A battery is not redundancy.

Plugging an NVIDIA DGX Spark into a portable power station is easy.

Building an AI service that survives a long outage is not.

A battery can prevent a reboot. It cannot survive a failed inverter, a dead switch, a corrupted model store, or a two-node model that needs both computers to answer. A generator can refill a battery. It cannot create software failover. Two computers can double capacity without adding redundancy at all.

The compute is large. The electrical load is small.

Each DGX Spark has 128 GB of unified memory, 10 Gb Ethernet, ConnectX-7 networking, and support for models up to 200 billion parameters on one unit. NVIDIA describes approximately 405-billion-parameter operation across two Sparks.

NVIDIA supplies a 240-watt adapter. Its regulatory documentation reports 233.2 watts at maximum load and 38 watts at idle for the complete system. Two Sparks therefore have a published combined ceiling of about 466 watts. Add the router, switch, local storage, monitoring, ventilation, and conversion losses, and 600 watts is a useful first planning load for a compact two-node critical bus.

Do not size from the Spark's processor figure. Size from the complete system, measured workload, supplied adapters, and every device required to keep the service usable. The nominal 48-volt device rating is also not permission to connect a battery directly; use NVIDIA's approved power adapter.

Size energy from average load. Size equipment from peak load.

Battery runtime begins with a simple planning equation:

Using a 600-watt critical load, 80% usable battery capacity, and 90% conversion efficiency:

These are planning calculations, not equipment specifications. Temperature, battery age, auxiliary loads, charger behavior, shutdown reserve, and measured duty cycle determine actual runtime.

Use measured average watts to calculate hours. Use maximum load, charger demand, startup behavior, wiring limits, and expansion headroom to size the inverter, protection, and generator.

The battery carries the outage. The generator carries the calendar.

The battery becomes the immediate source when utility or solar disappears. It bridges generator startup, absorbs load changes, and gives the control system time to decide what stays on. The generator then feeds the inverter/charger and restores battery energy.

That architecture is better than making the AI ride directly on raw generator transitions. For higher criticality, each Spark can sit behind its own online double-conversion UPS or protected downstream branch. An online UPS continuously rebuilds its output and has no transfer interval when input power disappears.

A little generator is enough—when it is doing the right job.

A representative 2.2 kW portable inverter generator may provide only 1.8 kW continuously. That can support a roughly 600-watt AI bus and limited charging. It cannot support 600 watts of live load plus an uncontrolled 1.5 kW charger with healthy reserve.

For a two-Spark node, a generator around the 3 kW continuous class creates a better envelope: critical load, controlled charging, conversion losses, and transient margin. Continuous output matters more than the larger surge number printed on the enclosure.

A recoil-start camping generator is an attended emergency charger, not autonomous infrastructure. Unattended endurance needs compatible electric or remote start, low-state-of-charge control, warm-up and cool-down logic, generator-aware input-current limiting, fuel monitoring, periodic exercise, and a documented manual path.

Portable generators remain outdoors, at least 20 feet from doors, windows, and vents, with exhaust directed away from occupied buildings. Fixed wiring needs approved transfer equipment and a qualified electrician. Never backfeed a building through an improvised cord.

Two Sparks can be redundant. A two-Spark model is not.

If the critical model fits on either Spark, keep a complete copy on both nodes and put them behind a health-aware local endpoint. One can be active and one warm, or both can serve requests.

If a large model is sharded across two Sparks, both are required to answer. The pair is one larger computer. Losing either machine loses the service. True redundancy for that workload requires two independent two-node groups—four Sparks—or a smaller fallback model that fits completely on either individual node.

NVIDIA's clustering support increases memory and throughput. It does not automatically create high availability. The service still needs replicated weights, health checks, routing, durable task ownership, and recovery logic.

Protect the complete service.

The router, switch, local DNS, storage, authentication, monitoring, and communications path all belong on protected power. Keep model weights, containers, retrieval indexes, source documents, recovery images, and operating procedures on site. Avoid one shared NAS as the only copy.

Agent work needs a durable queue outside the worker process. Tool calls should be idempotent or transaction-logged so failover does not repeat a payment, deletion, message, or equipment command. A warm spare is useful only when it inherits the right job state.

Cooling is also part of availability. The Spark's electrical load becomes heat inside the enclosure. A cabinet designed only for battery warmth can cook the compute; a ventilated cabinet can expose LFP batteries to low-temperature charging conditions. Treat the battery and compute as adjacent systems with different thermal requirements.

The node should get smaller before it goes dark.

At full charge, keep both replicas warm. As state of charge falls, stop fine-tuning, embeddings, and batch work first. Then reduce concurrency, camera frame rate, and context size. Switch to a smaller quantized model. At critical reserve, preserve only local retrieval, communications, security, and essential monitoring before checkpointing and shutting down cleanly.

Power policy should be deterministic. State of charge, runtime estimate, generator status, and service priority select the operating tier. Do not ask the language model whether the battery feels low enough to turn itself off.

Three reference architectures

Reference architectures show failure-domain logic, not a permit-ready electrical design. Final equipment selection depends on code, site conditions, ambient temperature, fuel strategy, and measured workload.

Run the dark test.

Complete the first Spark setup on stable power; NVIDIA warns against interrupting initial installation. Before remote deployment, verify automatic boot after power loss, model autostart, local routing, and recovery media.

Then pull the grid under load. Refuse the first generator start. Remove one Spark. Reboot the switch. Corrupt one local model copy. Disconnect the WAN. Restore service with queued work waiting. The test passes only when the critical endpoint remains available or fails into a known, honest state—and every interrupted job is visible afterward.

Track actual watts, battery runtime, generator starts, failed starts, recharge time, node readiness, failover time, queue age, model reload time, thermal alarms, and manual interventions. Redundancy exists only after the failure test proves it.

FAQ