AI infrastructure financing is not data center financing with GPUs in the racks. It is a distinct asset class with its own economics, depreciation curves, revenue models, and risk profiles that require specialized capital structures. A traditional data center rack draws 5-10 kW of power. An AI training rack draws 30-100+ kW. That single difference cascades through every aspect of the capital stack, from the building shell to the cooling system to the power infrastructure to the equipment itself. In 2024-2025, roughly $200 billion flowed into AI infrastructure globally. By 2026, that number is tracking toward $300 billion+. The capital is available. The challenge is structuring it correctly, separating the real estate from the equipment, matching financing terms to asset useful life, and building a capital stack where each layer is appropriately sized for the risk it carries. This guide breaks down how AI infrastructure deals actually get financed in 2026, from $10M GPU cloud deployments to $1B+ purpose-built training facilities. No theory, just the structures that are closing deals right now.
1Why AI Infrastructure Is Fundamentally Different from Traditional Data Centers
The comparison between AI infrastructure and traditional data centers is not a matter of degree. It is a difference in kind. Every assumption that works for traditional IT infrastructure breaks down when you apply it to AI compute. Understanding these differences is not academic. It directly determines which financing structures work and which ones fail.
| Variable | Traditional Data Center | AI Infrastructure | Financing Impact |
|---|---|---|---|
| Power per Rack | 5-10 kW | 30-100+ kW | 3-10x higher power infrastructure capex; utility capacity becomes the binding constraint |
| Cooling Method | Air cooling (CRAH/CRAC) | Liquid cooling mandatory (direct-to-chip, rear-door, immersion) | $2-5M+ additional capex per MW for liquid cooling infrastructure |
| Equipment Cost | $50-150K per rack | $500K-$2M+ per rack (GPU nodes) | Equipment value often exceeds building value; requires separate financing layer |
| Useful Life | 7-10 years (servers) | 3-5 years (GPUs) | Shorter depreciation = shorter loan terms; residual value risk is real |
| Revenue Model | Colocation leases ($/kW/month) | GPU-as-a-Service, training contracts, inference hosting | Revenue tied to compute utilization, not just space occupancy |
| Total Project Cost | $8-15M per MW (IT load) | $15-30M+ per MW (including GPUs and liquid cooling) | Larger capital stacks with more layers; equipment financing becomes the dominant component |
| Tenant Profile | Enterprise IT, SaaS, managed hosting | AI labs, hyperscalers, GPU cloud providers, sovereign AI programs | Different credit profiles; long-term compute contracts vs traditional colocation leases |
The Core Problem for Lenders
In a traditional data center, the building and its mechanical/electrical infrastructure represent 80-90% of total project value. The IT equipment belongs to the tenant. In an AI infrastructure deal (particularly GPU cloud or inference hosting) the operator owns the GPUs. A single H100 cluster can represent $200M+ in depreciating hardware sitting inside a $30M building shell. That inversion of the asset value ratio breaks traditional real estate underwriting. You cannot slap a 25-year CMBS loan on an asset where the most valuable component has a 3-5 year useful life.
Power density is the multiplier that changes everything. When a single rack draws 70 kW instead of 7 kW, you need 10x the power distribution infrastructure, 10x the cooling capacity, and 10x the backup generation, all within the same footprint. This is why AI-ready facilities cost $15-30M+ per megawatt of IT load, compared to $8-15M for traditional colocation. The building shell is roughly the same cost. Everything inside it costs dramatically more.
Liquid cooling is no longer optional. Air cooling physically cannot remove enough heat from GPU-dense racks. At 40+ kW per rack, you need direct-to-chip liquid cooling, rear-door heat exchangers, or full immersion cooling. The transition from air to liquid cooling adds $2-5M per megawatt in infrastructure capex, changes the ongoing maintenance profile, and introduces a mechanical system that most traditional data center operators have never managed. Lenders underwriting AI facilities need to evaluate cooling technology risk alongside credit risk.
GPU depreciation creates a financing mismatch. Traditional servers depreciate over 5-7 years with relatively stable residual values, a 3-year-old server still runs the same workloads. GPUs are different. Each new generation (H100 to H200 to B200 to GB200) delivers 2-4x performance improvements. An H100 purchased in 2024 for $30,000-$40,000 will compete against hardware that delivers equivalent performance for a fraction of the cost by 2027. This generational obsolescence risk is the single biggest challenge in AI equipment financing and the primary reason GPU financing terms are shorter and rates are higher than traditional IT equipment loans.
2Capital Structures by AI Facility Type
There is no single "AI infrastructure financing" product. The capital structure depends entirely on the facility type, who owns the GPUs, the revenue model, and the project scale. Here are the four primary categories and how capital flows into each.
1. Purpose-Built AI Training Facilities ($100M-$1B+)
These are the flagship projects, ground-up facilities designed specifically for large-scale AI model training. Think 50-200+ MW campuses with liquid cooling throughout, dedicated substation infrastructure, and long-term compute contracts with hyperscalers or major AI labs. Total project costs routinely exceed $500M and can reach several billion dollars for multi-phase campus developments.
Typical Capital Stack:
- Senior debt (50-60% of cost): Project finance from infrastructure-focused lenders. Non-recourse to the sponsor, secured by the facility and long-term contracts. 5-7 year construction-to-permanent structures with 15-20 year amortization on the real estate component.
- Mezzanine / preferred equity (15-25%): Infrastructure PE funds or private credit. Returns in the 12-18% range, structured as preferred equity or subordinated debt.
