Big tech committed roughly $400 billion to AI infrastructure in 2025, with another $650 billion lined up for 2026. But Sightline Climate’s tracker shows only about 5 GW of the 16 GW pipeline for 2026 is actually under construction. The gap between announcement and shovel is now the biggest signal for anyone building, financing, or investing in data center capacity.
Big tech’s combined capex hit roughly $400 billion in 2025, and Amazon, Microsoft, Meta, and Google have signalled a jump to around $650 billion next year. Bloomberg pegs the 2026 number at a 60% YoY increase. The press releases keep coming. The cranes don’t always follow.
For anyone bidding data center work, financing it, or building tech for it, the headline isn’t the spend. It’s the gap between what’s been announced and what’s actually got steel in the ground. That gap now has a number on it, and it’s bigger than most people in the AI conversation are admitting.
$400B was spent in 2025. ~$650B is committed for 2026. Only 5 GW of the 16 GW data center pipeline for 2026 is under construction. Three things are breaking the math: power infrastructure, energy costs, and aggressive depreciation accounting.
Table of Contents
1. The announcement-to-construction gap
2. Power is the real bottleneck, not chips
3. Energy costs are squeezing the operating model
4. The depreciation question nobody wants to answer
5. What this means for contractors and AEC tech
FAQ
The announcement-to-construction gap
16 gigawatts on paper. 5 gigawatts in concrete.
Sightline Climate’s first Data Center Outlook tracked 777 projects above 50 MW announced since January 2024. Their finding: at least 16 GW of capacity is slated to come online globally in 2026 across 140 projects. Only about 5 GW is under construction. The other 11 GW sits at “announced” with no visible site progress, despite typical build timelines of 12 to 18 months.
The 2027 picture is worse. Announced capacity hits 21.5 GW, under construction 6.3 GW. Beyond 2028, around 37 GW lacks a firm completion date.
The Stargate project tells you everything. In March, Oracle and OpenAI scrapped the planned 600 MW expansion of their Abilene, Texas campus from 1.2 GW to 2 GW after financing talks collapsed. The base build still proceeds. The expansion doesn’t. Sightline expects 30-50% of the 2026 pipeline to slip or cancel outright.
It is a pretty wild puzzle at the moment. If one piece of your supply chain is delayed, then your whole project can’t deliver.Sightline Climate analyst, Data Center Outlook 2026
Power is the real bottleneck, not chips
The grid component shortage that’s quietly killing schedules
Anyone procuring large power transformers in the last 24 months knows this already. Wood Mackenzie’s analysis shows lead times for power transformers averaging 128 weeks, generator step-up units at 144 weeks. Demand for substation transformers is up 116% since 2019. GSUs are up 274%.
Transformer prices are up 70 to 150% since 2020. The US imports more than 80% of its large power transformers, and 25% Section 232 tariffs on steel and aluminum (50% in some cases) hit transformer cores, switchgear housings, and structural components directly.
The result: developers are buying everything they can, as soon as they can, even when they have nowhere to put it. The classic bullwhip effect, which goes some way to explaining why hyperscaler capex looks so frantic at the same time so few sites are getting energised.
- Component lead times. Power transformers averaging 128 weeks. GSUs at 144 weeks. Switchgear and batteries similar.
- Tariff layering. Section 232 on steel and aluminum, Section 301 on Chinese components, recent copper actions all stacking onto critical electrical equipment.
- Community resistance. Moratoriums proposed in 10+ states including Maine, Michigan, New York, Ohio, Louisiana, and Virginia. The Maine House passed a moratorium 82-62 to halt large data centers until 2027.
The construction tech briefing your competitors are already reading
The deals, the data, and the deep takes on what’s actually getting built.
Join 2500+ ReadersEnergy costs are squeezing the operating model
What changed when behind-the-meter became the default plan
The grid won’t catch up fast enough, so hyperscalers are building their own. Meta’s Hyperion project in Louisiana is the cleanest example. Three Entergy gas plants approved, plus seven more proposed, totalling 10 plants generating roughly 7.5 GW for one data center. xAI is running gas turbines at its Memphis Colossus campuses. Google has a 933 MW gas plant in North Texas with Crusoe. Microsoft is buying Chevron-backed gas in West Texas.
The economics are getting harder. Wood Mackenzie projects gas turbine prices will rise 195% by end of 2026 versus 2019. US natural gas is up roughly 60% versus 2024. In Virginia, the world’s biggest data center cluster, electricity prices are up 267% over five years. Utilities requested over $29 billion in rate increases in the first half of 2025 alone.
Operators face higher capex (transformer and turbine prices up sharply), longer waits (multi-year lead times on critical gear), higher opex (gas and grid power both more expensive), and harder permitting (community pushback in 10+ states). All at once.
