Intelligence Report // Q2 2026 // CoreBastion Security Consulting
A practitioner analysis of the technical, economic, and operational considerations for colocation and hyperscale operators evaluating legacy facility upgrades to support AI inference and AI training workloads. This report presents the evidence on both sides without advocating a position.
Contents
01 // Market Context
AI workload demand is scaling faster than purpose-built capacity can be delivered. That gap is forcing operators to evaluate what legacy inventory can realistically absorb.
The density problem. Legacy colocation facilities were engineered for 3 to 8 kW per rack during the client-server and early cloud eras. Hyperscale cloud-era campuses pushed this to 8 to 15 kW. AI inference workloads today typically require 20 to 50 kW per rack. AI training clusters, particularly GPU-dense configurations using NVIDIA GB200 NVL72 or Vera Rubin, target 100 to 200 kW per rack, with next-generation platforms approaching 600 kW. The gap between legacy design parameters and current AI requirements is not incremental. It is structural.
Demand Driver
New greenfield capacity faces power interconnection queues of 12 to 36 months in major markets. In Northern Virginia, three major proposals were denied in a single month in 2025 due to power scarcity, zoning hurdles, and community opposition. Operators with existing legacy facilities are under pressure to evaluate whether those sites can be repositioned faster than new builds can be delivered.
Supply Constraint
Greenfield construction costs have escalated to $8 to $10 million per MW in constrained markets, up from $6 to $8 million pre-COVID. Most industry professionals expect further 5 to 15 percent cost increases in 2026. Combined with multi-year delivery timelines, purpose-built capacity is expensive and slow, creating a window for well-positioned legacy facilities to serve near-term demand if they can meet tenant specifications.
02 // Economic Analysis
The economic viability of a legacy DC upgrade for AI depends on workload type, site-specific power ceiling, lease structure, and tenant quality. Neither a blanket endorsement nor a blanket dismissal is supported by the evidence.
The Affirmative Case
Retrofit costs for a well-positioned legacy facility typically come in at $5 to $6 million per MW, compared to $8 to $10 million per MW for greenfield construction in constrained markets. That cost advantage, combined with a 6 to 12 month faster time to revenue versus new build, produces a meaningful margin benefit if tenant rates hold. A 30 percent increase in IT load at stabilized rates can produce a 50 to 100 percent EBITDA improvement per site when the upgrade attracts higher-margin AI tenants. For AI inference specifically, the density and power requirements are within reach for many facilities that complete targeted upgrades to cooling and power distribution. Brownfield reinvestment is standard practice in analogous capital-intensive industries: aviation, manufacturing, and healthcare all retrofit rather than replace when the core asset has remaining useful life.
The Negative Case
For full AI training workloads, the industry evidence is that retrofit economics fail in most cases. The density requirements for GPU training clusters exceed what the majority of legacy facilities can structurally or electrically support without investment that approaches or exceeds greenfield costs. More critically, the word "if" in the affirmative case is load-bearing: retrofit projects in live operating data centers almost never complete on budget or on schedule, and every cost overrun compresses the margin advantage. Major hyperscale and colocation tenants increasingly pre-qualify sites before negotiating leases, and facilities that cannot clear power density or cooling specifications lose access to the premium tenants needed to justify the upgrade investment. The risk is spending at retrofit prices to attract tenants who will only pay legacy colocation rates.
The opportunity cost factor. A legacy facility generating revenue from conventional IT tenants has a real alternative: continue serving that market rather than undergoing a disruptive, capital-intensive retrofit. The foregone revenue during construction, plus the risk of losing existing tenants during the upgrade, is a cost that frequently does not appear in retrofit pro formas but is material to the actual return calculation.
| Factor | Retrofit | Greenfield | Assessment |
|---|---|---|---|
| Construction cost per MW | $5M to $6M | $8M to $10M | Retrofit advantage |
| Time to revenue | 6 to 12 months faster | 18 to 36 months typical | Retrofit advantage |
| Budget certainty | Low; live DC complications | Moderate; construction cost inflation | Retrofit disadvantage |
| AI training compatibility | Limited; structural constraints | Designed to spec | Retrofit disadvantage |
| AI inference compatibility | Viable with targeted upgrades | Designed to spec | Context dependent |
| Power interconnection | Existing service; may expand | New interconnection; 12 to 36 month queue | Retrofit advantage |
| Tenant pre-qualification risk | High; must meet hyperscaler specs | Moderate; designed to spec | Retrofit disadvantage |
| Permitting | MEP modifications; faster than new | Full entitlement; slower | Retrofit advantage |
03 // AI Inference
AI inference is the more viable retrofit target. The workload characteristics align better with what legacy facilities can deliver after targeted investment.
