Expert Matilda Bailey is a distinguished networking specialist whose work lies at the intersection of cellular technology, wireless evolution, and next-generation power solutions. With a career dedicated to optimizing how data moves and how infrastructure scales, she offers a unique perspective on the modern data center as an “enclosed power grid.” In this discussion, she explores the groundbreaking potential of power orchestration to solve the industry’s most pressing bottleneck: the availability of electricity. By leveraging granular intelligence and high-speed control, Bailey explains how operators can reclaim massive amounts of “stranded” energy to fuel the generative AI revolution without waiting years for utility upgrades.
The following conversation delves into the mechanics of the Karman platform, the technical challenges of sub-millisecond power monitoring, and the strategic shift toward “intelligent kilowatts” in AI factories.
Data centers often maintain significant power headroom for redundancy, leaving a large portion of contracted electricity unused. How does orchestrating power at the rack level transform this “stranded” energy into usable compute, and what specific safety protocols ensure that redundancy isn’t compromised during a sudden hardware failure?
The traditional approach to data center design is incredibly conservative because the cost of downtime is so high. For instance, in a standard setup where a rack has four power feeds, the system is designed so that if one line fails, the remaining three can carry the entire load, which often means those lines are intentionally capped at 75% utilization. This creates a massive pool of “stranded” energy that sits idle just in case of a worst-case scenario. By implementing an orchestration layer like Karman, we can safely tap into that headroom by monitoring the environment with extreme precision and adjusting GPU utilization in real-time. If a hardware failure occurs, the system’s control response is faster than a human could ever react—under 20 milliseconds—allowing it to immediately throttle workloads or rebalance power flows to maintain stability. This ensures that redundancy is managed through intelligent, high-speed software rather than just leaving expensive electricity on the table.
Achieving response times under 20 milliseconds requires massive data processing at the edge. What are the technical hurdles of monitoring a million samples per second across a cluster, and how does this level of granularity change how operators manage the unpredictable power spikes common in AI inference workloads?
Monitoring at a rate of 1 million samples per second is a monumental data challenge because you are essentially trying to capture the “heartbeat” of the electrical current itself. The primary technical hurdle is processing that volume of information locally without creating its own massive power draw or latency, which is why utilizing custom modules, like those from Nvidia, is so critical for this embedded intelligence. This level of granularity is a game-changer for inference workloads, which are notoriously “bursty” and unpredictable compared to steady-state model training. When you can see a power spike the microsecond it begins, you can allow the system to run much closer to its theoretical limit because you have the visibility to “shave” those peaks before they threaten a breaker. It transforms power management from a passive, reactive task into a proactive, high-fidelity operation that treats electricity as a dynamic resource.
Grid connection delays are currently stalling new facility builds for years. For an operator with a fixed power limit, what are the step-by-step phases of retrofitting for power intelligence, and what metrics should they use to measure the return on investment compared to waiting for a utility upgrade?
The first phase of a retrofit involves installing embedded AI computers into the existing power infrastructure to gain visibility; you can’t manage what you can’t measure with high resolution. Once that intelligence layer is active, the second phase is integrating the software with the GPU orchestration layer to allow for dynamic workload throttling based on real-time power availability. The final phase is the “unlock,” where you actually scale up the number of physical servers in the rack to utilize that newly reclaimed 50% capacity boost. To measure ROI, operators should look at the “effective rack density” and the “time-to-market” for new compute capacity; if you can squeeze 130 MW of compute out of a 100 MW grid connection today, you are generating revenue years ahead of a competitor waiting for a substation upgrade. It effectively turns a fixed utility constraint into a flexible software-defined asset.
The industry is shifting from a focus on hardware availability to a focus on “intelligent kilowatts.” In what ways will an integrated power operating system become a standard requirement for AI factories, and how might this technology eventually migrate from massive data centers to smaller, edge-computing environments?
As we move toward “AI factories,” the power operating system becomes the essential fabric that links the electrical grid to the silicon, ensuring that every watt is doing productive work. We are seeing a shift where the smartest kilowatt wins, meaning the operator who can squeeze the most FLOPs out of a single megawatt will have the best margins. This technology is naturally starting with “neocloud” providers who are unburdened by legacy thinking, but it will inevitably migrate to the edge as AI becomes embedded in local infrastructure. In smaller edge environments, where power supplies are often even more constrained and less reliable than hyperscale centers, an orchestration layer is vital to prevent local grid overloads while running high-performance tasks. Eventually, this “power OS” will be as standard as a hypervisor is today, managing energy distribution across everything from massive rural clusters to localized urban edge nodes.
What is your forecast for AI power orchestration?
I believe that within the next five years, we will stop talking about data center capacity in terms of “total megawatts contracted” and instead talk about “orchestrated capacity.” My forecast is that power orchestration will become a mandatory requirement for any facility looking to host next-generation GPUs, as the cost of leaving 30% to 50% of your power stranded will be economically unsustainable. We will see a move toward “grid-safe” data centers that can actually communicate back to the utility, offering flexibility to the grid while maximizing their own internal compute density. Ultimately, the integration of power and compute will become so tight that the hardware itself will dynamically signal its energy needs millisecond-by-millisecond, effectively turning the entire global AI infrastructure into one giant, hyper-efficient, self-balancing energy ecosystem.
