GPU-as-a-Service Business Model

GPU-as-a-Service (GPUaaS) revenue comes in distinct quality tiers, each with different financing impact and operational characteristics.

Tier 1 (highest quality): Availability-based contracts where customer pays for guaranteed availability ($1M/month for 100 dedicated GPUs, available 24/7) regardless of utilisation. Revenue is highly predictable, collected monthly, decoupled from customer performance. Examples: Anthropic's internal cloud compute, large enterprises with SLA-backed contracts.

Tier 2: Take-or-pay contracts with minimum commitments ($200K/month minimum, customer pays for hours consumed above minimum or pays minimum regardless). Revenue contractually guaranteed up to minimum; customer absorbs utilisation risk. Examples: Meta, Google purchasing reserved capacity from CoreWeave, Lambda.

Tier 3: Fixed-term committed pricing (12-month reserved instance at $4.50/hour, per-hour billing). Revenue locked at agreed price; operator benefits from commitment but bears utilisation risk if customer underutilises.

Tier 4: Committed pricing with periodic resets (6-month commitment at $4.50/hour, auto-renews at then-current market rate). More flexible than Tier 3 but introduces repricing risk if rates decline sharply.

Tier 5: Standard on-demand pricing ($5.00/hour, no commitment, customer can stop any time). Revenue volatile and customer-driven; operator bears full utilisation risk.

Tier 6 (lowest quality): Spot/interruptible pricing ($2.50-$3.00/hour, customer accepts eviction with minimal notice). Lowest, most volatile revenue; suitable only for time-insensitive batch workloads.

Financing impact is dramatic: Tier 1 supports DSCR 1.40-1.60x at same gross margin as Tier 5, enabling lower borrowing costs and higher leverage. Tier 5 requires DSCR 1.00-1.20x due to utilisation volatility. Tier 6 spot requires DSCR 0.90-1.10x and is often excluded from senior debt covenant calculations.

Diversified portfolios mix revenue tiers strategically: 30-40% Tier 1-2, 30-40% Tier 3-4, 20-30% Tier 5-6. This structure creates a revenue floor (Tier 1-2 covers fixed costs and debt), upside from spot utilisation, and pricing flexibility. Tier 1-2 contracts require 24+ month commitments negotiated directly.

Tier 3-4 reserved instances are self-serve portals where customers choose terms (1-3 years) and payment methods. Tier 5-6 on-demand and spot are fully dynamic. Revenue assurance: prepayment/advance billing reduces payment risk; credit holds ($5-20K per account) back unpaid usage; audit rights verify logs; auto-scaling stops workloads at zero balance.

CoreWeave and Lambda require prepayment on on-demand; take-or-pay contracts allow net-30. Contract expiration timing matters. If 50%+ expire simultaneously, repricing triggers 20-30% utilisation drop. Staggered expiration with auto-renewal clauses (customer opts-out only) mitigates churn.

Gross margin is revenue per GPU/hour minus power, cooling, and networking costs. At $5.00/hour and $0.30-$0.40/hour opex, gross margin is $4.60-$4.70/hour (92-94%).

Gross margin is tier-agnostic (all tiers price $4-6/hour); Contribution Margin (gross margin less variable CAC) is more useful. On-demand CAC runs $1-2K per customer (~$0.33/hour over 2-year 6,000-hour lifetime).

Tier 1-2 contracts cost $50-100K annually in direct sales engineering and account management; for a $1M/month contract, this spreads to $0.05-$0.10/hour over term. CAC payback: 2-4 months on-demand; 6-12 months enterprise. Sales team composition drives margin: 100% enterprise requires 20-40% headcount (bleeds opex), 100% on-demand minimises sales cost but invites utilisation volatility. Optimal mix: 20-30% enterprise/committed, 70-80% on-demand, yielding 40-50% EBITDA margin.

CoreWeave (public S-1): Tier 2-3 focus, $5.50-$6.00/hour, 70-75% gross margin, ~50% EBITDA. B2B enterprise model (80%+ top 20 customers) with high pricing power from scarcity; limited on-demand mix avoids race-to-bottom. Lambda Labs: 40% Tier 3-4, 60% on-demand; $4.50-$5.50/hour, 75% gross, 35-40% EBITDA.

Bottoms-up developer model (long tail <$5K/month) with high CAC but developer loyalty retention. Together AI: 30% Tier 2, 70% on-demand; $3.50-$4.50/hour, 70% gross, 25-30% EBITDA. API bundling (GPU as COGS) requires low pricing; on-demand mix compresses margin.

Hyperscalers: 60-70% proprietary (internal), 30-40% external at $2-4/hour standard, $10-15/hour custom; 60-70% gross, 40-50% EBITDA on external (shared capex leverage). No single model dominates. CoreWeave trades scale for margin via enterprise scarcity.

Lambda prioritises developer adoption and stickiness. Together AI builds API consumption network. Hyperscalers monetise volume and shared infrastructure. Optimal depends on sales capability, customer access, and market conditions.

Operators misjudge when competing on headline pricing ($4.50 vs. $4.75/hour) whilst ignoring contract quality. A $5.00/hour cluster with 50% Tier 1-2 contracts (revenue floor, DSCR 1.35-1.50x, 50-60% LTV debt) outweighs $4.50/hour on-demand-only (utilisation 40-80%, DSCR 1.00-1.20x, 30-40% LTV).

Lenders advance 50-60% LTV at 300-400bps on Tier 1-2 portfolios vs. 30-40% LTV at 500-700bps on on-demand. Debt cost arbitrage is material.

Tier 1 at 300bps ($5.00/hour, $4.00/hour CFADS) beats on-demand at 600bps ($4.50/hour, $3.50/hour CFADS) by ~$6-9M annually on $75M capex. Contract quality creates the real moat.

Timing matters. Supply-constrained periods (2023-2024) let operators lock 2-3 year Tier 1-2 terms at premium pricing. As supply normalises, committed customers renegotiate at lower rates but renew; on-demand-first operators face rapid churn and undercut spirals. Optimal strategy: build Tier 1-2 base day one (highest margin, capital-efficient), then add on-demand/spot to absorb slack. Inverse (on-demand first) creates churn and repricing vulnerability.

Revenue tiers: Tier 1 availability-based (DSCR 1.40-1.60x) >> Tier 2 take-or-pay (1.30-1.50x) >> Tier 5-6 on-demand/spot (1.00-1.20x); contract quality determines financing capacity more than headline price

Optimal portfolio mix: 30-40% Tier 1-2 (committed), 30-40% Tier 3-4 (reserved), 20-30% Tier 5-6 (on-demand), balancing revenue floor with upside

Customer acquisition: enterprise contracts $50-100K/year sales cost (6-12mo payback), on-demand $1-2K/customer (2-4mo payback); CAC composition drives EBITDA margin structure

Contract quality arbitrage: $5.00/hour Tier 1 (50% of revenue committed) supports higher leverage and lower debt cost than $4.50/hour on-demand-heavy, creating competitive moat