Distributed and edge AI compute.

The practice covers fleets of small inference nodes, from a single box to a small multi-GPU system, placed in homes, small businesses, retail locations and energy sites. Nodes may run alone or alongside existing energy and network equipment on the same site.

We work for energy and hardware companies that are evaluating whether to build, deploy or operate such a fleet, and on what terms. The same independence applies as in our datacenter work: we hold no reseller quota and take no vendor incentive. A recommendation not to build is a valid outcome.

Each question below is answered per site type and per configuration, with the source of every figure stated.

• Power and circuits. Sustained and peak draw per node, the circuit it shares, breaker and panel capacity, and how the node behaves when the site load or tariff changes.
• Heat and noise by placement. Thermal envelope and acoustic level in a garage, utility room, living space, back office or equipment enclosure, including summer ambient temperatures.
• Backup power and ride-through. Whether the node needs a UPS, an existing battery or no backup at all, and what ride-through time each workload actually requires.
• Connectivity. Wired primary link, cellular backup, bandwidth per workload, and network isolation from the host's own devices and traffic.
• Physical security and remote management. Tamper resistance, encrypted storage, secure boot, out-of-band access and remote recovery for a node nobody on site will touch.
• Safety certification and residential use. Electrical safety listings, emissions class for residential environments (for example FCC Class B), and what installers and inspectors will require.
• Field service and spares. Failure modes, mean time between failures, swap-versus-repair policy, spares stocking, and the cost of a truck roll against the value of the node.
• Fleet orchestration and job routing. Provisioning, updates and monitoring across many nodes, and how jobs are routed to available capacity when individual nodes go offline.
• Uptime by workload type. Which workloads tolerate a node that is intermittently unavailable, which need redundancy across nodes, and which belong in a datacenter instead.

1. 01Requirements. Workloads, site types, placement constraints, service model and the commercial questions the fleet must answer, recorded as written requirements.
2. 02Market scan and requests for information. A structured scan of compute hardware, connectivity, power and fleet-software vendors, followed by written requests for information and quotes.
3. 03Evaluation against written criteria. Responses are scored against criteria fixed before the responses arrive. Gaps, assumptions and unverified vendor claims are stated as such.
4. 04Reference design and bill of materials. A node specification per configuration, with an itemized bill of materials, power and thermal budget, network design and management stack.
5. 05Pilot plan. A staged plan to validate the design on a small number of sites, with the measurements, acceptance criteria and decision points defined in advance.

• Written requirements and evaluation criteria
• Vendor landscape covering hardware, connectivity, power and fleet software
• Request-for-information package and a comparison of the responses received
• Reference design per configuration, with an itemized bill of materials
• Per-node power, thermal, acoustic and network budget
• Operating model: installation, remote management, field service and spares
• Cost model per node and per fleet, with every input sourced
• Pilot plan with acceptance criteria and a written recommendation

To discuss a distributed or edge compute fleet, describe the workload, the site types and the timeline at gpusmith.com/contact. Datacenter-scale work is described on the home page and about the firm.