Data Center Market Update: Week of August 10, 2026

This week’s news centers on grid reliability and local pushback rather than new capacity: PJM is weighing mandatory ride-through rules after a 3.8GW Virginia load trip, AWS withdrew a 500MW Maryland project amid community opposition, and three more US towns restricted data center development. Against that, North America colocation demand hit a record 25GW of H1 absorption and Thailand pushed ahead with a direct power-purchase pilot for data centers.

Key takeaways

• North America data center absorption hit a record 25GW in H1 2026 — double the year-ago pace and five times the level from two years prior, per JLL.
• AWS withdrew its 2.5 million sq ft, up to 500MW Calvert Cliffs, Maryland project after a 10,000+ signature petition and moves toward a county moratorium.
• Data Center Watch counted 75 projects worth over $130 billion delayed or cancelled in Q1 2026 — lenders are now pricing community opposition into financing terms.
• A 230kV transmission line failure on 22 July 2026 sent roughly 3,800MW of Northern Virginia data center load onto onsite generation; PJM is now drafting ride-through rules, and FERC has ordered NERC to file national computational-load reliability standards by 31 December 2026.
• Waxhaw, NC (pre-emptive 12-month moratorium), Charlestown, RI (outright ban on data centers and battery storage) and Dallas, TX (development-code hearing) all acted against data center growth in the same week.
• Thailand’s direct PPA pilot for data centers is moving into implementation, alongside BOI approval of four new hyperscale projects worth roughly $3.1 billion.
• B200 GPU rentals span $7.00-16.11/GPU-hour depending on channel — the widest hyperscaler-vs-marketplace spread of the current GPU generation.

Record demand meets a harder grid ceiling

JLL reported North America data center absorption hit an all-time high of 25GW in the first half of 2026, more than double the year-ago figure and five times the pace of two years ago. Demand isn’t the constraint anymore — the same week’s news shows the bottleneck has moved fully to grid interconnection, reliability and local permitting. See power density trends for how that demand translates into rack-level load.

AWS retreats from Calvert Cliffs, Maryland

AWS withdrew its application for a three-campus, 2.5 million sq ft project adjacent to the Calvert Cliffs nuclear plant, which would have drawn up to 500MW of power supplied by Constellation Energy. The company filed in May 2026; Calvert County confirmed the withdrawal on 4 August after a petition against the project gathered more than 10,000 signatures and officials began drafting a six-month moratorium. Neither AWS nor the county has formally tied the withdrawal to the backlash, but the pattern is now a market-wide signal: Data Center Watch counted 75 projects worth over $130 billion delayed or cancelled in Q1 2026, and Capacity Media reports lenders have begun pricing community opposition risk directly into data center financing terms. Buyers evaluating a site should treat local permitting risk as a diligence line item — see our due diligence checklist.

PJM moves toward mandatory ride-through rules

On 22 July 2026, a mechanical failure on a 230kV transmission line triggered the largest data center load-trip event in PJM’s history: roughly 2,970MW of Northern Virginia load transferred to onsite generation immediately, with another 1,099MW following as grid conditions shifted, for a combined 3.8GW. PJM is now weighing new reliability protocols covering voltage and frequency ride-through, onsite generation behavior during disturbances, and reconnection standards. The event compounds a broader federal push: on 16 July 2026, FERC directed NERC to file new or modified reliability standards for large computational loads by 31 December 2026. Operators relying on gas turbines or onsite generation as a grid hedge should note that “behind-the-meter” capacity is about to become a regulated asset class, not a workaround — relevant background in our SMR and nuclear power guide.

Local moratoriums spread beyond the usual markets

Three towns acted against data center development in the same week, none of them traditional primary markets:

Location	Action (Aug 2026)	Notes
Waxhaw, NC	12-month moratorium	Approved pre-emptively — no project yet proposed
Charlestown, RI	Outright ban	Covers both data centers and battery energy storage; second RI town to ban
Dallas, TX	Development-code hearing requested	Aims for clearer oversight rules, not a ban
Calvert County, MD	Moratorium in draft	Follows AWS withdrawal from Calvert Cliffs site

The spread from primary markets (Northern Virginia, Dallas-Fort Worth) into smaller jurisdictions with no live proposal (Waxhaw) shows opposition is now pre-emptive rather than reactive in some regions.

Thailand’s direct PPA pilot moves to implementation

Thailand’s pilot program letting data centers contract renewable power directly through power purchase agreements — rather than solely through the state utility — is moving from policy design into implementation, per a Dentons client alert published 14 August 2026. The move follows the Board of Investment’s approval of four new hyperscale projects worth roughly 100 billion baht ($3.1 billion). Direct PPAs matter for colocation buyers because they decouple new capacity from grid queue delays, the same constraint slowing projects in the US this week. See our Thailand data center market guide for how Bangkok compares to Singapore and Johor on power access.

GPU rental prices: the Blackwell spread widens

GPU	Rate range (Aug 2026)	Notes
H100	$2.00-2.70/GPU-hour (marketplace) up to $6.88-12.29 (hyperscaler)	AWS $6.88, Azure $12.29, OCI $10.75 vs Spheron/Jarvislabs marketplace rates
B200	$7.00-8.64/GPU-hour (Nebius, Lambda, RunPod) up to $16.11 (Google Cloud)	Widest hyperscaler-vs-marketplace spread of the current generation

Blackwell (B200/B300) access remains gated by enterprise contracts and waitlists at several clouds, which keeps published on-demand rates inconsistent. Full breakdown in our H100 rental price tracker and GPU cloud comparison.

Critical minerals: the next bottleneck behind power

Copper now accounts for roughly 83% of the modeled mineral mass required for AI data center buildout, with grid transmission and distribution equipment alone consuming 64% of that copper demand — a direct link to the same interconnection bottleneck driving PJM’s rule changes this week. Tin demand tied to HPC hardware has turned inelastic as roughly 190GW of new hyperscale capacity has been announced since early 2026, and processing bottlenecks — not raw ore volumes — are the binding constraint, per Data Center Knowledge’s 13 August analysis.

What this means for buyers

Grid reliability and local permitting risk, not raw demand, are now the swing factors in site selection: budget extra diligence time for interconnection queue position and county-level political risk, especially outside the top five primary markets where opposition is starting to appear pre-emptively. For power procurement, direct PPA structures like Thailand’s pilot are worth tracking as a way to bypass utility queue delays. On compute, treat $7-8.64/hour as the realistic B200 benchmark rather than list prices at the high end, and confirm GPU allocation terms before signing — cross-check any quote against the index and get a benchmark via our quote tool. Browse the underlying facility data in the data center catalog and track aggregate pricing in stats.