In short
AI data center colocation agreements are built around a master services agreement (MSA) with site specific service orders and a service level agreement (SLA) attached as a schedule. The MSA model lets a tenant add capacity at multiple sites without renegotiating the entire legal framework each time. Law firm analysis
AI workloads push power density to 10 to 30 plus kW per rack, and the most advanced GPU servers can demand 132 kW per rack today. Cooling alone accounts for 30 to 40 percent of a facility’s total electricity use. Providers now offer high density colocation with liquid cooling and granular SLA metrics beyond raw uptime. Service credits and termination rights are typically the sole remedies for an SLA breach. SEC filing, Data Center Knowledge
The U.S. colocation market is projected to grow from $38.80 billion in 2025 to $65.45 billion by 2030, an 11.0 percent compound annual growth rate. The AI and high performance computing segment within that is projected at a 26.3 percent CAGR. Market research Total global data center deal volume hit a record $61 billion in 2025. CNBC
How are colocation agreements for AI data centers structured?
A colocation agreement for an AI tenant is rarely a single document. It is typically built as a master services agreement (MSA) with three parts. The MSA sets the general legal terms. Site-specific service orders add the details for each location. An SLA is attached as a schedule or exhibit. This structure lets the tenant deploy capacity across multiple states or campuses without renegotiating the entire contract every time. Data center tier standards overview
The agreement is part real estate, part technology service
Practitioners describe these agreements as a hybrid. They carry the footprint of a property lease but layer on obligations for power delivery, cooling, and security that go far beyond a traditional landlord-tenant relationship. Law firm analysis
Twelve key provisions
One major law firm identified these as the provisions that anchor a colocation agreement for AI workloads. Law firm analysis
- Power
- Connectivity
- Ready for service timelines
- SLAs
- Pricing structure and escalators
- Term length and renewal
- Scalability and flexibility
- Security and compliance
- Liability and indemnification
- Disaster recovery and business continuity
- Financial stability of provider and customer
- Exit strategy and decommissioning
Each of these will be shaped by the extreme power and cooling demands of AI, and we walk through the most critical ones below.
What do SLAs for AI colocation include?
Uptime guarantees are the floor, not the ceiling
The Uptime Institute Tier Standard is the global benchmark. A Tier III facility guarantees approximately 99.982 percent uptime, which translates to about 1.6 hours of downtime per year. A Tier IV facility guarantees approximately 99.995 percent uptime, or about 26.3 minutes per year. Law firm analysis Tier III is the most common commercial choice. Tier IV is reserved for the most critical AI clusters, defense systems, and financial exchanges.
| Tier | Annual uptime | Approximate downtime per year | Typical use |
|---|---|---|---|
| III | 99.982% | 1.6 hours | Enterprise colocation |
| IV | 99.995% | 26.3 minutes | Mission-critical AI, exchanges |
For AI tenants, a simple uptime percentage is no longer enough. A leading law firm notes that the traditional five nines (99.999 percent) guarantee for power and cooling is now treated as the absolute minimum standard. Sophisticated SLAs add granular metrics for power quality (voltage and frequency stability), humidity and temperature tolerances, and required response times for the provider’s engineering teams. Law firm analysis
Providers already market SLAs that cover 100 percent uptime for power (including A+B redundant paths) and keep ambient room temperature at 72 degrees Fahrenheit plus or minus 5 degrees. Provider
How service credits work
When a provider misses an SLA metric, the tenant receives a credit against its monthly recurring charges. The credit is usually a percentage of the fee for the affected cabinet or service for that month. A representative colocation SLA from 2025 uses three metric areas (colocation space, power, connectivity) with tiered credits.
