Guide · Layer 3
Written for: seed / Series A scientist-founder
This is where the model answers the question that started the whole exercise: what does this cost, and can it win? The two earlier layers fixed the shape of the process and put physical numbers on the flows; this one puts dollars on them and collapses everything into one comparable figure — the levelized cost, the cost to make one tonne of product over the plant’s life.
Two streams of money feed that number. Capital is spent once, up front, to build the asset; operating cost recurs every year to run it. You can’t add a one-time stock to an annual flow until the capital is on an annual footing — so the spine of the layer is: build the capital cost, annualize it, add the yearly operating cost, divide by what the plant actually makes.
The defining discipline is knowing where the accuracy lives. A factored estimate here is a ±30–50% accuracy class, not a point, so almost all the real signal sits in a few big equipment items and a few input prices; the rest is bookkeeping. The classic early-TEA failure isn’t an arithmetic slip — it’s false precision: a twelve-significant-figure cost per tonne built on order-of-magnitude anchors. Spend effort on the inputs that move the answer; report to the precision the method can carry.
Six moves: cost the capital → roll up to total capital → annualize → cost the operating side → assemble the levelized cost → read it against price and credits. Each routes into the concept pages for the how; the reads below are about where to spend your effort.
Build capex bottom-up, equipment-first. For each unit operation, pick the 2–3 items that dominate its cost, size each to its duty, and scale from a known reference with the six-tenths rule — C₂ = C₁·(S₂/S₁)ⁿ — reading the per-class exponent from sizing scalars. Where a single item would be too big to build, stop scaling and start numbering up. Sum to a purchased-equipment cost, then gross up to installed ISBL with an installation factor.
🧭 Coach’s Read. Most of the capital signal is here, so concentrate here: get the 2–3 biggest items per unit operation well-sourced and sized, and factor the rest. Two places not to spend effort: don’t tune the six-tenths exponent (across the 0.5–0.7 band a 3× size change moves cost only ~15% — pick the class default and put that hour into the reference cost instead), and keep sizing scalars to ~5–6 classes. One trap that is worth attention: switch from scaling to numbering up once the size step exceeds ~2–3×, or you’ll claim an economy of scale that modular units (electrolyzer stacks) don’t deliver.
Installed ISBL is only the process units inside the fence. The figure that gets financed and annualized is total capex: ISBL + OSBL (off-sites — the battery-limits line decides what’s in) + indirects (engineering, fees) + contingency (the expected cost of what’s unresolved, sized to its accuracy class) + working capital.
🧭 Coach’s Read. When you quote “the capex,” quote this number — not ISBL, and not purchased-equipment cost. ISBL is often only ~half of total fixed capital, and delivered equipment a quarter to a third of installed cost, so quoting the inner figure can understate the plant ~2× and quietly halve your cost per tonne. Two own-goals: don’t drop contingency to look tighter (that reports an unlikely best case), and watch what each factor already contains — applying a Lang factor (which already bundles OSBL and indirects) and then adding OSBL again is the most expensive scope error in the build.
Convert total capex to an equivalent annual charge with the capital recovery factor: annual capital charge = CRF × total capex, where CRF is set by a discount rate i and a life N. For capital spent once and then run at steady state, this gives the same answer as a full discounted cash flow — without building one.
🧭 Coach’s Read. Use CRF, not DCF, at this stage — a DCF adds a price path, a ramp, a tax schedule, and a terminal value, none of which make an uncertain capex more accurate; it mostly buys false precision. But respect what CRF is sensitive to: it rides on two pure assumptions, and moving the discount rate from 8% to 12% lifts the annual charge by roughly a third with nothing physical changing. State
iandN, and keep them real-with-real or nominal-with-nominal — mixing a nominal rate with un-escalated costs is a classic silent error.
Build annual opex and split it by the fixed-vs-variable razor. The variable half is dominated by feedstock and energy: consumption per unit (from the mass-and-energy layer) × input price. The fixed half — maintenance (a few percent of ISBL/yr), labor, overheads — is factored, like capital.
