A colocation lease is priced in dollars per kilowatt of contracted IT power per month — and the unit is the first thing to understand, because it is not a rent on space. Two independent market guides define it the same way: GoDataCenters calls it "the monthly rate a data center operator charges per kilowatt of contracted power in a given metro … usually quoted as dollars per kW per month and inclusive of space, cooling and a base level of power," and Kurums' market statistics page states plainly that rates are "quoted per kilowatt of contracted IT power per month, because power — not floor space — is the scarce resource." You are buying a reserved slice of a substation, a chilled aisle to absorb its heat, and the floor to stand it on — in that order of scarcity. The floor area a kilowatt occupies is nearly a rounding error; the grid connection behind it is a multi-year project, which is why the price moved from $/sqft to $/kW and never moved back.

The second thing to understand is that the quoted per-kW number is one line of a three-line bill. The first-party proof is Digital Realty's own earnings table: in its Q2 2026 results the company books interconnection revenue separately ($20.5 million in the quarter) from base rent, because cross-connects are billed on their own meter. And the per-kW rent itself is charged on committed capacity — the kilowatts you reserved — not the kilowatts you actually consume. Electricity is then metered on top, at cost or at a contracted rate depending on the deal. Every comparison error in this market comes from collapsing those lines into one number.

Who publishes what, and what each number is worth, is the map for the rest of this piece. Brokers publish asking rents — the landlord's opening band. Public operators publish signed-lease economics in quarterly filings — the closest thing to contract prices anyone can check. Analysts publish trend rates — the direction and its durability. We use all three layers below, and the interesting finding is that in 2026 the layers disagree with each other in a specific, useful way.

What does the market actually charge per kW in 2026?

CBRE, the commercial real estate services firm, is the primary source for market bands, and its two 2026 flagship reports bracket the US market. From the North America Data Center Trends H1 2026 report (its Northern Virginia chapter, published August 2026): colocation vacancy fell to 0.24% in the half, all capacity delivering in 2026 has been committed, only 10.8 MW of supply is available, and "asking rates for 10-plus-MW requirements rose to $160 to $185 per kW" — while, in the same paragraph, "overall rental rates held steady." From the Global Data Center Trends 2026 report (Q1 2026 data, CBRE's own summary): Frankfurt is Europe's highest at $235 to $265 per kW-month and Singapore is the most expensive Asia-Pacific market at $330 to $475, with Northern Virginia vacancy at 0.3% and Atlanta at 1%. Three independent compilations of CBRE's data (ColocationScout, GoDataCenters, Kurums) agree on the smaller-deal band for Northern Virginia at $190–235 per kW-month for 250–500 kW requirements, Chicago at $200–230 — and they agree with CBRE's own Frankfurt and Singapore figures.

MarketBasisBand ($/kW-month)Source
Northern Virginia10 MW+ wholesale requirement$160–185CBRE H1 2026 NA Trends (first party)
Northern Virginia250–500 kW requirement$190–235CBRE Global Trends 2026 via three independent compilations
Chicago250–500 kW requirement$200–230same CBRE compilation record
Frankfurtcolocation asking rate$235–265 — Europe's highestCBRE Global Trends 2026 (first party)
Singaporecolocation asking rate$330–475 — APAC's highestCBRE Global Trends 2026 (first party)

Asking rents, however, are openings, not outcomes. The layer closer to a signed contract is the operators' own filings, and Q2 2026 delivered a unusually clean one from Digital Realty — one of the two largest data center REITs. Per its results release (also filed with the SEC): new leases signed in the quarter in the Americas cleared at $293 per kW-month for deployments of 0–1 MW (10.6 MW signed) and $156 per kW-month for deployments above 1 MW (44.2 MW signed); across all regions the same table shows $280 for sub-1-MW and $157 for above-1-MW, with EMEA at $269 and Asia-Pacific at $286 for the small band. Equinix, the retail-heavy operator, discloses per-cabinet revenue instead: $2,538 per cabinet-month globally ($2,743 in the Americas) in Q2 2026, up 6% year over year on 307,600 billed cabinets — figures compiled by ColocationScout from Equinix's Q2 2026 earnings presentation (page 24). Note what the two operators' disclosures measure differently: Equinix's per-cabinet figure has cross-connects and other charges inside it; Digital Realty's per-kW figures are base rent only, interconnection excluded. That is the three-line bill again, live in the data.

Why does the same megawatt cost 88% more when you buy it small?

Because the bands are real, and one operator's signed leases price the premium precisely. In Digital Realty's Q2 2026 Americas signings, 0–1 MW deployments paid $293 per kW-month while above-1-MW deployments paid $156 — a 1.88× premium per kilowatt for buying small. Kurums' reading of the broker data says the same thing in words: "Rates fall as deal size rises: wholesale leases of 10 MW or more price well below the 250–500 kW band, while small retail deployments price higher." CBRE's Northern Virginia numbers bracket it from the market side: $160–185 for 10+ MW against $190–235 for 250–500 kW.

The arithmetic on what that premium costs (ours, from the operator's own table): a 750 kW requirement priced on the sub-megawatt band rather than aggregated across the 1 MW line pays the $293-minus-$156 spread of $137 per kW-month — about $1.23 million a year on 750 kW. We are not arguing that every enterprise can or should buy wholesale blocks; a 750 kW requirement is a real and common size, and hyperscalers take the wholesale bands precisely because they sign for tens of megawatts. The honest use of this number is negotiating leverage: the size premium is now published quarterly by the operator itself, which means a buyer consolidating scattered cabinets into a single commitment has a signed-lease receipt for what consolidation is worth. JLL's midyear 2026 report names the squeeze the same way we'd frame it: enterprises with requirements of 500 kW to 3 MW — "standard corporate users" — are "completely squeezed out of the market" while hyperscalers leased nearly 10 GW in twelve months.