- Sponsor equity (20-35%): Developer equity, often co-invested with a hyperscaler or strategic partner who also provides the anchor compute contract.
- Equipment financing overlay: Separate facility for GPU hardware, sized to match 3-5 year refresh cycles. Often structured as a revolving equipment line that funds each generation of GPUs independently.
What makes these deals work: The anchor tenant. A 10-15 year compute contract from a creditworthy hyperscaler or AI lab is the foundation of the entire capital stack. Without it, project finance lenders will not touch a $500M+ ground-up facility. The contract provides the revenue certainty that supports non-recourse senior debt at reasonable rates (SOFR + 200-350 bps for investment-grade tenants). We have seen deals where the hyperscaler provides both the anchor lease and co-invests equity, aligning their interest in the facility with the developer's.
What kills these deals: Power. You can have $1B in committed capital and a signed hyperscaler contract, but if the utility cannot deliver 100+ MW to your site within 18-24 months, the deal dies. Interconnection queue timelines in many markets now stretch 3-5 years. Smart developers are securing utility capacity commitments before they have a single dollar of financing in place. The power comes first. Everything else follows.
2. GPU Cloud and Inference Hosting ($10M-$200M)
This is the fastest-growing segment, companies that lease data center space and fill it with GPU hardware to sell compute capacity on-demand or through reserved instances. The operator does not own the building; they own the GPUs. The business model is fundamentally an equipment arbitrage: buy GPUs, sell compute hours at a margin.
Typical Capital Stack:
- Equipment financing (60-80% of GPU cost): 3-5 year terms, 80-100% advance rate on new GPUs from tier-1 manufacturers. Monthly or quarterly payments. Often structured with 6-12 month interest-only period during deployment and ramp.
- Venture debt (supplemental): For VC-backed GPU cloud startups, venture debt provides non-dilutive capital to supplement equipment financing. Typically 20-35% of last equity round, 3-4 year terms.
- Data center lease: Not financing per se, but the facility lease is the second-largest cost. Triple-net colocation leases at $120-200+ per kW/month for AI-ready space with liquid cooling support.
- Equity: VC funding or growth equity. Increasingly, strategic investors (chip manufacturers, cloud platforms) are co-investing alongside financial sponsors.
What makes these deals work: Contracted revenue. Lenders underwriting GPU equipment loans want to see that the compute capacity is pre-sold through reserved instance contracts, not dependent on spot market pricing. A GPU cloud operator with 70%+ of capacity under 1-3 year contracts will get significantly better financing terms than one selling 90% on-demand. The contract backlog is the collateral story, not just the hardware.
What kills these deals: Residual value assumptions. If your financing model depends on the GPUs being worth 40-50% of original cost at the end of a 3-year loan, you are making an aggressive bet. Generational performance improvements mean that an H100 in 2027 competes against hardware that is 2-4x faster at a similar price point. Conservative lenders assume 10-20% residual value on GPUs at end of term. If your debt service coverage only works with rosy residual assumptions, the deal will not close.
3. AI-Ready Colocation ($20M-$500M)
These are existing or purpose-built colocation facilities that are being upgraded or designed to support AI workloads. The colocation operator provides the building, power, and cooling; tenants bring their own GPU hardware. The financing challenge is the power and cooling upgrade, converting a traditional 5-10 kW/rack facility to support 30-100+ kW/rack densities.
Typical Capital Stack:
- Traditional DC financing (base layer): CMBS, bank term loans, or life company debt on the stabilized real estate. 60-70% LTV, 5-10 year terms, 25-30 year amortization. This finances the shell, land, and base building systems.
- Equipment/infrastructure overlay (upgrade layer): Separate equipment financing for power distribution upgrades, liquid cooling systems, backup generation, and switchgear. 5-7 year terms matching the useful life of the mechanical/electrical equipment.
- C-PACE (where available): For qualifying energy improvements (particularly cooling system upgrades and on-site power generation) C-PACE provides long-term (20-30 year), fixed-rate financing that attaches to the property tax bill. Excellent for liquid cooling retrofits.
- Sponsor equity (25-40%): Higher equity requirement than traditional colocation because the AI-ready upgrade capex is substantial and not fully captured in the appraisal until tenants are in place.
What makes these deals work: Pre-leasing. AI-ready colocation space is in extreme demand. Operators who can demonstrate signed LOIs or leases from AI tenants at $150-250+ per kW/month can underwrite the upgrade capex with confidence. The spread between traditional colocation rates ($100-120/kW/month) and AI-ready rates ($150-250+/kW/month) justifies the infrastructure investment, if you have tenants committed.
What kills these deals: Power. Again. Upgrading a 5 MW facility to 20 MW requires a new utility feed, potentially a new substation, and 2-4 years in the interconnection queue. The building upgrade takes 12-18 months. The power upgrade takes 24-48 months. If you cannot align these timelines, you end up with an AI-ready building that cannot actually power AI workloads.
4. Edge AI Deployment ($2M-$50M)
Edge AI facilities are smaller-footprint deployments (typically 500 kW to 5 MW) designed for low-latency inference workloads close to end users. Think autonomous vehicle inference, real-time content moderation, telecom AI, and industrial IoT processing. These are not training facilities; they run pre-trained models at the point of use.
