# NM AI HPC Coordination Fabric **Status:** Concept brief / architecture control **Date:** 22 August 2026 ## Executive thesis New Mexico should treat advanced AI/HPC compute as strategic infrastructure: discoverable, schedulable, metered, governed and shareable across authorized public, research and commercial participants. The proposed control plane coordinates GB300-class compute, high-speed fabrics, storage and federated partner capacity. It brokers authorized compute airtime through auditable **compute-token allocations**. A compute token is an accounting and entitlement unit, not inherently a cryptocurrency or speculative asset. > Compute is the resource. Tokens are the allocation mechanism. Policy is the governor. The network is the fabric. AI assists the scheduler. New Mexico is the domain. ## Reference compute fabric The flagship tier uses GB300 NVL72-class systems. Inside each scale-up domain, GPUs communicate over NVLink/NVSwitch. Scale-out traffic between compute domains uses ConnectX-8 and XDR InfiniBand through Quantum-X800-class switching, with a non-blocking Clos/fat-tree target for tightly coupled AI/HPC jobs. The architecture separates three planes: 1. **Compute fabric:** NVLink scale-up plus 800 Gb/s-class InfiniBand scale-out for GPU east-west traffic. 2. **Data fabric:** high-throughput Ethernet/InfiniBand and NVMe-oF/parallel storage with GPUDirect-capable paths where appropriate. 3. **Management fabric:** physically/logically isolated OOB management, provisioning, telemetry, security and orchestration. ## Control plane Users request capability rather than a particular rack: GPU type/count, memory, duration, data classification, priority and locality. The control plane performs identity verification, authorization, policy evaluation, capacity discovery, placement, reservation, execution, metering and settlement. Core services include: - Identity, federation and Zero Trust policy - Slurm for HPC/batch scheduling - Kubernetes for services and inference - Capacity registry and provider federation - Policy-aware scheduler - Token ledger and allocation service - Usage metering and chargeback/showback - Audit/event ledger - Model and dataset governance - Energy-aware placement - Open APIs for authorized consumers and providers ## Compute-token model Tokens represent rights to measurable resources. Pricing or allocation weights can incorporate GPU model, GPU-hours, memory, CPU, storage, network, energy, priority and security requirements. Recommended classes: - **Mission:** critical government, emergency, cyber, water and infrastructure workloads - **Research:** universities, laboratories and approved scientific work - **Innovation:** startup and economic-development allocations - **Commercial:** paid surplus-capacity access - **Community:** education, tribal partnership and workforce programs - **Priority:** policy-controlled accelerated scheduling Commercial demand must never displace the **NM Sovereign Compute Reserve**. ## Sovereign compute reserve A defined percentage of capacity remains outside ordinary brokerage for wildfire, drought, water-resource forecasting, cyber incidents, emergency management, critical infrastructure, public safety and other sovereign missions. Emergency policy may preempt lower-priority workloads according to published rules. ## Settlement Every job creates an auditable record containing requester, provider, resource type, allocation, actual consumption, GPU-hours, storage/network/energy consumption, token debit, provider credit, security classification and immutable audit references. This supports grant accounting, interagency showback/chargeback, research allocations and commercial settlement without requiring a public cryptocurrency. ## Federation philosophy Institutions retain ownership and administrative control of their infrastructure and data. Approved excess compute may be advertised to the coordination fabric. Compute federation is separate from data federation: a provider can contribute compute without contributing datasets. The scheduler should bring authorized computation to protected data when feasible, enabling federated sovereign AI while preserving institutional boundaries. ## Design principles 1. **Mission before market.** Brokerage exists to strengthen public and research capacity, not consume it. 2. **Policy before placement.** No job reaches compute until identity, authority and data policy are resolved. 3. **Fabric before fleet.** Fast interconnect and topology are first-class architecture, not accessories. 4. **Measure everything.** Allocation, utilization, latency, energy and cost must be observable. 5. **Federate rather than centralize by default.** Coordination can create statewide capacity without requiring one owner of every GPU. 6. **Separate compute, data and management planes.** Performance and security both improve when responsibilities are explicit. 7. **Open interfaces, controlled participation.** Avoid architectural captivity while maintaining rigorous provider certification. 8. **Human authority remains explicit.** AI may optimize scheduling but policy, eligibility, emergency priority and economic rules remain governed decisions. ## Roadmap ### Phase I: Foundation Deploy initial GPU, fabric, storage, Slurm/Kubernetes, identity, telemetry and metering. ### Phase II: Accounting Measure GPU-hours, CPU, memory, storage, network and energy. Issue internal compute credits. ### Phase III: Federation Connect approved institutional capacity through standardized discovery, submission, policy and accounting APIs. ### Phase IV: Exchange Enable approved providers to advertise surplus capacity and authorized consumers to obtain allocations under transparent pricing and governance. ### Phase V: Autonomous coordination Use AI-assisted scheduling to optimize performance, cost, energy, locality and mission priority while preserving deterministic policy controls and human accountability. ## Strategic outcome The objective is not merely a supercomputer. It is a statewide coordination layer through which heterogeneous New Mexico AI/HPC resources can operate as an auditable, policy-governed computational commons.