What inference requires. AI inference serves production AI applications: chatbots, image recognition, recommendation engines, fraud detection. It prioritizes low latency, proximity to end users, and predictable, sustained load rather than the burst compute and extreme density of training. Inference racks typically run at 20 to 50 kW, with some high-throughput inference clusters reaching 60 to 80 kW. This is within the range achievable through targeted power and cooling upgrades in many legacy facilities without structural modifications.
Colocation Operator Scenario
A multi-tenant colocation facility with existing fiber diversity, cloud on-ramps, and enterprise proximity is well-positioned to attract inference tenants. The key upgrades are replacing legacy PDUs with high-density units, adding in-row or rear-door cooling to support 20 to 40 kW per rack, and upgrading power distribution to support redundant A/B feeds at higher amperage. The facility does not need to support 200 kW per rack. It needs to support a density tier that existing tenants do not fill and that inference operators require. Lease structures with 5 to 10 year terms at power-based pricing give the operator the cost basis recovery period needed to justify upgrade capital.
Hyperscale Operator Scenario
Hyperscale operators running owned legacy campuses have a different calculus. Their legacy facilities may already have robust power infrastructure, fiber diversity, and operational maturity. The upgrade path for inference focuses on cooling densification: adding direct-to-chip liquid cooling manifolds to a defined pod within the existing hall, upgrading power whips and branch circuits to match GPU server requirements, and deploying intelligent PDUs for capacity tracking. The hyperscale operator avoids construction risk on new sites while extending the productive life of existing campus capacity. Inference does not require the full campus; a dedicated pod approach allows phased investment.
Inference Upgrade Priority Stack
| System | Legacy Baseline | Inference Requirement | Upgrade Path | Complexity |
|---|---|---|---|---|
| Rack power density | 5 to 8 kW/rack | 20 to 50 kW/rack | PDU replacement, branch circuit upgrade | Moderate |
| Cooling capacity | Air-cooled, 5 to 8 kW/rack | In-row or rear-door for 20 to 50 kW | In-row cooling units; CHW piping to rows | Moderate |
| Power distribution (busway) | Legacy PDUs at 20 to 30A | 60A to 100A per rack | Overhead busway replacement or tap-off upgrade | Moderate |
| Network fabric | 10G switching | 100G to 400G spine-leaf | Full switch replacement; fiber plant upgrade | Moderate |
| Power redundancy (A/B) | Often single-path | Dual redundant feeds required | Second feed and transfer switching per row | Moderate |
| UPS capacity | Sized for legacy load | Rerating required for new IT load | Module addition or replacement; battery refresh | Manageable |
| Utility service capacity | Often at or near current load | Increased MW required | Utility load study; service entrance upgrade if needed | High if capacity constrained |
The inference window. JLL anticipates that by 2027, inference workloads will overtake training as the dominant AI requirement. Deloitte estimates inference already accounted for half of all AI compute in 2025 and projects that share growing to two-thirds in 2026. This near-term inference growth is the market window legacy facilities can realistically target before more purpose-built inference capacity comes online.
04 // AI Training
AI training presents a fundamentally different challenge. The density, power, and structural requirements of large-scale training clusters exceed what the majority of legacy facilities can accommodate at economically viable retrofit costs.
The density gap for training. AI training clusters using NVIDIA H100, H200, or GB200 NVL72 configurations require 100 to 200 kW per rack at scale, with next-generation Vera Rubin platforms targeting 600 kW per rack. Legacy facilities designed at 5 to 15 kW per rack face a structural gap of 10x to 40x. This is not addressable through PDU or cooling upgrades alone. It requires fundamental infrastructure replacement: substation-level power upgrades, chilled water plant expansion, structural floor loading modifications, and in many cases full power distribution bus replacement. The total cost of these interventions frequently approaches greenfield construction cost while delivering less usable capacity due to building constraints.
Colocation Operator Scenario
A standard multi-tenant colocation facility attempting to support a large-scale GPU training cluster faces compounding constraints. Floor loading for a fully populated NVL72 rack can exceed 300 pounds per square foot, which many legacy raised-floor environments cannot support without structural reinforcement. The substation or utility service may not have capacity for the required load increase. Chilled water infrastructure to support direct-to-chip cooling at 100 kW per rack requires a complete plant expansion. If the facility overcomes these constraints, it still faces the challenge that hyperscale training tenants pre-qualify sites to exacting specifications. Failing qualification after capital investment is a real risk.
