Data center power demand versus grid supply

A McGuireWoods legal alert added timeline detail to the Northern Virginia permit freeze: Loudoun County voted September 15, 2026 to begin a 12-month application pause process and had already ended by-right data center development before that vote.

A formal county resolution is expected in October, with a November 17 vote on the pause also referenced.

Governor Abbott on September 21 directed Texas's environmental regulator to halt data center air permits pending an ERCOT audit of grid and water impacts, stalling 8.7 gigawatts of planned capacity including Google, SpaceX, and Vantage sites.

Affected developers faced a notarized attestation deadline of October 12; observers noted the pause is unlikely to lift before midterm elections, and SemiAnalysis expects permits to resume around December 10 when ERCOT publishes its findings.

Amazon cancelled a $427 million campus, QTS walked away from a project with roughly $100 billion in planned spending, and Blackstone sold data center assets in Northern Virginia, the world's largest data center market.

Loudoun County scheduled a November 17 vote on a 12-month pause on new applications. Analysts attributed the pullback to land scarcity, power unavailability, and local political resistance rather than weakening AI demand, with Texas, Ohio, and the Midwest cited as the next build-out destinations.

NVIDIA published DSX AI Factory Platform adoption details naming GE Vernova, Hitachi Energy, Schneider Electric, Siemens Energy, Phaidra, and Emerald AI as partners, citing over $300 billion in power equipment backlog and 200 gigawatts in US interconnection queues.

Trump posted on Truth Social that large new data centers must generate their own electricity and pay host communities a cash dividend, going beyond the Ratepayer Protection Pledge that experts consider likely unenforceable. Michael Dell described the memory shortage as 'maybe a four year problem' given roughly four-year fab build times, placing relief no earlier than 2029 to 2030.

On SpaceX's Q3 2026 earnings call, Musk disclosed plans for 20 GW of new power and cooling in 2027, roughly sevenfold SpaceX's projected year-end 2026 capacity, guided Q4 capex to $35 billion to $40 billion, and said the bottleneck has shifted from chips to power and construction.

JPMorgan projected AI compute demand reaching 530 GW by 2031 against supply of only 250 GW, while TrendForce estimated a 180-million-fiber-kilometer global fiber shortage in 2026 with prices up roughly fivefold, expected to last through 2028.

Bank of America research finds AI capex running 50% above forecasts while revenue is only 20% above, with hyperscaler WACC more than doubling from 4% to 10%

The Financial Times counted at least 47 large US data center projects delayed, withdrawn, or blocked in 2026, with Bernstein rating only about 35% of the planned pipeline as credible.

That count places Microsoft's withdrawal from a planned Loughton, Essex data center after grid officials said power would not arrive until the 2030s within a wider, documented pattern of grid-driven project failures rather than an isolated case.

Microsoft canceled its data center in Loughton, Essex after grid officials told developer Nscale power would not be available until the 2030s.

Goldman Sachs projected five major tech firms will collectively spend $1.2 trillion on AI infrastructure in 2027 and $1.4 trillion in 2028, while ASML warned that AI chip-equipment demand will exceed its planned 2027-2028 supply and said it is weighing capacity expansions beyond an already-planned 30% increase. SpaceX was separately reported in early talks to borrow $40 billion to purchase Nvidia chips.

Eric Schmidt, Broadcom CEO Hock Tan, and Intel CEO Lip-Bu Tan each argued that capital, not power or chips, is AI infrastructure's binding constraint, with Schmidt estimating roughly $50 billion per gigawatt.

Morgan Stanley estimated a 76.5 GW shortfall against roughly 160 GW of projected US demand by 2029, and Amazon CEO Andy Jassy reportedly told an investor compute capacity is sold out beyond 2028. H100 rental prices held at roughly $2.70 per hour since July despite falling token costs, suggesting cheaper inference expands rather than reduces total compute demand.

Dell CEO told Wall Street Journal the memory shortage could last until 2029 or 2030.

A Bloomberg analysis published October 15 breaks down costs for a gigawatt-scale AI data center in Texas: servers account for at least half the total construction bill, with server and cooling installation alone running roughly $1.2 billion and DRAM costs rising fivefold year-over-year.

Power-connected land costs $584,000 per megawatt, and roughly $750 billion in hyperscaler capex competing for the same turbines and copper has pushed power-gear backlogs past 2030, with potential overruns on a single 1 GW site approaching $3 billion.

H100 GPU rental prices remain stable at roughly $2.70 per hour despite sharp declines in AI token costs since July

On October 12, NVIDIA, AMD, Google, Meta, and Microsoft published two open specifications at the OCP Global Summit covering AI factory power architecture and a grid resilience requirement that AI infrastructure restore to near-full load within two seconds of a disturbance.

Reporting the same day placed China at 50 GW of AI compute under construction or announced against a current 24 GW, closing on the US's 56 GW, with chip shortages as the key constraint on Chinese expansion.