🧭 Coach’s Read. For most commodity processes, feedstock and energy is the whole game on the operating side — get the dominant input and its price range right first. And note which error matters: the consumption figure is usually well-bounded, but the price swings more and faster, so it’s typically your largest operating uncertainty and the natural sensitivity axis (Layer 4’s job). Remember the structural linkage: the fixed half — including the annualized capital from move 3 — scales as 1/capacity factor, so for a capital-heavy or intermittent route the capacity factor can move cost per tonne more than any equipment detail in move 1. And watch the word free: a “free” waste feedstock rarely stays free — the day you make it valuable, the owner wants a cut, so model it as a revenue split, not a zero.
Bring the two streams together. The levelized cost is (CRF × total capex + annual opex) ÷ annual output, where the denominator is operating output — nameplate × capacity factor × time, not the nameplate rating. It decomposes into a capital share, a fixed-opex share, and a variable-opex share — and that mix is the route’s cost structure.
🧭 Coach’s Read. This is the headline number, and it’s meaningless without its basis — always state the system boundary, the capacity factor, and the CRF assumptions (
i,N). The most common way “X is cheaper than Y” goes wrong is comparing two levelized costs on mismatched bases. Two more: never divide by nameplate (you’ll spread cost over tonnes you never made), and resist reporting$805/twhen the honest answer is~$800/t ±30%. Keep the three shares visible — two routes at the same$/tcan respond oppositely to a price or utilization shock, and only the split shows which is fragile.
A cost of production is half the story. Set the levelized cost against revenue, pricing and byproduct credits: for a commodity you’re a price-taker, so the levelized cost is the floor you must beat, the price is what you receive, and the gap is margin. Then layer on carbon and policy credits — production subsidies, capture payments, carbon prices — gated by the product’s carbon-intensity basis, the emissions-side twin of the levelized cost.
🧭 Coach’s Read. Credits can be large enough to flip the ranking — a clean-hydrogen credit can rival the entire gross cost per tonne — which is why you must keep gross and net side by side. Reporting only the netted number hides how much of the result is policy rather than process, and policy is jurisdiction-specific, time-limited, and conditional on accounting rules (additionality, hourly-matched clean power) you must actually meet. Same discipline on the credit side: a big byproduct stream (the oxygen off electrolysis) is worth market price only up to the quantity a real local market can absorb — beyond that the marginal credit is zero. And treat carbon intensity with the same basis rigor as cost: for an electrified route the whole “green” claim lives or dies on the electricity emission factor you declare.
🪜 Leveling Up — DCF and full financial modeling. The single levelized number assumes one up-front outlay and flat operation. Climb to a year-by-year discounted cash flow when that genuinely breaks — staged capital, a multi-year ramp, project finance with a real debt/equity structure, or material tax and depreciation — or when a counterparty’s diligence wants NPV, IRR, and payback. Separately, a bottom-up, quote-driven capex (real vendor quotes against a FEED study) is a tighter accuracy class than the factored estimate. Both are worth it at the right moment, and both cost weeks of effort and a pile of new assumptions without making the physics any more accurate. At the maturity anchor, CRF + a factored capex is deliberately the right tool.
The six moves on green ammonia — every figure a round, illustrative anchor carried from the concept pages, where the full build (and its open validation items) lives:
n ≈ 0.62); the electrolyzer isn’t scaled but numbered up — tens of ~5–10 MW modules, because no single unit that size exists. Big items × an installation factor of ~3.5 → installed ISBL ≈ $0.5bn.i = 10%, N = 20 yr, CRF ≈ 0.117 → ~$117M/yr. Over ~330,000 t/yr (1,000 t/day × 0.90 capacity factor) → capital share ≈ $355/t.One revenue-side discipline, quietly violated all the time: price your product at a real, traded form, not an invented intermediate. End the model where the market sets a price and deduct whatever downstream cost it takes to get there. Pricing off an upstream intermediate nobody trades invents revenue the market will never confirm.
The pattern: the magnitudes and mechanics stayed on the concept pages; this layer sequenced them and said where the effort and the risk are. A late-stage guide could route the same pages to a full DCF — that swappability is the point.
Once the economics produce a number, the next layer pressure-tests it: Layer 4 — Analytical Layer, where sensitivities and scenarios find what actually moves the answer — usually one of the few input prices and the capacity factor this layer flagged.