Why datacenter capacity offers are hard to compare, and what a standard spec changes

Datacenter capacity offers, such as colocation quotes, are hard to compare because each one answers a slightly different question. One quotes total power, another IT power; one says “available Q3”, another gives a date that depends on a grid application. Offers become comparable when both sides describe the same few fields in the same units: IT capacity, grid power and its status, rack density, redundancy, cooling and a ready date. GPU compute adds the GPU model, count and interconnect.

The same words hide different things. The most common gaps:

• Power. A site's grid connection is not its IT capacity. IT capacity is what servers can use; the rest goes to cooling and building systems.
• Certainty. Power that is secured with the grid operator, applied for or only planned are three very different promises (see our guide to grid power status).
• Density. A hall built for a few kilowatts per rack cannot simply host AI racks that draw several times as much; power distribution and cooling decide it.
• Resilience. “Redundant” can mean N+1 on some systems or 2N throughout. The level matters more than the word.
• Timing. “Available” can mean ready to hand over, under construction, or waiting on power. A ready date is only useful with what it depends on.

None of this is bad faith. Each operator describes its site the way it is used to, and each buyer asks in its own template. The work of lining the answers up falls on whoever reads them.

A short list of fields settles most first conversations. On AICapacity, every request and every listing states them the same way:

• IT capacity in kW or MW, or for GPU compute, the number of GPUs.
• Grid power and its status: secured, applied or planned, for land and shells especially.
• Rack density in kW per rack.
• Redundancy and tier: N, N+1, 2N or 2N+1.
• Cooling: air, rear-door, direct liquid or immersion.
• Ready date: when the capacity can be handed over.

Location, certifications, renewable supply and heat reuse complete the picture. The glossary defines each term in plain words.

Once both sides use the same fields, a site can be checked against a request field by field. AICapacity does it in two steps.

First, hard filters remove pairs that cannot work: the wrong kind of capacity, a country outside the request, a site that is no longer available, or less than half the requested capacity. Only secured grid power counts toward that last check.

Then a fit score out of 100 weighs capacity, timing, density, redundancy and cooling, with a small bonus for renewable supply, certifications and heat reuse. Applied power still counts toward capacity, but if a request needs more than the secured power, the applied power's expected date becomes the ready date. Each operator can say which criteria matter most to it. A field the request leaves open is left out rather than guessed, and an operator's plan never changes the score.

The operator sees the score with its reasons, field by field: enough power, ready a quarter late, the right cooling. A number without the breakdown would only move the guesswork.

A standard spec gets the right parties talking sooner. It does not replace a site visit, technical due diligence or a contract. Those come after both sides have chosen to engage, usually under a mutual NDA. On AICapacity, names stay hidden until then: the operator says it is interested, the buyer chooses to engage, and only then do both see who the other is.