US data center grid power demand is forecast to rise 22% in 2025 to 61.8 GW and reach 134.4 GW by 2030, according to 451 Research. Bloomberg reported in October 2026 that building a gigawatt-scale AI data center in Texas now costs roughly $1.2 billion for server and cooling installation alone, with servers accounting for at least half the total construction bill after average prices more than doubled over two years and DRAM costs rose fivefold year-over-year. Power-connected land runs $584,000 per megawatt, private gas generation has tripled in cost, and roughly $750 billion in hyperscaler capex is competing for the same turbines, steel, copper, and fiber, pushing power-gear backlogs past 2030.
A RAND funnel analysis shows the supply constraint is severe: of 1,086 GW of planned US generation in the interconnection queue, only roughly 33 GW of net new capacity is realistically deliverable by 2030. Transformer News projects demand could reach 500 GW by 2035, exceeding total current US electricity consumption, and notes high-voltage transformer lead times exceed five years, with the manufacturing industry reluctant to scale because past boom-bust cycles led to overcapacity and factory closures.
Texas illustrates the bottleneck. ERCOT's large-load queue grew from 63 GW to 474 GW in 18 months, with roughly 90% tied to data centers, but only 4.3 GW is operational. Texas has paused new approvals while ERCOT audits credible versus speculative projects; transmission and substation approvals can take five to ten years. Galaxy Digital's Helios campus, holding over 1.6 GW of ERCOT-approved power, has contracted 526 MW to CoreWeave, projected to generate over $1 billion in average annual revenue.
SpaceX is building a foundry to cast turbine blades and vanes in-house; Musk said this could bring natural gas turbines online up to 18 months faster. Howmet Aerospace stock fell more than 7% following the announcement before partially recovering. The Financial Times reported established suppliers may expand capacity preemptively if they believe SpaceX can credibly manufacture these parts, and the competitive threat may ease the bottleneck even if SpaceX never produces at scale. SpaceX is projected to reach approximately 2.3 GW of compute capacity by late November 2026, up from 1.4 GW at end of Q2, as two 220,000-unit batches of NVIDIA GB300 GPUs come online. SpaceX is also in early talks to borrow $40 billion to purchase Nvidia chips, with Apollo Global Management leading a package of roughly $10 billion in bank loans and $30 billion in investment-grade debt expected to close in 2027, according to Bloomberg. Anthropic is contracted to pay SpaceX up to $84.5 billion through 2029 for compute access.
Goldman Sachs projects Amazon, Meta, Google, Microsoft, and Oracle will collectively boost AI infrastructure spending by more than 50%, reaching $1.2 trillion in 2027 and $1.4 trillion in 2028. Goldman said a $250 billion surprise in either direction next year would move S&P 500 earnings growth by roughly 6 percentage points.
China has 50 GW of AI compute under construction or announced against a current capacity of 24 GW, compared to the US's 56 GW, with one Inner Mongolia city alone planning 89 data centers, though chip shortages remain a key constraint.
Workforce and component bottlenecks add to the pressure. Ben Horowitz of Andreessen Horowitz said only 2% of US electricians hold DC power certification while AI data centers shift to 800VDC distribution. The AI infrastructure bottleneck is also shifting from compute to optical interconnects: Marvell is hiring an optics planning executive as connections scale from 800G to 1.6T, with its 1.6T Ara optical DSPs in volume production, and Credo targets the same market. Navitas completed its acquisition of Claros, adding vertical power delivery and integrated voltage regulator technology that places power conversion directly beneath AI chip packages.
At the OCP Global Summit, NVIDIA joined AMD, Google, Meta, and Microsoft to publish two open specifications: the Open Data Center Specification v0.7 covering power architecture, cooling, and physical layout for AI factories, and an OCP Grid Resilience Specification requiring AI infrastructure to stay connected during grid disturbances and restore to near-full load within two seconds. Visionbay introduced K-1, an AI factory platform using 800VDC architecture designed for NVIDIA Vera Rubin deployments.
Mizuho analysts project AI agents will drive agentic CPU sockets from 1.3 million in 2026 to 16 million by 2030, with the server CPU market potentially reaching approximately $209 billion by 2030. Agentic servers would require nearly 35,000 petabytes of DRAM and 542 exabytes of NAND annually by 2030, according to Mizuho.
Milk Road AI argues that cheaper token prices stimulate rather than reduce compute demand: open-weight models grew from a minority of token volume in June 2026 to 78.4% by mid-September on one AI gateway, and Ramp data shows token volumes rose roughly 7x from January to September while total AI spend fell. Because an open-weight token consumes roughly the same compute and energy as a closed-model token of similar size, Milk Road AI argues that cloud providers serving open-weight tokens retain high margins on the infrastructure side while the model layer faces margin compression.