.webp)
.svg)
Previous Post
The build-side economics of AI infrastructure
How compute capacity gets built and what enterprise buyers should understand about supply. Most AI infrastructure coverage focuses on what enterprises rent: per-GPU-hour pricing across neoclouds and hyperscalers. How that capacity gets built in the first place gets less attention, even though it's where most cost variation lives and where most supply-side decisions that affect enterprise buyers ultimately come from. For enterprise buyers evaluating whether to partner with an infrastructure provider, run workloads on a hyperscaler, or build internal capacity, the underlying economics matter. They shape what a provider can charge sustainably and the workloads they can support over time.
Regional construction costs vary widely
Data center construction costs are not uniform, even within regions. In APAC markets, construction costs range from $7.9 million per megawatt at the low end to $19.2 million per megawatt at the high end, more than a 2x spread inside a single region.¹ The global average ran $11.3 million per megawatt in 2026, up from $10.7 million in 2025, a 6% year-over-year increase driven by AI-specific design requirements like denser power and cooling.²
The North American market sees its own variation. The US holds 15.9 GW of the 23.1 GW of global IT capacity under active construction as of late 2025, but established hubs are tightening: Northern Virginia construction fell 29% in 2025 as power and land constraints became binding, and Hillsboro and Silicon Valley each contracted 14-15%. Nearly two-thirds of new North American capacity is now being developed outside legacy markets.³
European builds face higher power and compliance costs but also operate in a sovereignty-driven procurement environment. The EU AI Act and data residency requirements add scope to European builds in ways that can push costs above the global average.⁴
A provider's headline rate per GPU-hour reflects, in part, where their physical infrastructure was built and at what cost. A neocloud with most of its capacity in lower-cost regions can sustain lower pricing without margin compression, while one built primarily in the most expensive metros has fewer options.
GPU supply is allocation-bound
Established shortly after ChatGPT’s launch, with the support of Wistron, Foxconn, and Pegatron, Zettabyte emerged to combine the world’s leading GPU and data center supply chain with a sovereign-grade, neutral software stack.
Established shortly after ChatGPT’s launch, with the support of Wistron, Foxconn, and Pegatron, Zettabyte emerged to combine the world’s leading GPU and data center supply chain with a sovereign-grade, neutral software stack.
The hardware itself is the second constraint, and the GPU market in 2026 is structurally tight.
Lead times for NVIDIA H100 and H200 deployments are running 36-52 weeks.⁵ The constraint is upstream of NVIDIA: TSMC's CoWoS (Chip on Wafer on Substrate) packaging capacity is fully allocated through at least mid-2027, and HBM3e memory production at SK Hynix and Micron isn't meaningfully expanding before late 2026.⁵
Microsoft, Google, Meta, and Amazon placed multi-billion-dollar forward orders for Blackwell GPUs (GB200, B200) in 2025, consuming most of NVIDIA's available allocation capacity through the end of 2026 and into 2027.⁵ Industry analysts describe enterprise buyers as the lowest priority tier for NVIDIA's allocation decisions, behind hyperscalers and strategic regional partners.⁶ The practical effect is longer lead times and weakened pricing leverage, with 30-50% premiums on top-tier GPUs for buyers without preferred relationships.⁷
Procurement at this scale stops being a price-sheet exercise. Analyst guidance for enterprise buyers in 2026 consistently emphasizes the supplier relationship dimension: cultivating preferred customer status with OEMs and hyperscalers, locking multi-year framework agreements with fixed pricing and delivery terms, and treating queue position as a strategic asset.⁷
For infrastructure providers, the equivalent dynamic plays out one layer upstream. Providers with established supply relationships across multiple OEMs and system integrators can typically source hardware on better terms than providers buying through open market channels, because delivery speed and unit pricing both improve when the relationships are already there.
Power and government policy
Power is the third constraint, and arguably the dominant one in 2026. Across major US Independent System Operators, grid interconnection approval timelines now run five to seven years, compared to a previous average of roughly four.⁹ CenterPoint Energy in Texas reported a 700% increase in large load requests between late 2023 and late 2024, jumping from 1 GW to 8 GW.⁹ In the EU, wait times for securing a grid connection range from two to ten years depending on jurisdiction, with the Netherlands averaging ten years in key markets.¹⁰
The mismatch between build pace and power delivery is the constraint most likely to delay a project in 2026. Of 12 GW of US data center capacity announced for 2026, only 5 GW is under construction; 11 GW sits in the "announced" stage with no physical progress, and 25% of those projects have not disclosed a power strategy at all.¹¹ Roughly half of all global projects face delays attributable to power limitations and grid equipment shortages, with high-voltage transformer lead times stretching from a pre-2020 baseline of 24-30 months to roughly 5 years in 2026.¹¹
Government policy is moving in two directions at once: addressing the grid bottleneck and competing to attract investment. India's 2026 Union Budget introduced a 20-year tax holiday for data centers serving global clients, exempting foreign operators from taxes on overseas services through 2047.¹² Several US states (Indiana, Washington, Virginia, Iowa) offer sales and use tax exemptions on server equipment, power infrastructure, and construction inputs, with thresholds typically requiring $10M+ in investment and minimum lease terms of five years or more.¹³ The US Department of Energy issued a directive in October 2025 proposing federal jurisdiction over new loads above 20 megawatts to standardize interconnection procedures for projects accepting operational flexibility.⁹
For buyers, a provider's effective cost basis reflects more than construction costs and hardware sourcing. The policy environment of the markets where a provider operates also shapes their economics through tax holidays, sales and use exemptions, accelerated permitting, and power purchase agreements locked in years ago. A provider operating in a jurisdiction with a 20-year tax holiday and abundant industrial power has a meaningfully different cost structure from one operating in a power-constrained Tier 1 metro.
