AllStreet Capital · Market Insight
Data Center Pipeline Tracker
Announced → under construction → operational · the demand engine behind AI memory (Micron / HBM / DRAM)
Snapshot as of July 1, 2026 · sources: JLL, CBRE, Goldman Sachs, Sightline Climate, IEA, company filings · ask to refresh and this view updates
Of every 100% of announced AI data-center capacity, only about ~30% is actually under construction and roughly 50–60% ever comes online on schedule. The binding constraint is power, not compute or memory. That gap is the single best real-world governor on whether memory pricing holds or normalizes.
The announced → real funnel · US cohort with 2026 opening dates (Sightline Climate, ~140 projects)
Building · 5 GW
~31% actually under construction — poured concrete, not a press release
Not started · 11 GW
~69% announced with no sign of construction, despite 12–18 mo typical build time
On-time · ~50–60%
Goldman: only ~50–60% of near-term planned capacity lands on schedule · 30–50% delayed or cancelled
Announced
Under construction
Not started
Expected on-time
Global installed base → 2030 (JLL / CBRE)
103 GWOperational now (Q1'26)
~200 GWProjected 2030
14% CAGR2026–2030
New capacity added 2026–2030~100 GW
Under-construction already pre-leased77% demand real
Vacancy (16 largest markets)6.7% tight ↓ from 8.3%
Read: what is actually being built is spoken-for. The vapor sits in the announcement layer, not the construction layer.
Hyperscaler capex 2026 · the money behind demand
Big-4 combined 2026~$725B +77% YoY
Incl. Oracle & others$700–900B
Spending is committed; the question is how fast it converts to energized capacity.
The bottleneck · power, not chips (IEA / grid data)
US interconnection queue stuck~2,300 GW > entire US grid
Avg wait in queue → operational~5 years
Grid capacity for a new DC24–36 months
High-power transformer lead time~5 yrs was 24–30 mo
Global DC electricity by 2030 (IEA)945 TWh
2026-scheduled projects hit by power delays~50%
The shell takes 12–18 months; the power takes years. A project can sit "announced" indefinitely waiting on a grid connection and a transformer — that is why ~70% of the announced cohort is untouched.
Read-through to Micron / memory
- Base demand is durable — 77% of what's under construction is pre-leased, and memory is allocated via multi-year LTAs tied to committed builds, not press releases.
- Spike pricing is suspect — peak DRAM/HBM ASPs (and Micron's ~$100 EPS run-rate) are partly underwritten by announced capacity that converts at only ~50–60%.
- Two stacked bottlenecks — a memory shortage and a power shortage. Power is the slower one, so it meters how fast memory can actually be consumed.
- Watch the conversion rate, not the announcement flow. If power keeps conversion at ~50–60%, memory demand grows fast but not at the doubling pace baked into peak pricing → supports normalization over supercycle-forever.
Watch-list · leading indicators to track over time
~30%
Announced → building conversion
77%
Under-construction pre-lease
~5 yr
Grid interconnect wait
A rising conversion rate + falling transformer lead times = supply catching demand = pressure on memory ASPs. A stalling conversion rate = shortage persists = pricing holds. This is the dial that decides the memory thesis.
