Decode the blinking boxes.

One managed compute system; its identity is distinct from the rack containing it.

A structure that supports equipment and its power, cooling and cable connections.

A vertical mounting increment of 1.75 inches; rack elevation records where equipment belongs.

The verified hostname, serial number, rack and slot that tie a ticket to the intended physical device.

The set of components affected by a shared dependency; compare one device, one rack and wider scope before choosing a cause.

A component designated for service replacement by its manufacturer; replaceability does not imply live-removal permission.

General-purpose processor executing operating-system and application instructions.

Processor optimized for highly parallel work; compare expected hardware inventory with driver-visible inventory.

A removable memory module on systems that use sockets; not every server uses removable DIMMs.

Error-correcting memory can correct some errors and detect others; trends and uncorrectable events need context and escalation.

A storage protocol commonly used by PCIe SSDs; a missing namespace or device does not identify which part failed.

The peripheral interconnect used by many GPUs, NICs and drives; a path fault can resemble a failed endpoint.

Baseboard management controller providing hardware management independently of the host OS. Reachability does not prove the host is healthy.

Management through a path separate from the production OS/data path; it can remain reachable during a host outage.

Management-controller event history useful for correlating power, fan and temperature symptoms.

DMTF standard for management APIs; access and permissible actions remain site-specific.

Power-on self-test: early firmware checks before the operating system starts.

Power supply unit converting its input to power used by system electronics. Input loss and internal PSU failure are different diagnoses.

Power distribution unit; trace labeled outlets and feeds without opening or operating upstream electrical equipment.

Separate designed supply paths; two cords provide useful redundancy only when their upstream dependencies and capacity support it.

Uninterruptible power supply that supports continuity during input disturbances; actual facility design is site-specific.

Conductive distribution structure; vendor rack examples use it to deliver power from shelves to trays.

Rack-level assembly containing power supplies; distinct from a compute tray.

Controlled isolation of hazardous energy using the employer energy-control program and trained, authorized personnel.

Electrostatic-discharge controls protect electronics during approved handling; they do not make energized electrical work safe.

Approved steps, prerequisites, responsibilities, rollback and validation for a planned operation.

Site-specific emergency response; follow it rather than improvising around electrical, thermal or liquid hazards.

Coolant distribution unit; in liquid-to-liquid designs it manages a technology loop and exchanges heat with a facility loop. Other architectures exist.

Heat-transfer component in contact with a processor that transfers heat into coolant.

A supply or return distribution assembly serving multiple liquid paths.

Designed fluid coupling; disconnect only under the platform procedure and required isolation. Quick does not mean consequence-free.

Temperature difference. For a steady single-phase fluid path, heat transfer is approximately Qdot = mdot × cp × delta T: W = (kg/s) × (J/kg/K) × K. Use properties at the relevant conditions; this does not prescribe site limits.

The intended intake-to-exhaust route; blocked intakes or recirculation can create local thermal problems.

Equipment rejecting heat to outdoor air; the term alone does not determine whether server chips use liquid cooling.

Network interface controller linking a host to a network. Physical presence, link and IP reachability are distinct checks.

Interconnected network infrastructure that carries traffic among systems; identify which fabric a ticket concerns.

A high-performance fabric used in CoreWeave environments; the exact Cheyenne fabric remains unverified.

Remote direct memory access: data transfer between systems with reduced host CPU involvement; a capability, not a connector shape.

RDMA over Converged Ethernet; do not infer InfiniBand merely because RDMA is in use.

NVIDIA interconnect used between GPUs and switches in supported systems; separate from the external cluster network.

Data processing unit offloading infrastructure work; model and functions depend on platform.

Switch associated with a rack; its role may be management or data networking, so identify the actual path.

Whether an interface is enabled by configuration; IFF_UP is separate from physical carrier.

Driver-reported lower-layer link readiness. Missing carrier does not by itself identify cable, optic or port as the cause.

Rule selecting a next hop/interface for an IP destination; working link alone does not establish a usable route.

Name resolution. Failure to resolve a name differs from failure to reach a numeric IP address.

Fiber supporting a single propagation mode in its intended operating range; use the transceiver specification to choose compatible media.

Fiber supporting multiple modes; reach and compatibility depend on fiber grade and optical standard.

Small-form optical connector commonly used in duplex connections; connector shape alone does not identify media compatibility.

Multi-fiber connector family; MTP is a branded MPO implementation. Fiber count, keying, polish and polarity must match the design.

Module converting electrical and optical signals; match supported type, speed, wavelength, media and reach.

The end-to-end mapping that connects transmitter channels to intended receiver channels.

Transceiver telemetry such as received optical power and temperature where supported; interpret against vendor limits.

Direct-attach copper and active optical cable assemblies respectively; follow supported cable and port combinations.

Initial physical inspection and certification gate before automated provisioning proceeds.

Firmware preparation stage in CoreWeave Day 1.

Observation stage before the Day 1 Test stage.

Validated inventory eligible for customer delivery.

Workflow for investigating production problems and deciding the next recovery path.

Return merchandise authorization/vendor repair workflow; preserve identifiers and follow data handling policy.

Evidence that the repaired system meets required health checks before return to service.

NVIDIA driver error identifier: a clue requiring context, not a complete root-cause diagnosis.

Measured status, counters and events; correlate timestamp, scope and baseline before deciding what a signal means.

Confirm authority, affected services, maintenance scope and rollback before a disruptive action.

A concise record of identity, impact, time, observations, actions, validation and unresolved questions.

Developer confirms planned first-phase capacity and air-cooled chillers; exact GPU generation, rack layout and fabric are not established.

Exact installed GPU generation and OEM remain unverified; vendor reference hardware is a learning example.

Structure a real example as situation, task, your actions and results; add what you learned without inventing metrics.