Infrastructure & Fleet Platforms

GPU and server fleets at the scale that breaks most tools. Architected to hold there.

A fleet management platform that works cleanly at a hundred servers is a different system at twenty-five thousand. We architect and build the platforms that discover, monitor, configure, and control physical hardware at fleet scale: bare-metal servers, GPU estates, mixed and heterogeneous hardware, over the protocols the hardware actually speaks. This is our heaviest work, and most of it has no machine learning in it at all.

What is at stake

This is infrastructure other systems depend on, which means its failures cascade. A control plane that cannot keep up, a telemetry pipeline that falls behind, an access model that does not hold, and the blast radius is the entire fleet and everything running on it. The defining challenge is scale and heterogeneity together. The platform has to work across thousands of nodes that are not identical: hardware from different vendors, firmware that varies across an estate purchased as uniform, and baseboard controllers whose protocol implementations have their own quirks. Most platforms are never designed for the scale they are eventually asked to carry, and the cliff does not appear until the scale does.

RedfishIPMIBMCDCGM
designed for the fleetthe pilot architecturethe scale cliffnodescontrol-plane throughput
A discovery or polling model that is correct at a hundred nodes saturates at ten thousand. The cliff does not appear until the scale does.
The failure patterns we see here

Failure in a serious system is rarely random. These are the shapes we look for first.

01

The session-limit constraint

A baseboard controller supports only four to eight concurrent sessions. Naive fan-out polling across tens of thousands of nodes opens far more than that and crashes the very controllers it is trying to monitor. Polling has to be scheduled against that limit, per node, at fleet scale, and it is invisible until the fleet is real.

02

Protocol variance under a shared standard

Redfish is a DMTF standard, and a standard is a starting point rather than a guarantee. iDRAC 8 and iDRAC 9 differ, iLO 4 and iLO 5 differ, optional schema fields drop inconsistently, and error behavior is not uniform. Code that trusts the specification works in the lab and fails intermittently against the real estate.

03

Telemetry the controller does not expose

The GPU signals that matter most on a GPU fleet, ECC error counts, thermal throttling state, and XID errors, are not surfaced over Redfish or IPMI at all. Getting them requires in-band agents on the host, which is a foundational architecture decision about the deployment model rather than a late addition.

04

The scale cliff

Concurrency invisible at small numbers becomes the binding constraint at fleet scale. A discovery or polling model correct at a hundred nodes saturates at ten thousand, the architecture that worked in the pilot cannot carry production, and the rebuild arrives mid-flight and expensive.

How we help

The same method, in your language.

Architect

We architect fleet platforms from first principles for the scale and heterogeneity they will actually face: the control-plane design, the protocol strategy, the access model, and the observability. This is where the cost of the whole program is decided.

Build

We build them end to end: the drivers that talk to the hardware, the hardware abstraction layer over vendor dialects, the services that move telemetry and push verified control, and only then, if it belongs there at all, the intelligence layer on top.

Harden + Operate

We harden platforms against the real estate and stay on as it grows, because a fleet is not a fixed target and the assumptions that hold today are a function of the node count.

Diagnose

When an existing platform hits the scale cliff or breaks against heterogeneous hardware, we establish where it actually fails and engineer the path through it, including the honest answer about whether that path is a repair or a rebuild.

Control plane
fleet scale
RedfishIPMISNMPSSH
heterogeneous fleet
One control plane, four protocols that every vendor implements a little differently, twenty-five thousand nodes. The spec is a starting point, not a guarantee.
Why us, here

This is the clearest case of the highest-value work having no model in it. A fleet platform is overwhelmingly systems engineering: drivers, protocols, control planes, distributed telemetry, access control, and scale. Intelligence, when present at all, is one feature sitting on top of it. We architected one control plane over a mixed-vendor server and GPU fleet across three out-of-band protocols behind a single API, where the hard part was detecting the absence of a signal, because the node that goes quiet is the one that matters and naive polling never notices silence. Polling was scheduled against baseboard controller session limits so monitoring never degraded what it monitored, and in-band agents carried the GPU telemetry the controller does not expose. A vendor whose product is intelligence scopes the intelligence and assumes the platform. An integrator staffs the platform broadly and routes the firmware questions elsewhere. Both leave you owning the seam between the control plane and the hardware, which is where this class of program fails.

Explore another industry
How we engage

You tell us the situation.
We tell you what we see.

Tell us the system, the failure or the blank page, the stakes, and the date that matters. If you are in an urgent window, say so. We prioritize accordingly.

Start a conversation
mostafa@opulion.dev · Response within 24 hours · By inquiry