Preview v2 · Nvidia standardisation boundary

DRAFT PREVIEW v2 · NVIDIA STANDARDISATION BOUNDARY

Where does Nvidia's spec end?

V2 · POST-CHATGPT-REVIEW

ChatGPT flagged the v1 preview options as too static (Option 1) and too rack-focused (Option 3). The recommendation: redraw Option 2 as a dual-path event-to-action architecture that shows Nvidia's control plane crossing the rack/facility boundary via DSX event exchange, with critical safety actuation staying on the customer side.

The redesign incorporates new primary evidence: Nvidia's DSX Exchange event bus, its stateless async subscription model, and NICo's own documented split with BMS: "Critical safety actions remain with BMS; NICo handles orchestration and automation for non-critical conditions."

Option 2 (v2) · Dual-path control architecture

Two parallel sensor paths converge on NICo + Mission Control. Path A (Nvidia-managed) runs cold plate + inner-manifold sensors through the BMC. Path B (BMS-managed) runs facility-scale sensors (rope, spot, CDU, drip pan) through the customer BMS, but exchanges events with NICo via DSX Exchange. Below the boundary line, physical safety actuation (breakers, valves, facility escalation) stays customer-owned.

The boundary being drawn is logical/control-plane, not physical. Nvidia's orchestration plane crosses the rack/facility line via DSX; critical safety actuation does not.

Nvidia's leak-detection control path — dual sensor paths converging on NICo, with a critical-safety boundary at BMS OPTION 2 · v2 · CONTROL PATH · DUAL-SENSOR EVENT-TO-ACTION Nvidia's own docs draw a clean boundary. The event and orchestration path is Nvidia-defined and crosses the rack/facility line via DSX event exchange. Critical safety actuation stays with the customer's BMS. NVIDIA-DEFINED · EVENT AND ORCHESTRATION PATH PATH A · COMPUTE-TRAY Cold plate + manifold 2 sensors per tray BMC Redfish exposure PATH B · FACILITY-SCALE Rope · spot · CDU + drip pan sensors BMS facility-managed DSX event bus NICo + Mission Control Convergence: both paths land here Discovery · classification policy engine · workload API Health + allocation Block new allocations Migrate workloads workload plane Tray shutdown Graceful, gravity-aware of affected + downstream trays actuation plane · compute BMS event exchange Request rack isolation via DSX stateless async bus event plane NVIDIA · CUSTOMER · CONTROL INTERFACE CUSTOMER · CRITICAL SAFETY ACTUATION Rack AC breaker BMS opens circuit Site-picked breaker firmware critical safety, BMS-owned Coolant valve closure BMS actuates valve Site-picked valve topology critical safety, BMS-owned Facility-scale response Escalation, ticketing Onsite operator dispatch operator + BCM ticketing The boundary in one line Nvidia can standardise the event and orchestration path without owning the physical facility safety system. The sensor standardises; the economic control point moves into the event, policy and workload response layers. SOURCES · [P1] Nvidia Infra Controller · Key Capabilities docs.nvidia.com ("Critical safety actions remain with BMS; NICo handles orchestration") · [P1] docs.nvidia.com/dsx-exchange/architecture + stateless-async-bus · Mission Control 2.3 BMS integration · adikumar.co · 2026-09-21

Primary evidence: [P1] docs.nvidia.com/dsx-exchange/architecture (event bus, "BMS → NICo: coolant leak events so NICo can cordon nodes and migrate workloads in seconds"), [P1] docs.nvidia.com/dsx-exchange/stateless-async-bus (subscription model, five consumer categories), [P1] docs.nvidia.com/infra-controller/documentation/overview/key-capabilities (NICo/BMS split verbatim), [P1] Mission Control 2.3 BMS-BCM integration.

Reference · v1 Option 1 (split panel)

Kept for reference. ChatGPT verdict: too static, reads as a procurement boundary rather than a control architecture. Not recommended.

Nvidia standardisation boundary — what Nvidia owns vs what stays open OPTION 1 · SPLIT-PANEL · WHAT NVIDIA HAS AND HAS NOT STANDARDISED STANDARDISED BY NVIDIA Sensor placement Cold plate + inner manifold per tray Redfish metric paths rf_chassis_0_leakdetector_N_* NICo discovery BMC endpoint query pattern Severity taxonomy General / Severe / Critical + actions NICo → BMS handoff Escalation pattern documented Node-state annotation leak:large + 10-min timeout Scope: compute-tray sensing and Nvidia's own control plane from BMC through NICo to BMS handoff pattern. SPECIFIED. INTERFACE. SEQUENCE. OUT OF NVIDIA'S REACH The BMS itself Which BMS, which breaker firmware, which policy engine, which dwell timer, which interlock topology. Customer side. Facility-scale detection Rack rope, drip pan, CDU sensors. ~2/3 of total cable metres in the §04b topology math. BMS territory. Sensor supply Cable brand, fluid chemistry, false-alarm suppression algorithm. Design-in AVL. Multi-vendor fight. Site-specific integration Which valves, which breakers, which actuator topology. Nvidia says "BMS actuates." How is site choice. Where operators, specialists and cooling vendors continue to compete on qualification. CUSTOMER-SIDE. IMPLEMENTATION. QUALIFICATION. SOURCES · [P1] Mission Control 2.2 admin guide + Nvidia Infra Controller · leak-detection-handling docs.nvidia.com · adikumar.co · 2026-09-21

v1 Option 1 · comparison-first split panel.

Reference · v1 Option 3 (physical boundary map)

Kept for reference. ChatGPT verdict: risks implying "Nvidia = rack, customer = everything outside rack" which isn't the actual boundary. Nvidia's control plane crosses the physical rack/facility boundary through BMS integration via DSX. Not recommended as primary.

Nvidia standardisation as a physical boundary — inside the tray vs outside OPTION 3 · PHYSICAL BOUNDARY · INSIDE THE TRAY VS OUTSIDE Nvidia's reference architecture ends at a physical boundary: the compute-tray chassis. Inside is spec. Outside the chassis, on the facility side, everyone else still competes. FACILITY BOUNDARY · BMS-managed · outside Nvidia's spec Rack rope perimeter Specialist-supplied cable. Wires directly to BMS. Drip pan spots Under-floor + under-CDU. BMS integration only. CDU sensors Pressure, flow, dielectric. CoolIT / Motivair / Vertiv. BMS policy + actuator Breaker firmware, valves, dwell. Site-specific implementation. NVIDIA COMPUTE-TRAY standardised Cold plate idx 0 Manifold idx 1 Redfish path via BMC to NICo, then handoff to BMS. NICo→BMS Inside the tray: Nvidia's spec applies. Two sensors, Redfish exposure, NICo pipeline, severity taxonomy. Outside the tray: BMS-managed. Facility sensors, actuator topology, policy engine. Nvidia has no spec.

v1 Option 3 · physical rack/facility zones.

Recommendation: ship Option 2 (v2) as the primary visual for §04c, positioned to answer the "how much of the stack does Nvidia now own?" question directly. Keep Option 3 as an optional secondary illustration if the physical grounding helps a specific paragraph.

Thesis to sharpen in §12: current wording ("the response path and controls would continue to differentiate") is too broad. New wording that survives the Nvidia-standardises-everything critique: "The sensor can standardise; the economic control point moves into the event, policy and workload response layers."

Not shipping to the essay yet. Confirm and I'll integrate into §04c + tighten §12 in one PUT.