How to Test a Control-Plane Claim

Start with a failure that had nothing to do with software ambition. On July 10, 2024, a lightning arrester failed on a 230-kilovolt transmission line in Virginia. Automatic reclosing tried to restore the line several times, producing a series of voltage dips within about a minute and a half. The North American Electric Reliability Corporation, or NERC, later reviewed the event and found that roughly 1,500 megawatts of load disconnected at once. Most of it was data centers whose protection systems, reacting to repeated voltage disturbances, switched to backup power. Much of that load did not return for hours.

Nothing in that event was malicious or unusual on its own. Each facility's controls did what they were designed to do. The problem was structural. Many independent control systems, each rational in isolation, acted together in a way the grid operator had not anticipated. That is the risk profile of the control plane in one example. It is powerful precisely because it acts automatically, and automatic action at scale has system effects.

This gives a way to evaluate any control-plane claim. Ask five questions, in order.

First, which layer does it act on? A claim about cooling efficiency acts on the building management system and the cooling plant. A claim about utilization acts on the scheduler. A claim about grid flexibility must act on all layers at once, from the utility signal to the chip. A vague claim can conceal a single-layer product being sold as a whole-stack solution.

Second, what does it measure, and how fast? A control system can only act on what it sees. Accelerator telemetry from NVIDIA's Data Center GPU Manager, or DCGM, arrives per device and quickly. Facility telemetry from the electrical power monitoring system arrives at the switchgear. Utility signals arrive on their own timescale. A claim that a system responds to grid events within seconds should explain which data path makes that possible.

Third, what is the evidence type? There is a ladder. A reference design is a tested pattern. A demonstration is a single controlled event, such as the Phoenix test by Emerald AI, Oracle, NVIDIA, Salt River Project and the Electric Power Research Institute in 2025. A contract is a commitment with consequences, such as Google's 2025 demand response agreements with Indiana Michigan Power and the Tennessee Valley Authority. An operating record is repeated performance over time, such as the failure and recovery data Meta published in its July 2024 Llama 3 paper. Each rung is worth more than the one below. In Faraday's read, most marketing sits on the bottom rung.

Fourth, what is the baseline? Vendor announcements often describe gains as multiples, and NVIDIA's own material for Mission Control in March 2025 claimed up to five times higher infrastructure utilization and up to ten times faster job recovery. The phrase up to describes the best case. The useful question is: compared with what starting point, on what workload, measured how? A large improvement over a poorly run cluster says little about a well-run one.

Fifth, what happens when it fails? The Virginia event is the reason to ask. Every control system eventually receives bad data, loses a link or meets a condition its designers did not foresee. A credible claim describes the fallback: whether the system fails safe, whether a human can override it, and whether many sites running the same logic could act in unison. NERC's review is a reminder that correlated automatic behavior is a grid-level risk, not only a site-level one.

Apply the questions to something concrete. Suppose an operator says its new campus is grid-interactive. Ask which layers are integrated. Ask what signal it receives from the utility and how that reaches the scheduler. Ask whether the claim rests on a contract, a demonstration or a design document. Ask how much load it has actually shed, for how long and on which dates. Ask what its protection systems do during repeated voltage dips. An operator with a real control plane will answer each question with a specific. An operator with a brochure will answer with an adjective.

Faraday's read is that the control plane is becoming the most consequential and least visible part of an AI factory. It determines how much of the installed capital does useful work, how the facility behaves under stress and whether the grid will connect it at all. It rarely appears in announcements, which favor megawatts and chip counts. Learn to ask about it, and the rest of an AI infrastructure claim becomes much easier to judge.