Trade The News story
TTN Research Alert: AI powers Power Markets: The New GWs-Driven CapEx Cycle - from WULF–Fluidstack–Google to Microsoft–Constellation and AWS–Talen
- Artificial intelligence has moved from keynote flourish to load curve. Utilities and equipment makers now discuss interconnect queues, transformers and substation timelines, not “use cases.” Siemens Energy’s management captured the tenor in early August: “Despite the tariff situation, we currently see an extremely strong market in the United States for energy solutions.” The tell is where the orders land—grid gear and firm power—because the gating resource for AI over the next 24–36 months is not GPUs but megawatts.
- The contract syntax is evolving accordingly. On August 14, TeraWulf signed two 10-year high-performance computing colocation agreements with Fluidstack for more than 200 MW of critical IT load at Lake Mariner, representing roughly $3.7 billion of contracted revenue (with options taking the term value to about $8.7 billion). Google will backstop $1.8 billion of Fluidstack’s lease obligations to support project debt and receives warrants for ~41 million shares of TeraWulf (about an 8% pro forma stake), aligning capital structure with power availability. Phase one targets about 40 MW online in the first half of 2026, scaling to 200+ MW by year-end 2026. Note the site spec: dual 345-kV transmission, purpose-built liquid cooling, closed-loop water and low-latency fiber—electrical and thermal details have become the headline.
- Nuclear and utility tie-ups are becoming the backbone for schedulable megawatts. Microsoft’s long-dated agreement with Constellation to supply power for cloud load underwrites the restart pathway for Three Mile Island Unit 1, with operations targeted later this decade; Amazon Web Services paired site control with a large PPA tied to the Susquehanna nuclear station in Pennsylvania. Elsewhere, developers are locking utility-mediated blocks: QTS inked a 20-year subscription for several hundred megawatts with Georgia Power; Microsoft and Brookfield agreed a multi-region framework exceeding 10 GW of new renewables across 2026–2030. The common thread is long tenor and explicit mapping of power to data-center growth.
- On the ground, the bottlenecks are prosaic and stubborn: large-power transformers, grain-oriented electrical steel, high-capacity switchgear and long-lead HV cable. Substations—not shells—are the long pole. Lead times are being shortened the old-fashioned way by localising core steps: North American and Mexican facilities are ramping CRGO slitting and transformer-core fabrication specifically for grid, renewables and data-center demand. Until that capacity arrives in volume, a single 300-MVA transformer or missed switchgear lineup can idle a building designed for petaflops.
- System operators are rewriting interconnection playbooks to keep up with clustered load. PJM has fast-tracked rulemaking after projecting a sharp rise in peak demand this decade, much of it from data centers; ERCOT is triaging an unprecedented queue of large-load requests and haircutting exuberant forecasts in planning models. The planning question is no longer “if” but “how fast, where, and at what power-quality standard.” As campuses concentrate, congestion rents and nodal spreads widen; capacity constructs and bilateral offtakes increasingly price the value of flexible megawatts.
- Inside the fence, electrical architecture is being re-engineered. Legacy 12/48-volt distribution strains at AI-class densities; copper mass, losses and protection schemes do not scale. Designs are consolidating around 48-volt rails today and higher-voltage (up to ~800-volt) campus distribution next, which ripples from solid-state transformers and medium-voltage switchgear down to board-level regulators. Utilities, in turn, are tightening point-of-common-coupling requirements—harmonics limits, dynamic VAR support, fault-ride-through—so compliance and protection studies have become schedule-critical artefacts rather than back-of-the-book appendices.
- Thermals have turned into a supply chain in their own right. Direct-liquid systems are now mandatory at the top end, but the constraint has shifted to quick-disconnect standards, cold-plate machining, manifolds, chillers and closed-loop water treatment. Miss any one of those and tens of megawatts of compute slip a quarter. The consequence is that facility teams co-design hydraulics with the power train, and water engineering (source, treatment, rejection and, where viable, reuse) sits alongside PUE in board materials.
- The transmission path from “AI demand” to market outcomes is therefore mechanical. Data-center build raises peak load; peak load collides with scarcity in transformers, electrical steel and interconnects; scarcity lifts capacity value, favours fast-ramping thermal and storage, and pulls forward substation and HV gear capex; prices and capacity payments follow. To separate noise from signal, the research path is clear: track ISO interconnect queues and queue-reform milestones; monitor large-power transformer lead times and electrical-steel localisation in North America; map permitting timelines and power-quality compliance requirements at the point of common coupling; and watch how quickly campuses adopt higher-voltage distribution and direct-liquid cooling. The overlooked conclusion is simple: in this cycle, the tech stack begins at the substation.
