Reserving Data Center Capacity for GPUs on Order: Lead Times and Timelines

How far ahead should you reserve data center capacity for GPUs?

In 2026, reserve liquid-cooled capacity 12 to 24 months before your GPUs arrive: more than 80% of primary-market space under construction is already preleased (CBRE, Aug 2026).

CBRE's first-half 2026 survey found primary-market vacancy at a record low of 1.4%, 7,481.1 MW under construction, and commitments on 80.4% of it, up from 74.3% a year earlier. Less than 1,500 MW of future capacity remained available, "representing approximately six months of demand at the current absorption rate." Asking rates rose 8.3% for 3–10 MW deployments and 6.7% for 10 MW and up in the half (CBRE, Aug 2026).

JLL counts more than 66 GW under construction in North America, 77% of it in frontier markets, with vacancy at 1% for the third consecutive year, and its midyear report says "Most tenants securing space today are contracting for 2028 deliveries" (JLL, Aug 2026). Across the Americas, Cushman & Wakefield counts 37.7 GW under construction, "with 91.7% precommitted prior to delivery" (Cushman & Wakefield, Sep 2026).

For a tenant with GPUs on order, the space decision comes before the delivery window in the purchase order, not after it. A turnkey hall about a year out is the short end. Anything larger than a few megawatts, or anything that needs a new building, sits at the long end, and the long-lead steps behind it are laid out in how a data center is built.

How long do GB300 and Vera Rubin systems take to arrive?

In 2026 no NVIDIA partner publishes a GB300 NVL72 lead time, and NVIDIA expects supply to remain "a bottleneck at least through the end of fiscal year 2028" (NVIDIA, Aug 2026).

NVIDIA's fiscal 2028 ends in late January 2028. On the August 26, 2026 earnings call, CFO Colette Kress called the outlook "supply-constrained," said "We commenced production shipments of Vera Rubin earlier this month," and put Vera Rubin at about 20% of data center revenue in its fiscal third quarter, which ends in October (NVIDIA, Aug 2026). CoreWeave stood up a Vera Rubin NVL72 cluster for its customer Cognition in early September (CoreWeave, Sep 2026).

GB300 NVL72 is in volume production. Supermicro says it is "shipping volume SKUs across the GB300 NVL72" and took more than $60 billion of new orders in its June quarter (Supermicro, Aug 2026). Dell ended its July quarter with a $95 billion AI server backlog, and Futurum notes that supply "remains the main constraint" (Futurum, Sep 2026).

The parts behind the racks stay tight:

• Memory: Micron expects conditions "much tighter in calendar 2027 and 2028 than they were in 2026" (Computerworld, Oct 2026), and SK hynix expects demand to exceed supply through 2030 (The Next Web, Oct 2026).
• Packaging: TSMC's CoWoS supply "remains tight" (TrendForce, Oct 2026).
• The next rack: Rubin Ultra's Kyber NVL144 rack is on NVIDIA's roadmap for 2027; SemiAnalysis reported a slip to 2028, and NVIDIA answered "Our roadmap is intact" (Tom's Hardware, Jul 2026).

The hardware is often not the long pole. Supermicro's June-quarter revenue came in at $11.1 billion "due to some short-term customer delays in power, cooling, and networking": servers were ready before the sites were (Supermicro, Aug 2026). Jensen Huang said the same from the other side on NVIDIA's call: "You've got to go secure the land, power, and shell, which oftentimes is a couple, two, three years out." At GTC Berlin he put the cost of a three-month delay on a 1 GW AI factory at about $10 billion of missed revenue (The Next Platform, Oct 2026). Treat the ship window in an OEM order as a range, get it in writing, and plan for the hall to be commissioned before its earliest date.

How long from purchase order to power-on with a prefab data center module?

In 2026, vendors quote about 20 to 39 weeks to deliver an operating factory-built AI data hall, and none of those schedules brings new grid power (BladeRoom, 2026; Unite.AI, Oct 2026).

