Tier III+ SLA in AI datacenter colocation: what your contract must guarantee
A datacenter colocation SLA is not a marketing label. It’s a contractual document that defines what the operator owes you and what it pays if it fails to deliver. On a 6-to-9-year commitment covering 6 to 7 MW IT of AI inference workloads, the gap between a well-drafted SLA and a cosmetic one shows up the moment your infrastructure goes down, not before. This article spells out what an AI colocation contract has to contain to hold up in front of your procurement team, your CFO, and your board, and why most SLAs on the market fail that test.
TL;DR
A Tier III+ label isn’t enough. Only the contract truly protects your infrastructure. A defensible SLA defines availability component by component, covers N+1 redundancy on the DLC circuits, sets real penalties including on cooling, and puts the MTTR in writing. On a 6-to-9-year commitment, the strength of the operator (EQUANS) matters as much as the clauses.
What Tier III+ actually changes in your contract
The Uptime Institute’s Tier III+ certification is a recognized technical benchmark. It isn’t enough to protect your infrastructure if the commitments it implies aren’t translated into enforceable contract clauses. The label certifies design and construction. The contract is what binds the operator to actual operations, and that’s where your risk lives.
99.982% availability: the calculation your contract has to make enforceable
Tier III+ availability translates to 99.982% guaranteed uptime, or a maximum of 1 hour and 34 minutes of cumulative downtime per year. That figure only holds if it’s defined per critical component in your contract, not globally.
A global SLA at 99.9% can hide power running at 99.5% and a network at 99.95%. On 6 MW IT of GPU inference, power at 99.5% works out to 43 hours of potential downtime per year. On a GPU fleet of this size, where every hour of outage idles an asset whose depreciation keeps running, the difference between a precise SLA and a global one quickly adds up to significant annual exposure.
What your contract has to contain: a definition of availability per system (power, DLC cooling, network), with automatic monthly reporting and enforceable threshold values. If the operator refuses that level of granularity, the global SLA is a statement of intent.
| Metric | Tier II | Tier III | Tier III+ |
|---|---|---|---|
| Guaranteed availability | 99.749% | 99.982% | 99.982% + non-disruptive maintenance |
| Max downtime/year | ~22 hrs | ~1h34m | ~1h34m (online maintenance included) |
| Power redundancy | Partial | N+1 | N+1 concurrent |
| Cooling redundancy | No | N+1 | N+1 concurrent + redundant DLC circuit |
| Maintenance without outage | No | No | Yes |
| Uptime Institute certification | Not required | Tier III | Tier III+ Design + Constructed Facility |
N+1 redundancy: what it means for your GPU workloads
N+1 redundancy means every critical system has a standby unit on hand, ready to take over with no human intervention. On power and network, this mechanism is well documented in industry standards. On DLC liquid cooling, it’s less often handled with the precision it deserves.
A 100 kW GPU rack on native DLC generates a thermal load that air cooling can’t absorb as a fallback. If your contract guarantees N+1 redundancy on power but stays silent on the DLC circuits, you have a critical blind spot. When a liquid cooling circuit fails, switching over to air cooling isn’t an option at this density. Your GPUs shut down on thermal protection within minutes.
What your contract has to spell out: N+1 redundancy explicitly covered on the primary and secondary DLC circuits, with automatic failover that is tested and documented. The failover testing frequency (quarterly at minimum) has to be written into the contract, not into a technical appendix the operator can revise unilaterally.
The clauses most contracts leave out
Real financial penalties: the operator’s sincerity test
An SLA with no defined financial penalty is a statement of intent. This point is non-negotiable on a multi-year commitment covering critical workloads.
The penalty structure to demand: a service credit calculated on the actual duration of downtime, with progressive tiers. One example of a defensible structure: 10% of the monthly rent per hour of downtime beyond the SLA, capped at 30% of the monthly rent over the period. These figures aren’t universal standards, they’re negotiable, but their presence in the contract is the signal that the operator agrees to be held accountable for its commitments.
Two things to watch that often get left out of the negotiation. First: penalties have to apply to cooling incidents on the same footing as power outages. In a high-density DLC datacenter, a cooling incident is as critical as a power outage, and market SLAs frequently exclude it. Second: the definition of downtime has to be pinned down. Is a performance degradation that keeps your GPUs running at 60% of their rated capacity covered? If your contract doesn’t say so, the answer is no.
MTTR in writing: how long to restore a critical system?
The MTTR (Mean Time To Repair) is the maximum time to restore a system after an incident. It has to be contractually defined, not left to the operator’s judgment.
