Long-Lead Procurement & the End-to-End Equipment Supply Chain

At some point on every AI-campus program, the project manager stops drawing Gantt bars and starts reading a spreadsheet of delivery dates. On a power-bound campus the building rarely gates the schedule; the equipment does. Which track governs depends on the dependency network in Chapter 2.1: on a powered-shell fit-out, a brownfield conversion or a project with an open building permit, it can still be the building. You can pour the slab in eight months, top out the steel in twelve and pull the fiber in three. None of it matters if the load-serving substation transformer that steps utility HV down to site MV is quoted above 160 weeks (US; Wood Mackenzie, reported May 4, 2026), the generation-connected GSU at three to four years (US; SemiAnalysis, June 25, 2026) and the medium-voltage switchgear at 44 weeks (Wood Mackenzie, Q2 2025 survey). If you planned to bridge with gas turbines, note that GE Vernova said on September 16, 2026 that it had secured cash-backed reservations for multiple 2032 turbine deliveries; obtain the selected OEM's delivery commitment. The critical path of a 2026-vintage data center runs through a handful of factories on three continents, and the decision that governs it is when to place the order, against a design that is nowhere near finished.

Long-lead procurement is therefore a schedule discipline more than a purchasing one. This chapter covers the consolidated lead-time register, the list of what takes how long and so what to order first. It covers the contracting instruments that exist because the lead times are intolerable: reservation deposits, slot-reservation agreements, owner-furnished-equipment staging and buffer inventory. It covers the allocation dynamics where demand structurally exceeds supply (GPUs, HBM, CoWoS, transformers, copper), and the geographic-concentration and single-source exposures, with the 2025–2026 tariff and export-control regime on top. It ends with the control tower that turns a list of POs into delivered, accepted, energized equipment. → grid-side interconnection timing in Chapter 3.2; on-site generation in Chapter 3.4; the contracts that wrap these orders in Chapter 2.4.

Every serious owner keeps one artifact above all others before construction: a lead-time register. It lists each major equipment package, its current quoted delivery, the design maturity needed to release the order, and the latest date you can place it without slipping energization. The master schedule is reverse-engineered from it. The lead-times page is the live, dated version: every package below, resolved to its source claim, with a reverse-scheduling planner that turns a ready-for-service date into the PO dates it implies. The register encodes a hard truth: the long poles must be ordered before the design that specifies them is complete. The long poles can be released before the overall design is complete once their equipment ratings, utility requirements, interfaces and acceptance requirements are approved; a percentage of design completion alone does not establish readiness.

The register sorts naturally into tiers. The grid-tie tier (HV and generator step-up transformers, or GSUs, plus high-voltage switchgear and utility-side substation equipment) sits at the top, with the longest and most volatile lead times in the building. The power-chain tier (distribution transformers, medium-voltage switchgear, UPS and busway) is shorter but still measured in many months. The on-site generation tier (gas turbines and reciprocating engines for bridge or primary power, plus diesel gensets for backup) has been a critical path in its own right since the gas rush of 2024–2025. The cooling tier (chillers, cooling distribution units or CDUs, and dry coolers) gates the density the building can support. The compute tier (GPUs and, upstream of them, HBM and CoWoS advanced-packaging capacity) is governed less by manufacturing time than by allocation.

Once you accept that equipment gates the schedule, three contracting behaviors follow. They look reckless to a traditional procurement organization but are the rational response to the lead times, and each trades capital at risk for schedule certainty. The fork is how much of that trade you are willing to make.

Reservation deposits. The first behavior is paying a non-refundable or partly refundable deposit to hold a manufacturing slot before the design is frozen, and sometimes before financing closes. For a transformer or a switchgear lineup this might be 10–30% of the unit price; for a gas turbine it can be structured as a milestone toward a binding order. The deposit buys a place in the queue. The risk is plain: if the project slips or dies, you may forfeit the deposit or own equipment you cannot use. Not paying has an equally concrete cost. You lose the slot and rejoin the queue at the back, adding one to three years to energization.

Slot-reservation agreements as a distinct instrument. The most important contracting development of the 2024–2026 turbine rush is that the slot reservation has become its own legal instrument, separate from the equipment supply agreement. GE Vernova ended Q2 2026 with 116 GW of combined Gas Power equipment backlog and slot reservations and targeted at least 125 GW by year-end. On September 16, 2026, it reported cash-backed reservations for multiple 2032 deliveries and said its 12 GW of added 2030–2031 capacity was in some form of contracting. The combined total is not a reservation-only book. Obtain the selected OEM's delivery commitment. Siemens Energy showed a similar pattern in early August: a 69 GW firm gas-turbine backlog (15 GW booked, 6 GW shipped in the quarter), lead times of "three years or more," and a €51B Grid Technologies book with transformers the largest order contributor (Q3 FY2026, reported 2026-08-05). A slot-reservation agreement lets a developer lock a delivery window with a deposit and a schedule of conditions, then convert it to a full supply contract once the site, permits and offtake mature. It is a financial option on manufacturing capacity. Caterpillar–Hunt Energy and Rolls-Royce's 2027–2028 order book show the same pattern in engines. → the contract mechanics in Chapter 2.4.

