Lead-Time Tracker
Every long-lead package that sets an AI data center schedule, with dated, sourced figures from the guide's numbers register — and a planner that distinguishes defensible commitment dates from service-path uncertainty and quote gaps. The pattern to internalize: on a greenfield build, the schedule is set by the electrical chain, not the building or the GPUs — on a powered shell or a compute refresh, the compute and networking clocks take over.
Quick answers — the figures most visitors come for · updated 2026-Q1–Q2
~24–36 mo construction once power is secured; grid connection waits exceed 4 yr in primary markets and can overlap construction
Build an AI data center — site control to first tokens, counted from a secured power path. A new host-utility service path uses the actual local studies, facilities plan, tariff, and contract; it has no portable national duration.
~12–24 wk
GB300 NVL72 rack systems — accepted OEM order with a confirmed slot to dock — allocation decides when the clock starts, but the clock is real.
>160 wk
HV power transformers — usually the longest purchase order on a grid-fed project — ordered before design freeze.
40 wk or more (distributor-reported); EML capacity allocated past 2027 for buyers outside NVIDIA's agreements
800G optical transceivers (volume orders) — EML laser supply is the constraint; switches have no public weeks figure.
General estimates — your numbers will differ, in both directions. These figures come from published market surveys and reporting (dated and sourced in the register), which mostly reflect what a cold purchase order gets quoted. Framework agreements, vendor relationships, standard configurations, and stock positions routinely beat them; constrained markets, custom specs, and constrained suppliers miss them. Use this page to see the shape of the problem — then verify your own region, host-utility process, permits, vendors, and contracts before committing a schedule.
Before any PO: how fast can you get power at all?
Three procurement postures, three clocks. Everything below assumes you're building.
~24–36 mo construction once power is secured; grid connection waits exceed 4 yr in primary markets and can overlap construction
Self-build, greenfield
Site control to first token — but only counting from a secured power path. If new host-utility large-load service is required, its locally governed studies, facilities work, and contract can set total elapsed time; no national generator-queue duration is substituted.
~6–12 mo fit-out
Wholesale colo
Leased purpose-built shell for a live 50k+ GPU cluster.
days–weeks
Neocloud capacity
Renting a cluster someone else already built — the only fast lane.
If you started a project today
Select the project's power pathway, then pick a ready-for-service target — first tokens, not utility energization (the grid energizes months earlier; that's inside each package's on-site offset, shown per row). Applicable tasks with bounded public ranges are reverse-scheduled from the long end to their on-site milestone. Explicit predecessors can pull upstream need dates earlier. Region and block count define scope without inventing a duration multiplier; project maturity exposes unsatisfied release gates. Add a dated quote or confirmed delivery slot when you have one. Tasks outside the selected pathway are N/A and remain visible below the schedule, but they cannot enter lateness or critical-path arithmetic. Directional guidance, not quotes — validate locally before committing capital.
GPU & compute-system lead times: what an accepted order waits
Two clocks run in sequence. Allocation decides whether and when an OEM can accept your build slot — commercial, priority-based, and unquoted. But once the slot is confirmed, the order gets a delivery window like any other package: these are the windows public channel quotes and observed programs support today. The clock starts when the OEM confirms the slot, not when you first ask for a quote — and renting an already-built cluster (days–weeks, above) is a different product entirely.
~12–24 wk
GB300 NVL72 rack systems
New OEM-integrated liquid-cooled rack, accepted order with a confirmed slot → dock. 8–12 wk is the upside case with a named slot; custom CDU/network BOMs push the top.
~8–14 wk
HGX B300 / B200 servers (standard BOM)
8-GPU air- or liquid-cooled server on an OEM "golden" configuration. Custom BOMs exceed 20 wk — the config decision is a schedule decision.
~4–12 wk
H100 / H200 HGX systems
Prior generation, mature supply: H100 ~4–10 wk, H200 ~4–12 wk. Loose held-stock cards ship in days — but a card is not a tested system.
no public average
AMD MI355X / MI325X systems
No OEM publishes a lead time for these systems, and AMD's MI400-series Helios racks now lead its line. Get the slot and acceptance date in writing.
~6–12 mo
Multi-rack program, order → online
Agreement to commissioned capacity for a full deployment (IREN–Dell ~7–10 mo; Crusoe ~3–6 mo). Site work and integration, not the rack clock, set this number.
no public average
Hyperscale / framework allocation
No defensible public average exists — the negotiated allocation calendar and named ship slots govern. Report your contractual date; "sold out" is a demand signal, not a queue length.