Why is the renewal spread the real price discovery?

Here is the reversal this piece exists to flag. In the first half of 2026, CBRE's headline asking rents for 10+ MW deals moved +6.7% nationally across primary markets — and just +1.5% in Northern Virginia, with Silicon Valley and Hillsboro flat (the by-market and by-size breakdown is reported by two independent compilations of the H1 2026 report, GoDataCenters and Kurums, and is consistent with CBRE's own "overall rental rates held steady" language for Northern Virginia). In a single quarter of the same year, Digital Realty's renewal leases repriced +25.4% on a cash basis and +32.0% on a GAAP basis. Two layers of the same market, one moving single digits per half-year, the other moving mid-twenties per quarter.

We read the divergence structurally, not as an error on either side. With vacancy at 0.24% in Northern Virginia and 1% across North America for a third consecutive year (JLL), there is essentially no existing space being re-marketed — so the asking band mostly prices new builds and hypothetical blocks, and it is a landlord's opening position in a market where the landlord holds every card. The renewal spread prices the space incumbents already occupy, which is the only negotiation that happens at scale this year, and that is where the repricing shows up: +25.4% is what existing tenants, with their configured halls and their cross-connect meshes, actually agreed to pay rather than move (and with 0.24% vacancy, moving is not an option). JLL's trend line confirms the mechanism over years: North American rents up nearly 70% since 2020, averaging about 9% a year, with the firm expecting "this upward pricing momentum to hold firm through at least 2030."

The practical consequence for anyone signing a term in this cycle, in our view: budget your renewal at the operator-disclosed spread, not at the broker's asking band. CBRE's H1 2026 asking growth of 4–8% by deal size is what new capacity opens at; the 25% cash-basis renewal repricing at the largest US colocation operator is what your own lease renewal will negotiate against. The asking band is the brochure; the renewal spread is the receipt.

What is the arithmetic the quoted rate hides?

Utilization — and it is the biggest lever on effective cost that never appears on the quote. Rent is charged on committed kilowatts whether or not you draw them, so the price you actually pay per consumed unit of energy depends on how much of your reservation you use. Run the number (ours; assumptions stated): take 1 MW of committed capacity in Northern Virginia at the midpoint of CBRE's 10+ MW band, $172.5 per kW-month — that is $172,500 per month of base rent, before metered electricity and before any cross-connects. At 60% average utilization the facility actually consumes 1,000 kW × 0.6 × 730 hours = 438,000 kWh a month, so base rent alone works out to about $0.39 per consumed kWh. At 90% utilization it is 657,000 kWh and about $0.26 per kWh. Nothing about the contract changed — the quoted rate is identical — and the effective cost per unit of consumed energy fell by a third.

This is the quiet reason the wholesale market's biggest tenants run their halls hot and their utilization models harder. It is also why the per-kW rent and the per-kWh electricity bill cannot be compared as if they were the same kind of charge: the rent is an option on capacity, the electricity is a charge on consumption. A buyer comparing two quotes by headline $/kW while ignoring both utilization assumptions and the power pass-through is comparing brochures. The clean discipline — and the one thing every source we checked agrees on — is to make each provider split the quote into base rent per kW, metered power, and interconnection, then model your own utilization on top. Kurums' benchmarking note says exactly this; Digital Realty's reporting structure enforces it by construction; GoDataCenters' definition of the unit implies it.

What is the 2026 reality check?

The market's tightness is not a mood; it is inventory arithmetic. CBRE measured Northern Virginia at 0.24% colocation vacancy in H1 2026 with 10.8 MW available; JLL measured North America at 1% vacancy for the third consecutive year, with 25 GW absorbed in six months — double the year-earlier level and five times the level of two years ago — and 66 GW under construction, 95% of it pre-committed. CBRE's own numbers agree in kind: its H1 2026 report states all capacity delivering in 2026 in Northern Virginia is already committed, pushing leasing activity into 2027 and beyond, and (per GoDataCenters' compilation) 80.4% of the under-construction pipeline is preleased. Buyers "are often contracting for 2028 deliveries," per JLL. When the pipeline is pre-sold years ahead, asking rents stop being a discovery mechanism at all — there is nothing on the shelf to discover a price for.

The trend projections deserve the same skepticism every extrapolation deserves. JLL expects roughly 9% annual rent growth to hold through 2030; CBRE's Global Data Center Trends 2026 is titled around supply scarcity through 2030 driving pricing "to unprecedented highs." But the same record contains its own stress test: CBRE's Q1 2026 data shows Querétaro, Mexico — where hyperscale supply landed faster than tenants — swinging from 0.9% to 10.6% vacancy in a year (reported by Kurums from CBRE's data), with the new supply outpacing absorption. A 66 GW pipeline that is 95% pre-committed is a bet that the tenants' own demand forecasts hold; if AI capex pauses, the first place it shows up is renewal spreads, which — unlike asking bands — reprice every quarter. We hold the structural view (power-constrained primary markets stay tight for years) and the mechanical view (watch the renewal spread, not the headline) in the same head, and we think the second one is the more falsifiable of the two, which makes it the better metric to track.

And the transferable takeaway for readers who will never sign a colo lease: the structure generalizes. Any market where capacity is contracted years ahead, vacancy rounds to zero, and the quoted price is an opening band — GPU hours, top-of-rack power, even model API reservations — will show the same pattern this piece documented in the published record: headline rates move slowly, signed renewals move fast, and the buyer's real lever is utilization of what they committed to. The 88% small-block premium, the 25% renewal spread, and the $0.39-versus-$0.26 utilization gap are three views of one fact: in a capacity-scarce market, how you buy matters as much as what you pay.