The idea
What arrives next quarter was ordered long ago. How long ago depends on the layer. Hixson and Guliani give the span: order time "could be five years if you erect your own buildings," and "10 minutes" for a small cloud service with good automation [6].
| Layer | Lead time | Public record |
|---|---|---|
| Land, shell, utility power, large electrical gear | Years | "planned years in advance" [8]; grid connections and large transformers take years [60–62] |
| Long-haul network | Months to years | "months or even years to deliver" [63] |
| Machines and parts | Quarters | "over 6 months for some components" [8]; one accelerator vendor, "more than 12 months" in its own supply [64] |
| Turn-up | Weeks per job; months to staff | No published value |
A large new power load waits in the utility's own process, and little is published on how long [65].
Usual and cautious lead times. Late arrivals are common and early ones rare argued. So plan on a cautious lead time, near the slow end of what you've seen. For example, if machines usually take one quarter but one order in five takes two, plan on two.
Layers stack. The hall comes before the machines, the machines before turn-up. So a layer's need date is the next layer's start, not launch day. Microsoft's chief executive described the failure: "I don't have warm shells to plug into" [66].
Turn-up is a rate, not an order. Crews, dock space and test positions set it, like a checkout line. A slipped batch queues behind the next one, so a slip often shows up twice argued.
Contracts. Supply is bought through prepayments, take-or-pay commitments, non-cancelable orders and long power agreements. Once its cancel window closes, a contract's forecast becomes a liability. Give each one a register entry with its liability schedule, cancel window and owner. Operators name the risk: long commitments behind shorter customer contracts [68, 69].
A cancel window is rarely one date. What you owe usually rises in steps toward delivery. No published terms of that kind were found. Because the most common timing miss is a slip, the right to push a delivery out is often worth more than the right to cancel argued.
Options and postponement. An option is a small payment now for the right to buy later. Keep supply generic until the decision that says which shape [73]. Meta describes staging data center sites so it can "spring up capacity quickly," and partnerships that give "option value for future compute needs" [74]. Quantity-flexibility contracts let an order move less as delivery nears [70, 71].
Supply has step changes too. New sites, hardware generations, rationed parts and decommissions all move in steps. Supply firms up in its own stages: planned, contracted, under construction, delivered, live [68, 92]. Decommission is negative supply [2]. Score supply promises like demand declarations.
Where it shows up on the map
Each layer's commit point lands in the cadence its lead time forces. The table is the one on the map; timings vary by operator and market:
| Supply layer | Typical commit point | Lands in | The declaration is usually |
|---|---|---|---|
| Land, power, building shell | One to several years ahead | Annual | Intent, or none |
| Long-haul network; large contracts | Several quarters | Annual or quarterly | Intent to Dated |
| Machines and parts | One to three quarters | Quarterly | Dated |
| Machines from a stocked pool | Weeks; at configuration in the worked example | Monthly | Configured |
| Turn-up and allocation | Weeks per job; the rate is staffed months ahead | Monthly | Configured |
| Job placement and admission | Minutes to days | Weekly | Live |
The stage pages say what lands in each: annual, quarterly, monthly and weekly. After launch, supply promises are scored against actuals (post-launch).
How to use it
- List every layer a launch touches, with usual and cautious lead times. Use Appendix B of the seed paper as a sanity check, not as your number.
- Work back from the ready-for-service date: each need date is the next layer's start.
- Register every contract with its liability schedule, cancel window and owner.
- Send a vendor the base and hold the high as options. Padding at three layers of 10% each compounds to about 1.33 times real demand (illustrative), the "rationing game" [67].
- Check supply for shared drivers: four sites waiting on one transformer supplier slip together.
In the running example, high-memory machines take two quarters and standard machines come from a stocked pool in weeks (illustrative). The two need different commit points inside one launch. See commit points.
What goes wrong
- Planning on the usual lead time, then expediting in a bad year.
- Ordering the fast layer before the slow one it waits on.
- Liabilities discovered at quarter close.
- Taking supply promises at face value while discounting demand intents: half an outside view.
- A slipped decommission stalling an on-time delivery.
Where this comes from
Seed paper, Sections 6.2, 6.4 and 6.7, step 1 of Section 8.2, and Appendix B. Order times and planning horizon from Hixson and Guliani [6]. Public lead times from [8, 60–64]; the power wait from [65]. Contracts and the mismatch risk from [68, 69]. Staging sites and option value from Meta [74]. Postponement from Feitzinger and Lee [73]; quantity flexibility from Tsay [70, 71]. Supply stages from [68, 92]; decommission as negative supply from Flux [2].