What this means for the build vs rent decision
Established shortly after ChatGPT’s launch, with the support of Wistron, Foxconn, and Pegatron, Zettabyte emerged to combine the world’s leading GPU and data center supply chain with a sovereign-grade, neutral software stack.
Established shortly after ChatGPT’s launch, with the support of Wistron, Foxconn, and Pegatron, Zettabyte emerged to combine the world’s leading GPU and data center supply chain with a sovereign-grade, neutral software stack.
For enterprise buyers weighing whether to build internal capacity, partner with an infrastructure provider, or rent from hyperscalers, the underlying numbers have changed in 2026.
For high-utilization workloads, owned infrastructure breaks even against equivalent hyperscale cloud instances in under four months, with up to an 18x cost advantage per million tokens compared to Model-as-a-Service APIs over a five-year lifecycle.⁸ The breakeven assumes the buyer can source the hardware and complete the build, which is where the supply and power dynamics above start mattering.
Workload predictability. Sustained, high-throughput inference favors ownership because the per-token economics compound. Bursty experimentation favors rental because utilization stays low and capital tied up in idle hardware is hard to justify.
Capital availability. AI infrastructure builds run in the tens of millions per megawatt of capacity, and the timeline from contract to operational compute is now measured in years for greenfield projects. Buyers without that capital or that timeline are effectively in the rental market regardless of the TCO math.
Geographic and compliance constraints. Where workloads can run is increasingly constrained by data residency rules and export controls, alongside broader sovereignty requirements. For a buyer with strict residency requirements, the rental market has fewer options, and partnering with a regional infrastructure provider is often more practical than either pure rental or self-build.
Operational capacity. Owning compute means operating it, which means having or hiring engineering teams that can run a production-grade infrastructure stack. The cost of that team is often underestimated in build-side projections.
For most enterprise buyers, the practical answer is a partner relationship with an infrastructure provider whose regional footprint and supply chain depth are already in place, with operational maturity to match. The cost savings flow from the provider's underlying economics, because negotiating leverage on a published price sheet rarely moves the number much.
What this looks like from where we sit
For an AI infrastructure provider, the underlying economics determine what's possible for customers. The locations where capacity gets built and the hardware sourcing channels available to a provider both feed directly into the rates they can sustainably offer.
Zettabyte's infrastructure operations are anchored in regions with favorable construction and operational costs, and the supply relationships across hardware vendors and system integrators have been built over multiple build cycles. That combination is what makes the rates we can offer customers work.
What buyers should evaluate
A short framework for evaluating an AI infrastructure provider's underlying position:
What's the regional footprint? Providers concentrated in the most expensive metros face structural cost pressure. Providers with regional diversity can route workloads to lower-cost capacity where compliance allows.
What's the supply chain depth? Single-source providers are exposed when allocation tightens. Providers with established OEM relationships and integrator partnerships tend to have more flexibility when supply gets constrained.
What's the lead time on incremental capacity? A provider quoting 36-52 weeks for new GPU capacity is matching market reality. A provider quoting materially better is either pulling forward existing inventory, has a privileged supply relationship, or is being optimistic about queue position.
What's the cost transparency? Providers should be able to explain how their cost structure differs from hyperscaler list pricing, and which factors drive the difference.
What's the operational track record? Sourcing and deployment differ from operating at scale. References from production customers running sustained workloads matter more than marketing claims about supply chain depth.
Providers worth working with should be able to give specific answers across all five dimensions. The level of detail in those answers is usually a good signal of how well-developed their underlying position is.
Zettabyte builds AI infrastructure for enterprises, sovereigns, and AI-native companies. If the underlying economics matter to your team's evaluation, reach out: https://www.zettabyte.space/contact.
Sources
- Cushman & Wakefield. Asia Pacific Data Centre Construction Cost Guide 2026. https://www.cushmanwakefield.com/en/insights/apac-data-centre-construction-cost-guide
- JLL. 2026 Global Data Center Market Outlook. https://www.jll.com/en-us/insights/market-outlook/data-center-outlook
- Archdesk. 2026 Global AI Data Center Construction. https://archdesk.com/blog/global-ai-data-center-construction-2026
- Data Center Knowledge. New Data Center Developments: April 2026. https://www.datacenterknowledge.com/data-center-construction/new-data-center-developments-april-2026
- Spheron. GPU Shortage 2026. https://www.spheron.network/blog/gpu-shortage-2026/
- Network World. US approves Nvidia H200 exports to China. December 2025. https://www.networkworld.com/article/4103157/
- Uvation. H100 Availability: The Silent Crisis. https://uvation.com/articles/h100-availability-the-silent-crisis-threatening-enterprise-ai-plans
- Lenovo. On-Premise vs Cloud: Generative AI TCO, 2026 Edition. https://lenovopress.lenovo.com/lp2368-on-premise-vs-cloud-generative-ai-total-cost-of-ownership-2026-edition
- Hanwha Data Centers. Data Center Time to Power. March 2026. https://www.hanwhadatacenters.com/blog/data-center-time-to-power-how-campuses-speed-ai-deployment/
- Enlit World. AI growth vs power availability. 2026. https://www.enlit.world/library/ai-growth-vs-power-availability-a-perfect-storm-in-2026
- Tech Investments. Power Bottlenecks & The AI Data Center. 2026. https://www.techinvestments.io/p/power-bottlenecks-and-the-ai-data
- Bloomberg via Yahoo Finance. India Unveils 20-Year Tax Break. February 2026. https://finance.yahoo.com/news/india-unveils-20-tax-break-084140237.html
- Data Center Knowledge. U.S. Data Center Tax Incentives. September 2025. https://www.datacenterknowledge.com/regulations/u-s-data-center-tax-incentives-a-special-report