Hyperscale Operator Scenario
Hyperscale operators with owned campuses have more options. An owned campus with available land may be able to add a dedicated high-density training pod as a new building rather than a retrofit of existing halls. This is closer to greenfield construction on an existing campus than a true retrofit, but it benefits from existing utility interconnections, fiber, and operations infrastructure. Alternatively, some legacy hyperscale halls built for 15 to 20 kW per rack can be upgraded to support 50 to 100 kW per rack with aggressive direct-to-chip liquid cooling deployment, which is viable for certain older-generation training configurations. This is not a path to Vera Rubin densities, but it extends the useful life of the asset for current-generation training.
Training Upgrade Gap Analysis
| System | Legacy Baseline | Training Requirement | Gap Assessment |
|---|---|---|---|
| Rack power density | 5 to 8 kW/rack | 100 to 600 kW/rack | 10x to 75x gap; typically not addressable in legacy |
| Direct-to-chip cooling | Air-cooled only | Required at 100+ kW/rack | Full system installation; CHW plant expansion required |
| Utility power capacity | Sized for 5 to 8 kW/rack average load | 10x to 20x existing IT load | Substation upgrade or new service typically required |
| Floor loading | 125 to 150 lbs/sq ft typical | 250 to 350 lbs/sq ft for dense GPU racks | Structural engineering assessment mandatory; often cost-prohibitive |
| High-speed interconnect (InfiniBand/RoCE) | No native support | 400G to 800G GPU interconnect fabric | Full network replacement; specialized fabric required |
| Power distribution bus | 20 to 30A branch circuits | 100A to 200A per rack; busway rated for load | Full busway replacement in dense zones |
| Generator capacity | Sized for legacy N+1 load | Full IT load backup at new density | Additional gensets and fuel infrastructure required |
05 // Infrastructure Reference
Every major system category that must be evaluated in a legacy-to-AI upgrade. Organized from utility feed through to rack-level connections. Each item is a discrete project with its own lead time, permitting requirements, and cost.
Practitioner note. In a live operating data center, everything is a project. Adding cabling alone requires assessing tray capacity, calculating NEC fill compliance, and potentially replacing tray, hangers, and support structures before a single new circuit can be run. Operators who have not executed a significant infrastructure upgrade in a legacy facility frequently underestimate the cumulative complexity of concurrent system modifications in a live environment.
Utility and Substation
Primary Power Distribution
Secondary Power Distribution to Rack
Cabling and Cable Management
06 // Power Pathway Detail
The power delivery chain from substation through to individual rack connections is the most capital-intensive upgrade dimension and the one most commonly underestimated in initial feasibility assessments.
| Power System Layer | Legacy Condition | AI Requirement | Typical Upgrade Action | Lead Time |
|---|---|---|---|---|
| Utility service entrance | Sized for original load | Expanded service or second feed | Utility load study; service entrance modification | 12 to 36 months if grid-constrained |
| Main switchgear | Rated for existing load; potentially aged | Higher bus ampacity; current interrupting capacity | Bus replacement or parallel switchgear | 6 to 12 months |
| Transformers | kVA sized for original IT load | Higher kVA for dense AI racks | Replacement or parallel transformer addition | 6 to 18 months (long lead equipment) |
| UPS systems | Modular or monolithic; battery bank sized for existing load | Rerating; battery runtime recalculated | Module addition or system replacement; battery refresh | 3 to 9 months |
| Generators | N+1 for existing load | Full backup at new IT load density | Additional gensets; fuel system expansion | 6 to 18 months |
| Overhead busway | 100A to 225A bus rated for legacy density | 400A to 800A or higher for AI rack rows | Full busway replacement in upgrade zones | 3 to 6 months |
| Row PDUs | 20A to 30A, C13/C14 outputs | 60A to 200A, C19/C20 or proprietary | PDU replacement; circuit rebalancing | 1 to 3 months |
| Power whips to rack | C13/C14 at 10A to 15A | C19/C20 at 16A to 32A; GPU servers may use proprietary 60A connectors | Full whip replacement; rack power budget recalculation | Weeks; but requires concurrent PDU upgrade |
| In-rack power distribution (rPDU) | Basic strip PDUs | Intelligent rPDUs with per-outlet metering | Intelligent rPDU deployment; integration to DCIM | 1 to 2 months |
07 // Cooling Systems
Cooling is the second major constraint after power. Legacy air-cooled infrastructure was designed for 5 to 8 kW per rack. AI workloads require approaches that do not scale from that baseline without significant investment.