The depreciation question nobody wants to answer
Why Burry’s accusation matters even if his timing is off
In November 2025, Michael Burry of Big Short fame went public with the accounting argument that’s been quietly circulating among analysts. His claim: hyperscalers depreciate AI servers and GPUs over 5 to 6 years when the real economic life is closer to 2 to 3, given Nvidia’s 12 to 18 month product cycle. By his math, that’s $176 billion in understated depreciation across 2026-2028.
Microsoft uses 6-year schedules. Meta uses 5.5. Amazon shortened from 6 to 5 in early 2025. Nvidia issued a detailed memo to Wall Street analysts pushing back. CoreWeave’s CEO has defended the schedules, citing five-year customer contracts and 95% resale values on older A100 chips.
Where this matters for builders: if Burry is even partially right, the 2027-2028 wave of writedowns hits at exactly the moment the 2026 buildouts come online. That changes how serious operators should model the back end of these contracts. Infrastructure being financed today on aggressive depreciation assumptions might face an impairment cycle just as construction finishes.
What this means for contractors and AEC tech
Three plays the smart money is making
This isn’t an argument that data center construction is over. ConstructConnect’s data through November 2025 showed data center starts hit $53.7 billion year-to-date, up 138.6% YoY. The pipeline is just getting messier.
For contractors, Skanska’s playbook for hitting tech-company speed is the relevant reference: pre-qualified vendor pools, fast-track procurement during design, repeat crews who already know the standards.
For ConTech founders, the demand signal isn’t going to where the press releases say it is. MEP engineering capacity, electrical pre-fabrication, transformer logistics, and behind-the-meter power orchestration are where the binding constraints sit. OpenAI’s deal with Turner signals the hyperscalers know construction execution is now the rate-limiter on AI strategy.
For AEC investors, the question to ask any data center-adjacent thesis isn’t “is the demand real?” The demand is real. It’s “what does this look like if 30-50% of announced 2026-2027 capacity slips by 18+ months?” If the answer is “fine, we still win,” that’s a defensible thesis.
| Constraint | Status (April 2026) | Implication for builders |
|---|---|---|
| Large power transformers | 128-week lead time, prices up 70-150% since 2020 | Procure during design, not after permit |
| Generator step-up units | 144-week lead time, demand up 274% since 2019 | Behind-the-meter gas needs equally early procurement |
| Steel and aluminum tariffs | 25-50% Section 232, broader expansion in 2025 | Bid validity windows shrinking, escalation clauses standard |
| Community opposition | Moratoriums proposed in 10+ US states | Site selection now a regulatory risk question |
| GPU depreciation | 5-6 year schedules vs 2-3 year obsolescence | Tenant covenant strength may weaken in 2027-2028 |
| Hyperscaler announcements | $650B committed for 2026, 60% YoY growth | Demand real but spread across more sites than will get built |
FAQ
For 2026, around 5 GW of the 16 GW announced pipeline is under construction globally, per Sightline Climate. For 2027, it’s 6.3 GW out of 21.5 GW announced. Sightline expects 30 to 50% of the 2026 pipeline to be delayed or cancelled. (Source)
The bottleneck has shifted from chips to grid. Power transformers carry 128-week lead times. GSUs sit at 144 weeks. Steel, aluminum, and copper tariffs have layered onto already-tight markets. Community opposition is now driving moratoriums in 10+ US states. (Source)
Oracle and OpenAI dropped the planned expansion from 1.2 GW to 2 GW in March 2026 after financing talks collapsed. The base 1.2 GW build remains on track and Oracle’s broader 4.5 GW agreement with OpenAI is unchanged. Meta is in talks with Crusoe for the unused capacity, with Nvidia having put down a $150M deposit. (Source)
Burry argues hyperscalers depreciate AI hardware over 5-6 years when real economic life is closer to 2-3, given Nvidia’s annual product cycles. He estimates this understates depreciation by roughly $176B between 2026 and 2028. For builders and lessors, the relevance is tenant covenant risk: if writedowns hit during the 2027-2028 buildout completion window, hyperscaler willingness to absorb additional capacity could shift. (Source)
Grid interconnection queues are now multi-year in most major US data center markets. Behind-the-meter gas lets operators bypass the wait. Meta’s Louisiana Hyperion project alone has 10 gas plants approved or proposed, generating 7.5 GW. xAI, Google, and Microsoft are pursuing similar arrangements. The trade-off is climate commitments and rising turbine costs (Wood Mackenzie projects 195% above 2019 prices by year-end 2026). (Source)
Pre-qualified vendor pools, fast-track procurement that begins during design, and repeat specialised crews. Skanska’s approach of paying a premium for relationship-based subs across multiple states is now standard practice. Owner-furnished equipment can be 40-60% of the build, so vendor coordination beats cost discipline as the differentiator. (Source)
Related Articles
How Fast Can You Really Build a Data Center?
Inside OpenAI’s Construction Power Move
