AI data center SLA: the 5 missing clauses

SLA de centro de datos para IA: las 5 cláusulas que faltan

On an October night, a GPU cluster loses its training job at three in the morning. The room did not shut down: power was not interrupted, cooling did not fail and the generator did not start. Your contract records an availability of 99.98 % for that time slot. And still, eleven hours of work were lost. When operations asks who is responsible for those eleven hours, the SLA you signed does not say: it measures the room, not the workload.

That gap is the central problem of artificial intelligence workloads: infrastructure availability stopped guaranteeing the success of the job, and closing that gap in writing is contract work, not architecture work. This article gives you the five clauses a data center SLA does not include today, the Mexican legal source that already requires documenting that responsibility, and the deliverable you can demand.

What your data center SLA measures

For two decades the colocation model was built on a simple proposition: the operator brings the data center, the client brings its systems and both agree on service levels. The client bought space, power, cooling and connectivity and received reliable infrastructure, physical security and availability, with clearly defined responsibilities. SLAs measured facility availability, power availability and physical security.

What changed is the scale of the problem. An AI deployment spans power systems, liquid cooling, high-performance networking, GPUs, storage, software platforms and operations equipment: when a training job fails, the root cause can sit in several layers of infrastructure and of ownership. A data center is declared AI-ready when it meets four engineering requirements: rack density of up to 227 kW, direct-to-chip liquid cooling, readiness for the transient behavior of power and GPU cluster resilience.

Those four requirements are engineering requirements, not contract requirements. The manufacturer of the power and cooling infrastructure says it another way: Vertiv projects that the average density of an AI module goes from 40 to 100 kW to 100 to 250 kW in the coming years, with variable GPU power draws compared with CPU draws, which are far more stable. A contract written for stable loads and a plant with variable draws are two different problems. The technical context is already published here: what infrastructure your GPU cluster needs before the first H100.

The five clauses your AI SLA is missing

The starting point is a sentence from the infrastructure manufacturer: facility availability no longer guarantees the success of the workload. The question changed from “Was the facility available?” to “Did the infrastructure perform as expected for the workload?”. If the contract only answers the first one, the AI job runs at the client’s expense.

These are the five clauses that close that gap, with the evidence that must be requested in writing in each case:

  • Shared visibility with workload data, not just room data. The operator must share power utilization, cooling, environmental conditions, incident and asset health metrics. Today data center monitoring and monitoring of the client’s systems are separate systems: that is where the blind spots are born.
  • Workload performance variables and measurement method. Without definitions for power stability, thermal anomaly detection and workload performance, there is no applicable penalty: there is nothing to measure against.
  • Remediation and escalation times with an identified owner. A typical SLA returns a credit. What is missing is the maximum timeframe per severity, who escalates, at what time and what evidence is kept when the case is closed.
  • Resilience glossary with attached assumptions. N+1, concurrently maintainable, fault tolerant and distributed redundancy mean different things in different organizations. The clause is a signed annex with concrete definitions and assumptions: generator autonomy, cooling redundancy, UPS architecture and holdover capacity.
  • Allocation of responsibility by domain and limits of liability. An explicit model of what belongs to the provider, what belongs to the client and what is governed together. And the conversation almost nobody opens: the typical contract limits the operator’s liability with damage caps and excludes indirect losses. That is where it is decided who pays for a lost job.
  • ClauseWhat the SLA says todayWhat it should sayEvidence requested
    Shared visibilityData center availabilityPower, cooling, environmental, incident and asset health metricsPeriodic report with time series, not screenshots
    Workload performanceDoes not existDefined variables and measurement methodBaseline and threshold per variable
    Remediation and escalationCredit for breachTimeframe per severity, owner and escalation timeIncident log with timestamps
    Resilience glossaryLoose termsAnnex with definitions and assumptionsSigned contract annex
    Allocation of responsibility“The operator provides the data center, the client provides the compute”Matrix by domain and limits of liabilityMatrix plus contract review minutes

    The blind spot: resilience does not mean the same thing to both parties

    The underlying problem is not the wording: it is that both parties use the same vocabulary and are designing for different failure scenarios. The manufacturer says it this way: trust erodes as soon as the assumptions do not match, and each side’s assumptions are rarely aligned.

    The tier classification makes it visible. In the Uptime Institute system, Tier III is concurrently maintainable, with redundant components and routes and no downtime for maintenance; Tier IV adds fault tolerance and physically isolated systems. The standard itself warns that, with redundant components out of service for maintenance, the environment has a higher risk of disruption if a failure occurs. Two providers can sell the same tier with different maintenance windows, and that difference does not appear in any price sheet.