Typical Capital Stack:
- SBA 504 (owner-occupied): If the operator occupies the facility, SBA 504 provides up to 90% financing with below-market fixed rates on the CDC portion. The real estate component (building + land) qualifies for 504. Equipment may qualify separately. Maximum debenture of $5.5M ($5M for non-manufacturing).
- Equipment financing: Separate facility for inference GPUs and networking equipment. Smaller scale ($500K-$5M per deployment) but same 3-5 year structure as larger GPU financings.
- Conventional bank loan: Community and regional banks will finance small data center / edge facilities if the borrower has strong financials and the property has general-purpose alternative use.
- Sponsor equity (10-25%): Lower equity requirement than larger facilities, particularly with SBA 504 which can get total leverage to 90%.
What makes these deals work: The real estate has alternative use value. Unlike a 100 MW hyperscale campus that is essentially single-purpose, a 2 MW edge facility in a metro area can be repurposed as light industrial, telecom, or traditional IT. This gives lenders a fallback position and makes conventional financing viable.
What kills these deals: Unit economics at small scale. The per-MW cost of AI-ready infrastructure does not scale linearly downward. A 1 MW edge facility might cost $25-30M per MW, while a 50 MW campus achieves $15-20M per MW. The fixed costs of power infrastructure, security, and staffing spread across fewer megawatts. Edge deals need to demonstrate either premium pricing for low-latency compute or anchor contracts that cover fixed costs.
3Financing the GPUs vs. Financing the Building: The Dual Capital Stack
The single most important structural concept in AI infrastructure financing is the separation of GPU equipment financing from real estate financing. These are fundamentally different assets with different useful lives, depreciation schedules, risk profiles, and optimal financing structures. Combining them into a single loan is like financing a restaurant and its 30-year building with the same 5-year term, the mismatch destroys value.
The Dual Capital Stack Framework
Layer 1: Real Estate + Infrastructure
- Asset: Building shell, land, power distribution, cooling systems, backup generation
- Useful life: 15-30 years
- Financing: CMBS, bank term loan, life company, project finance
- Terms: 10-25 year terms, 25-30 year amortization
- LTV: 55-70%
- Collateral: First lien on real property + fixtures
Layer 2: GPU Equipment
- Asset: GPU servers, networking (InfiniBand/Ethernet), storage
- Useful life: 3-5 years
- Financing: Equipment loans, equipment leases, sale-leaseback
- Terms: 3-5 years, fully amortizing or with residual
- Advance rate: 80-100% of new equipment cost
- Collateral: UCC filing on equipment; no lien on real estate
GPU Equipment Financing: What Lenders Actually Underwrite
GPU equipment financing has matured rapidly since 2024. The market has moved from "we do not finance GPUs" to a structured lending market with clear underwriting criteria, though it remains more specialized than traditional IT equipment financing. Here is what lenders evaluate:
Contracted Revenue Coverage
Lenders want to see that the GPUs being financed are generating revenue under contract, not dependent on spot market pricing. The gold standard is a 1-3 year reserved instance contract from a creditworthy counterparty that covers 1.5x+ the debt service on the equipment loan. Deals with 70%+ contracted utilization close faster and at better rates than those relying on on-demand revenue.
GPU Generation and Manufacturer
Current-generation GPUs from tier-1 manufacturers (NVIDIA H100, H200, B200, GB200; AMD MI300X, MI350) get the best terms. Previous-generation hardware (A100) is financeable but at lower advance rates and shorter terms. The lender is assessing both the current market value and the likely value at end of term. NVIDIA hardware currently commands better financing terms due to deeper secondary markets and broader enterprise adoption.
Operator Track Record
Equipment lenders financing $10M-$200M in GPUs want to see that the operator has experience managing GPU infrastructure at scale. First-time operators face higher equity requirements (30-40% vs 10-20% for experienced operators), shorter terms, and more restrictive covenants. The operational complexity of maintaining a liquid-cooled GPU cluster is non-trivial. Lenders have seen inexperienced operators destroy hardware value through poor cooling management.
Residual Value Assumptions
This is where deals live or die. Aggressive lenders assume 30-40% residual value on GPUs at the end of a 3-year term. Conservative lenders assume 10-20%. The secondary market for enterprise GPUs is still maturing. There is no established "used GPU" market with the depth and transparency of used car markets or used aircraft markets. Lenders who got burned on crypto mining rigs in 2022 are particularly cautious about residual value assumptions on compute hardware.
Sale-Leaseback: Unlocking Capital from Existing GPU Clusters
Sale-leaseback structures are increasingly common for AI infrastructure operators who deployed GPUs with equity or venture capital and now need to free up capital for the next hardware generation. The mechanics: sell your existing GPU cluster to an equipment lessor or specialty finance company, then lease it back for the remaining useful life. You get cash today (typically 50-70% of current fair market value for GPUs with 2-3 years of remaining useful life) and continue operating the hardware under a lease.
When sale-leaseback works: You have a deployed GPU cluster generating contracted revenue, you need capital for next-generation hardware, and the lease payments are covered by existing revenue with margin to spare. The best candidates are operators with H100/H200 clusters under 18 months old with 2+ years of contracted revenue remaining.
When it does not work: If the GPUs are more than 24 months old, the sale-leaseback valuation drops significantly because the remaining useful life is limited. If your revenue is predominantly spot/on-demand, lessors will discount the valuation because the revenue stream supporting the lease payments is uncertain. And if you are using the proceeds to cover operating losses rather than fund growth capex, that is a red flag that will either kill the deal or result in punitive terms.