Every one of these schedules starts after the hard part. "Factory integration cannot create a power connection," reads a section heading from the modular vendor ModulEdge. Its illustrative schedule from design release runs 20 weeks of manufacturing and transport beside 28 weeks of site preparation, then 4 weeks of integration and acceptance, so site work, not the factory, sets the 32-week result (ModulEdge, Oct 2026). Futurum says the same of DSX POD: "onsite deployment can shrink from six months to 14 days once site utilities are ready," but "the binding constraint on AI factories in Europe is access to grid power, and DSX POD doesn't address that" (Futurum, Oct 2026). ABI Research makes the same point (ABI Research, Oct 2026).

What sets the date on a site that already has power?

GR1D analysis. Order to power-on for a factory-built AI hall on a site whose utility power is already energized:

On ModulEdge's illustrative numbers that path is 32 weeks from design release, about seven and a half months, plus whatever time the integrated systems test needs: inside the 6–12 months GR1D quotes for modular builds once power is in place (GR1D calculation). Without power in hand, the utility sets the date: JLL puts the average wait for a grid connection in primary US markets above four years (JLL, Jan 2026), and generator step-up transformer lead times passed 160 weeks by the first quarter of 2026 (Reuters via Kitco, Jul 2026). The full sequence for a conventional build is in realistic data center timelines.

Read as text

• Six paths to running GPUs, and what each needs in place
• Time from commitment to GPUs in service, 2026. Three paths fit inside about six months; each needs something that already exists.
• ABOUT 6 MONTHS OR LESS · LONGER · YEARS
• ABOUT 6 MONTHS OR LESS
• LONGER
• YEARS
• Existing liquid-cooled colocation
• Weeks to months · of fit-out, now stretching from weeks to several months
• of fit-out, now stretching from weeks to several months
• NEEDS · A provider with liquid-ready space free, at 1.4% vacancy
• NEEDS
• RISK · Scarcity; take-or-pay floors of 60–85% of allocated power
• RISK
• CBRE (Aug 2026)
• GPU cloud, reserved capacity
• Months · for a provider to stand up high-density capacity
• for a provider to stand up high-density capacity
• NEEDS · A provider with a spare cluster of the right GPU
• RISK · Multi-year take-or-pay; CoreWeave's contracts average about 5 years
• JLL (Aug 2025); CoreWeave 10-K (2026)
• Prefab hall on a powered site
• 20–39 weeks · vendor schedules, plus commissioning
• vendor schedules, plus commissioning
• NEEDS · Power, switchgear, permits and a pad already in hand
• RISK · Long-lead electrical gear; permits
• BladeRoom (2026); Flex via Unite.AI (Oct 2026)
• Converting an air-cooled hall
• 6+ months · delays "can extend beyond six months"
• delays "can extend beyond six months"
• NEEDS · An air-cooled hall with spare power
• RISK · CDU, piping and floor retrofit
• Conventional build, power in hand
• About 18 months · average for a 50 MW facility
• average for a 50 MW facility
• NEEDS · Power and permits
• RISK · Equipment lead times: 42 weeks on average in the US
• JLL (Jan 2026)
• New grid connection
• 4+ years · average wait in primary US markets
• average wait in primary US markets
• NEEDS · Nothing yet: land and an application
• RISK · Utility studies, transformers and regulators

Is prefab the only way to get AI capacity in six months?

No. In 2026 the six-month options are liquid-ready space that already exists, a GPU cloud contract, or a prefab hall on a site with power in hand (GR1D analysis).

The fastest builds on record reused something. NVIDIA says xAI and NVIDIA built Colossus in Memphis "in just 122 days," with 19 days from the first rack to training (NVIDIA, Oct 2024); Bisnow reports the site was a shuttered Electrolux factory, against about 15 months for other large training clusters (Bisnow, Sep 2025). Greenfield is slower: Crusoe began its Abilene, Texas campus in June 2024 and energized the first two buildings within a year (Crusoe, Sep 2025). IREN delivered Horizon 1 in Childress, Texas, 50 MW of direct-to-chip GB300 NVL72 capacity, to Microsoft in August 2026 under a contract announced in November 2025, about nine months later (GR1D calculation from IREN, Aug 2026).