Ramp data shows token volumes rose roughly 7x from January to September while total AI spend fell; H100 GPU prices held near $2.70/hour

Texas paused new data center grid approvals on October 7 as ERCOT audited a queue that grew from 63 GW to 474 GW in 18 months with only 4.3 GW operational.

Galaxy Digital's Helios campus, holding ERCOT-approved power, contracted 526 MW to CoreWeave at over $1 billion in projected annual revenue. The same day, SpaceX entered talks to borrow $40 billion to purchase Nvidia chips, with Anthropic contracted to pay SpaceX up to $84.5 billion through 2029 for compute access.

Reporting published October 6 identified high-voltage transformer lead times of five or more years as a hard ceiling on data center construction, distinct from turbine and grid bottlenecks, with only a third of 12 GW of campuses planned for 2026 under construction and manufacturers reluctant to expand due to past overcapacity cycles.

Ben Horowitz of Andreessen Horowitz added a workforce dimension on October 7, noting that only 2% of US electricians hold DC power certification as data centers shift to 800VDC distribution systems.

Marvell's VP of supply chain posted an opening for an optics planning executive on October 5, describing the AI infrastructure bottleneck as shifting from GPU compute to high-speed optical interconnects as clusters scale from 800G to 1.6T connections.

On October 6, Elon Musk confirmed Colossus 2 comprises 110,000 GB200s and 440,000 GB300s with a potential additional 220,000 GB300s by late December, and SpaceX is projected to reach 2.3 GW of compute capacity by late November 2026, up from 1.4 GW at end of Q2.

SpaceX projected to reach approximately 2.3 GW of compute by late November 2026, up from 1.4 GW at end of Q2

Qualcomm CEO Cristiano Amon projected on September 29 that global AI token demand will grow 40x by 2030, reaching 1.27 trillion tokens per 10 seconds, as AI shifts from human-paced interaction to agent-paced activity.

Amon noted that agentic AI carries hidden computational costs including context carrying, memory updates, and error checking, compounding the load on infrastructure. The projection adds a structural demand argument to existing forecasts that already show US grid supply falling far short of data center needs by 2030.

TrendForce projected global data center power demand will reach 490.7 GW by 2030, with growth above 30% annually through at least 2027

Goldman Sachs raised forecast for major tech AI infrastructure spending to $1.2 trillion in 2027 and $1.4 trillion in 2028

Anthropic reported in early talks to lease up to 1 GW of compute from Stream, majority-owned by Apollo Global Management

NVIDIA launched DSX Ready qualification program for power and cooling products; South Korea semiconductor exports surged 259.4% year-over-year to $34.1 billion in first 20 days of September

DRAM prices reported up roughly 171% year-over-year as Samsung, SK Hynix, and Micron shift production to HBM; Micron AI memory reportedly sold out through end of 2026

Several independent forecasters projected US data center grid power demand nearly tripling by 2030, with BloombergNEF at 118 GW and 451 Research at 134.4 GW from 61.8 GW in 2025.

A RAND analysis from April 2026 found only about 33 GW of net new capacity available after queue withdrawals. Hyperscalers including AWS, Google, Microsoft, Meta, and Alphabet signed nuclear power agreements to bridge that gap, and SpaceX confirmed plans to manufacture turbine blades in-house to ease a 60-to-90-week supply bottleneck.

NVIDIA and Emerald AI demonstrated flexible-load program with Silicon Valley Power throttling low-priority workloads in response to grid signals

Financial Times reports established turbine suppliers may expand preemptively in response to SpaceX competitive threat

Eric Schmidt argued capital, not energy, is the binding constraint on AI scaling, estimating $50 billion per gigawatt

Nomura projects global memory revenue reaching $3.675 trillion by 2030, with AI data center memory comprising nearly half

SpaceX confirms plans to build its own foundry for natural gas turbine blades and vanes; Howmet Aerospace stock falls more than 7%

SpaceX announces in-house turbine blade and vane casting to cut generator procurement delays by 18 months

Samsung reportedly considering repurposing NRD-K Line 2 facility to target 2nm base dies for HBM5

BloombergNEF raised its 2030 US data center capacity forecast 52% to 118 GW and its 2035 forecast 83% to 194 GW

Morgan Stanley projects combined capex of the Big Five cloud operators at approximately $1.2 trillion in 2027 and $1.4 trillion in 2028

Carnegie Endowment assessed hyperscaler nuclear commitments, noting Meta's Constellation PPA covers clean energy credits rather than direct electricity

RAND published analysis finding only ~33 GW of net available capacity from 1,086 GW US interconnection queue

Wells Fargo projects total hyperscaler compute capacity will surpass 125 GW by 2028, with capex nearly doubling to over $860 billion by 2027

Meta and Microsoft each signed 20-year PPAs with Constellation Energy for nuclear power.