The GPU-as-a-Service Revenue Model: How Lenders Underwrite It
GPU-as-a-Service (GPUaaS) is the dominant revenue model for AI infrastructure operators who own their hardware. The model is straightforward: purchase GPU capacity, sell compute hours to customers at a markup. Gross margins typically range from 40-65% depending on utilization rates, power costs, and whether the operator owns or leases the underlying facility.
Lenders underwriting GPUaaS businesses look at several key metrics: utilization rate (what percentage of GPU hours are sold, 70%+ is the threshold for comfortable underwriting), contract mix (reserved vs on-demand. Lenders credit reserved revenue at 100% and discount on-demand revenue by 30-50% in their underwriting), customer concentration (if one customer represents 40%+ of revenue, that is a single-tenant risk), churn rate (monthly customer churn above 5% signals pricing or service quality problems), and power cost as a percentage of revenue (above 30% squeezes margins and reduces debt service coverage).
DSCR Calculation for GPUaaS
Lenders typically require 1.25-1.50x DSCR on GPU equipment loans. The calculation uses contracted revenue only (not projected or on-demand revenue) minus operating expenses (power, facility lease, staff, maintenance) divided by annual debt service (principal + interest on the equipment loan). If your contracted revenue alone does not cover 1.25x debt service, you need either more contracts or less debt. It is that simple.
4Power and Cooling: The Variables That Make or Break AI Infrastructure Deals
Power availability is the single most important variable in AI infrastructure financing. Not the GPUs. Not the building. Not the tenant credit. Power. If you cannot secure committed utility capacity at a viable cost, nothing else matters. This section covers how power and cooling affect every layer of the capital stack.
Cost per MW of AI-Ready Capacity
| Component | Cost per MW (Traditional DC) | Cost per MW (AI-Ready) | Delta |
|---|---|---|---|
| Building Shell + Site | $2-4M | $2-5M | Modest increase (higher floor loads, reinforced structure) |
| Power Distribution | $2-3M | $4-7M | 2-3x, higher density busbars, switchgear, PDUs |
| Cooling Systems | $1.5-3M (air) | $3-8M (liquid) | 2-3x, liquid cooling piping, CDUs, pumps, heat rejection |
| Backup Generation | $1-2M | $2-4M | Higher capacity generators + fuel storage |
| Utility Interconnection | $0.5-1.5M | $1-5M+ | Often requires dedicated substation ($3-5M+ alone) |
| Networking + Fiber | $0.5-1M | $1-3M | InfiniBand fabric, high-bandwidth interconnects |
| Total per MW | $8-15M | $15-30M+ | Before GPU equipment |
These costs are for the facility infrastructure only. They do not include GPU hardware. Add GPU equipment at $15-40M per MW of IT load (depending on density and GPU generation), and a fully equipped 10 MW AI facility can easily exceed $300-500M in total project cost.
Power Purchase Agreements and Their Impact on Financing
A Power Purchase Agreement (PPA) (a long-term contract to buy electricity at a fixed or formula-based price) is becoming a standard component of large AI infrastructure deals. For facilities consuming 20+ MW, PPAs serve two critical functions: they lock in power costs for 10-20 years (protecting margins from rate volatility) and they demonstrate to lenders that the facility has a committed, long-term power supply.
Behind-the-meter PPAs (on-site generation, typically solar or natural gas) are particularly attractive because they can provide power independently of utility interconnection timelines. We have seen developers pair a 50 MW behind-the-meter gas generation facility with a 50 MW utility interconnection, using the on-site generation to begin operations while waiting for the utility feed. Lenders view this as de-risking the power timeline, which is often the longest lead-time item in the construction schedule.
How PPAs affect the financing: A long-term PPA at a favorable rate ($0.04-0.06/kWh in low-cost markets) directly improves the facility's operating margin and debt service coverage ratio. Lenders model power costs as the single largest operating expense, typically 40-60% of total operating costs for an AI facility. A PPA that locks in power at $0.05/kWh versus exposed utility rates that could fluctuate between $0.06-0.10/kWh provides the revenue stability that supports higher leverage and better pricing on the senior debt.
Utility Capacity and Interconnection Queue Realities
The interconnection queue (the process of connecting a new facility to the electrical grid) has become the primary bottleneck for AI infrastructure development. In markets like Northern Virginia (Loudoun County), which hosts the largest concentration of data center capacity in the world, utility providers are struggling to keep pace with demand. Dominion Energy has publicly stated that it needs to add generation capacity equivalent to multiple large power plants to meet projected data center demand through 2030.
Interconnection Timeline Reality Check
- Northern Virginia (Dominion): 36-60+ months for new 20+ MW feeds. Existing capacity is largely spoken for.
- Dallas-Fort Worth (Oncor): 24-48 months. Growing but still has available capacity in certain substations.
- Phoenix/Mesa (SRP/APS): 18-36 months. Water constraints are the bigger issue here.
- Columbus, OH (AEP): 18-30 months. Emerging market with available capacity.
- Midwest (various): 12-24 months. Lower demand means faster timelines, but fiber connectivity may be limited.
What this means for financing: Construction lenders will not fund a facility build until the interconnection agreement is executed (not just applied for, executed, with a committed delivery date). This means developers often need to spend $5-15M in pre-development capital (utility deposits, site acquisition, engineering) before any debt is available. This pre-development phase is typically funded with equity or bridge capital, and it represents real risk if the interconnection timeline slips or the utility cannot deliver the committed capacity.
