Data center Tier levels are Uptime Institute's four classes of site infrastructure: Tier I (Basic Capacity), Tier II (Redundant Capacity Components), Tier III (Concurrently Maintainable) and Tier IV (Fault Tolerant) [1]. A Tier III site can take any capacity component or distribution path out of service for planned work without affecting operations, though it stays exposed to an equipment failure or operator error [1]. A Tier IV site also keeps running through an individual equipment failure or path interruption [1].
No Tier carries an uptime percentage. Uptime removed expected downtime per year from the Tier Standard in 2009 and says it was never part of the Tier definitions [22, 23].
A Tier rates the facility, and the servers decide how much of its redundancy they can use [14, 24]. NVIDIA says a DGX B200 rack should be understood as receiving only N, even with dual feeds from discrete sources [14].
The four data center Tier levels
| Tier | Capacity components | Distribution paths | Planned maintenance | Equipment failure |
|---|---|---|---|---|
| I | N | 1 | Site-wide shutdown | Impacts the site |
| II | N+1 | 1 | Site-wide shutdown | Distribution failures impact the site |
| III | N+1 | 1 active, 1 alternate | No shutdown | Site still exposed |
| IV | N after any failure | 2 simultaneously active | No shutdown | One failure does not impact operations |
Sources: [1, 4], with component and path columns from the 2009 standard.
The Tiers are progressive, so each includes the requirements of the ones below [2]. A site's Tier is the lowest rating among its subsystems, with no fractional Tiers [4]. Uptime says terms such as "Tier III plus" have no basis in Tiers [23].
Tier III vs Tier IV
Tier III adds a redundant delivery path for power and cooling, so every component can be maintained without affecting IT [22]. Critical power distribution, from the UPS output down, must be active on both paths: if a rack receives power from two distributions, both must normally be active [23].
Tier IV adds fault tolerance: physically isolated systems and paths, continuous cooling and N capacity after any infrastructure failure [4]. All systems must be active/active [23]. Uptime also says Tier IV is not always better than Tier II, since the application's business value may be lower [22].
Component counts do not set the Tier. Uptime says Tier IV is possible "with just N+1 components, depending on how they are configured and connected to redundant distribution pathways" [23].
Is there a Tier III uptime percentage?
No. The figures attached to Tier III trace to Uptime's 2008 white paper. Its Table 2 is titled "Common Attributes Found in Data Centers that are Unrelated to Tier Requirements" and gives downtime "based on field data" from 16 data centers [5].
| Tier | 2008 field-data downtime (16 sites) | 2008 field-data availability | Typical ultimate kW per cabinet |
|---|---|---|---|
| III | 1.6 hours | 99.98% | Over 3 kW |
| IV | 0.8 hours | 99.99% | Over 4 kW |
Source: [5], Table 2, which Uptime titled as unrelated to Tier requirements.
The 99.982% figure is 1.6 hours out of 8,760 (derived from [5]). The paper says availability calculations do not determine the Tier [5]. Uptime dropped the downtime references in 2009 because operational behaviors can affect availability whatever the design [22]. The table's Tier III density, over 3 kW per cabinet, compares with approximately 120 kW for a GB200 NVL72 rack [5, 16].
WhiteFiber's investor presentation, filed with the SEC in September 2026, lists "99.982% Uptime and No More than 1.6 Hours of Downtime Annually" under Tier III, crediting Uptime Institute [20].
Utility feeds do not count toward a Tier
Tier Standard: Topology treats the engine-generator plant as the only reliable power source, because utility power is subject to unscheduled interruption [23]. The number of utility feeds, substations and grids neither predicts nor influences Tier level [23]. Tier III and IV generators must be able to run without a limit on consecutive hours at N load [4].
Uptime estimates GB200 NVL72 racks can jump from around 60 to 70 kW to more than 150 kW, and expects generator sets to take more of that stress than other power equipment [32].
Tier certified vs Tier designed
| Award | What it verifies | Validity |
|---|---|---|
| Design Documents | The design | Two years, for awards after 1 January 2014 |
| Constructed Facility | The built site, through demonstration tests | Until the facility is modified |
| Operational Sustainability | Operations, building and site | Gold three years, Silver two, Bronze one |
Each award is a prerequisite for the next [6, 7]. Uptime says more than 85% of data center designs are incorrectly executed in construction [22].