In-Row and Rear-Door Cooling
In-row cooling units deploy alongside racks in existing rows and capture heat at the source rather than relying on room-level airflow management. They require chilled water supply piping to the row, which is a construction project but not a facility-wide plant expansion. Rear-door heat exchangers attach directly to racks and provide passive or active cooling at similar densities. Both approaches are viable for inference-level workloads and can be deployed in phases. In-row units consume rack space (typically one unit per 4 to 8 racks), which reduces deployable density. Both require hot aisle/cold aisle containment to function effectively, which is often absent or incomplete in legacy facilities.
Direct-to-Chip Liquid Cooling
Direct-to-chip (DTC) cooling delivers coolant directly to the processor via manifolds and quick-disconnect fittings at each server. It is the required approach for AI training densities at 100 kW and above. The retrofit complexity is significant: chilled water supply and return piping must be routed to every rack in the upgrade zone, manifolds installed per rack, quick-disconnect fittings installed per server, and a leak detection system deployed throughout. Any coolant leak in a live data center is a critical incident. DTC has reached 47 percent market share in new deployments and the liquid cooling market hit $5.52 billion in 2025, which has standardized solutions and reduced costs, but the retrofit installation complexity in a live facility remains high.
08 // Decision Framework
Before committing capital to an upgrade program, operators should work through a structured site viability assessment. The answers to these questions determine which path, if any, makes economic sense.
| Assessment Area | Question | Implication if Constrained |
|---|---|---|
| Power ceiling | What is the maximum MW the utility can deliver to this site within a 24-month window? | If the power ceiling cannot reach the target AI load, the retrofit path is closed regardless of other factors |
| Target workload | Is the intended tenant running inference, training, or a mix? | Training at full density is rarely viable in true legacy facilities; inference is the more realistic target |
| Structural capacity | What is the floor loading rating, and can it support GPU rack configurations at target density? | Structural reinforcement is expensive and disruptive; if the floor cannot support the load, the path is likely not viable |
| Lease economics | What lease term and rate structure is required to recover upgrade capital at the cost of debt? | If tenant rates in the market do not support payback within the lease term, the model does not work |
| Tenant pre-qualification | Have target tenants specified their technical requirements, and has the facility been informally screened against those specs? | Investing before understanding what the tenant requires is the most common path to stranded upgrade capital |
| Fiber and connectivity | Does the facility have diverse carrier access and cloud on-ramp proximity appropriate for the target workload? | Inference workloads require low-latency access to end users; a facility without fiber diversity or cloud proximity is poorly positioned for inference |
| Existing tenant impact | What is the revenue at risk from existing tenants during the upgrade period, and is there a sequencing plan that protects it? | Disrupting existing tenants during upgrades is a real revenue risk; a phased approach that contains construction to a defined zone is essential |
| Alternative use case | If the facility cannot meet AI requirements, what is the alternative highest-value use for the asset? | Some legacy facilities are better positioned as enterprise IT, cloud archive, or government colocation than as AI infrastructure; the upgrade is not the only option |
The framing question. The productive question for any site evaluation is not "can we make this an AI data center" but rather "what AI-adjacent workloads can this site serve, and what is the minimum viable upgrade path to get there." That framing separates the opportunities where a targeted investment yields a sustainable return from the situations where capital is deployed against constraints that cannot be overcome at viable cost.
Sites Best Positioned for Retrofit
Facilities with existing substation capacity headroom or a confirmed utility upgrade pathway, location near dense fiber routes or cloud on-ramps, floor loading above 200 lbs/sq ft, existing chilled water plant with available capacity, and a tenant mix that can be protected in a defined construction zone are the most viable retrofit candidates. Colocation operators in these positions who can secure a long-term inference tenant commitment before committing upgrade capital have the strongest probability of a viable return.
Sites Where the Economics Are Unlikely to Work
Facilities at or near utility capacity with no credible expansion pathway, legacy raised-floor environments with floor loading below 150 lbs/sq ft, locations with limited fiber diversity, and facilities targeting large-scale training tenants who have not pre-qualified the site face the combination of constraints that the industry evidence suggests most commonly produces a failed retrofit. In these cases, the more durable strategy may be to continue serving the market the facility was built for while evaluating greenfield options for AI capacity.
CoreBastion Security Consulting // Intelligence Report // Q2 2026 // Restricted Distribution
This report is produced for practitioner use. Economic data sourced from ABI Research, JLL, BloombergNEF, Deloitte, and Data Center Dynamics. Infrastructure guidance reflects practitioner experience and industry reference standards. Site-specific conditions vary materially; all upgrade decisions should be validated by licensed mechanical and electrical engineers with direct facility assessment.