A retrofit is not a shortcut either. CBRE warns that "for air-cooled data centers requiring conversion to liquid cooling, delays can extend beyond six months," and that ready-for-service fit-out dates "are expanding from weeks to several months" (CBRE, Aug 2026). A GB300 NVL72 needs 132–142 kW and a direct-to-chip loop per rack; the full list is in GB300 NVL72 colocation requirements.

What goes into a data center capacity reservation contract?

In 2026 a capacity reservation fixes the kW, a ready-for-service date, a ramp and a take-or-pay floor, typically 60–85% of allocated power, plus remedies for delay (CBRE, Aug 2026).

Two of these moved this year. Delay penalties: The Information reported in October that tenants on large campuses now ask for more than $5 million a day for late delivery, up from about $1 million in the spring, and that one OpenAI lease gives a free day of rent for each day of delay, doubling after 90 days (AI Weekly, Oct 2026). Exit rights: termination for convenience in hyperscale leases is "starting to go away," a Vinson & Elkins partner told Facilities Dive (Facilities Dive, Aug 2026).

Agree which power figure each clause uses. A GB300 NVL72 is rated at 132 kW nominal and about 155 kW at peak (HPE, Sep 2026), so the kW commitment, the take-or-pay floor and any overage charge should all name the same basis; the powered-shell, turnkey and colocation trade-offs behind the choice of contract are in powered shell vs turnkey vs colocation.

What should you sign when for GPUs on order?

Work back from the GPU delivery date: shortlist sites with power in hand 18–24 months out, sign the reservation 12 or more months ahead, and commission before the racks arrive (GR1D analysis).

Read as text

• What to sign when, working back from GPU delivery
• Each step starts as early as the work it unlocks takes to finish. Months are counted back from the day the GPU racks arrive.
• 18–24 months out · Shortlist by power in hand · TENANTGet each site's energization and go-live dates in writing · PROVIDERShares utility, permit and equipment status, dated · Most tenants signing today contract for 2028 deliveries (JLL)
• 12–18 months out · Sign the reservation · TENANTkW, ready-for-service date, ramp, take-or-pay, delay remedies · PROVIDEROrders switchgear and transformers if not on site · Switchgear 44 weeks, power transformers 128 (Wood Mackenzie)
• 9–12 months out · Freeze the rack spec · TENANTkW per rack, coolant temperature and flow, floor load, network · PROVIDERReleases the hall or module design · GB300 NVL72: 132 kW nominal, about 155 kW at peak (HPE)
• 6–9 months out · Confirm the ship window · TENANTOEM delivery range in writing; order optics and cabling · PROVIDERBuilds modules and prepares the site in parallel · About 20 weeks of factory time (BladeRoom)
• 1–3 months out · Witness commissioning · TENANTLiquid-cooled load banks, flush and fill, integrated test · PROVIDERRuns Level 1 to Level 5 commissioning · Dirty coolant "can result in extensive project delays" (ASHRAE)
• Delivery day · Provision and burn in · TENANTReceives, provisions and burns in each rack · PROVIDERHands over power and cooling at the contracted spec · 9 hours to multiple days per new NVL72 rack (SemiAnalysis)
• Two dates to ask for · Every provider has a lead time from order and a site go-live date. The later of the two is yours; a reservation that names neither is a waiting list.
• Two dates to ask for
• Every provider has a lead time from order and a site go-live date. The later of the two is yours; a reservation that names neither is a waiting list.

Ask every provider for two dates: the lead time from order, and the site's own go-live. The later of the two is yours. A reservation that names neither is a waiting list.

What can push a capacity date in Texas?

In Texas in 2026 the dates most at risk are power dates: ERCOT paused energization approvals for computational loads of 75 MW or more on August 3, 2026 (ERCOT, Sep 11, 2026).

The national record says to expect slippage. JLL found that 57% of data center projects had a construction delay of three months or more in 2025 (JLL, Jan 2026); Sightline Climate expects 30 to 50% of the 140 projects (16 GW) planned for 2026 to slip to 2027 or later (E&E News, Feb 2026); and high-voltage breakers reached 125 weeks in the second half of 2025, against 77 weeks in 2023 (Reuters via Kitco, Jul 2026).

