Hyperscale vs Colocation: How Big Tech Builds Data Centers

Hyperscale means a company builds and fully owns a campus of 40 MW or more, at $11-20 million per MW of construction cost, in exchange for full control and a lower long-run unit cost. Colocation means a company leases capacity in a facility someone else built and owns, at roughly $196/kW/month wholesale, in exchange for speed and no construction risk. In 2026, every major cloud provider uses both models simultaneously, shifting the mix based on how fast they need capacity and how much balance-sheet risk they’re willing to carry.

Key takeaways

• Hyperscale threshold: ~40 MW and 5,000+ servers on a single site, per Synergy Research Group’s industry-standard definition. Hyperscale operators ran about 1,360 qualifying facilities worldwide at the end of 2025, holding 48% of global data center capacity.
• Build cost: hyperscale shell-and-core construction runs $11-14M/MW in 2026, rising to $15-20M/MW for AI-optimized liquid-cooled halls — before tech fit-out, which can add up to $25M/MW for AI infrastructure.
• Lease cost: wholesale colocation averages $196.25/kW/month ($2,355/kW/year) in primary North American markets, per CBRE H2 2025 — a record, up 6.6% year over year.
• Timeline gap is the real driver: self-built hyperscale campuses take 24-36 months of construction and 3-6 years end-to-end when new grid capacity is needed; colocation leases can go live in 6-18 months in facilities already under construction.
• Lease commitments are ballooning: aggregate hyperscaler future lease obligations passed $850 billion by mid-2026, up 63% year over year, led by Oracle ($250B), Microsoft ($197B), and Meta (~$183B), per Bloomberg’s filing analysis.
• No company picks one model exclusively. Meta leans hardest toward self-built AI campuses (Hyperion, Prometheus); Microsoft and Google blend owned regions with leased colocation capacity to hit AI capacity targets faster than they can permit and build alone.
• AI is compressing the decision window. McKinsey projects AI will account for roughly 50% of data center capacity demand by 2030, out of 219 GW in total projected demand — a growth curve no single builder can satisfy with owned construction alone.

For live market benchmarks, see the colocation price index, the global data center catalog, and current GPU rental prices.

What “hyperscale” actually means

The term gets used loosely in marketing, but the industry-standard definition is specific. Synergy Research Group and most analyst firms define a hyperscale data center as a single-operator facility with roughly 40 MW or more of critical power capacity and 5,000+ servers across 10,000+ square feet. The operator owns or fully controls the facility — there is no multi-tenant leasing inside it.

By that threshold, Synergy counted about 1,360 hyperscale data centers operating worldwide at the end of 2025, up from 1,136 a year earlier. Those facilities hold 48% of all data center capacity on the planet today, and Synergy projects hyperscale operators will control 67% of global capacity by 2031 — a share that keeps rising because hyperscale campuses are both multiplying and growing larger with each new build.

Colocation, by contrast, is defined by ownership structure rather than size: a third-party operator (Equinix, Digital Realty, STACK, Vantage, and hundreds of regional players) builds and runs the facility, then rents capacity — by the rack, by the cage, or by the multi-megawatt block — to any number of tenants. A colocation facility can be small (a few racks) or enormous (100+ MW, effectively hyperscale-sized); what makes it colocation is that the operator, not the tenant, owns the building and sells capacity to multiple customers, sometimes including the hyperscalers themselves.

Cost: building versus leasing

Model	Unit cost (2026)	What it covers
Hyperscale build-to-suit, shell-and-core	$11-14M/MW	Land, power infrastructure, building shell, base mechanical/electrical — before servers
Hyperscale build, AI-optimized (liquid-cooled)	$15-20M/MW	Higher-density power and cooling for GPU halls
Tech fit-out (AI infrastructure)	Up to $25M/MW additional	Servers, GPUs, networking, liquid-cooling loops
Wholesale colocation lease, primary US markets	$196.25/kW/month (~$2,355/kW/year)	Space, committed power, cooling, security — turnkey
Wholesale colocation, Ashburn (tightest market)	$215+/kW/month	Same, in the most supply-constrained primary market
Wholesale colocation, Singapore	$330-475/kW/month	Same, under APAC’s most rationed capacity market

Sources: JLL 2026 Global Data Center Market Outlook (construction costs); CBRE North America Data Center Trends H2 2025 (lease rates).

The comparison is not apples to apples. A build is a capital expenditure that becomes a depreciating owned asset and, over a 15-20 year facility life, typically produces a lower cost per kW than continuously renewing colocation leases. A lease is an operating expense with no construction risk, available on a timeline colocation build cycles allow rather than a company’s own permitting and procurement queue. Hyperscalers run the math on both and increasingly conclude that speed is worth the premium — which is why lease commitments are growing faster than owned-campus announcements.

