Even in a cloud-first world, someone runs the buildings. Colocation is a large and growing industry, it is where every network's physical reality lives, and it is a subject infrastructure candidates are expected to have at least a working grasp of — particularly in India, where data localisation rules and a construction boom have made it a genuine growth sector with real hiring.

The single most useful reframe: colocation is priced on power, not floor area. A rack is a 42U or 47U cabinet, but what determines its price is the kilowatts committed to it. A 2 kW rack and a 15 kW rack occupy the same tiles and cost very different amounts, because the constraint in a data hall is the electrical and cooling capacity, not the floor.

• Committed vs metered power. You commit to a kW figure and pay for it whether you draw it or not; overage above the commit is billed separately and expensively. Under-committing is a false economy and over-committing is dead money, so the sizing conversation is real work.
• A and B feeds. Two independent power paths to the rack, from separate UPS systems. Your equipment must have dual power supplies to use them — single-PSU kit in a dual-fed rack still dies when one path is maintained.
• Cross-connects. A physical cable from your rack to another tenant, a carrier, or a cloud on-ramp, billed per month per connection. At a dense exchange facility these accumulate quickly and are a meaningful part of the bill.
• Remote hands. Someone in the building who will reseat a drive or read a console at 3 a.m., billed hourly with a minimum. Worth budgeting rather than discovering.
• NRC. Non-recurring charges at install — racking, cabling, power commissioning. Always negotiated, frequently waived on a longer term.

The Uptime Institute tier system is the industry's shared vocabulary, and the important thing to know is what each tier promises rather than the availability percentage that gets quoted alongside it.

• Tier III is ‘concurrently maintainable’ — any single component or distribution path can be taken out for maintenance without shutting the load down. This is the commercial sweet spot and what most enterprise colocation is.
• Tier IV is ‘fault tolerant’ — it survives an unplanned failure of any single component, with physically separated systems. Substantially more expensive, and genuinely necessary far less often than it is specified.
• The certification distinction. ‘Tier III Certified Design Documents’ means drawings passed review. ‘Tier III Certified Constructed Facility’ means the building was physically tested. Marketing material routinely says ‘Tier III’ and means the first. Asking which is a very good question in a vendor conversation and an even better one in an interview.

PUE — power usage effectiveness — is total facility power divided by IT power. A PUE of 2.0 means you burn a watt of overhead for every watt of compute. Modern purpose-built facilities operate in the 1.2–1.5 range; hyperscalers report figures closer to 1.1. In India, ambient temperature and humidity make good PUE harder to achieve than in northern Europe, which is a genuine engineering constraint and one reason cooling innovation gets attention there.

Treat published PUE numbers with the same scepticism you would apply to any self-reported metric: it varies by season, by load factor, and by what the operator chooses to include in the boundary. A half-empty hall always reports a bad PUE because the fixed overhead is spread over less IT load, which is a measurement artefact and not an efficiency problem.

For twenty years an average enterprise rack drew 3–6 kW and data halls were designed around that. A rack of modern AI accelerators can draw 40–130 kW — an order of magnitude more in the same footprint. This is the single biggest structural change in the industry in a decade and it has consequences worth being able to name:

• Air cooling stops working somewhere around 30–40 kW per rack. Above that you need direct-to-chip liquid cooling or immersion, which means plumbing in a building designed for airflow — a retrofit most existing halls cannot take.
• The constraint moved to the grid. Data centre siting is now decided by where you can get tens or hundreds of megawatts connected, and grid connection queues in several markets run to years. This is why announcements are increasingly about power agreements rather than buildings.
• Vacancy collapsed and pricing rose in the major markets through the mid-2020s, reversing a long period of falling prices. Northern Virginia is the extreme case.
• India benefits structurally — data localisation requirements under the DPDP Act and RBI's payment-data rules mean certain workloads must be hosted domestically, which underwrites demand independently of the AI cycle.

A question you may well be asked: when does owning hardware in colocation beat renting from a cloud? The method matters more than the answer.

The honest shape: cloud wins decisively on elasticity, on time-to-start, and on anything with a spiky or unknown load. Colocation wins on steady, predictable, high-volume workloads — and the two places it wins most dramatically are bulk egress, where the pricing gap is enormous, and sustained GPU capacity, where cloud rental over three years can exceed the purchase price of the hardware several times over. Which is precisely why a number of large AI companies have moved to owning.

The mature answer to “cloud or colo?” is therefore not a side, it is a method: “Split the workload. Elastic, unpredictable and new things belong in cloud; large steady-state things with predictable load and heavy egress belong where you can buy them outright. Then compute the three-year total for each, including the people, and be honest about the elasticity you are giving up.”