| Availability range | Space/power credit (% of MRC) | Connectivity credit (% of MRC) |
|---|---|---|
| 99.5% to 99.95% | 5% | 5% |
| 99.0% to 99.5% | 10% | 5% |
| Below 99.0% | 15% | 15% (combined logic applies) |
Aggregate credits across all three areas are capped at 50 percent of the monthly recurring charge. Provider SLA document
OVHcloud publishes an AI-specific SLA that covers AI Training, AI Notebooks, and AI Deploy. Monthly availability below 99.50 percent but at or above 99 percent triggers a 10 percent credit of the hourly cost per full hour of unavailability. Below 99 percent but at or above 95 percent triggers 25 percent. Total credits are capped at 30 percent of the monthly invoiced amount. Unavailability is defined as all AI Training, AI Notebooks, or AI Deploy jobs submitted by the Customer for all regions combined per minute returning error code 500 or 503. Provider SLA document
Liability allocation and caps
Service credits are almost always stated to be the sole and exclusive remedy for an SLA failure. An older but still instructive SEC filing from 2003 shows a provider capping total monthly credits under the Premium Bandwidth SLA at 50 percent of the monthly invoice for that service, excluding power and co location fees. The customer had 30 days to notify the provider of any breach of the Service Level Targets. SEC filing
For AI tenants, the work being interrupted can be worth many multiples of the monthly colocation fee. Liability allocation must therefore be written so each party takes responsibility for losses from its own sphere. The provider carries the risk for the physical environment and core infrastructure. The tenant carries the risk for its own servers and AI models. Law firm analysis
Practical note. Many colocation providers treat their SLA terms as non negotiable. One trade publication reports that providers have placed their SLAs on lock down because the terms are deeply embedded in their systems and processes. Trade press The most a tenant can often do is negotiate the scope of what counts as unavailability and the size of the credit tiers for the specific workload.
How do AI power densities change the physical deal?
Rack power has jumped from single digits to triple digits
Standard enterprise IT equipment draws roughly 3 to 5 kW per rack. High-density colocation for AI now supports 10 to 30 plus kW per rack. A single NVIDIA H100 GPU draws 700 watts under full load. A server with eight H100s plus system overhead consumes about 6,500 to 7,500 watts. Racks that hold multiple such servers can draw 15 to 30 kW total. Provider
The ceiling is moving fast. Schneider Electric states that the latest NVIDIA-based GPU servers require 132 kW per rack when fully loaded. The next generation, expected within a year, will require 240 kW per rack. Forbes analysis, Trade press, Vendor analysis, Industry report, Trade press The AFCOM 2025 survey found average rack density had already doubled from 7 kW in 2021 to 16 kW, with 79 percent of respondents expecting further growth due to AI and high performance workloads. Trade press The Uptime Institute 2025 survey found that average server rack power densities continue to rise slowly, driven by greater adoption of racks in the 10 kW to 30 kW range, and few facilities exceed 30 kW. Uptime Institute Global Data Center Survey 2025
Cooling becomes the bottleneck
Cooling consumes 30 to 40 percent of an AI data center’s total electricity. Liquid cooling can reduce that cooling energy consumption by up to 30 percent compared with air cooling. Trade press Providers are responding. Digital Realty offers high-density colocation with up to 150 kW of cooling per cabinet and supports both air-assisted liquid cooling (AALC) and direct liquid cooling (DLC). Provider Equinix reported in its 2025 annual report that 22 percent of new deployments demanded liquid cooling, up from 9 percent in 2024, and allocated a $1.2 billion global retrofit budget. Market research
For a real project, a semiconductor firm built a hardware verification environment for NVIDIA DGX H100 in a Digital Realty AI-enabled data center, doubling density and relying on the provider’s 99.999 percent SLA for uptime. Provider case study, Provider uptime SLA
How is colocation pricing structured for AI workloads?