Liquid Cooling: Capex vs. Opex Tradeoffs
The transition to liquid cooling is the single largest technical shift in data center design in two decades. For AI infrastructure, it is not optional. It is mandatory. The question is not whether to deploy liquid cooling, but which technology and how to finance it.
Direct-to-Chip (Cold Plate) Liquid Cooling
Water or coolant circulates through cold plates mounted directly on GPU chips. This is the most common approach for new AI deployments and is supported by all major GPU OEMs. Capex: $500K-$1.5M per MW for in-row cooling distribution units (CDUs), piping, and manifolds, plus the server-level cold plates (typically included in server pricing).
Financing angle: CDUs and piping are fixtures that attach to the building. They can be financed as part of the real estate/infrastructure layer with longer terms. Cold plates are part of the server and depreciate with the GPU equipment.
Rear-Door Heat Exchangers (RDHx)
Liquid-cooled heat exchangers mounted on the rear door of server racks that capture exhaust heat before it enters the room. Can handle 30-50 kW per rack. Useful for retrofitting existing air-cooled facilities. Capex: $300K-$800K per MW, lower than direct-to-chip but less effective at the highest densities.
Financing angle: RDHx units can be classified as building equipment or IT equipment depending on how they are attached. This classification matters for financing, building equipment qualifies for longer-term real estate financing; IT equipment goes in the shorter-term equipment facility.
Immersion Cooling
Servers are submerged in non-conductive dielectric fluid (single-phase or two-phase). Eliminates the need for fans, reduces PUE to near 1.0, and can handle 100+ kW per rack. Capex: $1-3M per MW for tanks, fluid, heat rejection systems. Higher upfront cost but lower ongoing operating expense and significantly better energy efficiency.
Financing angle: Immersion cooling is the most expensive upfront but offers the best PUE, which means the lowest ongoing power costs. Lenders evaluating operating projections will see lower opex, which supports higher debt service coverage. The tradeoff is a larger initial capital requirement. C-PACE financing is particularly well-suited for immersion cooling due to its long useful life and energy efficiency benefits.
5Who Is Financing AI Infrastructure in 2026
The capital sources for AI infrastructure span from traditional real estate lenders to specialized equipment financiers to infrastructure-focused private equity. Different sources play different roles in the capital stack, and understanding who does what prevents wasted time approaching the wrong lenders.
| Capital Source | Role in Stack | Typical Size | Best For |
|---|---|---|---|
| Infrastructure PE | Equity + mezzanine | $50M-$1B+ | Purpose-built training facilities, large campus developments |
| Private Credit Funds | Senior + mezzanine debt | $25M-$500M | Flexible structures, non-bank senior debt, construction + term |
| Equipment Lessors | GPU hardware financing | $5M-$200M | GPU-as-a-Service operators, inference hosting, equipment refresh |
| CMBS | Senior debt on stabilized RE | $20M-$500M+ | Stabilized, leased colocation facilities with credit tenants |
| Venture Debt | Supplemental debt | $5M-$50M | VC-backed GPU cloud startups, growth-stage AI infrastructure companies |
| Life Companies | Long-term senior debt | $25M-$300M | Stabilized AI-ready colocation with 10+ year leases to investment-grade tenants |
| Bank Term Loans | Senior debt | $10M-$150M | Smaller facilities, recourse structures, relationship-based lending |
| Government Programs | Incentives + subsidized debt | Varies | CHIPS Act funding, state incentives, SBA 504 for smaller owner-occupied facilities |
Infrastructure Private Equity
The largest AI infrastructure deals are backed by infrastructure-focused private equity firms that have pivoted from traditional energy and telecom infrastructure into digital infrastructure. These firms bring two things that pure financial lenders cannot: operational expertise in infrastructure development and the ability to write equity checks in the hundreds of millions. They typically invest at the platform level (backing a developer who will build multiple facilities) rather than single-asset project finance.
What infrastructure PE looks for: Management team with data center development track record, secured power capacity across multiple sites, anchor tenant relationships (ideally signed LOIs or contracts), a development pipeline that justifies the platform investment, and a clear path to portfolio-level scale (typically $1B+ in total development cost across 3-5 facilities).
Private Credit for AI Infrastructure
Private credit funds have become the most flexible capital source for AI infrastructure deals that fall between traditional bank lending and infrastructure PE. They can structure construction-to-permanent facilities, provide mezzanine layers, and move faster than traditional lenders. Pricing is higher (SOFR + 400-700 bps for senior, 12-18% for mezzanine) but the structural flexibility is worth the premium for deals that do not fit neatly into traditional lending boxes.
Private credit is particularly well-suited for AI-ready colocation conversions where the operator needs to upgrade power and cooling infrastructure to capture the AI workload premium. The deal size ($20-100M) is often too small for infrastructure PE but too complex for traditional bank lending. Private credit fills that gap.
Government Incentives: CHIPS Act and State Programs
The CHIPS and Science Act of 2022, while primarily focused on semiconductor manufacturing, includes provisions that benefit AI infrastructure. Facilities that qualify as "semiconductor-adjacent", including advanced packaging, testing, and compute facilities that support domestic chip utilization, may be eligible for investment tax credits of up to 25% of qualified capital expenditure. The application process is competitive and requires detailed economic impact analysis, but for facilities in the $100M+ range, a 25% ITC can meaningfully reduce the equity requirement.
