A fusion plant costing study usually carries power supplies as one line, and the line hides a count. For a magnetic confinement plant with high-temperature superconducting (HTS) coils, that count is in the tens of thousands per plant. The supply chain question for a fleet is therefore not “can we build the supplies” but “can we build them at a rate the fleet will actually order,” and the answer depends on arithmetic that rarely appears in a cost estimate.
Counting units at one plant
Take the example I use in the power supply market analysis for a 2,000 MW plant: nine coil types, 14 individual coils, and a total coil current of 391.2 MA [3]. A 4 mm HTS tape carries about 110 A at the margins used for design, so a coil winding is thousands of parallel tapes, and the current is split into parallel branches. At about 20 kA per regulated supply, each branch needs its own unit for current sharing, fault isolation and quench protection. Dividing 391.2 MA by 20 kA gives about 19,560 units. The analysis carries roughly 19,500 once individual coils are rounded, before any spares.
Redundancy raises the count. Critical coil systems are specified N+1 or N+2, and every unit includes its own quench detection and fast energy dump. The unit count is therefore a design choice as much as a physics result. Doubling the per-unit rating to 40 kA halves the number of units and roughly doubles the stress on each one, and the right answer depends on tape current sharing and fault isolation, not on a count target.
In the cost account structure, this is Account 22.1.7, and within it the coil supplies are the largest single share of the magnetic confinement power supply cost [3][4]. A line that large, made of that many identical items, behaves more like a manufactured product than a construction item.
Most of the learning happens inside the first plant
Wright’s Law says unit cost after N units is C(N) = C1 · N^b, with b = log2(1 − L) for a learning rate L per doubling of cumulative production [1]. The useful quantity is the number of doublings, and here the counts are so large that the first plant does most of the work.
One plant’s 19,560 units is log2(19,560) = 14.3 doublings. A fleet of 100 plants is about 1.96 million units, or 20.9 doublings. The step from the first plant to the hundredth adds only log2(100) = 6.6 doublings. At a 15% learning rate, the learnable part of unit cost falls to about 10% of its starting value by the end of the first plant, and to about 3% by the end of the hundredth. Going from plant 1 to plant 100 improves the learnable part by a factor of about three.
That calculation assumes the supplies are one design, with experience accruing to one manufacturer. It also leaves out the cost floor, which for power electronics is real: semiconductors, magnetics, copper and enclosures follow materials and commodity learning, which is far slower than assembly labor. With half of unit cost on the floor and a 15% rate on the rest, unit cost is about 55% of the first unit after one plant and about 52% after a hundred. The costing in the market analysis takes the NOAK coil supply cost at half of FOAK [3], which is consistent with a floor-dominated curve and not with continued decline.
The implication is easy to miss. For coil supplies, fleet size beyond the first plant buys little additional unit cost reduction. What the fleet changes is the rate at which units must be delivered.
Fleet scale is a capacity problem
Consider a fleet that completes ten plants per year, a deliberately modest number for a mature industry. That is about 195,600 units per year. Over 250 working days it is about 780 units per working day, or roughly 49 per hour on a 16-hour, two-shift schedule. Each unit is a regulated 20 kA converter with protection, controls and a factory acceptance test. The test stations, the burn-in time and the power semiconductor supply become the constraint well before the design does.
Three things follow from the arithmetic, and each can be checked against a spreadsheet in an afternoon.
First, test and burn-in capacity scales with units per day times hours per unit, not with unit cost. A supplier that spends a day per unit on acceptance testing needs several hundred test positions to ship at fleet rates.
Second, the supply of power semiconductors and magnetics is shared with every other industry buying them. A fusion fleet demanding millions of units competes in that market and cannot assume its own pricing.
Third, the unit design must be frozen early. A change to the 20 kA unit interface resets the experience count that justified the cost curve. With this many units, a mid-program redesign costs a large part of the learning already earned.
Pulsed concepts count shots, not coils
Pulsed concepts replace the coil count with a replacement flow. The literature I reviewed for the market analysis puts commercial capacitor lifetimes at roughly 10^5 to 10^7 shots at full energy, and plant lifetimes at 10^8 to 10^9 shots [3]. At 0.1 to 1 Hz and 80% availability, a plant fires about 2.5 million to 25 million shots per year. A bank rated for 10^7 shots at 1 Hz needs replacing in under half a year. The same bank at 0.1 Hz lasts about four years.
So the number of capacitor banks a pulsed fleet buys depends on repetition rate and rated lifetime as strongly as on plant count. Replacement volume is repeat business that scales with the installed base, and it accumulates experience for the supplier in the same way new construction does. A design that makes banks and switches replaceable without disturbing the rest of the assembly turns that flow into a standard product.
Practical takeaway
When someone hands you a power supply line in a plant estimate, ask for four numbers: units per plant, units per year at fleet rates, the learning rate and floor fraction behind the NOAK price, and the design freeze date for the unit. If the count is missing, the NOAK price is a guess about a factory nobody has sized. If a supplier cannot say how many units it can test per day, the fleet schedule is optimistic whatever the unit price.
If you are working through the unit count and capacity for a specific magnet or pulsed power system, get in touch through woodruffeng.com.
Also published on LinkedIn: https://www.linkedin.com/pulse/how-many-power-supplies-does-fusion-fleet-need-simon-woodruff-qog5e/
References
1. T. P. Wright, “Factors Affecting the Cost of Airplanes,” Journal of the Aeronautical Sciences 3(4), 122–128 (1936). https://doi.org/10.2514/8.155
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
3. Fusion Advisory Services Ltd, “Power Supplies for Fusion: Market Analysis and Technical Review,” March 23, 2026 (unit counts, NOAK/FOAK ratio and capacitor lifetime ranges as cited in the report).
4. S. Woodruff, “A costing framework for fusion power plants,” arXiv:2601.21724 (2026). https://arxiv.org/abs/2601.21724