Before you sign in Texas, ask for:

1. The energization date in an executed agreement, not a queue position, and whether the site's load is above or below ERCOT's 75 MW line (ERCOT large-load interconnection).
2. Which permits are issued and which are applied for. Under the TCEQ pause, an application on file is not a permit.
3. The long-lead equipment on order, with the factory's delivery dates.
4. The site's evidence against the four-rung ladder in AI data center site selection in Texas.

Frequently asked questions

How far in advance should I reserve colocation for GPUs?

In 2026, 12 to 24 months before the hardware arrives. More than 80% of primary-market space under construction is preleased (CBRE, Aug 2026), and most tenants signing now are contracting for 2028 deliveries (JLL). Turnkey space has been preleased about a year before completion, build-to-suit about two and a half years ahead.

How long does it take to get a GB300 NVL72 after ordering?

No NVIDIA partner publishes a standard lead time. GB300 NVL72 ships in volume, Vera Rubin production shipments began in August 2026, and NVIDIA expects supply to stay a bottleneck through at least its fiscal 2028, which ends in January 2028. Get the ship window in writing and have the hall commissioned first.

How long does a prefab data center module take from order to power-on?

Vendors quote about 20 to 39 weeks to deliver an operating factory-built hall, and NVIDIA's DSX POD cuts on-site deployment to as little as 14 days, but only at a site whose power, cooling and permits are ready. With power in place, modular builds take 6–12 months; without it, the grid sets the date.

Can I get AI data center capacity in six months?

Yes, if something already exists: liquid-ready colocation space, a GPU cloud contract, or a powered site for a prefab hall. xAI's first Colossus cluster took 122 days in a converted factory. Converting an air-cooled hall to liquid cooling can add delays beyond six months (CBRE, Aug 2026).

What is a take-or-pay clause in a colocation contract?

A minimum the tenant pays whether or not it uses the power it reserved. CBRE says these floors are becoming common in the US, at 60% to 85% of allocated power (Aug 2026). Agree whether the floor applies from the ready-for-service date or only after the ramp.

What happens if the data center isn't ready when my GPUs arrive?

That depends on the contract. Colocation agreements can carry tiered credits and termination rights for missed ready-for-service dates (Morgan Lewis, May 2026), and tenants on large campuses were reported in October 2026 to be asking for more than $5 million a day for late delivery. Without those clauses, the tenant carries the idle hardware.

