Data Center Power
The constraint on AI data centers is no longer chips but power infrastructure. We read the move to 800 VDC for 1 MW racks, GaN/SiC adoption, and the transformer / grid-interconnection bottleneck through primary sources such as NVIDIA, OCP and EPRI.
As the AI data center bottleneck shifts from semiconductors to power infrastructure, this report maps the move to 800V DC delivery for 1 MW racks, GaN/SiC adoption, power-cooling integration, and the triple wall of transformers, grid interconnection, and regulation — all from primary sources. It covers a supplier player map with product comparisons, three-scenario market sizing, implications for Japanese companies, role-based playbooks for investment/procurement/design, risks with leading indicators, and case studies across 13 chapters plus an appendix (based on public information from NVIDIA, OCP, EPRI, NERC, DOE, and others).
NVIDIA's 800VDC rollout timeline: Vertiv ships products in H2 2026, Foxconn leads deployment at its Kaohsiung K-1 site, and Kyber racks ship in 2027.
DOE has published a draft 2026 National Transmission Needs Study that identifies data centers, manufacturing expansion, and large industrial loads as key causes of transmission constraints. U.S. data-center electricity use is projected to rise from 176TWh, or 4.4% of total power, in 2023 to 325–580TWh and as much as 12% by 2028. This article maps a grid bottleneck that is becoming as material as semiconductors.
AI supercomputing is shifting from performance in FLOPS to power and efficiency. TOP500 leader LineShine reached 2.198 EF at 42.2 MW with a CPU-only design, while Microsoft Fairwater points to distributed AI superfactories. Primary data shows the move from FLOPS to megawatts.
On July 14, 2026, New York issued the first U.S. pause on state environmental permits for new hyperscale data centers above 50MW, for up to one year. About 12GW of data-center load is queued at NYISO. The previous day, Meta announced an expansion of Hyperion to 5GW and more than USD 50 billion. This article examines a U.S. market where regulation and expansion now proceed at once.
As liquid cooling becomes a baseline for AI servers, designers need operational requirements, not just cooling-method selection. This article reviews TCS loop pressure of 140-450kPa, ASHRAE coolant classes W1-W5, IEC 62368-1 leak testing, and sensor requirements.
NERC has identified computing loads, including data centers, crypto assets, and hydrogen electrolysis, as posing reliability issues distinct from conventional industrial loads. With 50MW blocks changing in 250ms and standards planned by end-2026, thresholds, requirements, and timing are becoming central to site strategy.
FERC issued a large-load interconnection ANOPR in October 2025 and plans action by June 2026 after reviewing more than 3,500 pages of comments. PJM summer peak demand is projected to rise 70GW over 15 years.
OCP is standardizing data-center liquid cooling - cold plates, CDU, UQD connectors, coolants - with NVIDIA, Meta, Dell, and Intel involved.
A guide to data center efficiency metrics: The Green Grid's PUE, WUE, Partial PUE, waste heat reuse, and the xUE family including PUE/WUE/CUE.
Compare 48V, +/-400V, and 800V DC for AI data-center racks by current, losses, power density, OCP direction, and procurement risk.
Primary sources explain why transformer lead times delay AI data-center energization. Large transformer lead times doubled from about 50 weeks in 2021 to an average 120 weeks in 2024, prices rose 80%, and one U.S. manufacturer disclosed a five-year wait for new orders. CISA, DOE, and NERC show the three barriers: custom production, interconnection queues, and new large-load rules.
Quick evaluation guide to 800V DC power for AI data centers: efficiency gains, fewer PSUs, density benefits, DC protection, and sourcing risks.
AI data centers are adding onsite generation, BESS, and demand response as grid-only power becomes constrained.
A quick guide to data-center liquid cooling, split into direct liquid cooling(DLC/direct-to-chip) and immersion cooling(single-phase and two-phase). It explains cold-plate local heat removal, dielectric immersion, applicability, maintenance differences, and selection criteria based on primary and academic sources.
As high-density AI racks shift toward liquid cooling, cooling choices reshape power supplies, SiC/GaN devices, Tj margins, and efficiency.
OCP's Diablo sidecar power-rack concept moves from in-rack 48V DC to ±400/800V DC to support 100kW to 1MW AI racks.
800V DC power for AI data centers: NVIDIA's 1MW-rack roadmap, up to 5% efficiency gains, DC protection hurdles, and supplier timing.
AI data-center bottlenecks are moving to power infrastructure. EPRI projects 9-17% of U.S. electricity by 2030 and up to 10 years to energization.
onsemi's GaNEXUS pairs 650V GaN with GlobalFoundries and 40-200V GaN with Innoscience, with H1 2026 samples for AI data-center power.
Common trends from onsemi Q1 2026 and Infineon Q2 FY2026 earnings. As power demand for AI data centers expands rapidly, automotive applications remain in recovery. A breakdown of the two-axis structural shift that design and procurement teams should understand.
Generative-AI data centers are driving SiC/GaN UPS and PSU adoption as efficiency rules tighten. 2026 power-device demand trends and what they mean.
Power consumption for a single rack of servers running GPT-4 has surged from an average of 10-15kW in 2020 to over 100kW for racks equipped with the latest AI accelerators, with NVIDIA H100-dense racks approaching 70kW individually, and further increases anticipated with the next-generation Blackwell architecture.