OFE staging and buffer inventory. The third behavior decouples the order date from the install date. Owner-furnished-equipment (OFE) staging means taking delivery early and holding the equipment; buffer inventory means keeping spares for items that fail or arrive damaged. You order the long-lead transformer to arrive when the pad is ready, not when you need it energized. You also carry critical spares (a spare transformer for a campus, CDU pumps, switchgear breakers), because one failure of an item with a two-year lead time is an existential schedule risk, not a maintenance event. The cost is carrying inventory and storage. The alternative is a two-year hole in the schedule when a unit arrives damaged.

For a transformer or turbine, the quoted N weeks combines queue position, approved drawings, component supply, production, factory testing and logistics; it is not all manufacturing work. For a compute package without a secured factory slot, the governing variable is allocation. When the vendor rations a sold-out GPU configuration, willingness to pay does not establish delivery: the buyer needs an allocation for the specified SKU and rack system. The real gate sits upstream of the accelerator, at HBM (high-bandwidth memory) and CoWoS (TSMC's advanced 2.5D packaging, which bonds the logic die to the memory stacks). GB200 and GB300 use HBM3E, while Vera Rubin uses HBM4. Micron said on September 30, 2026 that agreements cover the vast majority of its calendar-2027 HBM bit supply, with significant year-over-year price increases. TSMC's 2026 disclosures describe continued CoWoS capacity expansion, and on July 16, 2026 it said packaging capacity was limiting customers' growth. A quoted ~6–9-month order-to-delivery window must identify the allocation and the starting event before it can enter the IMS. → the silicon and memory supply detail in Chapter 7.6; advanced-packaging engineering in Chapter 7.7.

Allocation creates a power dynamic that bleeds into the rest of the deal. Large buyers secure GPUs through long-term commitments, prepayments and, increasingly, vendor financing: the chip vendor or a partner extends credit, takes equity or co-invests in the buyer so the buyer can afford the order the vendor will then fulfill. This circularity (the supplier financing its own demand) is both a genuine mechanism for clearing supply and a structural risk the finance chapters scrutinize closely. For procurement, the practical consequence is that the GPU order is a relationship and a capital commitment, negotiated quarters ahead, and the buyers inside the allocation are the ones who committed capital and volume early. → the circular-financing debate and GPU-cloud demand underwriting in Chapter 2.5; the residual-value and depreciation stress in Chapter 1.8.

A cloud bridge must separately prove contiguous capacity, topology, data transfer, workload acceptance, price and exit date. Renting an available accelerator recovers schedule only if the job can run on the offered system.

The lead times above are bad because the supply base is thin and concentrated. Several of the most critical packages come from a handful of factories, often in a handful of countries, and a disruption at a shared source can move several supposedly independent project dates together. This structural risk sits underneath the schedule risk, and it forces a sourcing fork on the items where it bites hardest.

Transformers are the canonical example: a small number of OEMs, a global shortage of grain-oriented electrical steel, and bushing and tap-changer sub-suppliers that are themselves single points of failure. A fire, a labor action or an export restriction at one shared plant can move delivery dates across several regions. Grid-scale switchgear and HV breakers are similarly concentrated. HV switchgear, breakers and complete substations require separate dated delivery commitments; a historical breaker lead time is not a lineup or whole-substation duration. The copper crunch sits underneath both. An NVL72 rack alone carries thousands of in-rack copper NVLink cables, the power chain is copper-intensive end to end, and every electrification and grid-replacement project on the planet wants the same copper, a demand collision that the 2025 tariff regime then taxed. And the HBM/CoWoS gate concentrates dependent accelerators into packaging sites in Taiwan and memory fabrication and assembly sites in South Korea, Taiwan and Japan. As of October 5, 2026, planned US HBM packaging does not provide current supply diversification: SK hynix's Indiana facility targets mass production in the second half of 2029. Count independent failure paths, not bidder logos: record the awarded factory, critical upstream source, shipping port and qualified recovery option. A second assembler using the same transformer core or HBM supply does not separate that failure.

On top of thin supply and long lead times, 2025–2026 added a volatile trade-policy overlay that changes the price, and sometimes the legality, of imported equipment. You can no longer treat a quoted price as the price. Model the tariff and export-control regime that applies on the day the goods clear customs, which may differ from the day you placed the order.

