Due Diligence

Comprehensive GPU infrastructure due diligence spans five evaluation domains, each scored 1-5 (1=critical deficiency, 5=institutional strength). Hardware/Procurement domain: GPU specification (matching workload requirements), supplier diversification (single-source vendor risk), equipment age (deployment cohort and refresh cycle implications), bulk purchasing efficiency (demonstrating cost discipline).

Networking/Interconnect domain: fabric architecture (fat-tree vs rail-optimised decision alignment), switch redundancy (N+1, N+2 configurations), latency performance (production vs specification benchmarking), upgrade path clarity (topology scalability). Cooling/Power domain: cooling technology selection (immersion vs closed-loop), power distribution redundancy (N+1 electrical delivery), backup power capacity (UPS hold-up duration, generator run hours), thermal monitoring (temperature variance and anomaly detection).

Software/Operations domain: orchestration platform maturity (Kubernetes, custom schedulers, monitoring dashboards), automation depth (manual intervention rate, runbook coverage), customer onboarding processes (time-to-production, training provided), incident response procedures (MTTR targets, escalation paths). Security/Compliance domain: data isolation (multi-tenant segregation, memory scrubbing), access controls (RBAC implementation depth), audit trails (compliance logging, regulatory proof), physical security (perimeter access, internal movement controls). Domain scores below 2.5 indicate operator viability risk; below 3.0 suggest material capex investment required within 12-24 months.

Hardware procurement review validates capex assumptions against actual contract terms and supplier credibility. Key assessment points: Contract volume commitments (non-cancellable vs renegotiable); Allocation certainty (supplier supply commitments vs customer demand); Pricing locks (fixed pricing duration vs escalation clauses); Payment terms (COD vs extended financing); Lead times (deployment timing credibility). Red flags: NVIDIA GPU allocations dependent on neocloud operator direct sales (circular dependency); long-lead components (switches, power supplies) with 16-24 week delivery creating deployment slippage risk; single-source suppliers without secondary options; spot market purchasing indicating short-term capital constraints.

Recent engagements uncovered operators maintaining 8-12 week GPU inventory buffer masking customer churn; high inventory accumulation with flat revenue indicates demand shortfall masked by prior-period over-purchasing. Networking fabric assessments frequently reveal engineering over-specification: operators procuring dual-fabric redundancy (two complete 800Gbps switching layers) creating 25-35% capex waste compared to N+1 switch redundancy. Credible procurement strategies show: 24-36 month GPU allocation agreements with escalation clauses tied to market pricing, diversified switch sourcing (Nvidia, Arista, Cumulus), and strategic inventory (4-8 week buffer, not 12+).

Revenue quality assessment decouples gross revenue from sustainable cash contribution. Five critical metrics: Top-10-customer concentration (target <60% for infrastructure players; >75% indicates concentration risk); Customer cohort churn (monthly 25-35% normal, >40% indicates product-market deficiency); Contract term length (30-day commitments vs 12-month agreements indicate stickiness); Pricing realisability (contracted vs realised rate; >8-12% misalignment indicates discount creep); Customer acquisition cost payback (8+ months indicates poor unit economics). Red flags: month-to-month contracts with 2+ week cancellation windows (highly elastic demand); aggressive promotional pricing offering 35-50% discounts; customer concentration in single industry vertical; recent customer wins from direct hyperscaler competition.

A recent engagement revealed operator reporting 92% YoY revenue growth concealing 58% customer cohort churn; each cohort retained at 5.2 month LTV paid back CAC but acquisition velocity required ongoing price competition eroding margins. The operator faced cash flow cliff when customer acquisition decelerated: prior cohorts retained insufficient base to absorb acquisition investment drop. Revenue quality assessment requires 24-month rolling cohort analysis, not headline growth rates.

Infrastructure readiness assessment maps capex work-in-progress against customer demand and financing runway. Four key flags: (1) Facilities under construction with customer commits covering <40% capacity—speculative capex creating stranded assets; (2) Power infrastructure upsizing on 18+ month lead time but capacity deployment on 8-10 month cadence creates utilisation ceilings below projections; (3) Cooling system commissioning delayed >4 weeks relative to schedule compounds thermal management risk; (4) Networking fabric deployment phased across 6+ months introduces operational complexity and customer experience inconsistency. Financing structure risks manifest in three patterns.

Refinancing cliffs emerge when construction-stage debt matures into permanent debt requiring higher refinancing rates if utilisation misses targets. Covenant tightness—base case assumptions of 85% utilisation and 3% annual pricing decline—leaves minimal headroom to 1.20x DSCR thresholds. Subordination layers grant mezzanine debt holders blocking rights on asset sales or leverage adjustments, severely limiting operational flexibility.

The most dangerous structure pairs asset-backed financing with speculative capex expansion, locking leverage at 60%+ debt-to-capital on facilities still ramping. Customer acquisition slowdowns and pricing declines then leave operators with no flexibility to adjust capex or absorb losses. Review structures for: senior debt covenants <1.25x DSCR threshold (permitting modest utilisation variance), capex deployment phased to actual demand signals (not speculation), and equity retaining >8 quarters of cash runway below base-case breakeven.

Five-domain technical scoring: hardware/procurement, networking/interconnect, cooling/power, software/operations, security/compliance; scores <2.5 indicate viability risk

Hardware red flags: NVIDIA allocation circularity, >12-week component lead times, single-source suppliers, spot market purchasing masking capital constraints

Revenue quality decouples headline growth from sustainable cash: top-10 concentration >75%, churn >40%, LTV payback >8 months, pricing realisability variance >8-12% signal deficiency

Infrastructure readiness risks: speculative capex <40% customer commits, thermal commissioning delays, refinancing cliffs, covenant tightness <1.25x DSCR threshold on permanent debt