N, N+1 and 2N are shorthand for how much backup a data center builds into its power and cooling systems. The number tells you what happens when something breaks.
What N means
N is exactly what you need and nothing more. If your load requires four cooling units, you have four. Any one of them fails and you are already over capacity. N is fine for a lab or a development environment. It is not a basis for a production commitment.
What N+1 means
N+1 adds one spare on top of what you need. Four cooling units required, five installed. One fails, the other four carry the load and a technician schedules a repair without shutting anything down. For power, N+1 typically means one extra transformer, UPS (uninterruptible power supply) module or generator beyond the minimum. The site stays up through a single failure anywhere in the chain.
The catch is that N+1 only tolerates one failure at a time. If a second unit fails before the first is repaired, you are back to bare minimum, or below it.
What 2N means
2N doubles everything. Two complete, independent power paths from the utility feed to your rack. Two independent cooling loops. Each path is sized to carry the full load on its own. One path fails entirely and the other takes over without touching the other. Tier III already requires concurrent maintainability, which is explained below.
For a team running workloads where an hour of downtime has a direct dollar cost, 2N is the standard worth asking about.
What "concurrent maintainability" adds
A site can have N+1 on paper and still require a maintenance window to swap a failed component. Concurrent maintainability means any part of the infrastructure can be repaired or replaced while the site runs at full load. No planned downtime, no "we need a four-hour window Saturday night."
Uptime Institute's Tier III standard requires concurrent maintainability. Tier II does not. When you are evaluating space and power for hardware you plan to run for three years or more, this distinction matters more than almost any other spec. A contract that locks you in for 36 to 60 months at a site that cannot be maintained without a shutdown is a contract that will produce at least one painful conversation.
See Tier III vs Tier IV: What the Tiers Mean for AI for how those standards map to AI workloads specifically.
Why cooling failures hit faster when racks run hot
Power redundancy gives you time. A UPS bridges the gap, a generator starts, and the transition takes seconds to minutes. Cooling redundancy gives you much less time, and the margin shrinks as rack density rises.
A standard server rack running 5 to 10 kW has some thermal mass. If cooling drops, temperatures rise over minutes, giving operators time to respond. A single H100 or H200 server draws around 10 kW. Fifty of those servers, 400 GPUs, draw roughly 500 kW in a cluster. A GB300 NVL72 rack draws up to 142 kW on its own. At that density, a cooling failure can push the temperature inside the GPU chips to shutdown thresholds in a very short time, not minutes, and GPUs will throttle or shut down before a human can intervene.
This is why the cooling redundancy question is not academic for high-density AI infrastructure. The failure mode is fast and the hardware is expensive.
Questions to take into a call with an operator
- What is the redundancy level for power, and is that N+1 or 2N at every point in the chain, from utility feed to rack?
- What is the redundancy level for cooling, separately from power?
- Is the site concurrently maintainable, and is that written into the contract?
- What is the recovery time if a cooling unit fails at full rack load?
- Is there a single point of failure anywhere between the utility and my cabinet?
- If one power path is taken offline for maintenance, what does the other path carry, and what is its rated capacity?
- How does the cooling design change if I bring in liquid-cooled racks?
That last question matters because liquid cooling, which rack-scale Blackwell systems such as the GB300 NVL72 and all Rubin systems require, adds a facility water loop to the picture. That loop has its own redundancy requirements, separate from the air-handling systems the rest of the site may rely on.
Colocation, which is space and power for hardware you own, is a long commitment. The redundancy architecture you accept at signing is the one you operate under for the life of the contract.
Sources
Redundancy tier definitions are drawn from Uptime Institute's Tier Standard: Topology, which is publicly available. The H100 and H200 figure of approximately 10 kW per 8-GPU server is a widely used planning figure. The GB300 NVL72 maximum of 142 kW per rack and the individual B300 GPU rated power of 1,400 W are from NVIDIA's published reference architecture. The GB300 NVL72 figure of 132 to 140 kW in operation is from Supermicro's datasheet.
Share your requirements. We'll canvas the market.
Tell us what you need (for colocation space and power: market, kW or MW, cooling and timing; for reserved GPU capacity: region, GPU type, cluster size and timing) and we'll canvas colocation and neocloud operators in parallel. Shortlist in 48 hours. We don't sell hardware. We broker colocation space and power, and reserved GPU compute capacity.
Earlier conversations get better terms. When you engage early, we have time to negotiate with vendors before you need to commit. We're paid only when a deal closes, and the fee structure is agreed in writing before any introduction.