By Djellal Djouad
JP Morgan's latest research note (May 2026), built on satellite imagery analysis from SynMax, makes a quiet but important admission: more than 60% of the AI data-center capacity scheduled to come online in 2027 has not broken ground, and another 7% is already running behind schedule.
For CrossVol readers, this number is not a surprise. It is the physical confirmation of a thesis we have been tracking since last year — The China AI Disruption Thesis — whose entire foundation rests on one observation: the US grid is structurally late on what the hyperscalers have promised. JPM's data, independently sourced, says the same thing.
This post lays out the triple divergence now visible between hyperscaler CAPEX announcements, actual physical construction, and credit-market pricing.
1. The number JPM is pointing at — 60% of announced 2027 capacity is empty land
The data comes from SynMax, a geospatial analytics firm using satellite imagery to measure earthworks, foundation pours, and structural progress, then reconciling those measurements against hyperscaler announcements.
Announced 2027 AI data-center capacity: 21.5 GW
Actually under construction as of May 2026: 6.3 GW
That is a 1:3.4 gap between announcement and physical build — the widest in the modern data-center cycle. The gap between announcing capacity and building it is widening, not closing.
2. PJM — the grid side already crossed its critical line
PJM (Pennsylvania–Jersey–Maryland Interconnection) is the geographic core of US AI infrastructure. It is also where the physical electricity constraint has now become public:
Apr 2026: PJM Market Monitor stops new data-center grid connections — a formal admission of capacity scarcity.
May 2026: PJM capacity prices +75% in 12 months. The market is pricing scarcity, not narrative.
Apr 10, 2026: Emergency backstop procurement of 14.9 GW, off-cycle and outside the normal auction.
Ongoing: Interconnection queue at 7+ years post-approval. A 2026 approval = grid hookup in 2033–34.
2025: $11.8B transmission expansion plan approved — but transmission build is a decade-scale process.
PJM is not alone. High-voltage transformer lead times have stretched to 24 months (from 6 months in 2022). New transmission lines need ~10 years for permitting and construction. 40–70% of switchgear and battery components are sourced from China — a constraint China is in no hurry to relax.
Now overlay the spending. The four hyperscalers — Microsoft, Alphabet, Amazon, Meta — are guiding to roughly $725B of combined 2026 CAPEX (+77% YoY). Yet US data-center capacity under construction fell −5.7% from end-2024 to end-2025 — the first contraction since 2020.
Capital is going to silicon. Silicon is waiting on concrete. Concrete is waiting on the grid.
3. Hyperscaler CDS — the credit market is six months ahead
Here is what equity investors are missing: the credit market has already started repricing the CAPEX–grid divergence.
Bank of America: single-name hyperscaler CDS volume is up ~10× vs early 2025 (per Matt Mandell, BoA Single-Name CDS desk).
Saba Capital (Andrew Weinberg): "This is the best AA credit-default-swap opportunity in a long time."
S&P Dow Jones Indices added Meta, Alphabet and Microsoft to the CDX Investment Grade index this year.
JP Morgan structured a bespoke 5-hyperscaler CDS basket.
The structural buyer here is the bank CVA desk (credit valuation adjustment). To keep underwriting bonds and trading derivatives with hyperscalers, banks need to hedge the counterparty exposure their books are absorbing — that means buying protection. FX-denominated bond issuance hedged back to USD through cross-currency swaps also drives CVA-mediated CDS demand, structurally and persistently.
Translation: hyperscaler CDS demand is not speculative — it is structural. It puts a floor under price while it rises.
4. The triple divergence — what equities don't want to see
Three time series are now moving in opposite directions:
CAPEX announcements — accelerating (+77% YoY)
Physical construction — decelerating (under-construction capacity −5.7%, 60% unstarted)
Credit pricing — repricing forward (CDS volume ×10, CDX inclusion, bespoke basket)
The equity narrative is still anchored in the simple equation CAPEX = future revenue. The satellite imagery, PJM's regulatory actions, and CDS market flow are drawing a different picture:
A meaningful portion of announced capacity will not be physically or electrically operational by 2027.
The window CrossVol watches for the first credit signal to surface is Q4 2026 → Q1 2027. That is when (a) the next PJM capacity auction prices print, (b) hyperscaler 10-Q disclosure begins separating "built GW" from "announced GW", and (c) the ROI evaluation pressure on the NVIDIA chip inventory accumulated in 2024–2025 starts to bite.
5. CrossVol's positioning — observation, not recommendation
This is not a trade recommendation. CrossVol publishes observations, not trades.
That said: when CAPEX, physical construction, and credit pricing all diverge in the same direction at the same time, history says we are usually in the evening before an infrastructure-bubble inflection. Two signals worth watching — that is all we are saying.
This thesis is developed in full in the CrossVol Research book The China AI Disruption Thesis, available on Amazon Kindle (English).
Read the full thesis
→ The China AI Disruption Thesis — full explainer on crossvol.com (free, no paywall)
→ Book hub on crossvol.com with chapter summaries and case studies
→ The China AI Disruption Thesis — Kindle on Amazon (B0H11WH3R9)
→ Beyond Gamma Exposure — Paperback on Amazon (B0H2QSF3X1)
CrossVol Research publishes non-consensus research on derivatives, volatility, dealer positioning, and macro flow — across equities, FX, futures, futures options, rates, credit, and commodities. Eleven languages. Cross-asset by default.
Author: Djellal Djouad — CrossVol Research
— CrossVol Research
CrossVol Team & Djellal Djouad
Further reading