On Uptime's public list, viewed 7 November 2026, 869 of 1,562 Tier entries show a design award as their only Tier award, and 673 show a constructed-facility award [9]. The list gives no dates, so lapsed design awards cannot be counted. Corvex's 2025 annual report describes a "tier III-designed facility" [21], a phrase that names no award.
TIA-942 Rated 3 and EN 50600 Class 3
Uptime and the Telecommunications Industry Association agreed in 2014 to separate their systems, and a TIA rating may not use the word Tier [10, 23]. TIA-942 has Rated-1 to Rated-4, certified by TIA-licensed bodies [11, 12]. EN 50600 has Availability Classes 1 to 4, and the CEN, CENELEC and ETSI coordination group says self-declaration is all that is available, with no formal certification [13].
The wording differs at the rack. TIA's Rated-3 says "Typically, only one distribution path serves the computer equipment at any time" [12]. Uptime requires both critical power paths to a rack to be normally active at Tier III [23]. A TIA-942 facilities certification lasts three years, with surveillance audits after years one and two [12].
NVIDIA's GB200 and Vera Rubin NVL72 reference architectures accept any of the three: Uptime Tier 3, "TIA942-B Rated 3 or EN50600 Availability Class 3" [18, 19]. The rack power explainer covers what those racks draw.
N+1 vs 2N redundancy and outage rates
Uptime defines N as capacity sized to the expected peak load, N+1 and N+2 as one or two additional components, and 2N as the full capacity duplicated in two independent paths or systems [26]. Its October 2026 analysis pooled over 1,200 survey responses from 2025 and 2026 [26]. The share reporting at least one impactful outage in three years was 70% with N cooling against 39% with 2N, and 68% with N power against 42% with 2N [26]. N+1 and N+2 fell between [26]. The authors caution that correlation is not causation and that N samples are small [26]. In Uptime's 2025 survey, 42% of respondents named 2N as their usual level of redundant power equipment, and 41% N+1 [27].
Why a dual-fed DGX B200 rack gets N
A DGX B200 has six power supplies and needs five energized [14]. NVIDIA says rack provisioning "should be understood as delivering ONLY N, even if the provisioning is comprised of dual feeds from discrete sources" [14]. In its two-source layout, each source feeds half the rack, so a UPS failure affects half the systems [14]. Feeding each server from both sources "increases the probability of disruption rather than reducing it" [14]. Split three and three, one lost source leaves three supplies, two short of five (derived from [14]).
NVIDIA shows two routes to N+1. It has not validated any static transfer switch with DGX B200 systems and does not currently recommend that design [14]. Uptime says a single large static transfer switch does not meet Tier III or IV criteria [24]. The second route, "6 to make 5", feeds each system from six discrete UPS sources and may limit a rack to two systems [14].
Newer systems differ. The DGX B300 has 12 power supplies in N+N, with six required [15]. A GB200 NVL72 rack has eight power shelves described as N+N [16]. A Vera Rubin NVL72 rack has four shelves with three required, also called N+N [19]. Ask how each maps onto the facility's A and B feeds.
Questions to settle before reserving capacity
A CBRE footnote describes its wholesale rents as "asking rates for 250+ kW at N+1/Tier III requirements" [35]. It names no Uptime award [35]. When capacity is reserved ahead of hardware bought under a forward agreement for GPU systems, these questions come before signing:
- Which Tier award the site holds, its date and the phase or room on the foil [8].
- For a TIA-942 or EN 50600 claim, who certified or assessed it and when it lapses [12, 13].
- Whether critical power is active on both paths to the rack [23].
- How many power supplies each GPU system needs energized, and from which sources [14, 15].
- The power and cooling redundancy, in N terms, of the hall that hosts the racks [26].
Rillor's capacity agreements reserve power and space in 250 kW, 500 kW and 1 MW blocks, liquid-cooled or rear-door cooled, on a 24-month standard term priced per kW-month and sized to the GPU systems by power draw and cooling grade.
Read the full report
The full report adds the 2009 Tier requirements table, the award tally by Tier, TIA's Rated-1 to Rated-4 definitions, outage reductions for N+1 and N+2, and NVIDIA's five DGX B200 power layouts.