Build, Buy, or Colocate in 2026: How the Math Has Actually Changed

Amazon and Meta are looking at the same market in 2026 and making opposite bets. Meta added $79 billion in new data center lease commitments this year alone. Amazon added just $10 billion, preferring to keep building and owning its own facilities. Both companies have access to the same capital, the same vendors, and the same power-constrained market — and they’ve reached different conclusions about whether to build, buy, or colocate. If the two most sophisticated infrastructure operators on the planet don’t agree, it’s worth working through the actual numbers before assuming the answer is obvious for your own organization.

What it costs to build, in real 2026 numbers

Shell-and-core construction now averages $11.3 million per megawatt globally, ranging from $6.5 million/MW in India to $16 million/MW in Singapore. That’s before you add anything AI-specific. An AI-optimized, liquid-cooled facility runs $20 million or more per megawatt for the building alone, and once you add the GPU fit-out itself, an all-in build lands at $30–40 million per megawatt. Hyperscalers collectively are guiding to $635–670 billion in combined capex for 2026, with roughly $240 billion of that going specifically to physical infrastructure rather than chips.

Those are numbers only a handful of organizations on earth can absorb as an ongoing line item — which is exactly why the build-vs-lease decision splits so sharply once you’re not one of the four or five companies capable of writing a nine-figure check without blinking.

What it costs in time — and this is where building really hurts

Cost is only half the picture. Timeline risk has gotten measurably worse, not better, over the last two years. Average hyperscale build time is now 18–24 months, up from roughly 12 months before 2025. The full development cycle including permitting runs 3–6 years. And grid connection — the single biggest bottleneck discussed across every part of the industry right now — runs 4–10 years depending on region, with London facing an 8-year queue and Amsterdam a 10-year queue.

Even once you clear permitting and power, the equipment itself has its own lead-time problem. Large power transformers now take 128 weeks to arrive from order. Generator step-up units take 144 weeks. Switchgear runs 45–80 weeks. Average equipment lead time across the board sits at 33 weeks — roughly 50% longer than before 2020. Sightline Climate’s tracking found that of the 16 GW of AI data center capacity announced for delivery in 2026, only about 5 GW — roughly one in three announced megawatts — actually had a spade in the ground. Power availability and permitting, not construction speed, are what actually determine whether an announced project ships on schedule.

A framework that actually holds up

The clearest decision framework treats this as a function of scale and time horizon rather than a philosophical build-versus-lease debate. Colocation is the right default for near-term AI capacity needs — 0 to 3 years out — particularly for requirements in the 1–5 MW range where a provider can deliver capacity within 6–12 months. Own-build only starts to make sense once you can commit to 10+ MW of sustained demand and can absorb a 24–36 month lead time before that capacity is usable. Workload profile matters too: inference and user-facing applications, which need to be geographically distributed and responsive, tend to fit colocation well; heavy training and batch processing, which can tolerate more centralization, are where own-build’s economics improve. Specialized compliance requirements (FedRAMP, StateRAMP, CJIS) can sometimes be met through purpose-built colo, but in some cases require the level of control only ownership provides.

The other useful discipline is normalizing any comparison to $/GPU/month and $/MWh actually delivered to the GPU, rather than comparing headline lease rates to headline construction costs — a raw per-kW number hides meaningful differences in power efficiency, cross-connect fees, and remote-hands costs that only show up once you do the full unit-economics math.

The real-world test playing out right now

2026 has turned into an unplanned natural experiment on this exact question. US technology companies committed $850 billion in aggregate data center lease obligations in Q1 2026 alone — up 204% year-over-year and 31% quarter-over-quarter. Oracle now holds roughly $250 billion in future lease obligations, the largest of any company. Meta added $79 billion in new lease commitments this year (a 76% quarter-over-quarter jump), pushing its total to roughly $183 billion. Microsoft added $41 billion, bringing its total to about $197 billion. All three are leasing aggressively rather than building, largely because Northern Virginia vacancy fell to just 0.3% in Q1 2026 and locking in space now — even at a leased premium — beats waiting years for owned capacity to come online.

Amazon has taken the opposite path, adding only $10 billion in new leases and continuing to build and own roughly 24 million square feet of its own data center space (leasing a similar amount separately), on the stated logic that owned infrastructure is cheaper over a facility’s full 20-year lifespan. Alphabet is pacing its own commitments more conservatively, tying capacity growth to visible cloud revenue growth rather than pre-committing years ahead.

The market has already rendered a verdict on the aggressive leasers, at least so far this year: Oracle is down 28% year-to-date, Microsoft down 17%, and Meta down 10% — while Amazon is up 7% and Alphabet up 14%. That’s not proof either strategy is objectively correct, but it’s a clear signal that capital markets are actively penalizing balance-sheet risk from aggressive leasing commitments right now, which is worth knowing before you assume “just lease it, that’s what everyone’s doing” is a strategy without its own risk.

What this means if you’re not a hyperscaler

Almost none of this changes the practical answer for a mid-market buyer evaluating 1–5 MW of capacity on a normal enterprise timeline: colocation remains the pragmatic default, not because building is a bad idea in principle, but because the timeline risk of building — a 4–10 year grid queue, 33-week average equipment lead times, an 18–24 month build cycle even once power is secured — is simply not a risk most organizations can absorb, regardless of how the per-MW cost math compares on paper. The record-tight colocation pricing covered in our 2026 pricing analysis is real and it’s frustrating, but it’s still faster and less risky than the alternative for the vast majority of buyers reading this. Own-build only becomes a serious option once you’re planning 10+ MW of sustained, multi-year demand and can absorb multi-year timeline uncertainty as a cost of doing business — which is a decision closer to Amazon’s playbook than Oracle’s.

This exact framework — including the specific cost, timeline, and workload factors that should drive your build-vs-buy-vs-colocate decision — is the opening chapter of the Data Center Buyer’s Guide, precisely because it’s the decision every other chapter assumes you’ve already made correctly.

Written from 15+ years running data center design, operations, and project management.

Sources

• Gain America — Colocation vs. Own-Build: AI Data Center Strategy for Enterprises in 2026
• Archdesk — AI Data Center Construction 2026: Capex, Cost per MW, Delays
• Tech Insider — Data Center Leases Hit $850B: Meta, Microsoft Lead (202

Written by

Raajeev Ratra

Data Center Infrastructure Expert | 15+ Years in DC Design, Operations & Project Management

Raajeev is a seasoned data center professional with hands-on experience in hyperscale facilities, colocation design, power & cooling infrastructure, and global DC operations. He shares practical insights to help engineers and IT leaders build better infrastructure.

Connect on LinkedIn →