State-level incentives vary widely but can be substantial. Several states have introduced AI infrastructure-specific incentive packages that combine property tax abatements (10-20 years), sales tax exemptions on equipment purchases, utility rate discounts for large loads, and workforce development grants. The most aggressive states (Texas, Ohio, Indiana, Georgia) are competing directly for AI infrastructure investment with incentive packages that can reduce total project cost by 5-15%.
6Risk Factors That Lenders Evaluate in AI Infrastructure Deals
AI infrastructure lending is higher-risk than traditional data center lending. Lenders know this, and their underwriting reflects it through shorter terms, lower leverage, and more restrictive covenants. Understanding what they worry about (and being able to address it in your financing package) is the difference between a deal that closes and one that dies in committee.
1. Technology Obsolescence (GPU Generational Risk)
This is the number one risk in every AI infrastructure credit memo. GPU performance doubles roughly every 18-24 months. An H100 purchased today will be outperformed by hardware costing the same or less in 2-3 years. Lenders model this as depreciation acceleration risk, what happens to the collateral value if the GPU market moves faster than expected?
How to mitigate: Short equipment financing terms (3-4 years, not 5-7), conservative residual value assumptions (10-20%, not 40-50%), contracted revenue that covers debt service regardless of hardware resale value, and a documented GPU refresh strategy showing how you will fund next-generation hardware without triggering a liquidity crisis.
2. Tenant and Customer Concentration
Many AI infrastructure deals rely on one or two anchor tenants for 60-80%+ of revenue. If that tenant downsizes, defaults, or shifts workloads to their own infrastructure, the operator is left with a highly specialized facility and a massive debt service obligation. This risk is amplified by the fact that the largest AI compute buyers (hyperscalers) are simultaneously building their own infrastructure.
How to mitigate: Diversify the customer base (no single tenant above 40% of revenue for the best financing terms), secure long-term leases with penalty clauses for early termination, and demonstrate that the facility has features (location, power, cooling) that make it attractive to multiple potential tenants, not just the current anchor.
3. Power Availability and Interconnection Certainty
A signed interconnection agreement with a committed delivery date is table stakes for construction financing. But lenders also evaluate the utility's ability to actually deliver. In congested markets, utilities have missed committed delivery dates by 12-18 months. If your construction loan matures before power arrives, you have a finished building generating zero revenue with a loan coming due.
How to mitigate: Build construction loan maturity buffers (6-12 months beyond expected power delivery), secure behind-the-meter generation as backup, and include utility delay provisions in the construction loan agreement that extend maturity if power delivery slips due to utility actions.
4. Water Rights and Environmental Risk
AI workloads generate significantly more heat per square foot than traditional IT, which increases cooling water consumption for facilities using evaporative cooling. In water-stressed markets (the American Southwest, parts of Texas, and increasingly parts of the Southeast) water availability is a genuine constraint on facility operations. Some municipalities have begun restricting water allocations for new data center developments.
How to mitigate: Deploy water-free cooling technologies (direct liquid cooling, air-cooled chillers, immersion cooling), secure long-term water rights or municipal water agreements, and document water usage projections with WUE (Water Usage Effectiveness) metrics that demonstrate responsible consumption.
5. Regulatory and Export Control Risk
The regulatory environment for AI infrastructure is evolving rapidly. U.S. export controls on advanced AI chips (BIS Entity List restrictions, NVIDIA A100/H100 export limitations to certain countries) affect both the supply chain and the customer base. Sovereign AI initiatives are driving demand for domestic compute infrastructure in multiple countries, but they also create regulatory complexity for operators serving international customers.
How to mitigate: Document compliance with all applicable export control regulations, understand which customer segments are affected by restrictions, and build the regulatory compliance infrastructure (ITAR, EAR, CFIUS) into your operating model before seeking financing. Lenders will ask about this, have the answers ready.
6. Construction and Execution Risk
AI-ready facilities are more complex to build than traditional data centers. Liquid cooling systems require specialized contractors. High-density power distribution requires custom switchgear with long lead times. GPU delivery timelines from NVIDIA and AMD are subject to allocation constraints. Any delay in the construction or equipment delivery timeline directly impacts revenue commencement and debt service coverage.
How to mitigate: Use experienced data center general contractors with AI facility track records, secure GPU allocations before closing on construction financing, build schedule contingencies into the construction budget (10-15% contingency is minimum), and phase the buildout so early phases generate revenue while later phases are under construction.
7Deal Structures That Are Closing in 2026
Theory is useful but structures that are actually closing deals are more useful. Here are the four deal frameworks that are getting to the finish line in 2026, with enough detail to model your own deal.
Structure 1: Construction-to-Permanent with Hyperscaler Pre-Lease
This is the cleanest structure for large AI training facilities. A developer secures a 10-15 year pre-lease from a hyperscaler or major AI lab, then uses that contract as the foundation for a construction-to-permanent financing.
Example Framework (Anonymous): 60 MW AI Training Campus
- Total project cost: $800M (facility infrastructure only, tenant installs own GPUs)
- Anchor lease: 15-year triple-net from investment-grade hyperscaler covering 100% of capacity
- Senior construction loan: $480M (60% LTC), SOFR + 250 bps, 30-month construction period, converts to 10-year permanent at completion
- Mezzanine debt: $120M (15% of cost), 14% current pay, 3-year term with extension option
- Sponsor equity: $200M (25% of cost)
- Permanent DSCR: 1.65x on senior debt, 1.25x on total debt
Why this works: The hyperscaler credit (typically investment-grade) de-risks the revenue stream. The 15-year lease exceeds the loan term, so there is no lease rollover risk during the financing. The developer does not own the GPUs, eliminating technology obsolescence risk from the lender's perspective. The facility is essentially a long-term infrastructure lease backed by a credit tenant, which is a well-understood structure that institutional lenders can execute on.
