Hydrogen cost forecasts often borrow one of the clean-energy transition’s most useful empirical tools: the experience curve. If a technology has historically fallen by a given percentage in cost every time cumulative deployment doubles, analysts can estimate how many doublings lie ahead and compound the learning rate into a future cost. That works best when a doubling of installed capacity corresponds reasonably well to another round of manufacturing and installing comparable units. As I argued in my earlier analysis of hydrogen learning rates, electrolyzers and complete hydrogen plants belong to a very different economic reference class from solar modules and batteries. The newer evidence exposes an additional problem: installed gigawatts can rise partly by making stacks, modules, shared process equipment and whole plants larger, so a capacity doubling does not necessarily represent a comparable doubling of manufacturing experience. The distinction shows up clearly in a 2025 European electrolyzer-project study that assembled project capital-cost and capacity data going back to 2005. The raw data produce striking experience rates: costs fall 23.3% across all projects, 32.1% for PEM and 22.9% for alkaline electrolysis for each doubling of cumulative installed capacity. Those numbers look comparable with, or better than, historical curves for solar and batteries. But when the researchers normalize project costs for estimated project-size economies, the rates fall to 13.3%, 17.6% and 7.3% respectively, with the remaining alkaline relationship no longer statistically significant. Costs really did decline; the raw curve was simply attributing several different mechanisms to “learning.” The deeper problem is what a doubling actually represents. The full TFIE Strategy Briefing analysis follows stack-size growth, chemical-plant scale, manufacturing repetition and the system boundary that a defensible hydrogen-cost forecast needs to use. Chemical-plant scale economies are substantial but front-loaded. A 100 MW hydrogen plant does not require one hundred times the compressors, transformers, water-treatment systems, cooling equipment, gas purification or engineering of a 1 MW plant. Larger projects share equipment, spread fixed engineering across more output and move common systems toward economically sensible sizes. That can produce dramatic reductions as demonstrations become industrial plants. Once compressors, transformers and other process equipment approach practical scale, however, further capacity increasingly comes from replicating optimized process trains. The next plant still benefits from better procurement, standardized designs and more experienced contractors, but it does not repeatedly capture the entire gain from moving out of demonstration scale. The denominator becomes more problematic when the stacks themselves grow. Suppose cumulative electrolysis capacity rises from 5 GW to 50 GW. That is a tenfold increase, or 3.32 doublings of installed capacity. If the average finished stack remains 1 MW, completed stack count also rises tenfold, from 5,000 to 50,000, so capacity and stack-count doublings track one another. If average stack size rises from 1 MW to 5 MW during the same expansion, only 10,000 stacks are required at 50 GW. Installed capacity still records 3.32 doublings, but completed stack count has doubled only once. The factories are gaining experience, but not at the rate implied by cumulative gigawatts. Finished stack count is not a perfect manufacturing denominator either. Larger stacks contain repeated cells, membranes, plates and electrode area, and those components have their own manufacturing experience. Better materials and higher current density can also increase output without requiring proportional increases in material. The narrower point is enough: during rapid equipment upscaling, installed megawatts can materially overstate the repetitions occurring at several important manufacturing levels. An experience curve based only on cumulative GW hides that changing physical architecture. The system boundary creates another mathematical constraint. According to the IEA’s 2025 electrolyzer cost breakdown, the stack represents only about 15–20% of installed capital cost. Roughly 25–30% sits in balance-of-plant equipment such as power electronics, piping, compressors and gas treatment, while engineering, procurement, construction and contingency can account for more than half. Even a 20% cost reduction in a component representing one fifth of total CAPEX removes only about 4% from the installed project cost. The rest of the project can improve too, but compressors, civil works, electrical connections and construction have their own scale and productivity trajectories rather than inheriting the stack factory’s learning rate. None of this implies that electrolyzer projects will remain as expensive as today’s first-generation installations. Manufacturers can improve stack designs and materials, EPC contractors can standardize layouts, procurement can become more efficient, shared equipment can reach sensible scale and repeat construction can eliminate first-of-a-kind mistakes. The mistake is taking all of those mechanisms, collapsing them into one historical learning rate per installed-capacity doubling and compounding that number mechanically toward 2035 or 2050. Several of the largest early reductions occur because the industry is making the one-time transition from small demonstration projects to properly scaled industrial facilities. Electrolyzer CAPEX is also only part of the eventual hydrogen price. Electricity dominates the variable cost, and manufacturing learning cannot eliminate it. Very cheap wind and solar electricity tends to be intermittent, while capital equipment benefits from high utilization. Compression, storage and distribution come afterward and have their own infrastructure economics. Better equipment will reduce costs, but it cannot turn every part of the production and delivery chain into a mass-manufactured module with a solar-style experience curve. A defensible hydrogen forecast therefore has to separate manufacturing improvement, electrochemical performance, stack-size effects, chemical-plant scale, repeat engineering and construction from electricity, utilization, financing and logistics, then model each on its own terms. Hydrogen will get cheaper than many early projects. The newer evidence gives much less reason to expect a steep, persistent cost curve simply because global electrolyzer capacity keeps doubling. Low-carbon hydrogen should be planned around the price complete systems can realistically achieve and directed toward uses where the molecule is valuable enough to pay that price. For the deeper analysis of what the apparent learning rates are actually measuring — including stack-size growth, project-scale effects, manufacturing repetition and the system boundary a defensible hydrogen-cost forecast needs to use — read the full analysis in TFIE Strategy Briefing.