Formula 1 is the pinnacle of real-time engineering.
A modern F1 car runs roughly 300 sensors. Tire temperature. Brake temperature. Suspension travel. Ride height. Steering angle. Fuel flow. Battery performance. Aerodynamics. Hydraulic pressure. Vibration. G forces. Every one of those sensors feeds data back to the pit wall at about 1.1 million data points per second.
A room full of engineers watches every data point in real time. They are not just monitoring the car. They are making economic decisions. Every tenth of a second of lap time is worth real money in constructor points, sponsorship performance, and prize fund allocation. The engineers know exactly what each sensor reading costs if it moves in the wrong direction.
That is why Formula 1 teams win races. Not because the car is fast. Because they understand the economic consequence of what the data is telling them while the race is still happening.
Now consider a 1 gigawatt AI factory.
The sensor is not the insight. The economic consequence of the sensor is the insight.
A campus at that scale operates somewhere between 250,000 and one million physical sensors across the full stack. You are monitoring: 50,000 to 100,000 GPUs. Tens of thousands of CPUs. Thousands of power distribution units. Hundreds of switchgear breakers. UPS systems. Transformers. Cooling towers. Chillers. Pumps. Liquid cooling distribution units. CRAH and CRAC units. Valves. Flow meters. Pressure sensors. Humidity sensors. Network switches. Fire systems. Weather stations. Gas generators. Battery energy storage systems.
Each of those assets exposes its own telemetry. A single GB300 rack can surface hundreds of individual data points. One CDU exposes 50 to 150 points. A chiller commonly exposes 100 or more. A generator controller often exposes 200 to 500. A modern UPS may expose hundreds of measurements.
The data is not the problem. AI factories generate tens to hundreds of millions of telemetry points per second. The infrastructure exists. The sensors are there. The data is flowing.
Operators know when a chiller has an alarm. They know when a power circuit trips. They know when a GPU cluster throws an error. What they do not know is what any of those events costs in real time across the full stack. They are reading individual sensor values. They are not reading economic consequence.
Formula 1 figured this out decades ago. An engineer does not look at tire temperature in isolation. They look at tire temperature in the context of lap time, fuel load, gap to the car ahead, pit stop window, and constructor standings. Remove any one of those connections and the decision degrades. The sensor value alone is not actionable. The economic consequence of that sensor value — in context — is.
AI factory operators are in the position of knowing their tires are warm. They are not running the race.
The $1 million per megawatt figure is not a round-number estimate. It is derived from two independently measurable economic layers in every AI campus.
The tenant layer is where most of the value lives. An H100 GPU rents at $2.23 to $3.99 per hour on committed contracts from major cloud providers.[2] NVIDIA's own DGX H100 reference design puts power consumption at 10.2 kW per 8-GPU server, yielding approximately 800 H100s per megawatt of compute load.[3] At a conservative $2.50 per hour blended rate, that is $17.5 million per megawatt per year at 100 percent utilization.
Published surveys put actual GPU utilization in production AI environments at 50 to 60 percent for large optimized facilities — and considerably lower for most enterprises.[4] The gap between realized utilization and potential utilization is the economic loss. Recovering 4 to 5 percentage points of that gap generates approximately $750,000 per megawatt per year — without adding a single GPU.
The facility layer adds the rest. At the national wholesale average of $196 per kilowatt per month,[1] one megawatt of delivered capacity generates $2.35 million per year in facility revenue. PUE inefficiency, cooling degradation, and power quality events typically consume 10 to 15 percent of that before it reaches the compute load. Recovering that layer adds approximately $250,000 per megawatt per year.
One million dollars per megawatt per year. Sitting in the gap between what the sensors are telling you and what you are doing about it economically.
At Synestria, we built our platform around a metric called Economic Availability. EA measures the full operational stack in real time. Power generation. Electrical distribution. Cooling capacity. Network throughput. Compute utilization. Workload execution. Six domains, one number, updated continuously.
EA is not a monitoring metric. When EA drops, the system shows you exactly where in the chain the degradation is occurring, exactly what the causal path looks like, and exactly what it is costing in compute output and dollars if nothing changes. When EA recovers, the improvement is directly measurable in production output. We bill on gain-share. Nothing until we deliver a documented improvement.
The data center industry has operated for twenty years on PUE as its primary efficiency metric. Power Usage Effectiveness tells you how much of the total power drawn actually reaches the IT load. It is a useful number. It has driven real progress on cooling and power infrastructure.
PUE tells you nothing about whether the IT load is doing anything useful.
A campus can run a PUE of 1.15 and have GPUs sitting idle while jobs queue upstream waiting for a network bottleneck to clear. PUE does not see that. The ESG report does not see that. The tenant sees it in their monthly utilization statement.
EA sees it. In real time. Before the queue backs up. Before the throttle fires. Before the revenue disappears.
PUE became a household name in data centers because it gave operators a single number to optimize. Before PUE, efficiency was a conversation. After PUE, it was a number.
EA needs to do the same thing for economic output.
If you are operating an AI campus and you do not know your EA number, you are running a Formula 1 car without a pit wall. The sensors are all there. The data is all flowing. You are just not reading what it costs.