Your briefing

Stargate is a nationwide, $500 billion data‑center program led by OpenAI with Oracle and SoftBank that targets seven US sites totaling more than 9 GW of planned facility power — roughly the scale of New York City’s single‑hour peak demand — and an aggregate compute capacity the authors estimate at ~20 million H100‑equivalents. Abilene, TX is furthest along (~0.3 GW operational as of Apr 22, 2026; 250k H100‑eq; four of eight buildings live) and is expected to reach 1.2 GW by Q4 2026. The six other sites (Shackelford 2.0 GW, Doña Ana 2.2 GW, Milam 1.2 GW, Port Washington 1.3 GW, Saline Township 1.4 GW, Lordstown <0.3 GW) are mostly slated for Q4 2028 handovers. Developers are favoring on‑site natural‑gas microgrids and closed‑loop liquid cooling to shorten grid interconnection timelines and limit water evaporation, but those choices raise cost and regulatory scrutiny; procurement, financing, and local opposition remain key risks to the 2029 build‑out target.

DG Matrix's whitepaper (published 17 March 2026 by Data Centre Dynamics) analyzes how 'stranded power' within facilities limits AI data-center growth and shows that conventional distribution can't match dynamic AI workloads. It proposes software-defined power infrastructure to dynamically route energy across grid, storage and distributed resources, unlocking additional deployable compute within current utility limits and improving economics.

Greencoat Renewables’ Paul O'Donnell outlined how the yieldco model that dominated renewable finance for a decade has moved into a new phase of active operations, hybridization and growth into digital infrastructure. He traced the firm’s origins to 2012/13 and its listing in 2017, and reported total invested capital of about €13.5 billion, with just under 8 GW under management across roughly 445 assets. That long-term accumulation of government‑backed, contract-stable assets is giving way to a more complex market: higher interest rates, shifting policy after Europe’s energy crisis, and new demand drivers such as AI and data centre build‑out are forcing owners to become trading-savvy operators.

Paul described the practical response: long-term PPAs directly with tech companies, hybridizing wind/solar sites by adding batteries for arbitrage and grid services, and partnering to scale capabilities rather than trying to internalize every function. He cited a joint venture with a battery/storage provider (referred to in the interview as 'Coato') to access pricing and innovation — the partner, he said, had recently won roughly 80% of certain Italian opportunities. On storage he emphasized falling capex and attractive incremental costs when batteries are added to existing sites, and argued revenue stacks (day‑ahead arbitrage, balancing markets, capacity/flexibility services) now underpin investment cases. Turning to data centres, Paul said they already account for an estimated 22–23% of Ireland’s electricity and present both demand and policy tailwinds: governments want data centres to source renewables and provide grid backup/flexibility. He estimated a 100 MW data centre costs ~€1.5 billion to build plus €300–500 million of enabling energy infrastructure, and argued truly off‑grid, gas‑backed data centres are unlikely in Europe. Hosts agreed the strategic pivot — which they labelled 'yield to growth' — makes sense and welcomed energy investors entering the digital infra space, while also flagging structural risks for listed yieldcos (market marking, investor alignment and competition from fixed income). The market reaction to Greencoat’s March strategy — a roughly 20% share price rise — suggested investors rewarded the clearer growth and integration story.

Engineering the future of cooling — DCD streamed a three‑part broadcast on Tuesday, 10 March 2026 (9am, 10am, 11am ET) featuring Siemens, Nortek, Ark, Tillion, Fleet, Johnson Controls and Armada speakers. Sessions emphasized engineering methods (physics‑based simulation, digital twins, advanced controls), integrating liquid cooling from grey to white space, and modular/prefabricated solutions to scale high‑density AI deployments.

High-voltage transmission is surging in the U.S. after a decade-long lull: 888 miles of 345 kV+ lines were completed in 2024 (vs. ~4,000 in 2013). Planned spending jumped—PJM from $920M (2021) to ~$12B (2025)—and utilities plan $1.4T by 2030 (about half T&D). Some former boosters resist public cost-sharing even as private demand (data centers, Permian) is ready to pay.

AI chips sit at the center of modern AI progress: specialized accelerators (Nvidia Blackwell/Hopper, Google TPU, Amazon Trainium, Huawei Ascend) are designed by a few firms but mostly fabricated by TSMC, with critical upstream suppliers such as ASML (EUV) and HBM vendors Samsung, SK Hynix, and Micron. By Q4 2025 roughly 20 million AI chips had been shipped by the five largest designers, with Nvidia supplying ~50% of units and >66% of deployed compute capacity. Metrics matter more than sticker price: although flagship chips rose from $5,700 (2016) to $34,000 (2022), computation per dollar improved dramatically (H100 ≈17× P100), and compute-per-dollar has doubled ~every 2.5 years. Chips consume ~1,000 W each and total datacenter power reached tens of gigawatts by late 2025; energy efficiency has improved ~40% annually, but aggregate electricity use still grows because deployment outpaces per-chip gains.

Epoch AI's AI Chip Owners explorer (launched 2026-04-06) maps who owns the world's leading AI chips by distributing previously estimated sales volumes (Nvidia, Google TPU, Amazon Trainium, AMD, Huawei) across major owners using analyst estimates, company disclosures, capex data, and frontier data-center analysis. The study finds US hyperscalers control over 70% of AI compute, led by Google (≈5 million H100-equivalents, ~25% of global compute) with Google TPUs contributing nearly 4 million H100-equivalents (CI 3.1M–4.5M). Mainland China holds just over 5% of cumulative compute as of end-2025; illicit imports may have added tens of thousands to >100,000 A100/H100s in 2024 per prior reports but are unlikely to close the gap. The explorer includes interactive visualizations, a methodology and dataset for researchers and policymakers.

Across techniques they find typical trade spans of order 1–2 OOM in each axis: (a) Chinchilla-style scaling can trade ~1.2 OOM training for ~0.7 OOM inference; (b) MCTS exhibits an S-curve where at low performance ~1 OOM training ↔ ~1.6 OOM inference (and the trade shrinks as models near perfect play); (c) IMP pruning to ≈10% density saved ≈1 OOM inference at ~0.7 OOM extra training because of multi-round re-training; and (d) resampling with cheap verification (e.g., code generation pass@k) extrapolates to much larger spans (authors estimate up to ~3 OOM training savings vs ~6 OOM inference increase) but caution low confidence in extrapolation. The authors categorize three tradeoff archetypes—saturating/constant span (e.g., Chinchilla), saturating/decreasing span (e.g., MCTS, limited n@k), and non-saturating/increasing span (resampling with effectively unlimited trials)—and show that some techniques combine roughly additively but often with diminishing returns, so realistic combined spans are ~2–3 OOM. Methodologically they fit empirical curves (smoothly broken power laws, log relations) and prioritize Pareto frontiers for optimal TC/IC choices. The paper emphasizes governance implications: because aggregate inference cost often exceeds training cost in deployed systems, commercial models will be biased toward low-IC configurations, but researchers or well-resourced actors can invest IC to simulate larger-model capabilities for evaluation, internal use, or small-customer deployments—affecting safety assessments and policy proposals aimed at controlling frontier capabilities.