Sources

1. North America Data Center Trends H1 2026 — CBRE, 2026-08-27
2. North America Data Center Report, Midyear 2026 — JLL, 2026-08-11
3. Americas Data Center Update, H1 2026 — Cushman & Wakefield, 2026-09-14
4. CBRE: Record Data Center Construction Fails to Ease Capacity Crunch — Data Center Frontier, 2026-08-28
5. DFW Data Center Demand Remains Hot as Pre-Leasing Hits 95% — Dallas Innovates, 2026-09-03
6. TD Cowen: AI demand caused 2.1GW of US data center leases in 90 days — DatacenterDynamics, 2023
7. NVIDIA Q2 Fiscal 2027 Earnings Call Transcript — NVIDIA, 2026-08-26
8. CoreWeave Delivers NVIDIA Vera Rubin NVL72 Performance at Production Scale, Starting With Cognition — CoreWeave, 2026-09-30
9. Fourth Quarter Fiscal 2026 Prepared Remarks — Super Micro Computer, 2026-08-11
10. Dell Technologies Q2 FY 2027: AI Orders Fuel Server and Storage Growth — Futurum, 2026-09-09
11. Memory squeeze set to tighten through 2028, Micron says — Computerworld, 2026-10-01
12. SK Hynix says memory shortage will last to 2030 — The Next Web, 2026-10-28
13. TSMC Reportedly Eyes 20% CoWoS Price Hike Amid AI Demand Surge — TrendForce, 2026-10-27
14. Nvidia's Kyber rack for Rubin Ultra reportedly delayed to 2028 — Tom's Hardware, 2026-07-06
15. Huang shrugs off protestors to kick off GTC Berlin with DSX pods, robots — The Next Platform, 2026-10-21
16. NVIDIA Introduces DSX POD AI Factories to Accelerate Europe's Infrastructure Deployment — NVIDIA, 2026-10-21
17. Modular Data Centres — BladeRoom, 2026-06-01
18. Flex Adds Factory-Built AI Modules to Modular Data Center Portfolio — Unite.AI, 2026-10-12
19. Eaton Expands Modular Data Center Offering for Rapid Deployment of AI Factories From Grid to Chip — Eaton via Business Wire, 2026-01-28
20. Vertiv Industrializes AI Deployment with Digitally Orchestrated Infrastructure — Vertiv via Nasdaq, 2026-02-26
21. Schneider Electric unveils new prefabricated reference design — Schneider Electric via GlobeNewswire, 2026-10-21
22. Modular data centers: Duos Edge AI and LG CNS — IEEE Spectrum, 2026-03-30
23. NVIDIA DSX POD: Modular AI Factory Deployment — ModulEdge, 2026-10-21
24. At GTC Berlin, NVIDIA Puts the AI Factory in an Actual Container — Futurum, 2026-10-22
25. NVIDIA's DSX POD: Closing the Compute Infrastructure Gap — ABI Research, 2026-10-26
26. Transformers in 2026: Shortage, Scramble, or Self-Inflicted Crisis? — POWER Magazine, 2026-01-02
27. Commissioning and Performance Validation, AI Data Center Framework — ASHRAE
28. How AI is reshaping data center power testing and commissioning — DatacenterDynamics
29. ClusterMAX 2.0: The Industry Standard GPU Cloud Rating System — SemiAnalysis, 2025-11-06
30. 2026 Global Data Center Outlook — JLL, 2026-01-08
31. US power companies scramble to secure equipment as surging data center demand strains supplies — Reuters via Kitco, 2026-07-09
32. The rise of Neocloud in the AI landscape — JLL, 2025-08-26
33. CoreWeave, Inc. Form 10-K, fiscal year 2025 — CoreWeave via SEC, 2026
34. NVIDIA Ethernet Networking Accelerates World's Largest AI Supercomputer, Built by xAI — NVIDIA, 2024-10-28
35. Musk's xAI Building Massive Memphis Data Center Hub At Record Speed — Bisnow, 2025-09-18
36. Crusoe Announces Flagship Abilene Data Center Is Live — Crusoe, 2025-09-30
37. IREN Delivers Horizon 1 to Microsoft and Achieves NVIDIA Exemplar Cloud Status on GB300 NVL72 — IREN, 2026-08-13
38. Activating the Colo Model: What Really Matters in Data Center Colocation Agreements — Morgan Lewis, 2026-05-18
39. Data Center Leasing: 5 Strategies for Navigating Today's Market — Vinson & Elkins, 2026
40. SB Energy, Inc. Form S-1 — SB Energy via SEC, 2026-08-31
41. Data Center Late Fees Jump to $5M a Day, Marsh Rolls Out Cover — AI Weekly, citing The Information, 2026-10-25
42. The Data Center Lease Is the Bond — Clifford Chance, 2026-07
43. Neocloud Storm Gathers as Data Center Deals Stall Over Credit Risk — Data Center Knowledge, 2026-04-06
44. Landlords flex rental power as data center viability becomes 'everybody's problem' — Facilities Dive, 2026-08-31
45. Batch Zero Update, ERCOT Board Item 14 — ERCOT, 2026-09-11
46. Texas Large Load Interconnection Update: ERCOT Batch Zero Pause and Verification Process — Baker Botts via Mondaq, 2026-08
47. Governor Abbott Orders TCEQ to Pause All Permit Issuance for Data Center Projects — Bracewell, 2026-09-23
48. TCEQ Backtracks, Suspends Data Center Permits Issued After Investigation Shows It Ignored Abbott's Pause — Public Citizen, 2026-10-22
49. Texas data center permits: TCEQ suspends approvals — Texas Tribune, 2026-10-21
50. Report: Data center construction delays grow — E&E News, 2026-02-24
51. HPE NVIDIA GB300 NVL72 QuickSpecs — HPE, 2026-09-08