The Infrastructure Bottleneck: Why Scaling AI Requires More Than Just Power Plants

The core thesis of this conversation is that the energy crisis surrounding AI is not a failure of generation, but a failure of transmission and conversion efficiency. While the public focus remains on building more power plants, the real bottleneck is the outdated, passive, and mechanical infrastructure that routes electricity from the grid to the chip. The hidden consequence of this calcified supply chain is that we are wasting massive amounts of potential compute power as heat. The advantage belongs to those who shift from mechanical, 50-year-old hardware to active, software-defined power electronics. This is essential reading for infrastructure architects and energy investors who need to look past the more power narrative to understand the systemic efficiencies that will define the next decade of AI scaling.

The Hidden Cost of Passive Infrastructure

The electricity grid has functioned with minimal load growth since the 1980s, leading to a stagnant supply chain. Utilities have been incentivized to prioritize capital expenditure on maintenance rather than innovation, resulting in a reliance on heavy, oil-filled, mechanical transformers. Baglino notes that these systems are not just inefficient; they are a boat anchor on the path to electrification.

If you are not in a growth industry what are you doing a lot of the talent left the states and went to areas that were growing... the whole supply base became siloed and a little bit uncompetitive.

-- Drew Baglino

When we attempt to scale AI data centers, which are essentially massive heat-conversion engines, using this legacy infrastructure, we lose significant energy in the conversion process. Baglino argues that by replacing passive, 60Hz mechanical switches with wide-bandgap semiconductor technology like Silicon Carbide and Gallium Nitride, we can shift to high-frequency power management. This allows for galvanic isolation and voltage conversion at hundreds of kilohertz, shrinking equipment by 100x and recapturing roughly 35 megawatts of compute power for every gigawatt of grid capacity.

The Parking Lot Problem of Battery Scaling

The physical limitations of current compute infrastructure are mirrored in how we store energy. Baglino introduces the parking lot analogy to explain why charging speeds plateau. In a lithium-ion battery, lithium ions must find parking spots in the anode and cathode. As the battery fills, the remaining spots become harder to access, creating a bottleneck.

If you imagine that the parking lot at a stadium... has exactly the number of spots as there are people going to the stadium... if the parking lot were set up as a line and the person could literally just drive along until they see the hole they could find their spot really quickly.

-- Drew Baglino

Most current battery architectures struggle with this 2D-to-3D transition. The implication is that increasing charging speed is not just a matter of more power from the charger; it is a fundamental materials science challenge. Unless we optimize the physical layout of the battery to resemble a linear parking lot, we will continue to hit diminishing returns on charging speed, regardless of how much power we feed into the system.

Why Data Centers Could Actually Lower Your Electric Bill

Conventional wisdom suggests that AI data centers are a drain on the grid, driving up costs for residential consumers. Baglino flips this: data centers are the best utility customers because they provide base load demand. In a system where utility costs are amortized over total kilowatt-hours served, data centers inflate the denominator, which can drive down the per-unit cost for everyone else, provided the utility does not overbuild or unfairly shift the infrastructure costs to residential ratepayers.

This creates a systemic opportunity: if data centers are equipped with energy storage to manage training ripples, they cease to be a grid liability and become a grid asset. They can stabilize frequency and voltage, effectively acting as a grid support network. The downstream effect is that regions with high data center density could see lower electricity rates over time, provided the regulatory and incentive structures allow the data centers to pay for their own infrastructure.

Key Action Items

• Audit your infrastructure dependencies: Identify long-pole mechanical components in your supply chain, such as switchgear or transformers, that have not seen innovation in decades. These are your primary points of failure and inefficiency. (Immediate)
• Decouple ego from engineering work: Adopt a fail fast culture where technical solutions are discarded the moment better data arrives. If you are defending a design because you spent three months on it, you are slowing your own progress. (Immediate)
• Prioritize Grid-Positive Data Center Design: If building or investing in compute facilities, prioritize integrated energy storage. This transforms the data center from a volatile load into a stabilizing asset. (12 to 18 months)
• Shift to High-Frequency Power Electronics: For new builds, replace passive, oil-filled transformers with active, solid-state power electronic solutions. This reduces footprint by 100x and improves efficiency, creating a lasting competitive advantage in compute-per-watt. (12 to 24 months)
• Re-evaluate the Long Pole in Project Schedules: When managing hardware projects, avoid letting the longest lead-time item dictate the entire project timeline. Use soft tooling or prototyping to compress the schedule, even if it adds short-term cost. (Ongoing)