    Worse: the vocabulary is in dispute. Uptime Institute and TIA published a joint statement in which TIA commits to modifying the ANSI/TIA-942 standard to eliminate the word “Tier”, so that the Uptime tier system and the TIA scale end up separated. When your provider says “Tier III” and the counterparty says “Rated-3”, they do not contradict each other: they use two reference systems that the industry decided to separate. The glossary with assumptions is not a legal ornament: it is the clause that prevents the argument after the incident. The context is in what a Tier I, II, III and IV SLA is.

    In Mexico the facility contract is already a data contract

    If your AI workload processes personal data —and in Mexico almost always does: customer records, employee files, telemetry with identifiers—, the data center contract is not only a services contract: it is also a data processing contract.

    The Federal Law on Protection of Personal Data Held by Private Parties, in its current text —published in the Federal Official Gazette on March 20, 2025, with the latest reform on November 14, 2025, and whose authority is now the Secretariat of Anti-Corruption and Good Government, not the INAI—, defines in article 2, section XII, the “data processor” as the person who processes personal data on behalf of the controller. The operator that hosts the client’s data falls, in practical terms, into that role. And four obligations derive from it:

  • Article 18. The controller must maintain administrative, technical and physical security measures, and may not adopt measures less than those it maintains for its own information, considering the risk, the consequences and the sensitivity of the data. In the contract, that means stating the security level the operator commits to maintain: access control, logs, video surveillance. And it cannot fall below the client’s standard.
  • Article 19. Security breaches that significantly affect the rights of data subjects are communicated immediately. The SLA must define what is considered a breach, who reports it, within what timeframe and with what evidence. “Immediate” is the legal deadline; “in the monthly report” is the practice: the contract decides which one is met.
  • Article 20. Those who take part in any stage of processing must keep confidentiality, and that obligation survives even after the relationship ends: it covers the operator’s staff and the remote hands that work on site. In the most regulated sector, that contract is already required with the CNBV requirements: the 4 CNBV requirements when using a remote hands service.
  • Articles 2, 35 and 36. A transfer is any communication of personal data made to a person other than the data subject, the controller or the data processor; the receiving third party assumes the same obligations, and article 36 allows transfers without consent when the transfer is necessary by virtue of a contract executed in the interest of the data subject. In a data center, that is a clause almost nobody has: what data is processed, who the processor is, which subprocessors are involved and in which countries they are accessed.
  • None of this is legal advice: the final wording is reviewed by the client’s legal area. The point is another one: Mexican law already requires documenting in writing what your SLA takes for granted, and that is the clause with the most leverage in a negotiation.

    What you can demand in writing

    The audit you need is not a list of good intentions: it is a document with the table of the five clauses, the evidence, the glossary and assumptions annex and the responsibility matrix by domain. It is the same exercise the industry already runs on the infrastructure side: the Uptime Institute maintains a tier gap analysis service and has accumulated more than 4,300 certifications. What not everyone can do is apply it to the contract of an AI workload in Mexico.

    The Noxtel point: we quote the contract and SLA gap audit for AI workloads: the five clauses with their evidence, the resilience and assumptions annex, the responsibility matrix and the personal data annex with articles 18, 19, 20, 35 and 36 verified against your current contract. We deliver it as a document you can send to your provider. If your data center does not have an SLA yet, the starting point is the data center assessment.


    Sources

  • Schneider Electric. “Colocation and AI: Why trust, transparency, and standards become design requirements” (Vance Peterson, October 30, 2026) — https://blog.se.com/datacenter/2026/10/30/colocation-and-ai-why-trust-transparency-and-standards-matter/
  • Vertiv. “Transforming colocation facilities to meet demand for AI compute” (technical document, 2026) — https://www.vertiv.com/495e4c/globalassets/documents/white-papers/transforming-colocation-facilities-to-meet-demand-for-ai-compute_white-paper_english.pdf
  • Uptime Institute. “Tier Classification System” (tier classification system) — https://uptimeinstitute.com/tiers
  • Uptime Institute. “Uptime Institute and TIA agree on clear separation” (joint statement) — https://uptimeinstitute.com/about-ui/press-releases/uptime-tia
  • Uptime Institute. “Tier Gap Analysis” (tier gap analysis service) — https://uptimeinstitute.com/tga
  • Want to master this?

    Noxtel Academy →

    Also in Facility Management

    ← Back to categories