A modern accelerator is a 700-watt space heater that prefers cold air. Everything else about edge AI infrastructure follows from that one fact, and most of the field hasn't internalized what it implies, because the implications come out of electrical engineering and not software engineering.
So here's a practical view of power budgeting for AI deployed away from a data center. This is the part of the conversation everyone glosses over when a vendor wheels a ruggedized box onto a trade-show floor. It's also the part that decides whether the box works in the environment the customer actually needs it in. That customer is usually a logistics officer, a plant operator, a field engineer, or a colocation manager. Not one of them signed up to design an electrical room around an AI deployment, and that's exactly what the deployment is quietly asking them to do.
The math, stated plainly
Start with the power draw of the silicon. A current-generation accelerator at sustained workload pulls somewhere between 350 and 700 watts, depending on the variant. I'll plan around 700, because the workload that justifies buying the thing is hardly ever the idle workload.
A node typically carries eight accelerators. Eight times 700 is 5,600 watts. Now add CPU, memory, storage, networking, and the supporting power-conversion losses. A realistic node draws between 7 and 10 kilowatts under sustained load. Plan around 8.
A single rack with three nodes draws 24 kilowatts. A four-rack pod draws close to 100 kilowatts. In a colocation facility those numbers are routine. At the edge they're extraordinary.
The edge runs on a different power envelope. A US 120V 20A circuit delivers 1.9 kilowatts. A 208V 30A drop delivers roughly 5. A typical commercial 480V three-phase service lands somewhere around 75 kilowatts. A military FOB generator, the B-1 unit, puts out around 100 kilowatts on a good day, and less than that when the temperature is extreme or the fuel has degraded.
Here's the first practical takeaway. One current-generation AI node eats roughly half of a typical commercial three-phase drop. Two nodes is the practical ceiling on a single drop if you want any margin left for the rest of the building. None of that is what the slide deck implies.
The scenarios you will actually encounter
Four sketches. Each one brings its own failure modes.
Scenario one. Saudi Aramco diesel. A field deployment in the Gulf states usually runs on whatever generators are locally available. The diesel is contaminated by regional standards. The frequency drift runs past North American ranges. The voltage is nominally 380V three-phase but sags under load to levels that would shut down a US data center. The accelerator's power supply has to auto-range across all of that. Data-center-grade supplies generally won't tolerate these conditions, while industrial-grade supplies will, and the cost gap between them is real. The marketing literature doesn't always mention it.
Scenario two. CONUS shore-based stadium event. A sports-broadcasting deployment in a US city has shore power, but the contracted electrical service was sized for the broadcast crew, not for an AI inference cluster bolted on after the fact. The transformer feeding the truck park is rated for 50 kilowatts. The deployment wants to add 30. That leaves 20 kilowatts of margin for the entire rest of the broadcast. The math closes on paper and then falls apart the first time a lighting change draws a transient that shoves the transformer past its rated load. The deployment trips offline at the exact moment it was needed.
Scenario three. Tactical FOB B-1 generator. Military forward operating base power is the harshest environment you'll meet in the field. The B-1 generator is rated for 100 kilowatts, but real-world derating for ambient temperature, fuel quality, and continuous load drags the practical capacity down to 70 kilowatts or less. The AI deployment shares that with everything else the FOB has to keep running. A two-node AI cluster takes 16 kilowatts off the top, leaving 54 for the rest of the base. The FOB commander has to make a tradeoff, and it has to be made consciously, with a full read on the AI cluster's power profile. Make it blind and the FOB loses something far more important than AI.
Scenario four. Industrial colocation. A manufacturing plant decides to host an AI inference cluster in its existing motor control center. The MCC has spare capacity, but the power quality is awful, because the manufacturing equipment throws harmonic distortion the AI cluster's power supplies were never designed to ride out. The cluster runs for three weeks and then fails in a way the failure symptoms don't explain. Diagnosis eats another two weeks. The lesson, learned the hard way every time, is that AI infrastructure has to sit away from heavy industrial loads, or behind isolation transformers the original deployment plan never budgeted for.
Those four cover a meaningful slice of edge AI deployments. The common thread is that the power assumption baked into the AI hardware design gets met badly by the real environment.
The fuel logistics nobody mentions
A 100-kilowatt deployment at the edge burns roughly 8 to 10 gallons of diesel an hour, at a generator efficiency of 12 to 15 kilowatt-hours per gallon. Round-trip, sustained, that's about 200 gallons a day for a 100-kilowatt load.
200 gallons a day is a logistics burden. One fuel truck per week, per site. The fuel has to come from somewhere. The route has to be secured if the deployment sits in a hostile environment. The storage has to exist on site. An AI cluster drawing 100 kilowatts isn't a software deployment at all. It's a logistics commitment that happens to run inference.
Deployments that ignore this find one of two failure modes waiting for them. Either they run dry and go dark right when the operation needs them, or they burn more fuel than anyone budgeted and turn into a financial problem that gets the whole thing cancelled before it ever produces operational value.
The right answer is to size the deployment to the fuel logistics rather than the reverse. Your AI capability budget sits downstream of your logistics budget. Vendors who pretend otherwise aren't being honest about what they're selling you.
The failure modes in production
I'll name three I've watched recur, and that deployment plans almost never anticipate.
Failure mode one. Thermal cascade. The accelerators run hotter than the cooling system was sized for. The cooling system works harder, its own power draw climbs, and now the upstream power supply sits closer to its limit. The next environmental excursion trips the breaker and the cluster goes offline. Diagnosis isn't obvious, because what you see is a tripped breaker, not a thermal limit. The fix means resizing thermal and electrical headroom together, as one problem.
Failure mode two. Brownout-induced corruption. The local utility, or the local generator, drops voltage for a moment. The accelerator's power supply rides through the dip. The CPU's power supply doesn't. The system reboots, the in-flight inference is corrupted, and the audit log records an inconsistent state. The fix is a UPS sized for the worst brownout you've actually observed, which is generally not what the deployment plan budgeted.
Failure mode three. Silent derating. The accelerator's firmware detects sustained high temperature and throttles quietly. The cluster's measured throughput slides 30 percent over two weeks and nobody notices, because the workload still finishes, just slower. The customer concludes the cluster was underspecified. The real cause is thermal. The fix is to monitor and alert on accelerator clock frequency, not just utilization.
What to specify
A procurement document for an edge AI deployment should pin down the following with numbers instead of adjectives.
Sustained power draw at peak workload, in kilowatts. Maximum brownout duration the system tolerates, in milliseconds. Operating temperature range at full performance, in Celsius. Operating humidity range, in percent relative humidity. Power supply input range, in volts and Hertz, with explicit auto-ranging behavior. Fuel consumption rate at peak workload, if generator-powered, in gallons per hour. Cooling capacity required to hold the deployment in spec, in watts of heat rejection.
Vendors who can answer those in writing, with measured numbers from a representative deployment, know what they're selling. Vendors who answer in adjectives are selling you a slide deck.
A closing observation
Edge AI deployment is an electrical engineering problem wearing a software engineering label. The deployments that succeed treat it that way from the start. The ones that fail discover the engineering halfway through, after the budget's gone and the schedule has already slipped.
The customers most exposed to this are the ones who never had to treat power as a primary constraint before. That's most enterprise IT teams, a lot of SOC operators, and a real fraction of the defense procurement community. The right preparation is to learn the math before the procurement, not after the deployment starts falling over.
The math isn't hard. It just has to get done.