Power and cooling drive the bill, not the square footage. A provider may charge per kilowatt, per circuit, or based on a committed power allocation. If the tenant’s GPU power needs exceed that commitment, overage penalties apply. Provider
Full cabinet pricing is the most common model for GPU colocation, rather than pricing by individual rack unit, because GPU clusters need stable power and cooling to avoid throttling during long training runs. Provider
Retail and wholesale rates
Retail colocation in tier 1 metros averaged $150 to $250 per kilowatt per month in 2025. That rate bundles cross connects, 24 by 7 remote hands support, and access to carrier neutral meet me rooms. Market research
Wholesale colocation for commitments of 250 kW to 4 MW averaged roughly $130 to $190 per kW per month as of Q4 2024. Large scale requirements are seeing greater price increases because capacity in mature markets is very tight. Market research U.S. colocation vacancy hit an all time low of 1.6 percent nationally in Q4 2024. Northern Virginia vacancy was 0.3 percent, and several other markets were below 0.5 percent. Market research
The fees that make up the monthly recurring charge
Monthly recurring charges (MRCs) include the base space charge, power and utility costs, connectivity fees, taxes, and any ancillary services. Non-recurring charges (NRCs) are one time costs such as account setup and installation. Annual escalators may be a fixed percentage or tied to an index. Law firm analysis
What are the major AI colocation deals and market trends?
Record deal volume
U.S. AI data center financings doubled from $30 billion in 2024 to an expected $60 billion in 2025. Trade press Global AI data center deal volume hit a record $61 billion in 2025. CNBC
Landmark leases
- Applied Digital and a U.S. investment-grade hyperscaler signed a 15 year lease for 200 MW at the Polaris Forge 2 campus in North Dakota, valued at $5 billion. The tenant holds a right of first refusal on an additional 800 MW. Primary
- Applied Digital and a third hyperscaler signed a 15 year $7.5 billion lease for 300 MW at Delta Forge 1 in Louisiana. Together with Polaris Forge 1, total contracted lease revenue across three campuses exceeded $23 billion. SEC filing
- CoreWeave signed three 15-year leases, with three five year options, for a combined 400 MW of IT load at Applied Digital’s Polaris Forge 1 campus. Primary
- Cipher Mining entered a 15 year approximately $5.5 billion agreement with AWS to deliver 300 MW of AI grade capacity at a Texas facility. Cipher Mining press release, Cipher Mining 10-K
- PowerHouse Data Centers executed a long term hyperscaler lease at its Arcola campus in Northern Virginia for 120 MW across 615,000 square feet. News
Infrastructure investors are buying the power plants and platforms that support these workloads. Blackstone acquired the 774 MW Potomac Energy Center power plant in Loudoun County, Virginia, for $1 billion to supply nearby hyperscaler campuses. Market research Blackstone also acquired Airtrunk for $16 billion in 2024, at the time the largest data center transaction on record. Market research
How workloads are moving
The Uptime Institute 2025 AI Infrastructure Survey found that colocation facilities already host 34 percent of AI inference workloads, second only to on-premises central locations at 46 percent. Another 45 percent of owner operators plan to host AI inference applications. Primary A separate IT leader survey found that 65 to 91 percent plan to move AI workloads from the cloud to colocation, with generative AI applications at the high end. Market research
What happens when an SLA is breached?
A landmark failure
In November 2025 a cooling system failure at a CyrusOne facility in Aurora, Illinois, near Chicago, halted CME Group’s futures and options markets for more than 10 hours, freezing a large share of global derivatives trading. The industry now cites this as a failure that drove demand for real time SLA monitoring. Data Center Dynamics
The lender’s view
A single second outage can trigger a full month’s rent credit. A longer or repeat outage can give the tenant a right to terminate the lease early. For a single tenant AI data center asset, losing the major tenant slashes income to near zero. Lenders should stress test loan scenarios against hypothetical SLA breaches to see if the loan stays viable. Trade press
Lease cancellations are not hypothetical. In early 2025 Microsoft canceled leases equivalent to two entire data centers because of facility and power delays. Meta has taken similar steps. Trade press
SLA insurance and risk transfer
New insurance products allow a provider to transfer SLA risk to a carrier. The policy mirrors the contractual SLA terms and pays out automatically when a performance failure occurs. Trade press Lockton launched an AI data center SLA insurance product designed for this purpose. Trade press
The cost of downtime is high even in ordinary circumstances. The Uptime Institute’s 2024 survey found that 54 percent of respondents said their most recent significant outage cost more than $100,000, and one in five saw costs above $1 million. Primary
Dispute resolution
Arbitration is increasingly the preferred mechanism for AI data center disputes. It offers privacy, confidentiality, and the ability to pick neutral decision makers with real technical expertise. One law firm notes that the FIDIC model contracts, widely used for international construction projects, adopt a multi tiered approach that starts with a dispute avoidance board, moves to amicable settlement, and ends in final and binding arbitration under ICC Rules. Law firm analysis
How do agreements allow for future technology?
