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

There is a moment on every AI-campus program when the project manager stops drawing Gantt bars and starts reading a spreadsheet of delivery dates, because the realization has landed that the building does not gate the schedule — the equipment does. 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 generator step-up transformer that connects you to the grid is quoted at 144 weeks, the medium-voltage switchgear at 44, and the gas turbines you planned to bridge with are sold out into 2029. 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.

This chapter treats long-lead procurement as what it actually is: a schedule discipline, not a purchasing one. We build the consolidated lead-time register — the canonical list of what takes how long and therefore what must be ordered first. We then work through the contracting instruments that exist precisely because the lead times are intolerable: reservation deposits, slot-reservation agreements, owner-furnished-equipment staging, and buffer inventory. We trace the allocation dynamics that govern the items where demand structurally exceeds supply — GPUs, HBM, CoWoS, transformers, copper. We name the geographic-concentration and single-source exposures, overlay the 2025–2026 tariff and export-control regime, and close on 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 maintains one artifact above all others in the pre-construction phase: a lead-time register that lists each major equipment package, its current quoted delivery, the design maturity required to release the order, and the latest date you can place it without slipping energization. It is the document from which the master schedule is reverse-engineered. The hard truth it encodes is that the long poles must be ordered before the design that specifies them is complete — you commit to a transformer's MVA rating, impedance, and winding configuration while the single-line diagram is still at 60%, because if you wait for 100% the slot is gone.

The register sorts naturally into tiers. The grid-tie tier — HV and generator step-up (GSU) transformers, high-voltage switchgear, and the utility-side substation equipment — sits at the top, with the longest and most volatile lead times of anything 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 become a critical path in its own right since the gas rush of 2024–2025. The cooling tier — chillers, cooling distribution units (CDUs), and dry coolers — gates the density the building can actually support. And 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 the equipment gates the schedule, three contracting behaviors follow that look reckless to a traditional procurement organization and are in fact the only rational response to the lead times. Each one trades capital-at-risk for schedule certainty — and the fork is how much of that trade you are willing to make.

Reservation deposits. The first behavior is paying a non-refundable or partially-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 you a place in the queue. The risk is obvious: if the project slips or dies, you may forfeit the deposit or own equipment you cannot use. The consequence of not paying it is equally concrete — you lose the slot and re-enter 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 told investors it expected to end 2026 with at least 110 GW of combined gas-turbine backlog and slot-reservation agreements, and that it could be effectively sold out through 2030 — meaning much of that book is reserved capacity, not yet converted to firm turnkey orders. 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 is decoupling the order date from the install date by taking delivery early and staging equipment — owner-furnished-equipment (OFE) staging — and by holding buffer inventory of the items that fail or get damaged. You order the long-lead transformer to arrive when the pad is ready, not when you finally need it energized, and you carry critical spares (a spare transformer for a campus, CDU pumps, switchgear breakers) because a single 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 most of the register, the lead time is a manufacturing fact: it takes N weeks to wind a transformer or cast a turbine. For the compute tier, the governing variable is different and stranger — it is allocation. You do not get GPUs by being willing to pay; you get them by being allocated them, because demand structurally exceeds supply and the vendor rations. The real gate sits upstream of the accelerator itself, at HBM (high-bandwidth memory) and CoWoS (TSMC's advanced 2.5D packaging that bonds the logic die to the memory stacks). HBM3E was effectively sold out through 2026 with a supply gap on the order of 30% and quarterly price rises of 15–20%; TSMC's CoWoS capacity for 2026 was essentially fully booked with NVIDIA holding roughly half of it. The accelerator's ~6–9-month order-to-delivery window is almost irrelevant next to whether you sit inside someone's allocation. → 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 that 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 the procurement organization, the practical consequence is that the GPU order is not a purchase order — it is a relationship and a capital commitment, negotiated quarters ahead, and the buyers who are 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.

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 to any one of them ripples across every program on Earth at once. This is the structural risk that 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 plant moves delivery dates worldwide. Grid-scale switchgear and HV breakers are similarly concentrated, which is why HV substations quote 3–5 years while medium-voltage gear has eased. 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 copper is simultaneously demanded by every electrification and grid-replacement project on the planet — a demand collision that the 2025 tariff regime then taxed. And the HBM/CoWoS gate concentrates the entire compute tier into a few fabs in Taiwan and a few memory plants in Korea and the US.

On top of thin supply and long lead times, 2025–2026 layered a volatile trade-policy overlay that changes the price and sometimes the legality of imported equipment. The procurement organization can no longer treat a quoted price as the price — it must model the tariff and export-control regime that applies on the day the goods clear customs, which may differ from the day the order was placed.

The most material change for electrical equipment was the US Section 232 action on copper: from August 2025, a 50% tariff on semi-finished copper products and copper-intensive derivatives (cables, connectors), recalibrated in April 2026 to apply to the full value of semi-finished products. Crucially for builders, electrical grid equipment received a 15% transitional rate through end-2027 to avoid kneecapping the very buildout the policy was meant to protect — a carve-out worth knowing, because it changes the math on imported switchgear and transformers versus domestic. Steel and aluminum Section 232 duties stack on top of the 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: qualifying a domestic or allied-country transformer and switchgear OEM to escape both the tariff and the single-source exposure, and accepting a higher unit price or a different lead-time profile in exchange for tariff certainty and supply resilience. The fork is the familiar one — the cheapest imported unit may carry a tariff and a concentration risk that the slightly-more-expensive nearshore unit does not, and the right answer depends on the tariff trajectory you believe and the schedule you can tolerate. → the broader policy and permitting environment in Chapter 3.3; the contractual allocation of tariff and change-in-law risk in Chapter 2.4.

Who actually buys the long-lead equipment — the owner, or the EPC contractor — is a decision with direct schedule and risk consequences, and it is decided 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, most rapidly-depreciating asset in the building, and the owner who controls the allocation relationship is not going to delegate it. Transformers and gas turbines are frequently OFE on AI campuses precisely because the owner wants to reserve the slot and place the order long before an EPC is even selected — 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 — critically — 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 interface-warranty exposure is the hidden cost of OFE, and it is contracted around 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. The gap between the two — manufacturing slippage, failed factory acceptance, a transformer that cannot physically reach the site — is closed by a supply-chain control tower: 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 — not waiting for the supplier's status email, but visiting the factory, verifying that the GO-steel arrived and the windings are on schedule, and escalating 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; catching it at the factory is far cheaper than catching it at the gate. → the systems-integration and commissioning view in Chapter 11.2.
• Logistics of oversized and heavy loads. The most under-appreciated control-tower job. A large power transformer can weigh 200–400 tonnes; a fully-built NVL72-class rack ships at roughly 1.4 tonnes with a point load far above a standard raised floor's rating. These are super-loads requiring route surveys, permits, escorts, sometimes bridge reinforcement, and a confirmation — done early — 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.