Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.
The debate over hyperscaler self-build has consumed an enormous amount of analytical energy over the past eighteen months.
It is, for the most part, the wrong debate.
The question is not whether colocation survives alongside hyperscaler-owned campuses.
It does.
The question is which segment of the colocation market survives at premium economics and which segment begins, quietly, to compress.
The answer has significant consequences for every fund with exposure to digital infrastructure assets.
What the market currently believes
The dominant read is structural coexistence. Self-build grows. Colocation grows. Both are absorbing a demand wave too large for any single model to capture.
The data supports this at the aggregate level: ~$52.4 billion global wholesale colocation market in 2025, projected at ~$118.6 billion by 2034. Vacancy in North America’s primary markets fell to 1% for the second consecutive year at end-2025. JLL projects colocation at a ~20% CAGR through 2030.
Against numbers like these, the threat to the colocation model appears minimal.
This read is not wrong. It is incomplete.
The mechanism the consensus underweights
Colocation is not a single asset class. It never was. But the AI infrastructure buildout is now large enough to make the internal distinctions matter at the portfolio level.
Consider what hyperscalers are actually doing with their procurement.
Microsoft’s cancellation of approximately 2 GW of planned US and European leases in 2025 was widely interpreted as a self-build pivot.
The more precise read is narrower: Microsoft exited early-stage agreements in markets where standard leased capacity could be replicated by self-build. It continued leasing in markets where time-to-power was the binding constraint. The distinction is important. The cancellations were not a statement about colocation. They were a statement about undifferentiated colocation.
CyrusOne’s Q4 2025 earnings confirm the directional trend. Wholesale deals comprised 72% of bookings, with average sizes above 20 MW and terms averaging approximately 12 years.
Hyperscalers are committing long and committing large. But a 12-year fixed-rate wholesale contract signed at today’s compressed vacancy premiums is not necessarily the asset a fund underwrites at current multiples. Over a 12-year term, the negotiating leverage shifts. Operators whose product is replicable will face renewal discussions on the hyperscaler’s terms.
The counterargument deserves direct engagement.
At 1–2% vacancy, every operator has pricing power today. That is factually correct. The bifurcation thesis is directional, not immediate.
But infrastructure funds do not underwrite assets against today’s vacancy, they underwrite against the cash flows that materialize over a 10–15 year hold. Pricing the future correctly requires identifying which assets carry structural protection and which carry structural exposure.
What is structurally protected
Two sub-segments within the colocation market are structurally protected from hyperscaler margin compression, and the protection mechanism in each case is the same: their product cannot be replicated with capital alone.
The first is interconnection-anchored colocation.
Equinix’s IBX campuses are network-effect ecosystems with over 10,000 co-located customers. No hyperscaler campus can replicate this. A self-built AWS facility cannot offer a customer simultaneous private peering to Azure, Google Cloud, and 50 carriers. The interconnection fabric is not real estate.
Equinix’s reported xScale utilization exceeded 90% across Northern Virginia, Frankfurt, and Singapore through 2025, and the $15 billion joint venture with GIC and CPP Investments closed in February 2026 to fund additional hyperscale campuses in Tokyo, Sydney, and Frankfurt at 10–15 year wholesale terms. Interconnection density generates platform-effect pricing power that square footage cannot.
The second is BTS and powered-shell operators controlling pre-entitled, power-ready land. This is the structural insight that the self-build vs. colocation debate consistently obscures.
NVIDIA’s current DGX Blackwell systems draw approximately 120 kW per rack, roughly 10 to 20 times the power density of traditional enterprise infrastructure. No standard multi-tenant colocation hall accommodates this. BTS and powered-shell structures, delivering the structural shell, grid connection, and completed permitting stack without prescribing the interior, are the dominant procurement model for new large-format AI workloads.
The scarcity these operators control is not capital. Capital is abundant. Grid connections in Northern Virginia now require 24–36 months of advance planning. In 2025, 25 planned data center projects were cancelled due to community and regulatory opposition, with the rate quadrupling from 2024 levels.
Operators who hold completed entitlements are selling time-to-power. That is the genuinely scarce resource in this market. No capital, regardless of its cost, can recover the time already spent in the permitting queue.
The $30 billion Blue Owl–Meta joint venture in 2025, funding a ~2 GW campus in Louisiana, pre-leased by Meta and potentially scaling to ~5 GW, illustrates where premium returns are being structured.
Developer capital. Hyperscaler tenant credit. Pre-entitled land. Long-duration lease.
This is not a colocation trade in the traditional sense. It is a developer-operator model that has been reshaped by AI-scale demand
The counterargument
The strongest version of the opposing view is this: at current vacancy levels, even commodity colo operators are effectively setting prices. Bifurcation is a 2028–2030 concern, not a 2025–2026 underwriting consideration. Funds with 5-year hold periods may never see the compression.
This is a reasonable objection for certain hold structures. It does not address the valuation question. If commodity multi-tenant colocation assets are currently priced at multiples that reflect interconnection-anchored or BTS-equivalent structural durability, the pricing error is embedded in today’s acquisition, not in the future cash flows. The correction, when it comes, will be felt at exit which is precisely when hold period structures expose the mismatch.
The forward implication
Funds evaluating colocation assets now need to make a sub-segment call before applying sector comps. Three questions determine which side of the bifurcation an asset sits on.
Does the operator control pre-entitled, power-ready land in a constrained market? If not, how long would it take a well-capitalized competitor to replicate the product?
Is the asset interconnection-anchored? Is co-location density and network-effect presence a structural feature of the asset, or incidental to it?
What is the BTS and powered-shell capability of the operator? Can the operator design and deliver dedicated single-tenant campuses to hyperscaler specifications, or is the product standardized multi-tenant square footage?
Assets that score well on these three questions belong in one category. Assets that do not belong in another. Both categories exist in today’s market. Both are being underwritten with similar multiples. That is the mispricing.
The open question
As BTS and powered-shell become the dominant procurement model for new large-format hyperscaler demand, does the traditional multi-tenant retail colocation model retain pricing power at the enterprise and mid-market level or does hyperscaler demand pressure on one end and private infrastructure buildout on the other eventually compress the entire undifferentiated middle?
Totally agree! The rise of self-build options is definitely shaking things up. But instead of killing colocation, it’s pushing for more innovation and specialization. It’ll be interesting to see how providers
I remember when I first got into the data center game, it felt like a gold rush. Now with the self-build trend, it's like watching two worlds collide. Some love the freedom of DIY, while others swear by colocation's reliability. Can't wait to see how this evolves!