As AI systems scale, the infrastructure challenge is no longer just about chips, models, and software performance. It is increasingly about the physical systems that allow computation to happen at all: power delivery, cooling, water access, and the speed at which new capacity can be brought online. Power conversion is becoming a much more important design question. As racks move from conventional power densities toward megawatt-scale configurations, every inefficiency in the electrical pathway becomes more consequential. For stakeholders across the energy sector, that makes AI infrastructure more than just a datacenter story. It is also a story about grid constraints, industrial load growth, thermal management, and how developers can design facilities that are efficient enough, flexible enough, and resilient enough to operate at the scale AI now demands.

Host Sylvia Leyva Martinez is joined by Nick Wright, Vertical Solutions Manager at Siemens. Their conversation explores why the growth of the AI factory is pushing operators to rethink traditional electrical architecture, especially the number of conversion steps required to move power from the grid to the chip. Nick explains why conventional AC-heavy setups are under pressure as compute loads become denser, more dynamic, and more power-intensive, and why more direct AC-to-DC pathways are drawing increased attention. The episode also examines what that shift means in practice: less energy lost in conversion, less excess heat to manage, different implications for cooling design, and a growing role for higher-voltage DC systems, digital twins, monitoring technologies, and new protection equipment. Along the way, the discussion widens beyond the building itself to consider how AI facilities may evolve into more grid-aware assets, capable of interacting more intelligently with the broader energy system rather than functioning simply as passive loads.

For developers, IPPs, utilities, financiers, and infrastructure planners, the episode offers a clear signal that power architecture is becoming a strategic decision much earlier in the project lifecycle. One of the key takeaways is that this is not a simple story of DC replacing AC. The more relevant point is that as racks scale, reducing unnecessary conversion steps can improve efficiency and system performance in ways that matter economically at very large scale. But the conversation also makes clear that conversion efficiency is only one part of a much broader infrastructure equation. Access to reliable power, water availability, cooling strategy, workforce readiness, supply chain bottlenecks, equipment lead times, and safety considerations all shape whether a new AI facility can be delivered on time and scaled over the long term. The players most likely to succeed will be the ones that stop treating power as a late-stage procurement issue and instead plan holistically across energy, compute, operations, and grid interaction from the beginning.

This episode is brought to you by twentytwo & brand -- a marketing and PR agency built specifically for energy leaders.

Lots of agencies say they work with energy companies, twentytwo & brand was built for them. They've partnered with more than 120 companies driving the energy transition — from growth-stage startups to globally recognized industry leaders. Media relations, brand design, video, paid advertising, and community engagement — they cover it all under one roof. No onboarding lag, no industry crash course -- they speak your language on day one.

If you're ready to sharpen your story and supercharge your marketing, find them at twentytwoandbrand.com/woodmac.

See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

A year ago the data centre conversation was about scale. Increasingly it is about what happens when the announcements meet the physical grid. New capacity is being announced at roughly 435 megawatts a month, enough to power a city of 400,000 people, but two-thirds of that committed load tends to disappear the moment utilities ask for a financial commitment behind it. AI training facilities create load profiles that drop 30% in five minutes, or 190 megawatts in three. The grid's mechanical inertia is retiring just as hyperscaler ambition accelerates, and regulation is lagging on both sides of the meter.

In this episode, interim host Bridget van Dorsten returns the microphone to host Sylvia Leyva Martinez. The two recap six months of the show through clips from Chris Seiple (Wood Mackenzie), Tom Falcone (Large Public Power Council), Akhil Batheja (Bloom Energy), Kay Aikin (Dynamic Grid), Kristina Carlquist and Christian Payerl (ABB), Shannon Miller (Mainspring Energy) and Nick Chaset (Octopus US), and map out the questions that will shape Sylvia's return.

The central tension of the past six months: announcements are racing ahead, but utilities, regulators and the physical grid cannot move at hyperscaler speed. A data centre can be built in two years. New generation takes five to ten. That mismatch is why developers have stopped waiting on the grid: 35% of US data centre project capacity announced in 2025 was planned with around-the-meter generation, and 92% of bridge prime power deals are now struck before the end tenant is signed, inverting the usual logic of infrastructure development. But collocated power is still, in Bridget's framing, a science project. The load behaviour demands a coordinated portfolio: supercapacitors and UPS catching millisecond swings, synchronous condensers supplying inertia, fuel cells and linear generators offering modular, fuel-flexible bridging power as a hedge against demand risk. Nick Chaset's intervention cuts the other way: the UK already hosts the world's largest residential virtual power plant, and the cheapest megawatt is the one you don't build. The episode closes on duelling forward views, Tom Falcone's cooperative optimism against Kay Aikin's affordability death spiral, with the question of who ultimately holds the bag if the announcements don't materialise still unresolved.

Sylvia sets out what she wants to explore next: the regulatory contradiction where utilities support bring-your-own-generation but cannot guarantee protection from curtailment; renewables supply constraints and the transformer and labour bottlenecks that affect every fuel source equally; the return of energy security as a framing now that "energy transition" has fallen out of political favour in the US; and the community opposition data centre developers are only beginning to grapple with.

See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

The energy transition conversation focuses on what connects to the grid. Far less attention goes to whether anyone is coordinating what those assets do once connected. AI training runs swing hundreds of megawatts in seconds as GPUs checkpoint and restart a profile that looks like a generator tripping offline. At distribution level, millions of inverter-based resources create localised variability that overwhelms individual circuits even when aggregate models look healthy. The planning tools in use today were designed for neither problem.

Host Bridget van Dorsten is joined by Kay Aikin, CEO and Founder of Dynamic Grid, energy engineer, grid architecture advisor to the DOE-supported GridWise Architecture Council, and contributor to the UN Environmental Program's building decarbonisation work. Kay unpacks what an AI training facility actually does to the grid with full GPU load for hours or days, then a drop to ten percent in seconds during checkpointing. She talks about how at the scale now planned, the Stargate project in Texas alone could represent ten percent of ERCOT disappearing in four seconds. The behaviour is stochastic and cannot be modelled with traditional statistical tools. At distribution level, virtual power plants responding to wholesale signals without circuit-level visibility can create competing oscillations, the kind of emergent dynamics that contributed to the Spanish grid failure.

The proposed fix is an AI controller at the substation, sending price-based signals and flexible operating envelopes to large assets and VPP operators, giving them twenty-four-hour forecasts and real-time circuit visibility. Total cost: under a hundred thousand dollars installed. The reason it isn't everywhere is cost-of-service regulation. Utilities earn returns on deployed capital, so a million-dollar transformer replacement is more profitable than software that eliminates the need for it.

Without new approaches, rebuilding the US distribution grid could cost up to ten trillion dollars by 2040. Kay is developing grid utilisation metrics with regulators in Maine, Virginia, and Maryland to incentivise extracting more from existing infrastructure. The episode closes on the need for distribution system operators and the affordability death spiral that looms if the structural incentives don't shift.

See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.