Structure 2: Equipment Finance Overlay on Existing Data Center Shell
For operators who already own or lease a data center shell and need to finance the AI-ready upgrade (power distribution, liquid cooling, and GPU hardware) this layered approach separates the base facility from the AI equipment.
Example Framework (Anonymous): 5 MW Colocation-to-AI Conversion
- Existing facility value: $40M (stabilized colocation, partially leased)
- Existing mortgage: $24M (CMBS, 60% LTV, 3 years remaining on term)
- AI upgrade capex: $35M (power distribution + liquid cooling + backup generation for 5 MW AI-ready capacity)
- Equipment financing: $28M (80% of upgrade capex), 5-year term, monthly payments, UCC lien on M&E equipment
- Equity contribution: $7M (20% of upgrade capex)
- Post-upgrade NOI: $12M (AI tenants at $200/kW/month vs $110/kW/month for traditional colocation)
- Combined DSCR: 1.45x covering both the existing CMBS and the new equipment loan
Why this works: The existing CMBS stays in place, no costly defeasance or refinance. The equipment overlay is sized to the infrastructure upgrade only, with a term that matches the useful life of the M&E equipment. The AI-ready space commands a 80-100%+ premium over traditional colocation rates, so the incremental revenue comfortably covers the incremental debt. The lender on the equipment layer has a UCC lien on equipment that is physically installed but legally separable from the real property, so there is no lien conflict with the CMBS.
Structure 3: SPV Isolating GPU Economics from Real Estate
The Special Purpose Vehicle (SPV) structure is increasingly used by GPU cloud operators who want to separate the equipment risk from the facility risk, and use different capital structures for each. The real estate sits in one entity with long-term financing. The GPUs sit in a separate entity with shorter-term equipment financing. The operating company leases the facility from the RE entity and owns or leases the GPUs through the equipment entity.
Example Framework (Anonymous): GPU Cloud Operator, $150M Total Deployment
- RE SPV: Owns the facility ($50M value). Financed with a $32.5M bank term loan (65% LTV, 7-year term). Leases facility to OpCo at market rates.
- Equipment SPV: Owns the GPU clusters ($100M in hardware). Financed with $80M equipment facility (80% advance, 3-year term with annual refresh tranche). Revenue from GPUaaS contracts flows to this entity first, covering equipment debt service.
- OpCo: Operating company that manages the facility, operates the GPUs, and contracts with customers. Holds the customer contracts, employs staff, and pays facility lease + equipment lease from operating revenue.
- Waterfall: Revenue to OpCo → facility lease payment to RE SPV → equipment debt service to Equipment SPV lender → operating expenses → remaining cash to equity.
Why this works: It isolates the risks. If the GPU market shifts and the equipment value declines, the RE SPV and its lender are protected, the facility has value independent of the GPUs. If the real estate market softens, the Equipment SPV lender is protected. Their collateral is the hardware, not the building. The structure also allows different capital sources to finance their area of expertise: real estate lenders finance the building, equipment lenders finance the GPUs. Each lender underwrites what they understand.
Structure 4: Sale-Leaseback of GPU Clusters to Fund Next Generation
This structure is specifically designed for AI infrastructure operators facing the GPU generational refresh cycle. You deployed $100M in H100 GPUs in 2024 using equity. Those GPUs are generating strong revenue but you need capital for B200/GB200 hardware. Rather than raising dilutive equity, you sell the existing cluster to a specialty finance company and lease it back, freeing capital for the next generation.
Example Framework (Anonymous): H100 Cluster Sale-Leaseback
- Original GPU cost: $100M (deployed 14 months ago)
- Current fair market value: $65-75M (based on secondary market pricing and remaining useful life)
- Sale-leaseback proceeds: $55-65M (75-85% of FMV, the lessor takes a haircut for residual risk)
- Lease term: 24 months (aligned with remaining primary useful life of H100 hardware)
- Monthly lease payment: Structured so that total lease payments + residual value purchase option = approximately 110-115% of sale proceeds (implied 10-15% annual cost of capital)
- Use of proceeds: $55-65M deployed toward next-generation GPU hardware, supplemented by $35-45M in new equipment financing for the balance
Why this works: The operator maintains uninterrupted use of the existing GPUs while freeing capital for the next generation. The lessor acquires an income-producing asset with a known revenue stream (the lease payments from a creditworthy operator). At end of term, the operator either purchases the GPUs at a pre-negotiated residual value or returns them to the lessor, who sells them into the secondary market. This is not a distressed transaction. It is capital recycling, and it is becoming a standard part of the GPU infrastructure lifecycle.
Each of these structures can be combined. A large AI infrastructure platform might use Structure 1 for their anchor facility, Structure 2 for expansion into existing shells, Structure 3 to isolate GPU risk across the portfolio, and Structure 4 to manage generational hardware transitions, all within the same corporate family.
8What Kills AI Infrastructure Deals: Lessons from the Market
Not every AI infrastructure deal closes. The failure rate on deals that enter the financing market is higher than traditional data center deals because the asset class is newer, the risks are less well understood, and many sponsors are first-time infrastructure operators. Here are the patterns that consistently kill deals.