Timeline: the actual reason the mix is shifting

Cost alone doesn’t explain why hyperscalers are leasing more colocation capacity even as they announce record capex. The deciding factor is time.

Path	Typical timeline
Self-built hyperscale campus, construction only	24-36 months
Self-built campus, full cycle (site selection, permitting, grid interconnection)	3-6 years when new transmission capacity is required
New high-voltage transmission/generation capacity alone	24-48+ months
Colocation lease in a facility under construction	12-24 months to delivery
Colocation lease in standing or near-complete inventory	6-18 months to live capacity

Sources: JLL 2026 Global Data Center Market Outlook; Data Center Frontier construction-timeline coverage.

AI demand is compounding on a timeline shorter than most grid-interconnection queues. A hyperscaler that identifies a capacity gap today cannot permit and energize a new substation in time to meet a 2027 training run — but a colocation operator who started site work two years ago, anticipating exactly this demand, can offer power now. That asymmetry is the core reason lease commitments hit $850 billion by mid-2026, a 63% jump from a year earlier, even as Meta, Microsoft, and Google simultaneously announce tens of billions in new self-built campuses.

How the four largest cloud providers actually split it

• Meta has leaned hardest toward owning AI-first infrastructure outright. Its Hyperion and Prometheus campuses pair on-site or behind-the-meter power generation (two 200 MW sites building toward 400 MW combined) with fully owned compute halls — a bet that vertical control over power and design pays off at Meta’s scale. Meta still leases for general-purpose cloud capacity and added roughly $79 billion in new lease commitments in a single 2026 quarter, a 76% jump, showing the owned/leased split moves quarter to quarter even at the most build-focused hyperscaler.
• Microsoft runs the more balanced model: self-owned, high-density AI compute hubs at core sites, plus leased capacity in strategic markets to fill gaps faster than its own construction pipeline allows. Microsoft’s disclosed lease commitments sat near $197 billion after adding $41 billion in a recent quarter.
• Google leases more than either Meta or Microsoft relative to its footprint — it doesn’t always own the buildings it operates from, sometimes renting a facility a third party owns and runs power and cooling for, while retaining control over the server and networking layer inside.
• Oracle leads all hyperscalers in aggregate future lease commitments at roughly $250 billion, reflecting its later, faster entry into AI cloud capacity — leasing is the only way to scale a cloud footprint on a multi-year-compressed timeline.

This is consistent with the broader pattern Data Center Frontier and other analysts describe: hyperscale and colocation capacity are growing together, not substituting for each other, because AI demand outstrips what any single construction pipeline can deliver.

Decision framework: when each model fits

Build (hyperscale campus) when:

• Your capacity need is durable and large enough (100+ MW) to amortize the higher unit construction cost over 15+ years.
• You need non-standard power or cooling design (behind-the-meter generation, custom liquid-cooling loops) that a multi-tenant operator won’t build to your spec.
• You have 3+ years of runway before capacity is needed and can absorb permitting and interconnection risk.
• Vertical control over site security, power procurement, or expansion rights matters strategically, not just financially.

Lease (colocation) when:

• You need capacity inside 6-24 months, faster than your own construction and interconnection timeline allows.
• Demand is uncertain enough that a depreciating owned asset is a worse bet than a cancellable or renegotiable lease.
• You need geographic diversity (multiple markets, multiple countries) faster than building a campus in each one.
• Capital discipline matters more than long-run unit cost — leasing keeps the obligation off the balance sheet as capex, even though the multi-year rent commitment shows up in disclosures.

Most hyperscalers run both simultaneously, using owned campuses for the durable core of their footprint and colocation leases as the flexible edge that absorbs demand spikes and speed-to-market pressure. Track live lease-rate benchmarks by market in our colocation price index or compare specific facilities in the global data center catalog.

What this means for buyers below hyperscale

If you’re not negotiating at gigawatt scale, the build-vs-lease calculus still applies in miniature: a large enterprise deploying multiple megawatts still faces the same trade-off between colocation speed and self-built control, just at a fraction of the size. For nearly all deployments under roughly 20-40 MW, colocation dominates — the construction economics of a self-built hyperscale campus simply don’t work below the volume threshold that makes $11-20M/MW capex pay for itself.

The practical takeaway: benchmark any wholesale colocation quote against the price index before signing, request our quote comparison service if you’re evaluating multiple markets, and if you’re weighing a build-to-suit or powered-shell deal instead of a standard lease, see our guide on powered shell vs turnkey development models for how those contracts differ from either pure model. For due-diligence steps before committing to either path, see our data center due diligence checklist.