A GPU colocation lead time is the elapsed time between signing a colocation agreement and running production AI workloads in the space. It covers three things in sequence: the operator confirming it has the power and cooling your racks need, that capacity being energized and commissioned, and your hardware landing, racking and cross-connecting. In 2026 the first step dominates the other two. Where power is already built and unclaimed, buyers move in weeks. Where it is not, the wait runs into quarters, and for anything depending on a new grid connection, into years.
What actually sets a GPU colocation lead time?
Four inputs decide almost all of it. First, whether the operator has energized, unreserved capacity at your density today. Second, whether your racks are air cooled or liquid cooled, because a liquid fit-out adds commissioning work that an air-cooled cabinet does not. Third, the size of the request, since a single cabinet fits into space an operator already has while a multi-megawatt block usually does not. Fourth, the market, which mostly acts as a proxy for the first input.
Hardware delivery matters less than buyers expect. Most schedules that slip in 2026 slip on the facility side rather than on the GPUs. Allocation letters and delivery dates from the hardware vendor are worth confirming early, though they are rarely the binding constraint. Our GPU colocation guide covers the density and cooling side of the requirement in more detail.
How long do primary markets take in 2026?
Primary market vacancy hit a record low of 1.4 percent in the first half of 2026, down from 1.6 percent a year earlier, and Northern Virginia sits at 0.2 percent in the same report (CBRE North America Data Center Trends H1 2026). Available capacity across all primary markets is under 1,500 MW, roughly six months of demand at current absorption rates. Of everything under construction, 80.4 percent is already preleased, against 74.3 percent a year prior, so the pipeline is largely spoken for before it is finished.
For a buyer that arithmetic turns into a queue. QuoteColo’s September 2026 guidance puts liquid-cooled deployments in power-constrained primary metros at three to six months, and Ashburn specifically at a six to twelve month queue. If the requirement depends on new power rather than existing power, the horizon changes entirely: JLL’s 2026 global outlook reports that the average wait for a grid connection in primary data center markets now exceeds four years.
How fast are secondary markets?
Considerably faster, when the power is already there. QuoteColo puts typical windows in Colorado, Phoenix, Atlanta and Ohio at four to eight weeks. A single air-cooled rack running inference in a secondary market with available power can be live in two to six weeks. A GB200-class cluster drawing 80 kW or more per rack runs 60 to 90 days, because the liquid loop and its commissioning sit ahead of everything else on the schedule.
The speed comes from slack, not from anything structural about a smaller market. Atlanta makes the point well. CBRE counts it as a primary market and records 2,882 MW under construction there in the first half of 2026, up 52.3 percent year over year, enough to pass Northern Virginia as the largest construction market in North America. Atlanta still quotes secondary-market lead times in QuoteColo’s September survey. Supply is what moves the number, not the label on the market. Two metros in the same tier can sit a full quarter apart on availability.
A secondary market with no spare power is slow for the same reason. Checking availability before committing to a market avoids restarting a search six weeks in, which is roughly how long it takes for a shortlist built on reputation to come back empty. Markets such as Columbus are worth screening on current availability rather than on last year’s reputation.
What do the typical windows look like side by side?
| Deployment | Typical window | What sets it | Source and period |
|---|---|---|---|
| Single rack, air cooled, secondary market with power available | 2 to 6 weeks | Existing capacity at standard density | QuoteColo, September 2026 |
| Multi-rack in Colorado, Phoenix, Atlanta or Ohio | 4 to 8 weeks | Unclaimed energized capacity | QuoteColo, September 2026 |
| GB200-class cluster at 80 kW or more per rack | 60 to 90 days | Liquid loop fit-out and commissioning | QuoteColo, September 2026 |
| Liquid cooled in a power-constrained primary metro | 3 to 6 months | Queue for energized capacity | QuoteColo, September 2026 |
| Ashburn, Northern Virginia | 6 to 12 months | 0.2 percent vacancy in H1 2026 | QuoteColo September 2026; CBRE H1 2026 |
| New build dependent on a new grid connection, primary market | Over 4 years on average | Utility interconnection queue | JLL 2026 global outlook |
Why is the gap so wide?