What could move these ranges
The signals behind the windows above — demand, supply commitments, ramps, and risks. These are status, not durations: none of them belongs in a project schedule, but each explains why a quoted range might shift under you.
The allocation layer
demand signal
“cloud GPUs are sold out” (Nov 2025)
Cloud GPU capacity, all generations
NVIDIA's own words, Q3 FY26: “Blackwell sales are off the charts, and cloud GPUs are sold out.”
earnings-derived
~17.4 mo
Dell AI backlog coverage
Dell's AI backlog divided by its quarterly AI-server revenue: months of shipments at the current run rate. That measures staged programs, not your server's queue; a standard accepted order still ships in weeks.
supply commitment
$279B, through FY2029
NVIDIA supply-commitment horizon
$279B of supply and capacity commitments at July 2026, mostly memory, scheduled through NVIDIA's fiscal 2029.
as of 2026-07-26
historical
~12 mo
Blackwell backlog (B200 / GB200)
The last cleanly reported backlog figure (Morgan Stanley, Oct 2024) — historical context; NVIDIA has talked allocation, not weeks, ever since.
forecast
Blackwell ~71% of 2026
2026 shipment mix
Blackwell raised to ~71% of NVIDIA's 2026 high-end shipments; Rubin cut to 22% on HBM4 validation, ConnectX-9 migration, power and cooling steps.
as of 2026-04
forecast
leads NVIDIA's 2026 shipments
GB300 / B300 (Blackwell Ultra)
The volume platform of 2026 — GB300-based systems drive most GPU-server shipments; AI servers +28% YoY.
as of 2026-10
milestone
shipping since June 2026
Vera Rubin NVL72 (VR200)
Shipping since June 2026 (Dell to CoreWeave); NVIDIA reports production shipments from fiscal Q3 2027. The HBM4 ramp sets volume.
as of 2026-08
milestone
In production; initial systems installed (Oct 2026)
AMD Helios (MI400 series)
In production; initial systems running at OpenAI, Meta, Microsoft and Oracle (AMD, October 2026). No shipment volumes published; anchor commitments absorb early supply.
as of 2026-10-12
demand signal
nearly fully subscribed
AWS Trainium3
The custom-silicon path is allocation-bound too: Trainium2 largely sold out, Trainium3 volumes nearly fully subscribed, Trainium4 planned for 2027.
as of 2026-08
Memory & networking — the constraints around the GPU
supply commitment
sold out through 2026
HBM: SK hynix
NVIDIA's lead HBM supplier pre-sold its entire planned 2026 memory supply — announced at its record Q3 2025 earnings.
as of 2025-10
supply commitment
HBM4 high-volume shipments
HBM: Micron
Micron's calendar-2026 HBM is committed under contract, and HBM4 is in high-volume shipments for its lead customer's platform (fiscal Q3 2026 results).
as of 2026-06
milestone
volume ramping 2026
HBM4 generation
Samsung first to volume (>$1B within ~4 months); SK hynix's meaningful volume shifts to 3Q26 — the gate behind Rubin.
as of 2026-06
trend
lead times lengthening
Switch & optics supply
Dell'Oro: lead times for advanced substrates, optics, and high-speed networking components lengthening into 2026. The weeks-level 800G optics figure is in the planner above.
as of 2025-12
milestone
volume ramp Mar 2026
Broadcom Tomahawk 6 (102.4 Tb/s)
Production volume from March 2026 on “unprecedented” demand; Arista's TH6-based 1.6T switches ship from Q4 2026.
as of 2026-03
schedule risk
Rubin ramp risk
ConnectX-9 NICs
The ConnectX-8→9 migration is a named reason TrendForce cut Rubin's 2026 shipment share.
as of 2026-04
Why these numbers rule the schedule: the integrated master schedule (Ch 2.1) · the long-lead supply chain (Ch 2.3) · the power-bound era (Ch 16.1). Every underlying figure, with sources: the numbers register.
Embed this tracker
The headline lead times as a live widget — free to embed with attribution:
<iframe src="https://aidatacenterguide.com/embed/lead-times" width="420" height="330" style="border:1px solid #ddd;border-radius:8px" title="AI data center lead times"></iframe>