No Committed Power
We see this repeatedly: a sponsor has a site, a building plan, GPU commitments, and even tenant LOIs, but no executed interconnection agreement with the utility. Without committed, contracted power with a delivery date, no construction lender will proceed. An interconnection application is not the same as an interconnection agreement. Until the utility has committed to deliver specific MW capacity by a specific date, the power risk is unquantified and unfinanceable.
Equipment Value Exceeds Real Estate Value with No Separation
When $200M in GPUs sits inside a $30M building, and the sponsor is trying to finance the whole thing with a single loan, the deal does not work. Real estate lenders will not underwrite to the equipment value. Equipment lenders will not take real estate collateral risk. The dual capital stack approach (separate financing for real estate and equipment) is not optional at this ratio. It is structural necessity.
Revenue Projections Based on Spot GPU Pricing
GPU cloud spot pricing is volatile. H100 hourly rates dropped 50-60% from peak 2023 pricing to mid-2025 as supply caught up with demand. If your debt service coverage only works at peak spot rates, no lender will underwrite it. They will stress test against contracted revenue only, discounting on-demand revenue by 30-50%. Build your financial model on committed contracts, not spot market projections.
First-Time Operator, Mega-Scale Ambition
A team with no data center operating experience proposing a $500M AI training facility is a non-starter with institutional lenders. The operational complexity of managing liquid-cooled GPU infrastructure at scale is substantial, and the consequences of operational failure (destroyed hardware, data loss, SLA penalties) are severe. Lenders want to see a management team that has operated at the proposed scale, or at least at 25-50% of it. If you are a first-time operator, start with a $20-50M deployment, build the track record, then scale.
No GPU Refresh Strategy
If your business plan shows GPU hardware deployed in year 1 and still generating revenue at the same rates in year 5, the deal will not survive due diligence. Every lender in this space understands generational obsolescence. You need a documented refresh strategy that shows how you will fund next-generation hardware (sale-leaseback of existing clusters, revolving equipment facilities, reinvested cash flow) and how revenue rates adjust as new hardware delivers better performance at lower cost.
9Building Your AI Infrastructure Financing Package
AI infrastructure is the most capital-intensive asset class to emerge in commercial real estate and equipment financing in decades. The deals that close efficiently share a common framework: they separate real estate from equipment, match financing terms to asset useful life, demonstrate contracted revenue coverage, and address technology obsolescence head-on.
Here is what your financing package needs:
1. Power documentation: Executed interconnection agreement (not just application), utility capacity study, PPA terms (if applicable), behind-the-meter generation plans (if applicable).
2. Revenue documentation: Signed leases or compute contracts, LOIs from prospective tenants, pricing analysis showing contracted vs on-demand revenue mix, customer credit profiles.
3. Technology specification: GPU hardware specification and manufacturer, cooling system design and technology, power distribution architecture, network topology. Lenders need to understand what they are financing.
4. Financial model: 5-10 year projections with GPU refresh cycles modeled, sensitivity analysis on utilization rates and pricing, DSCR analysis on each financing layer separately, residual value assumptions documented and justified.
5. Management team: Track record in data center operations at scale, technical expertise in liquid cooling and GPU infrastructure, relationships with GPU manufacturers (allocation access matters), customer relationships and contracted pipeline.
6. GPU refresh strategy: Documented plan for next-generation hardware procurement, sale-leaseback or secondary market disposition of current-gen hardware, financing structure for refresh cycles (revolving equipment facility, retained cash flow, etc.).
The AI infrastructure financing market is deep but specialized. General-purpose commercial lenders, even those with data center experience, are often not equipped to underwrite GPU economics, model technology obsolescence risk, or structure the dual capital stack that these deals require. You need capital sources (lenders, equipment financiers, and equity partners) who have closed AI infrastructure deals specifically, not just data center deals generally.
PeerSense maintains relationships across a curated network of capital sources across CMBS, equipment financing, private credit, infrastructure PE, and venture debt, including lenders who specialize in AI infrastructure specifically. Referral fee at closing. Tell us about your AI infrastructure deal and we will connect you with the right capital.
For background on traditional data center financing structures that serve as the foundation for many AI infrastructure deals, see our companion guide: Data Center Financing: Capital Structures for Hyperscale, Colocation, and Edge Infrastructure.
The Bottom Line
AI infrastructure is not a niche within data center financing. It is becoming the dominant use case driving capital deployment in the sector. The facilities being built today are purpose-designed for AI workloads: 30-100+ kW per rack, liquid cooling throughout, dedicated substations delivering tens or hundreds of megawatts, and GPU hardware worth more than the buildings that house it. Financing these projects requires separating real estate from equipment, matching terms to useful life, building revenue documentation around contracted compute rather than speculative demand, and addressing technology obsolescence as a feature of the business model rather than pretending it does not exist. The capital is available, $200B+ flowed into AI infrastructure in 2024-2025, and the pace is accelerating. The structures outlined in this guide are not theoretical; they are closing deals right now, from $10M edge deployments to $1B+ training campuses. The deals that succeed are the ones that come to the market with committed power, contracted revenue, experienced operators, and a capital stack where each layer is appropriately sized and structured for the asset it finances. PeerSense connects AI infrastructure projects with the specialized capital sources (equipment lenders, private credit, CMBS, infrastructure PE, and venture debt) that understand this asset class. Referral fee at closing.