A tenant signing a 15 year lease today cannot know exactly what cooling or power technology it will need in year seven. Colocation agreements now commonly include technology-proofing clauses. These let the tenant install next-generation cooling, such as direct to chip liquid cooling, or draw down significantly higher power densities in the future, subject to a commercial framework that shares the upgrade costs between the parties. Law firm analysis
The Open Compute Project (OCP) launched an open letter in 2025, signed by nearly 70 companies including Google, Meta, Microsoft, and NVIDIA, calling for collaborative standards across four workstreams. Power (standardizing high voltage DC interfaces for racks over 500 kW). Cooling (common Coolant Distribution Unit designs and commissioning standards). Mechanical (standardized aisle widths, rack dimensions, and weight limits). Telemetry (a common secure protocol for monitoring and autonomous control). Primary These standards, once adopted, will be referenced in SLAs and colocation agreements.
What security provisions do AI colocation agreements need?
AI tenants often require security provisions that go into a dedicated schedule. Providers describe the standard as fortress grade. The schedule specifies multi layered security zones, advanced biometric authentication, anti climb perimeter fencing, vehicle crash barriers, and 24/7 surveillance with advanced threat detection analytics. Law firm analysis
How does the FERC co-location proceeding affect power costs?
A colocation agreement does not just cover the physical space and the cooling. It also governs how the tenant gets power. When an AI data center sits behind the same point of interconnection as a generator, the question arises whether the data center must pay full transmission and capacity charges or can bypass some of those costs. That question directly affects the power price in the colocation agreement.
In February 2025 the Federal Energy Regulatory Commission (FERC) issued a show cause order in Docket No. EL25-49-000. The order opens a proceeding to examine co-location issues for large AI loads in the PJM region. Chairman Mark Christie said FERC needs to act and act soon to address these issues. FERC news release The outcome, still pending as of this writing, will help determine what grid charges a co located AI data center must pay, and that number will flow directly into the power pricing and SLA commitments of future colocation agreements.
Key takeaways
- The standard AI colocation contract is an MSA with service orders and an SLA attached. It is both a real estate lease and a technology service agreement.
- AI tenants need SLAs that go beyond uptime to cover power quality, temperature tolerances, and engineering response times. Credits usually cap at 50 percent of monthly fees and are the sole remedy.
- Power densities are climbing from 10 to 30 kW per rack to 132 kW today and a projected 240 kW in the next generation. Liquid cooling is no longer optional.
- Pricing is driven by power and cooling, not square feet. Retail rates run $150 to $250 per kW per month in top metros, and wholesale rates run $130 to $190 per kW per month.
- The U.S. market saw $60 billion in deals in 2025, with individual AI factory leases reaching $7.5 billion. Vacancy rates are at historic lows, and lenders now stress test loans for SLA breaches.
- Technology-proofing clauses let tenants upgrade cooling and power in the future under a cost sharing framework. OCP standards will soon give those clauses a common technical benchmark.
- Security provisions are now detailed, fortress-grade schedules.
- The FERC co-location proceeding will determine grid charges for behind the meter AI loads, which will flow into the power price of future colocation agreements.
