IndexData Center Demand Forecasting in 2026: How Developers Read Conflicting Market Signals
Demand forecasting for data centers now has to reconcile leasing demand, power availability, workload mix and delivery capacity before a site is worth pursuing.
Data center demand forecasting in 2026 is no longer a broker absorption exercise. It is a development underwriting discipline.
The headline numbers are enormous. JLL's 2026 Global Data Center Outlook says the global data center sector is likely to add 97 GW between 2025 and 2030, effectively doubling global capacity and expanding at a 14% CAGR. The same report estimates up to $3 trillion of infrastructure investment by 2030, including $1.2 trillion of real estate asset value and another $1 trillion to $2 trillion of tenant IT fit-out.
Those numbers are real enough to change capital allocation. They are not specific enough to buy land.
For institutional developers, the question is narrower: which demand is executable in a specific market, under a specific utility regime, with a specific delivery window? That is where most generic market forecasts fail.
Leasing demand and buildable demand are different numbers
Leasing demand measures appetite. Buildable demand measures what can actually be delivered.
A hyperscaler can sign leases, reserve capacity and announce capex faster than utilities can energize a campus. That gap is now the defining feature of the market. JLL reports that average grid connection waits in primary data center markets now exceed four years. It also notes that operators are moving toward behind-the-meter power, battery storage and direct energy generation because normal interconnection timelines cannot keep pace.
CBRE's H1 2026 North America data center research points in the same direction: AI demand remains strong, but power constraints are shaping where growth can land. That is the key underwriting distinction. Demand is not disappearing. It is being filtered through power.
A credible forecast separates four layers:
Tenant demand: signed leases, active RFPs, cloud region expansion plans and AI workload growth.
Deliverable capacity: sites with utility service, equipment procurement visibility and credible energization dates.
Power-adjusted pipeline: announced projects discounted for interconnection, substation, tariff and equipment risk.
Capital-ready supply: projects that can secure debt and equity after power, cost and entitlement risk are priced.
The third layer is where developers should spend the most time. Announced pipeline is noisy. Power-adjusted pipeline is investable.
Workload mix changes where demand lands
AI demand is not one thing.
JLL estimates AI represented about a quarter of data center workloads in 2025 and could represent half by 2030. It also expects inference workloads to overtake training as the main AI requirement around 2027. That matters because training and inference do not produce the same real estate demand.
Training clusters want massive power blocks, high-density cooling and specialist campuses. Inference demand is more distributed. It rewards latency, regional placement and network proximity. A market forecast that treats every AI megawatt as the same megawatt will misread both site criteria and lease depth.
For developers, this changes the diligence question. The old question was, 'Is there enough demand in this market?' The better question is, 'What kind of demand can this site serve better than competing sites?'
A 200 MW campus near constrained transmission may work for a dedicated training tenant if power can be privately procured. A 20 MW infill facility with strong fiber and low latency may fit enterprise inference better. Both count as data center demand. They are different products.
Cost escalation belongs inside the demand forecast
Demand that cannot support replacement-cost rent is not executable demand.
JLL reports that average global data center construction cost rose from $7.7 million per MW in 2020 to $10.7 million per MW in 2025, a 7% CAGR. It forecasts $11.3 million per MW for 2026 before tenant IT fit-out. That cost base changes which tenants can absorb new supply and which markets can justify speculative development.
This is where some demand forecasts overstate opportunity. They count required capacity without testing whether the rent, power price, tariff exposure and delivery cost clear the return threshold.
A developer's demand model should carry at least three cost scenarios:
Base case shell-and-core cost per MW.
AI-density premium for liquid cooling, higher electrical capacity and reinforced floor loading.
Delay case for transformer, switchgear, interconnection or permitting slippage.
If demand only works in the base case, it is not demand. It is an option on perfect execution.
AI improves forecasting by connecting the evidence
AI does not make demand forecasts more useful by producing prettier market summaries. It helps when it connects fragmented evidence that humans miss or update too slowly.
A real demand forecast should ingest utility queue data, tariff filings, power availability letters, substation upgrade history, land control records, leasing announcements, construction starts, zoning activity and local opposition signals. None of those inputs is decisive alone. Together, they show whether demand can become capacity.
This is where an AI-native operating partner like Build fits the development workflow. The value is not a generic market memo. The value is a live demand model tied to real sites, real power constraints and real delivery risk.
Human judgment still owns the final call. Developers need to decide whether to trust a utility's timeline, whether a tenant's expansion plan is bankable and whether a local approval path is politically durable. AI can narrow the field and keep the evidence current. It cannot turn weak evidence into conviction.
The developer's forecast should answer one hard question
The useful forecast is not 'How much demand exists?' It is 'Which demand can this team deliver against, at this site, before the window closes?'
The best forecasts rank demand by executable fit, not just size. They show where power is real, where construction cost is financeable, where inference demand needs regional capacity and where announced supply should be discounted.
Data center developers do not need more market excitement. In 2026, demand is abundant. Deliverable demand is scarce.