Cluster Economics: Key Assumptions

GPU cluster utilisation drives revenue and debt service capacity. Contracted (take-or-pay) utilisation typically reaches 80-95%; spot-heavy utilisation ranges 60-75% due to interruptible, price-sensitive workloads. Large AI/ML training customers (OpenAI, Anthropic, Databricks) run at 85-95% on reserved capacity; inference-heavy customers (Hugging Face, API providers) operate at 70-85%; opportunistic research jobs occupy spot at 40-60% utilisation.

Blended utilisation typically settles 75-85% in steady state (year 2-3), declining modestly to 70-80% in years 4-5 as newer clusters launch and older Blackwell transitions to lower-margin inference. Conservative models assume 80% utilisation (year 1-2), declining 2% annually; aggressive models sustain 85%.

Utilisation assumption drives 15-25% EBITDA margin swing. Moving from 75% to 85% utilisation at $5.00/hour adds $4.4M annual revenue on a 10MW cluster, improving EBITDA margin 3-5 percentage points.

GPU hardware follows a three-phase lifecycle with distinct pricing and margin profiles.

Phase 1 (0-12 months, current-gen premium): Blackwell B200 commands $5.50-$6.50/hour on-demand, with training workloads (large model training, fine-tuning) consuming capacity at premium pricing. Phase 1 margin: 70-75% gross, 40-50% EBITDA.

Phase 2 (12-36 months, inference transition): As training completes and inference scales, pricing declines to $4.50-$5.50/hour. Mix of training and inference customers creates blended revenue. Gross margin 65-70%; EBITDA margin 30-40%.

Phase 3 (36+ months, long-tail multi-tenant): Older Blackwell or H100 workloads (low-latency inference, data processing) accept $3.50-$4.50/hour. Workloads are low-volume and low-margin. Gross margin 60-65%; EBITDA margin 15-25%. Previous-gen H100 pricing has historically declined 15-25% per year, trending toward $2.50-$3.50/hour by 2026-2027.

Models should assume: year 1 blended pricing $5.00-$5.50/hour (Phase 1-2 mix), year 3 declining to $4.50-$5.00/hour (Phase 2-3 mix), year 5 stabilising around $4.00-$4.50/hour (Phase 3 dominant). Pricing decay of 8-12% annually is conservative; faster decay (15-20% annually) is plausible if supply exceeds demand.

GPU clusters require capital refresh every 3-6 years to remain competitive. Current-gen Blackwell accelerators cost $30-40K per unit in bulk; a 10MW cluster requires $48-64M refresh capex. Hyperscalers (Google, Meta, Microsoft) typically amortise over 4-5 years; neoclouds (CoreWeave, Lambda, Crusoe) typically use 3-5 year depreciation. Faster depreciation (3 years) reflects accelerating innovation but increases annual expense; slower depreciation (6 years) reduces expense but risks obsolescence.

For financial models, assume 4-year depreciation (straight-line 25% annual charge) for current-gen clusters. This aligns with market observation that Blackwell will transition to inference-only by year 3-4, supporting slower refresh in years 4-6 as next-gen hardware (Rubin) launches. Accelerator salvage value typically ranges 40-60% of original capex for secondary market H100 SXMs (older H100 units reselling at approximately $18,000-$22,000, about 50-60% of $30-35K original cost in strong secondary markets), creating material secondary market revenue.

Models should assume: year 2 capex 5-10% of revenue (incremental capacity, network upgrades); year 4 capex 20-30% of revenue (major hardware refresh); year 6 capex 15-25% of revenue (refresh #2 or partial next-gen replacement). This pattern creates working capital and financing needs in years 4-6, necessitating equity reserves, refinancing, or reduced distributions.

Power costs typically represent 30-40% of total GPU cluster opex. Baseline: $0.08-$0.12/kWh for US colocation (Virginia, Texas, Oregon); $0.10-$0.15/kWh for European facilities; $0.06-$0.10/kWh for sovereign/subsidised facilities (Saudi Arabia, UAE, Iceland). A 10MW cluster costs $576K-$1.2M annually in electricity.

Per-GPU cost: at 700-1,200W thermal per GPU cluster-wide (Blackwell B200 standard), costs approximately $0.30-$0.50/GPU/hour in power. Stress scenarios: +20% power cost (renewable PPA expires) increases opex $115-240K annually; +40% scenario (grid constraints) increases opex $230-480K annually.

Facility opex (rent, cooling, networking, management) typically adds $500K-$1M annually for mid-size clusters, or $0.35-$0.70/hour equivalent cost. Liquid cooling reduces cooling opex 10-15% versus air cooling but adds $1-3M capex and requires vendor lock-in. Most models conservatively assume air cooling with incremental liquid cooling evaluated in sensitivity.

Staffing for a 10MW cluster typically requires 5-15 FTE (systems engineers, support, sales ops) at $80K-$150K/year (including benefits), totaling $400K-$2.25M annually. Larger clusters (20-50MW) benefit from sublinear staffing scaling, declining to 40-60% of revenue for largest players. Separate fixed costs (3-5 core engineers) from variable costs (customer support, sales ops scale with customer count).

Networking and connectivity (transit, DDoS mitigation, monitoring tools, software licensing) add $100K-$300K annually. Depreciation of non-accelerator capex (racks, power distribution, switches, cabling) typically adds 10-15% to accelerator depreciation. Combined, non-power opex typically adds 20-30% to power opex.

Include contingency reserve (5-10% of total opex) for equipment failure, supply chain disruption, or unplanned maintenance. Conservative models use 10% contingency; mature operators with proven systems use 5%.

Utilisation: contracted 80-95%, spot-heavy 60-75%; blended baseline 75-85% in steady state, declining 2% annually

Three-phase GPU lifecycle: Phase 1 (current-gen, $5.50-$6.50/hr on-demand, 70-75% gross margin), Phase 2 (transition, $4.50-$5.50/hr, 65-70% margin), Phase 3 (long-tail, $3.50-$4.50/hr, 60-65% margin);pricing data current as of Q1 2026

Hardware refresh: 4-year depreciation for current-gen; capex spikes in years 4-6 (20-30% of revenue) for major refresh cycles

Power costs: $0.08-$0.15/kWh baseline (30-40% of opex); stress scenarios (+20-40%) material to EBITDA sensitivity