Current market standards: 4 hours for a power incident, 2 hours for a network incident. These values are negotiable downward if the operator runs a permanent on-call team and keeps critical spare parts on site. For high-density GPU inference workloads, 2 hours of downtime on 6 MW IT represents a direct, calculable loss. Bring that figure into your MTTR negotiation, not as an emotional argument but as the basis for calculating penalties.
This is exactly where EQUANS, a Bouygues subsidiary and the operational operator of Voltekko’s infrastructure, brings a guarantee few operators can put on the table. With 20 years of datacenter operations and a permanent 24/7/365 on-call team, EQUANS carries operational responsibility for the MTTR across the entire duration of the client commitment. This isn’t a startup’s promise. It’s the commitment of an industrial operator whose continuity doesn’t hinge on the next funding round.
Dedicated capacity blocks: the commitment that protects your planning
The difference between shared colocation and a dedicated block is fundamental for an AI neocloud in scaling mode. A dedicated 6-to-7 MW IT block means your IT power isn’t shared, your power supply is sized exclusively for your needs, and your scaling capacity is contractually preserved for the duration of the commitment.
What your contract has to spell out on this point: guaranteed IT power per block (in kW, not in rack equivalents), the option to expand with a defined notice period (90 days is an acceptable standard), and the terms for revising capacity upward without renegotiating the entire contract.
For your board, this argument is concrete: a dedicated block eliminates the risk of resource contention. With a shared operator during a period of strong GPU demand, your access to IT power can be degraded without any formal breach of the SLA, as long as that SLA doesn’t guarantee the exclusivity of your block. This risk isn’t theoretical. GPU capacity demand in Europe has grown sharply since 2023, and shared operators are under heavy capacity pressure.
Why the operator matters as much as the contract
A good contract with a failing operator is worth nothing at 3 a.m. when a DLC circuit goes down. The quality of the operational operator is the variable most selection processes underweight, until the first incident.
EQUANS: 20 years of datacenter operations, not a construction subcontractor
EQUANS is often presented as the construction and design partner for Voltekko’s infrastructure. That’s accurate, but it’s the visible part. The part that matters for your 9-year commitment is operations.
EQUANS, a Bouygues subsidiary, is the contractually committed operational operator on Voltekko’s sites, in France (greater Paris area) and Portugal (Alcochete, under construction). In concrete terms, that means: permanent 24/7/365 on-call coverage, industrialized preventive maintenance based on processes proven over 20 years of datacenter operations, a clear chain of responsibility when an incident hits, and escalation procedures defined in the contract, not left to the discretion of a technical team whose makeup can change.
The question your board will inevitably ask: “Who’s operating in 7 years, at 3 a.m., when a critical system goes down?” The answer is EQUANS. Not a team of five engineers hired 18 months ago. A first-tier industrial operator whose operational continuity is structurally guaranteed by its membership in the Bouygues group.
The responsibility structure: who answers for what?
The clarity of the chain of responsibility is a selection criterion your procurement team will examine. Here’s how it’s structured at Voltekko.
Voltekko is the operator and the client’s single contractual point of contact. Voltekko signs the colocation contract, carries the SLA commitments, and is the first-line contact for any contractual or commercial question.
EQUANS is the operational operator, responsible for meeting the technical SLAs on power, DLC cooling, and network. Operational responsibility rests with an industrial player whose financial footprint and continuity don’t depend on Voltekko’s funding trajectory.
REED, a Société Générale group subsidiary, is the institutional investor in Voltekko’s capital. REED’s presence isn’t a marketing argument. It’s a risk structure your board can analyze. An operator backed by Société Générale carries financial credibility your procurement team can verify and a durability that independent operators with no institutional anchor can’t offer.
This three-tier structure (contractual operator, industrial operator, institutional investor) is exactly what makes it possible to answer the most common objection in decision committees: “What if the operator no longer exists in 5 years?”
What your next negotiation session has to cover
A Tier III+ SLA is only defensible in front of your board if it contains five precise elements: availability defined per critical component (not globally), N+1 redundancy explicitly covered on the DLC circuits, progressive financial penalties applicable to every critical system, an MTTR set in the contract per incident type, and a definition of dedicated capacity in guaranteed IT power.
If your current colocation contract, or the one you’re negotiating, doesn’t contain these five elements, you’re carrying an operational risk your board hasn’t approved, because it was never presented to them in this form.
Evaluating a colocation contract for your AI inference workloads? Our technical teams analyze your exposure to operational risk and the clauses to demand in your negotiation.