Capacity is a membership

The cleanest template sits in an SEC filing. The 2021 GlobalFoundries–Cirrus Logic capacity agreement has exactly the anatomy in question: a $50 million non-refundable capacity reservation fee — a pure capacity right, separate from any unit — a $175 million prepayment credited against future wafer orders, and a roughly $1.6 billion minimum purchase commitment with shortfall payments that cut both ways: the customer pays for volume it does not order, and the fab pays if it fails to deliver capacity. At portfolio level GlobalFoundries disclosed over $20 billion of such long-term commitments plus about $4 billion in advance payments and fees. TSMC collected $3.8 billion of capacity prepayments in a single quarter of 2021, over $3 billion of it from one customer. Cloud runs the same logic at retail: AWS Reserved Instances trade a one-or-three-year commitment for discounts up to roughly 72%, and the zonal variant carries an actual launch guarantee during shortage — the whole structure exists because it lets the party with demand knowledge sell predictability to the party with capital risk. And Costco is the consumer-scale proof that margin can live in the membership: fees are about two percent of revenue and the large majority of profit, while goods move at roughly eleven percent gross margin, for decades, with famous retention.

Search for this contract in manufacturing services and the space is nearly empty. Fictiv's Premium membership — billed as digital manufacturing's first paid subscription — turns out to sell software features and preferred pricing, not a capacity right. Xometry's marketplace model deliberately avoids holding capacity risk at all. Inside EMS, capacity reservation exists everywhere — as bespoke contract clauses: reservation fees deductible against firm orders, minimum-volume guarantees, take-or-pay terms. And roughly eighty percent of EMS agreements already price on some open-book or cost-plus basis, which means units-at-documented-cost is not the radical part; it is close to industry normal, worn quietly. The closest operating analogue to a productized capacity right is the air-cargo block space agreement — forwarders paying for guaranteed lane capacity whether they fill it or not, renegotiated in every soft market; there is even a startup building a marketplace to trade those rights. The conclusion cuts two ways: productizing the capacity subscription for contract manufacturing is genuinely novel — and no one in this industry has yet absorbed, at product scale, the lesson of what happens to it in a downturn.

Operations research spent twenty-five years on exactly this structure, and its results read like a design review. Erkoc and Wu showed that the coordination knob is partial deductibility: a reservation fee that is partly credited against real orders and partly kept aligns both sides' incentives across a wide range of prices. Cachon's push-pull framework says a pure prepaid commitment dumps all utilization risk on the customer while a pure on-demand model dumps it all on the manufacturer — and that the intermediate splits are what actually coordinate a supply chain. The make-to-order pricing literature treats contract customers with guaranteed lead times and spot customers with quoted ones as distinct classes — precisely the subscriber-versus-spot split. And Jordan and Graves proved the result that disciplines the portability promise: qualifying each product at just two plants, chained across a network, captures nearly all the protection of qualifying everywhere. Even the vocabulary exists — electricity-market research formally calls this design a capacity subscription, and describes what it buys as priority under scarcity, not a dedicated line. The theory's verdict on a pure form — zero unit margin, hard SLA, annual fee — is that it is an unusually sharp risk split that should be blended. Which is what the published model does.

The precedents also document the deaths. GlobalFoundries' iron-clad take-or-pay fees were, in the 2023–24 downturn, renegotiated rather than litigated — you do not sue customers you want to keep — and its dual-sourced customers used their portability to extract price cuts, a warning that "produce at any qualified hub" hands every customer structural leverage at renewal. Cloud shows the reputational failure: roughly 29% of committed cloud spend is wasted, an entire FinOps industry now exists to audit commitment breakage, and the behavioral-economics literature on gym memberships quantified how much customers resent paying for capacity they did not use. A procurement team will divide the subscription fee by units actually shipped; if utilization is 60%, a "no margin on units" claim reads as a hidden 40% markup. This is why Model B on the economics page carries its mechanics in bold rather than in fine print: at most 85% of a hub's nameplate is subscribable, half the fee is deductible against handling on units actually shipped, unused reserved units roll over one quarter, the commitment resizes ±20% at renewal, the two-week promise penalizes us through service credits, and one second-hub qualification is included because two chained hubs buy nearly all the protection of many. Every one of those lines is a named failure mode, answered.

AWS priced its commitments off years of demand telemetry; TSMC and GlobalFoundries priced theirs off decades of cycle data and still mispriced 2023. A subscription sold with a hard SLA and no utilization history is insurance underwritten without an actuarial table. So the conditions are published with the model: at least two certified hubs with chained qualifications before portability is a promise, twelve to twenty-four months of utilization telemetry before hard take-or-pay pricing — until then the fee anchors to the cost of capacity actually held — and a written, auditable definition of "cost," because the model's best sentence is otherwise its biggest dispute. One-year terms come first; multi-year discounts only after the data exists, in the same order AWS introduced them. The register will hold the dates when the first contract signs.

Cirrus Logic 10-K — the GF capacity agreement · GlobalFoundries 20-F FY2023 — LTA portfolio · GlobalFoundries 20-F FY2024 — the renegotiations · Tom's Hardware — TSMC prepayments · AWS — Reserved Instances · Money — Costco's membership economics · Fictiv Premium — what it actually sells · VentureOutsource — EMS agreement structures · The Loadstar — block space agreements · Erkoc & Wu — capacity reservation contracts · Cachon — push, pull, and advance-purchase risk sharing · Jordan & Graves — chained flexibility · Doorman — capacity subscription in power markets · Flexera — cloud commitment waste · DellaVigna & Malmendier — paying not to go to the gym