Three forces compound. Preleasing takes capacity off the table before it exists, which is what the 80.4 percent figure describes. Construction economics slow new entrants, with costs reaching 10.7 million dollars per MW in 2025 and forecast at 11.3 million in 2026 (JLL 2026 outlook). And the interconnection queue sets a floor that no amount of capital shortens, which is the four-year figure above.
None of that is going to loosen quickly. JLL projects 97 GW of new capacity between 2025 and 2030, which roughly doubles the sector, but that supply arrives across five years while AI demand is arriving now. Planning against 2027 and 2028 availability tends to be more useful than waiting for the market to loosen. The same squeeze is what pushes mid-size requirements into difficulty, as we covered in buying 250 kW to 3 MW.
When is a primary market still worth the wait?
Often, and the reasons are specific rather than sentimental. Inference serving end users has a latency budget, and the population you are serving may sit in a primary metro. Carrier density and cloud on-ramps are concentrated in those markets, so a hybrid architecture with heavy egress to a public cloud can cost more to run from a cheap secondary site than the rent saves. Some regulated workloads carry data residency terms that name a region, which removes the choice of market entirely.
Training is the clearer case for moving. It is latency tolerant, it runs in bulk, and it cares about power price and availability far more than about proximity. A training cluster that waits nine months for Ashburn when Columbus or Salt Lake City could have it running in six weeks is usually paying for an attribute it does not use. The exception is a cluster that will later serve inference from the same site, where the network position earns its keep on the second workload. Rate differences between markets are set out in our colocation pricing by market breakdown.
How do you shorten the wait?
- Ask for energized capacity by name rather than asking what is available, because an operator’s answer to the second question often includes space that is not powered yet.
- Separate the air-cooled and liquid-cooled parts of the requirement and let the air-cooled portion go first, so inference or preprocessing can start earning while the dense racks wait.
- Size the first phase to what you can actually fill in a quarter, and put the rest on a ramp schedule with priced options rather than committing to it up front.
- Run two or three markets in parallel to the same deadline instead of sequentially, since a serial search adds the full evaluation cycle to every rejection.
- Get the cross-connect and carrier requirements into the first conversation, as network provisioning quietly adds weeks after the space itself is ready.
- Confirm who holds the utility relationship and whether the capacity is contracted or merely requested. Contracted capacity supports a firm date; a request sitting in a utility queue does not.
Running markets in parallel is easier with a single view of who operates where, which is what the provider directory is for.
FAQ: GPU colocation lead times
Q: How long does GPU colocation take to deploy in 2026?
Between two weeks and twelve months, depending almost entirely on whether energized capacity exists at your density. Secondary markets with available power typically run four to eight weeks, liquid-cooled deployments in constrained primary metros run three to six months, and Ashburn runs six to twelve months (QuoteColo, September 2026).
Q: Are secondary markets always faster?
No. They are faster when they have unclaimed energized capacity, which many currently do. A secondary market that has just absorbed a large tenant can quote the same wait as a primary one. Check current availability before choosing a market, rather than relying on how the market was described last year.
Q: Does liquid cooling add to the lead time?
Yes, generally. An air-cooled cabinet can often drop into existing space, while a liquid loop needs fit-out and commissioning. QuoteColo puts GB200-class deployments at 80 kW or more per rack at 60 to 90 days even where capacity exists.
Q: Why does Northern Virginia quote the longest waits?
Because it has the least room. CBRE records 0.2 percent vacancy there in the first half of 2026, the lowest of any primary market it tracks, alongside 2,420.2 MW under construction that is mostly preleased.
Q: What is the single most common cause of a slipped date?
Capacity that was requested from a utility but not yet contracted, presented to the buyer as though it were firm. Asking whether power is contracted or requested resolves most of it in one question.
Get a read on what is actually available
Lead times in 2026 are a supply question before they are a market question, and the supply picture changes month to month. Tell us the density, the size and the deadline, and we will come back with the markets and operators that can hit it. Request a quote and we will screen current availability against your date.