Frequently asked questions
Q:What is the difference between a colocation agreement and a traditional AI data center lease?
A:A traditional lease gives the tenant space and base building power. A colocation agreement adds continuous obligations for power delivery, cooling, connectivity, and security. It is written as an MSA with site-specific service orders and an SLA, and is a hybrid real estate and technology service contract. Law firm analysis
Q:How much uptime should an AI colocation SLA guarantee?
A:Most commercial colocation offers Tier III (99.982 percent uptime, roughly 1.6 hours of downtime per year). For the most critical AI workloads, Tier IV (99.995 percent, roughly 26 minutes per year) is used. Beyond raw uptime, AI tenants should look for metrics on voltage stability, temperature, and response times. Law firm analysis
Q:What limits are placed on SLA service credits?
A:Credits are a percentage of the monthly fee for the affected service. They are typically tiered based on how far the metric fell below the guarantee. Total credits are typically capped at 50 percent of the monthly invoice, and they are declared the sole and exclusive remedy. SEC filing
Q:Why are power density and liquid cooling so central to AI colocation?
A:A standard enterprise rack draws 3 to 5 kW. A rack of NVIDIA H100 GPU servers can draw 15 to 30 kW, and the latest hardware reaches 132 kW per rack. Air cooling cannot handle that load efficiently. Liquid cooling becomes necessary, and it can use 30 percent less energy than air cooling. The cooling system’s capability is now an essential factor in colocation pricing and SLA commitments for AI workloads. Industry pricing guide
Q:How are AI colocation deals priced?
A:The typical model is per kilowatt or per full cabinet, not per rack unit. Retail prices in top U.S. metros range from $150 to $250 per kW per month. Wholesale prices for 250 kW to 4 MW run $130 to $190 per kW per month. The price bundles power, cooling, cross connects, and remote hands. Overage penalties apply if the tenant exceeds the committed power allocation. Provider
Q:Can a tenant terminate the lease if the SLA is breached repeatedly?
A:Yes, a serious or repeated outage can give the tenant a right to terminate. For a single-tenant AI data center, losing the tenant can drop income to near zero. This risk is why lenders now stress test loans against SLA failure scenarios. Trade press
Q:Are SLA terms negotiable with large colocation providers?
A:Many providers have standardized their SLA programs and resist changes. One trade report says SLAs are on lock down. The terms are very difficult to change across the board. Trade press
Q:What is a technology-proofing clause?
A:It is a provision in a long-term colocation agreement that permits the tenant to install new cooling systems, draw more power, or adopt future standards that did not exist when the lease was signed. The clause establishes a process for sharing the upgrade costs between the tenant and the provider. Law firm analysis
Q:What did the FERC show cause order on co-location do?
A:The order opened a proceeding to examine whether the PJM tariff adequately addresses the rates, terms, and conditions of service for AI data centers co-located at generating facilities. The outcome will affect the power pricing embedded in colocation agreements for co-located AI facilities. The proceeding was still open at research time. FERC eLibrary
Q:How big is the market for AI colocation in the U.S.?
A:The entire U.S. colocation market is projected at $38.80 billion in 2025 and $65.45 billion by 2030. The HPC & AI segment within it is growing at a 26.3 percent CAGR. Market research
Subscribe to The Compute Law Brief
The Compute Law Brief is a free weekday newsletter on the law of AI infrastructure across tax, real estate, construction, power, and deals. The big US build markets and federal law. Three minutes a morning. No paywall, and no email gate to read the blog. Subscribe if you want it in your inbox.
Junde Liu, JD, LL.M. (Taxation) candidate at UF Law. Originally published on Compute Law Blog. This article is general information and does not constitute legal advice. Reading it does not create an attorney client relationship. The reader should not act on the basis of any content here without first consulting a licensed attorney in the relevant state. Last reviewed for accuracy May 23, 2026.