The four places buyers look, and how each one bends#
A buyer researching colocation has essentially four sources available: the operator's own site, a third-party data-center directory, an AI assistant or search engine, and a person who sells for a living. Each is useful. None is neutral, and the failure modes are different enough that stacking them does not cancel the bias out.
| Source | Bends toward | Still the best source for |
|---|---|---|
| Operator's own site | Counting generously, publishing planned capacity beside delivered capacity, and self-describing certifications | Site codes, addresses, product structure, and what the operator will commit to in writing |
| Third-party directories | Counting network nodes as data centers, and carrying listings years after an operator has left a market | Cross-checking an operator's own roster, and finding buildings an operator does not market |
| AI assistants and search | Naming the largest brands regardless of fit, because those brands dominate the text they learned from | Orientation and vocabulary — knowing what to ask about |
| A salesperson | The operators they are paid to sell, and toward a transaction happening at all | What is actually deliverable, at what price, on what date — which nothing else can tell you |
What operator sites get wrong#
Operator marketing is rarely false. It is optimistically counted, and the counting convention is almost never disclosed. Three patterns account for most of it, and we hit all three while building our own directory.
Counting conventions that inflate a footprint#
Operators count campuses, buildings, site codes, metros and network locations inconsistently, and pick whichever unit produces the better number. One operator's own colocation page advertised roughly three times as many edge data centers as its own facility directory listed as colocation-sellable — a contradiction inside a single website. Another's marketing described more than forty facilities across eighteen markets, while an independent rebuild from its own site codes found closer to fifty-nine buildings across nineteen metros; that operator counts campuses where a rebuild counts buildings, so the two numbers disagree in both directions at once.
The most consequential version we found was an operator whose audited filing disclosed materially fewer sites, metros and sellable megawatts than its own marketing claimed — and the filing named several sites slated for closure without itemising which. The regulated document and the marketing page describe the same company and do not agree. When those two disagree, the filing wins.
Pipeline presented next to delivered capacity#
Announced campuses, permitted parcels and buildings under construction sit on the same page, in the same visual treatment, as buildings you can move into next month. The megawatt figures are frequently the planned build-out rather than the delivered load. We excluded a number of sites from our own directory for exactly this reason: a permitted campus phased to 2027, a first-building figure whose operator never actually stated the building was operational, an eighteen-megawatt figure explicitly described in the source as long-term potential. Planned capacity is not capacity you can buy, and a directory that blurs the two will eventually quote a building that does not exist yet.
Certifications that are self-declared#
This is the one that costs buyers real money, because it survives procurement review. The Uptime Institute Tier classification is a genuine third-party certification, awarded per facility. A great deal of what appears next to the word "Tier" on operator sites is not one.
- "Tier III design" or "Tier 3 standard" means the operator says it built to that standard. No third party has audited it.
- "Tier III equivalent" and "Tier III quality" are marketing constructions with no certifying body behind them at all.
- "Tier 5" does not exist. One large operator markets a "Tier 5" standard administered by a standards organisation that operator founded. That same operator separately holds genuine, independently-audited Uptime Tier IV certification on two of its older buildings — so it is simultaneously the holder of a real certification and the publisher of an invented one.
The same pattern runs through compliance. A logo on a trust page is not an audit letter. Across every major operator we hold, we could not find a downloadable auditor letter or certificate number for their SOC 2 and PCI DSS claims. Compliance is also usually scoped to specific facilities rather than the portfolio: one operator's federal authorisation covers three named buildings out of a fleet of seventy-six, and another's federal authorisation covers its software products and not its data centers at all — its own web copy invites the misreading.
What third-party directories get wrong#
Directories are genuinely useful and we use several. They fail in two specific ways, both of which follow from how they are built: they ingest listings, and a listing is easier to add than to remove.
Network nodes counted as data centers#
A point of presence — a rack of an operator's own network gear inside somebody else's building — is not colocation you can buy. Ingested by address, it looks exactly like a facility. We found one operator whose aggregate directory listings ran to roughly thirty-six locations against sixteen the operator itself published as data centers, with the difference being network nodes. Another operator appears in listings at nine locations of which exactly two are colocation-sellable — the rest are bare-metal compute sites and a network node.
The trap is worse than a wrong count, because these entries have precise, real addresses. In our Phoenix work, one operator's network node sits inside a competing operator's building. An ingest keyed on address produces a facility that is real, correctly located, attributed to the wrong company, and not for sale.
Listings that outlive the operator's presence#
Colocation directories decay quietly. Two examples from our own Phoenix reconciliation: one operator was still listed across multiple directories at a downtown Phoenix building although its own current facility directory lists no Phoenix site anywhere, with a second Phoenix site flagged decommissioned by one aggregator and absent from the operator entirely — the pattern of a quiet market exit that the listings have not caught up with. Another operator was carried in listings at a Phoenix building it sold out of around 2016. That listing was roughly a decade stale and still returned by search.
What AI assistants get wrong#
Ask an assistant who to consider for colocation in almost any US metro and you will get the same three or four global brands, plus trade press. This is not the model being careless — it is the model being accurate about its training data. The largest operators generate most of the published text about colocation, so they dominate the answer regardless of whether they fit your requirement or even operate in your market.
The clearest illustration in our own data: the operator most often named first for any US colocation question has had no facility in Phoenix since 2012, when it divested its Phoenix site. It is a completely reasonable answer to "who are the major colocation providers" and a wrong answer to "who should I get a Phoenix quote from." A buyer who does not already know this loses weeks discovering it.
The second failure is quieter and costs more. Regional and mid-tier operators are close to invisible in AI answers, because comparatively little is written about them. For a two-to-twenty-cabinet requirement in a metro like Phoenix or Dallas, those are frequently the operators who will actually compete hardest for the business. An answer that omits them is not just incomplete — it points a price-sensitive buyer directly at the operators whose own qualification guidance says they are the wrong fit for that deal.
What a provider-independent view actually looks like#
"Independent" is not a claim you can check, so it is worth defining by its outputs. These are the properties that distinguish a genuinely provider-independent view from a marketing surface, and they are testable — you can hold any source, including this one, against them.
- It names operators it cannot sell. A market view that happens to contain exactly the operators its author earns from is an advertisement. Our metro guides name operators we have no commercial route to, and say so on the page.
- It says who each operator is wrong for. Fit is symmetric. An operator's own qualification guidance is usually clear that certain deals are bad for them too; publishing only the positive half is where most "neutral" content quietly fails.
- It attributes every number. Operator-published, third-party estimate, or audited filing — these are three different grades of evidence and a page that flattens them into one voice is hiding its weakest claims among its strongest.
- It marks what it does not know. An honest gap is more useful than a confident guess, because a gap tells you where to ask a question and a guess tells you not to.
- It refuses to state availability. Whether a specific operator has space in a specific building at a specific density is knowable only by asking that operator, and the answer decays in weeks. Any source that answers it from a page rather than from a conversation is guessing.
- It is dated. A colocation fact without a date is not a fact, it is a rumour with good posture.
That last pair is why we publish no availability anywhere on this site and date every page. It is also, honestly, why the advisory model exists at all: the useful version of this information is not a page, it is the set of answers you get when you ask every relevant operator the same questions at the same time and write down what they said. The sample reports show what that produces, and how to run that process yourself is written out in full — including the parts you do not need us for.