The most material change for electrical equipment was the US Section 232 action on copper. The August 1, 2025 measure imposed a 50% duty on covered copper content. From April 6, 2026, listed metal articles generally attract 50% and listed derivatives generally 25%, assessed on full customs value. For listed Annex III aluminum and steel derivatives, ordinary Column 1 duty plus Section 232 duty generally totals 15% through December 31, 2027; where ordinary duty is already at least 15%, no additional Section 232 duty is due under that provision. Apply the amended HTS coverage and the origin and content rules, including the June 8, 2026 changes. Steel and aluminum Section 232 duties also apply to structural-steel and enclosure costs. On the compute side, export controls on advanced accelerators and the equipment to make them reshape which chips can ship where, a procurement-defining constraint for any operator serving, or sourcing through, a controlled jurisdiction.

The strategic responses are nearshoring and dual-sourcing. Qualify a second factory to reduce shared-source exposure. Domestic and nearshore sourcing change tariff exposure according to the finished product's HTS classification, origin and metal content; allied-country sourcing does not by itself remove the duty. Check the arithmetic first: Wood Mackenzie estimated a 30% US power-transformer supply deficit in August 2025, with imports supplying about 80% of power transformers and unit costs 77% above 2019. A domestic-only requirement narrows the available supply base, so compare qualified domestic and import delivery commitments alongside their entry-date tariff costs. The tradeoff is the familiar one: the cheapest imported unit may carry a tariff and a concentration risk that a nearshore unit may not, and the right answer depends on the tariff trajectory you believe and the schedule you can tolerate. Accepting a higher unit price or a different lead-time profile can buy tariff certainty and supply resilience. → permitting and the regulatory critical path in Chapter 3.9, and export controls and sovereignty in Chapter 3.12; the contractual allocation of tariff and change-in-law risk in Chapter 2.4.

Whether the owner or the EPC contractor buys the long-lead equipment has direct schedule and risk consequences, and you decide it package by package, not once for the whole project. Owner-furnished equipment (OFE) means the owner places the order, takes title, carries the deposit and the inventory risk, and hands the equipment to the contractor to install. Contractor-furnished pushes the procurement, the float and the supplier-management burden onto the EPC, who prices that risk into the contract.

The logic divides cleanly by item. GPUs are almost always OFE. They are the most valuable, most allocation-constrained and fastest-depreciating asset in the building, and the owner who controls the allocation relationship will not delegate it. Transformers and gas turbines are frequently OFE on AI campuses because the owner wants to reserve the slot and place the order long before an EPC is chosen; waiting for the contractor to procure the long pole would forfeit the schedule. Chillers and CDUs are split: owners increasingly furnish them to lock density-critical capacity early, but they sit closer to the contractor's mechanical scope. The price of OFE is that the owner now owns the expediting, the factory acceptance, the logistics and the interface warranty gap. When an owner-furnished transformer fails to integrate with a contractor-installed switchgear lineup, the finger-pointing between supplier and installer lands on the owner. That exposure is the hidden cost of OFE, and you contract around it explicitly. → OFE vs contractor-furnished in the delivery-model framing of Chapter 2.2; the interface-warranty mechanics in Chapter 2.4.

A purchase order is a promise, not a delivery. A supply-chain control tower bridges the gap between the two (manufacturing slippage, failed factory acceptance, a transformer that cannot physically reach the site). It is a dedicated function that tracks every long-lead package from order to energization and intervenes when reality drifts from the plan. On a single-building project this might be one expeditor; on a multi-gigawatt campus it is a standing organization with three core jobs.

• Expediting. Active, on-the-ground tracking of manufacturing progress against milestones. Don't wait for the supplier's status email: visit the factory, verify that the GO-steel arrived and the windings are on schedule, and escalate the moment a sub-supplier slips. For an item with a two-year lead time, discovering a six-week slip in month four is recoverable; discovering it at the promised ship date is a disaster.
• Factory acceptance testing (FAT). Witnessing the equipment pass its acceptance tests at the factory before it ships: a transformer's impulse and heat-run tests, switchgear functional checks, a CDU's flow and pressure verification. A unit that fails FAT and goes back into the queue is a schedule event measured in months, and catching it at the factory is far cheaper than catching it at the gate. → FAT's place on the commissioning ladder in Chapter 13.1.
• Logistics of oversized and heavy loads. Programs most often overlook this job. A large power transformer can weigh 200–400 tonnes and needs its own heavy-haul route and permit plan. For racks, name the generation and use the OEM's packed shipping weight and handling plan: GB200/GB300 NVL72 wet racks are roughly 3,000–3,300 lb (1.36–1.50 tonnes), and NVIDIA describes Vera Rubin NVL72 at roughly 4,000 lb (1.81 tonnes, March 16, 2026). That mass triggers route and structural verification, but acceptance depends on the OEM foot/wheel and rolling reactions, temporary rigging states and the complete floor assembly, not on calling the cabinet mass one point load. These are super-loads requiring route surveys, permits, escorts, sometimes bridge reinforcement, and an early confirmation that the equipment can physically turn onto the site road and through the building opening. Programs have discovered at the gate that the transformer cannot make the final turn. The control tower verifies the route before the order ships, not after.