- The contract syntax is evolving accordingly. On August 14, TeraWulf signed two 10-year high-performance computing colocation agreements with Fluidstack for more than 200 MW of critical IT load at Lake Mariner, representing roughly $3.7 billion of contracted revenue (with options taking the term value to about $8.7 billion). Google will backstop $1.8 billion of Fluidstack’s lease obligations to support project debt and receives warrants for ~41 million shares of TeraWulf (about an 8% pro forma stake), aligning capital structure with power availability. Phase one targets about 40 MW online in the first half of 2026, scaling to 200+ MW by year-end 2026. Note the site spec: dual 345-kV transmission, purpose-built liquid cooling, closed-loop water and low-latency fiber—electrical and thermal details have become the headline.
- Nuclear and utility tie-ups are becoming the backbone for schedulable megawatts. Microsoft’s long-dated agreement with Constellation to supply power for cloud load underwrites the restart pathway for Three Mile Island Unit 1, with operations targeted later this decade; Amazon Web Services paired site control with a large PPA tied to the Susquehanna nuclear station in Pennsylvania. Elsewhere, developers are locking utility-mediated blocks: QTS inked a 20-year subscription for several hundred megawatts with Georgia Power; Microsoft and Brookfield agreed a multi-region framework exceeding 10 GW of new renewables across 2026–2030. The common thread is long tenor and explicit mapping of power to data-center growth.
- On the ground, the bottlenecks are prosaic and stubborn: large-power transformers, grain-oriented electrical steel, high-capacity switchgear and long-lead HV cable. Substations—not shells—are the long pole. Lead times are being shortened the old-fashioned way by localising core steps: North American and Mexican facilities are ramping CRGO slitting and transformer-core fabrication specifically for grid, renewables and data-center demand. Until that capacity arrives in volume, a single 300-MVA transformer or missed switchgear lineup can idle a building designed for petaflops.
- System operators are rewriting interconnection playbooks to keep up with clustered load. PJM has fast-tracked rulemaking after projecting a sharp rise in peak demand this decade, much of it from data centers; ERCOT is triaging an unprecedented queue of large-load requests and haircutting exuberant forecasts in planning models. The planning question is no longer “if” but “how fast, where, and at what power-quality standard.” As campuses concentrate, congestion rents and nodal spreads widen; capacity constructs and bilateral offtakes increasingly price the value of flexible megawatts.
- Inside the fence, electrical architecture is being re-engineered. Legacy 12/48-volt distribution strains at AI-class densities; copper mass, losses and protection schemes do not scale. Designs are consolidating around 48-volt rails today and higher-voltage (up to ~800-volt) campus distribution next, which ripples from solid-state transformers and medium-voltage switchgear down to board-level regulators. Utilities, in turn, are tightening point-of-common-coupling requirements—harmonics limits, dynamic VAR support, fault-ride-through—so compliance and protection studies have become schedule-critical artefacts rather than back-of-the-book appendices.
- Thermals have turned into a supply chain in their own right. Direct-liquid systems are now mandatory at the top end, but the constraint has shifted to quick-disconnect standards, cold-plate machining, manifolds, chillers and closed-loop water treatment. Miss any one of those and tens of megawatts of compute slip a quarter. The consequence is that facility teams co-design hydraulics with the power train, and water engineering (source, treatment, rejection and, where viable, reuse) sits alongside PUE in board materials.
- The transmission path from “AI demand” to market outcomes is therefore mechanical. Data-center build raises peak load; peak load collides with scarcity in transformers, electrical steel and interconnects; scarcity lifts capacity value, favours fast-ramping thermal and storage, and pulls forward substation and HV gear capex; prices and capacity payments follow. To separate noise from signal, the research path is clear: track ISO interconnect queues and queue-reform milestones; monitor large-power transformer lead times and electrical-steel localisation in North America; map permitting timelines and power-quality compliance requirements at the point of common coupling; and watch how quickly campuses adopt higher-voltage distribution and direct-liquid cooling. The overlooked conclusion is simple: in this cycle, the tech stack begins at the substation.
