) example in that paper, direct costs (Account 20) are 71% of total capital cost. Within them, structures and site (Account 21) are 12% and reactor plant equipment (Account 22) is 31%. Turbine plant, electric plant, miscellaneous plant and heat rejection add roughly another 15% between them. On the indirect side, owner’s cost, supplementary costs and financial costs are 8%, 10% and 8% [1].
Those figures are a NOAK example, not a prediction for any particular machine. What matters is the shape. A large share of plant capital sits in accounts that are not the fusion power core at all, and a debate that treats FOAK as a single number averages together accounts that behave very differently.
Accounts that learn
Some accounts fall steeply with repetition because their cost is mostly engineering hours, process development and risk. They do not depend much on the mass of material in the product.
Indirect costs are the clearest case. Earlier ARPA-E work on buildings, turbine halls, heat exchangers, site works and construction management found that non-power-core costs often dominate total plant cost, and that modularization measurably changes both direct and indirect cost [1]. Indirects are a design outcome, not a fixed percentage. A plant designed so that large assemblies are built in a factory and set on site carries less field labor, less schedule risk and less owner’s cost than one built in place. Contingency is the same story: in the framework’s NOAK mode, the contingency on buildings is switched off, because repetition is exactly what retires it [1].
Bespoke fabrication also learns. A one-off magnet winding or a first-run power supply carries tooling, qualification and rework that the second hundred units do not. This is ordinary experience-curve behavior, and it is why the framework treats modularization, centralized manufacturing and learning as explicit cost-reduction levers [1].
Accounts that do not
Other accounts carry a floor set by physics and commodity prices, and no amount of production experience takes them below it.
A resistive or superconducting magnet needs a certain amount of conductor to carry a certain current around a certain volume, and that amount follows from Ampère’s law and the field you want. A capacitor bank needs stored energy of at least the pulse energy divided by its round-trip efficiency, and the dielectric volume per joule is set by the material. A steam cycle needs a heat sink, and the heat to be rejected follows from the thermal power and the cycle efficiency. Learning shaves the labor and margin around those quantities. It does not shave the quantities.
Buildings behave similarly. The framework scales them linearly with gross electric power [1], so a more compact core does not shrink the turbine hall. This is one reason the retrofit pathway in the same paper is interesting: reusing an existing steam cycle, turbine island and cooling infrastructure could avoid much of the balance-of-plant cost for a FOAK plant [1]. Reuse attacks the floor directly, where learning cannot.
Where the component level meets the plant level
This is where hardware choices matter, and where I think costing conversations go wrong most often. A component choice moves an account in two ways: it changes the floor, and it changes how fast the account learns.
Take power supplies, Account 22.1.7, which is 5% of capital in the worked example. In that paper the account is scaled from an ITER reference with a learning credit factor [1]. The reference is a bespoke, low-volume design. A pulsed concept built from standard capacitor modules, switches and chargers that are used across many units starts on a different curve, because the unit being repeated is small and the repetition count is large. Two designs with the same FOAK price can have very different NOAK prices for this reason alone.
The same logic applies to magnets. A modular, stackable construction makes the repeating unit a coil section, so production experience accumulates in the factory and not in the field. A one-piece custom winding hides the learning inside a single expensive build. The FOAK quote for each could be identical, and the NOAK quotes will not be.
Practical takeaway
When you review a plant estimate, do three things before you argue about the total. First, ask whether the estimate is FOAK or NOAK, since the framework treats the two as separate modes with different contingency [1]. Second, for each major account, separate the part set by material mass and physics from the part set by labor, tooling and risk. Third, ask what the repeating unit is, and how many of them the plant, and then the fleet, will contain.
An estimate that cannot answer those three questions is not yet an estimate of NOAK cost, however carefully the FOAK number was built. If you want to work through these questions for a specific magnet or pulsed-power design, get in touch through woodruffeng.com.
References
[1] S. Woodruff, “A costing framework for fusion power plants,” arXiv:2601.21724 (2026). https://arxiv.org/abs/2601.21724
[2] S. Woodruff, A. Durham, A. Higginbottom, C. Raastad, “Extension of the fusion power plant costing standard,” arXiv:2602.19389 (2026). https://arxiv.org/abs/2602.19389
Also published on LinkedIn: https://www.linkedin.com/pulse/foak-cost-wrong-number-argue-simon-woodruff-7hrre/
By Dr. Simon Woodruff, CEO — Woodruff Engineering Inc.