Markets, demand and the site search

Lesson 1 set out the criteria by which sites are judged. Before those criteria can be applied, a developer must know what it is building and for whom, and must organize a search that finds good candidates without wasting money or revealing its plans too early. This lesson looks at the types of data center developers and the requirements each brings, the geography of data center markets, latency and when it matters, the stages and data sources of a site search, confidentiality, and how to make a decision that stands up when the weights and facts change.

Who builds data centers, and what they need

The requirements for a site depend strongly on the type of facility:

Large AI training campuses have changed site selection. Because a training cluster can sit far from users, the search is driven by where hundreds of megawatts can be energized soonest, which has pushed developers to rural areas with transmission capacity or gas supply, retired power plant sites with existing interconnections, and regions with surplus hydro or nuclear generation. Inference and cloud capacity still need to be close to network hubs and population centers.

The geography of data center markets

In the US, Northern Virginia, centered on Loudoun and Prince William counties, is the largest data center market in the world, built on dense fiber, early internet exchange points and a supportive tax regime, but now constrained by transmission capacity and growing community resistance. Other major markets include Dallas-Fort Worth, Phoenix, Chicago, Atlanta, Silicon Valley, Portland and central Oregon, Columbus (Ohio) and the New York-New Jersey area. Newer markets, many attracted by AI training demand, include parts of Texas outside Dallas, Georgia, Indiana, Pennsylvania, Wisconsin and Louisiana, where large tracts of land and power can be assembled.

Market analysts track inventory (operational megawatts), absorption (megawatts leased per year), vacancy, pipeline (megawatts under construction and planned) and rents. Low vacancy and high absorption signal demand; a large pipeline relative to absorption can signal future oversupply, though in recent years power constraints have limited how much of the pipeline actually gets built.

Latency: when distance matters

Light travels in optical fiber at about 200,000 km/s, two-thirds of its speed in vacuum, so each kilometer of fiber adds about 5 µs of one-way delay. Fiber routes are rarely straight; a route factor of 1.3 to 1.5 times the straight-line distance is typical. Network equipment adds further delay, as does every routing hop.

Synchronous replication between data centers in a region typically requires them to be within about 100 km of fiber, which is why cloud availability zones are clustered. For AI training, latency within the cluster matters enormously, but latency between the cluster and the outside world hardly at all, which explains the shift of training campuses to remote, power-rich locations.

Organizing the search

A site search starts with a written requirements brief, agreed with the business: the target capacity and phasing, the date the first capacity must be energized, the latency and network requirements, the acceptable range of energy prices, water and cooling strategy, minimum land area, risk tolerances for hazards, and any sustainability commitments such as carbon-free energy. The brief turns vague goals into criteria that can be screened, and identifies fatal flaws: conditions that rule out a site whatever its other merits, such as no realistic power before a deadline or location in a floodway.

The search then narrows in stages, as described in Lesson 1, from regions to communities to parcels. At each stage it uses spatial data layers, increasingly combined in geographic information systems (GIS):

• Transmission lines and substations with their voltage and, where available, available capacity from utility studies or hosting capacity maps.
• Natural gas transmission pipelines, for on-site generation.
• Long-haul and metro fiber routes, internet exchanges and carrier hotels.
• Flood maps, seismic hazard, wildfire risk, tornado and hurricane exposure (Lesson 1).
• Zoning maps, parcel boundaries and ownership from county GIS, prime farmland from USDA soil surveys, wetlands from the National Wetlands Inventory, and protected areas.
• Water supply and wastewater capacity, airports and their height restrictions, and the distance to homes, schools and other sensitive uses (Lesson 6).

Economic development agencies, utilities and brokers are important sources. Many utilities and regions now market "site-ready" or certified sites with known power, water, zoning and environmental status, which can cut months from due diligence.

Confidentiality

Large data center projects are sensitive. If a landowner learns that a major technology company is interested, the asking price rises; if a community learns of a project before the developer can explain it, opposition can form around rumors. Developers therefore usually work through brokers and special-purpose companies with neutral names, use project code names, and sign non-disclosure agreements with utilities, landowners and economic development agencies.

Confidentiality has limits. Public bodies are subject to access-to-information laws, council decisions must be made in public, and secrecy that lasts too long breeds distrust. Many communities now resent learning about a project only when it is presented as nearly decided, and several have adopted policies requiring disclosure of the end user before approvals. A good practice is to plan the moment of public disclosure deliberately, usually once land is secured and before formal applications, and to arrive with credible information on power, water, noise and benefits (Lesson 10).

Making a robust decision

The weighted scoring of Lesson 1 combines many criteria into one number, but the weights are judgments and the scores are uncertain. A robust decision examines how the ranking responds to different weights and to new information.

import numpy as np

rng = np.random.default_rng(410)
criteria = ["power", "energy cost", "network", "water", "hazards", "land", "approvals"]
base_w = np.array([0.30, 0.15, 0.15, 0.10, 0.10, 0.10, 0.10])
scores = {"Site A": [9, 6, 8, 5, 7, 6, 5], "Site B": [5, 9, 6, 8, 8, 7, 8], "Site C": [7, 7, 5, 9, 6, 9, 6]}
S = np.array(list(scores.values()), dtype=float)
print("weighted scores with the base weights:", ", ".join(f"{k} {v:.2f}" for k, v in zip(scores, S @ base_w)))

# Perturb the weights: Dirichlet draws around the base weights (concentration 50)
wins = np.zeros(3)
for _ in range(20000):
    w = rng.dirichlet(base_w * 50)
    wins[np.argmax(S @ w)] += 1
print("share of weight draws in which each site ranks first:", ", ".join(f"{k} {v / wins.sum():.0%}" for k, v in zip(scores, wins)))

# How much would the power weight have to fall for Site A to lose first place?
for pw in np.arange(0.30, 0.0, -0.01):
    w = base_w.copy(); w[0] = pw
    w[1:] = base_w[1:] / base_w[1:].sum() * (1 - pw)  # spread the remainder in proportion
    tot = S @ w
    if np.argmax(tot) != 0:
        print(f"Site A loses first place when the power weight falls to {pw:.2f} (scores {np.round(tot, 2)})")
        break

weighted scores with the base weights: Site A 7.10, Site B 6.85, Site C 6.90
share of weight draws in which each site ranks first: Site A 64%, Site B 21%, Site C 15%
Site A loses first place when the power weight falls to 0.25 (scores [6.96 6.98 6.89])

Site A ranks first in about two-thirds of the weight draws, so it is the leading candidate but not a certain one, and its lead depends on power being weighted at 25% or more. That points to the real question: how confident is the team in Site A's power advantage? If its energization date is firm, Site A is a sound choice; if it depends on a transmission upgrade not yet approved, its power score should be lower, and the ranking could flip. Sensitivity analysis thus directs due diligence to the facts that matter most.

Other techniques complement scoring:

• Total cost of ownership over the project life (land, construction, energy, taxes, water, network) for the finalists, converted to a present value, separates cost from qualitative factors.
• Scenario analysis tests the finalists against plausible futures: higher energy prices, a carbon price, a tightened water allocation, a delay in a transmission line.
• Portfolio thinking: experienced developers advance several sites in parallel, with land options and early utility applications, because some will fail for reasons no screening could predict. The cost of options and studies on sites that are dropped is the price of speed.