Colocation / Shared responsibilities
AI Governance for Colocation: Provider and Tenant Duties
In colocation, one company owns the computing workload and another often operates the building. Their systems can interact without sharing the same control rights or legal duties.
Sections in this article
A colocation customer may manage GPU job scheduling while the facility operator manages the electrical supply, cooling equipment and shared building controls. These responsibilities meet at defined service interfaces: power delivery, rack inlet conditions, coolant supply, metering and incident communications.
Problems arise when contracts describe only average capacity while the software changes the rate of demand, or when each party assumes the other is documenting and controlling an AI application. The answer is a specific responsibility matrix, supported by measured service boundaries and clear intervention procedures.
Start with the technical and contractual boundary
GPU workloads, in-rack equipment where agreed, tenant networks and software-level workload limits.
Reserved power, environmental conditions, telemetry access, alerts and escalation.
Building mechanical plant, common electrical distribution and protective control systems.
This is a common commercial pattern, not a universal legal allocation. In managed colocation, hosted infrastructure or a bespoke high-density agreement, the provider may operate some tenant equipment or the customer may maintain designated coolant components. The contract and as-built drawings determine the actual handoff.
As a default security posture, tenant software should not receive unrestricted write privileges on the provider's BMS or EPMS. Where coordination is needed, specify an authenticated, purpose-limited interface; a read-only signal is often sufficient for scheduling. Any proposed automation across the boundary deserves its own safety and cyber review.
Liquid cooling: why the CDU does not always settle responsibility
A coolant distribution unit (CDU) can be a practical boundary between a facility-water system and a technology-cooling loop, but ownership is not inherent in the equipment label. One provider may supply and maintain the CDU; another may require the tenant to do so. Manifolds, quick disconnects, leak detection, flow assurance, water quality and response times all need explicit allocation.
| Technical interface | What the agreement should specify |
|---|---|
| Facility-side cooling supply | Temperature, flow/pressure service envelope, sensor location and measurement period |
| CDU and heat exchanger | Owner, maintenance rights, alarms, parts and fault escalation |
| Technology cooling loop | Coolant specification, filtration, hoses, leak response and monitoring ownership |
| High-density rack | Permitted equipment density, power allocation and transient load behavior |
A universal claim that servers will fail within ten seconds of every liquid-cooling fault is not supportable: thermal response depends on workload, thermal mass, server protections and loop design. Validate response windows for the specific deployed equipment instead.
GPU ramp rates need an agreed measurement point
A customer can remain under its contracted peak kW yet create load changes that matter for shared infrastructure. Rather than insert an arbitrary “500 kW per minute” clause, engineers should review the tenant's observed and expected ramp behavior, the upstream electrical design and the utility's requirements.
- Scope: identify the PDU, branch or agreed metered point.
- Permitted profile: set tested ramp and peak limits, with sampling interval and exceptions.
- Notifications: state when a customer must signal a planned high-load event.
- Remedy: use agreed escalation, mitigation and service-continuity measures.
- Evidence: retain synchronized measurements from both sides of the handoff.
Responsibility for damage or SLA credits does not transfer automatically because one party owns the software that initiated a load change. Causation, contract wording, protective equipment behavior and applicable law all matter.
Put intervention and evidence into the SLA
A workable agreement names the contacts who can call an incident, request workload throttling, or suspend automation under defined conditions. Emergency power disconnection must follow the site's approved electrical safety procedures and contractual authority; it should never be presented as an automatic remedy a provider can exercise without qualification.
| Event | First coordination step | Evidence to retain |
|---|---|---|
| Cooling supply outside agreed range | Provider operations verifies its supply boundary; informs affected tenants | Supply sensors, alarms, maintenance state and timeline |
| Unexpected GPU power spike | Review agreed meter data; notify customer's operations lead | Time-aligned ramp profile, limits and protection events |
| AI control recommendation rejected | Owner of the affected system checks the safeguards | Model version, requested change, rejection and follow-up |
Hosting AI is not the same as using or supplying the AI system
Under the EU AI Act, a colocation company does not automatically become the provider or deployer of every customer's AI software merely by housing its servers. But if the facility operator uses an AI cooling optimization product for its own operations, that is a separate use case requiring an appropriate role and risk assessment. High-risk classification depends on the system's intended purpose and the regulation's criteria, not simply on the presence of GPUs.
EU data-centre energy reporting is a separate framework. The Energy Efficiency Directive and Delegated Regulation (EU) 2024/1364 set out reporting requirements and indicators for qualifying data centres. Those rules should not be presented as AI Act obligations or as requiring every colo provider to publish tenant-specific PUE or water metrics. Review the reporting entity, thresholds and available measurement data.
Takeaway: document the power and cooling service boundary, control privileges, evidence-sharing rules and incident authority before assigning responsibility for AI-related incidents. A precise contract is more useful than a blanket statement that one side is always liable.
Sources and further reading
Primary frameworks and reference materials. Engineering decisions require site-specific validation; legal classifications require assessment of the actual use case.