Preemptible VM Compute Market Maker: A Full-Dimensional Deep Analysis
Chapters 一、做市商的核心逻辑 / I. Core Logic of Compute Market Makers
Compute Market Maker = Price Arbitrage + Interruption Fault Tolerance + Multi-Cloud Redundancy + Standardized Access ① 以 Spot/Preemptible 折扣价（60-90% off）采购GPU算力 Procure GPU compute capacity at Spot/Preemptible discount prices (60-90% off).
Absorb interruption risks through Checkpoint/Restore + automatic multi-cloud switching.
Resell with an “Enterprise-grade SLA” premium (yet still remaining 80%+ cheaper than public cloud On-Demand).
Earn risk premium — essentially underwriting the “interruption risk”. 二、四层做市商模型 / II. Four-Tier Market Maker Model
Margin
Tier 1
Aggregator: P2P matchmaking, 5-15% commission Vast.ai, RunPod Comm. 5-15%
Tier 2
Smart Broker: ML predicts interruption + auto multi-cloud switching Spot by NetApp 3-5%
Tier 3
Vertical Integration: Self-built GPU DCs, SLA guarantees CoreWeave, Lambda 30-50%
Tier 4
DePIN: On-chain reverse auction + Token incentives Akash, io.net
Token Protocol fee <5% • • • • 🦞 Preemptible VM Compute Market Maker Full-Dimensional Deep Analysis 三、H100 GPU价格分层（关键洞察） / III. H100 GPU Price Stratification (Key Insights)
Tier 1
Public Cloud On-Demand (Anchor Price) $72 - $98 / hr Tier 2 (Source)
Public Cloud Spot (Procurement source) $24 - $49 / hr Tier 2 (Specialist) 专业GPU云 On-Demand Specialized GPU Cloud On-Demand (CoreWeave/Lambda) $2.06 - $2.99 / hr Tier 2 (Specialist Spot) 专业GPU云 Spot Specialized GPU Cloud Spot $1.50 - $1.80 / hr Tier 3
Decentralized Marketplace (Vast.ai/io.net) $0.80 - $2.00 / hr Tier 4
Consumer-Grade Batch Processing (Salad) $0.02 - $0.10 / hr
Core Finding: Public cloud pricing is up to 47 times higher than independent GPU cloud providers. The market maker’s pricing anchor should target the CoreWeave tier instead of AWS.
H100 Price Trend: From 2024 Q1 to 2025 Q2, Spot prices dropped from $3-5/hr to $1-1.50/hr due to capacity expansion and the B200 substitution effect. 四、单位经济学（以H100为例） / IV. Unit Economics (H100 Example) 成本端（每GPU小时） / Cost Side (per GPU Hour) 收入端 / Revenue Side
Weighted procurement: $0.80-$1.50/hr (Hybrid setup)
Platform fee + storage + O&M: $0.15/hr 总成本 / Total Cost: ~$1.47 / hr
Enterprise SLA Package
Standard Package
Batch Package 混合收入 / Blended Revenue: $2.08 / hr 毛利 / Gross Profit: $0.61 / hr → 毛利率 / Gross Margin: 29.3% 规模效应预测 / Scale Effects Projections: GPU 规模 / GPU Scale 年化总毛利 / Annual Gross Profit Projections 100 GPU $374,000 / 年 (year) 1,000 GPU $3,740,000 / 年 (year) 10,000 GPU $37,400,000 / 年 (year) 🦞 Preemptible VM Compute Market Maker Full-Dimensional Deep Analysis 五、期权定价理论类比 / V. Option Pricing Theory Analogy
Market makers are essentially underwriting interruption risk — conceptualized through the Black-Scholes framework: Spot Price = OnDemand × (1 − P(中断) × L(损失率)) Spot Price = OnDemand × (1 − P(Interruption) × L(Loss Rate))
Delta
Price Sensitivity
Dynamically balance Spot/On-Demand positions Gamma
Interruption Probability Shock
Fast rebalancing during new GPU releases or large client onboarding Vega
Volatility Sensitivity
Monitor NVIDIA production lines, export controls, and crypto mining waves Theta
Time Decay
Lock in low volatility for long-term reserves, use Spot for short-term elastic demand Rho
Substitution Cost
Monitor price changes of new channels and alternative GPUs
Portfolio Hedging Strategy: Long Spot (low cost) + Short On-Demand (reselling SLA premium) + Checkpoint capability (insurance) + Cross-cloud dispersion (reducing correlation). 六、技术栈核心 / VI. Core Technology Stack
- 中断容错流水线 / Interruption Fault-Tolerant Pipeline AWS Spot Interruption Notice (提前2分钟预警 / 2-minute warning) → 自动触发 Checkpoint 保存 / Auto-trigger Checkpoint Save (PyTorch DCP / DeepSpeed ZeRO) → 停止接收新请求 / Stop accepting new requests → Karpenter 在替代区域启动新节点 / Karpenter provisions new nodes in alternative zones → 从 Checkpoint 恢复 / Restore from Checkpoint → 总中断时间 / Total downtime: 3-10 分钟 (minutes)
- Kubernetes Spot 节点池 / Kubernetes Spot Node Pools
High-performance Spot node auto-scaling + automatic interruption replacement.
Huawei open-source, Gang Scheduling ensures distributed training consistency. Kueue: K8s原生ResourceFlavor，Spot/On-Demand自动优先级切换 K8s native ResourceFlavor, automatic Spot/On-Demand priority switching. • • • 🦞 Preemptible VM Compute Market Maker Full-Dimensional Deep Analysis 3. GPU Checkpoint 性能 / GPU Checkpoint Performance
Llama 70B takes approx. 1-3 minutes (140GB parameters).
Llama 405B takes approx. 5-10 minutes.
GPU Live/Hot Migration remains unresolved in the industry — mainstream is “graceful interruption + rapid recovery”. 七、市场规模 / VII. Market Size 指标 / Metric 2025E 估计 / Estimate 全球Spot GPU市场 / Global Spot GPU Market $18 B 全球DePIN GPU市场 / Global DePIN GPU Market $2.5 B 做市商可服务市场 (SAM) / Market Maker Serviceable Addressable Market $5 - 10 B 做市商可得市场 (SOM，5-10%渗透) / Market Maker Serviceable Obtainable Market (5-10% penetration) $500 M - 1 B 全球闲置GPU占总GPU比例 / Global Idle GPU Ratio 40 - 60% 八、中国市场（特殊机会） / VIII. Chinese Market (Special Opportunities) 中美核心差异对比 / Global vs. China Comparison 维度 / Dimension 全球 / Global 中国 / China
High-end GPUs
Full NVIDIA lineup accessible
A100/H100/B200 banned; restricted to H20/ downgraded versions
Substitution
Almost non-existent
Huawei Ascend 910B (~80% of A100 performance) rising fast
Idle Sources
Data Centers + Consumer + Mining
Massive mining remnants post-ETH PoS (e.g., large volume of RTX 3090)
Players
Vast.ai and others are maturing
<10 players, ultra-early blue ocean market DePIN
Compliant operations
Strictly restricted by policy and regulation • • • 🦞 Preemptible VM Compute Market Maker Full-Dimensional Deep Analysis 中国市场的核心机会 / China’s Core Opportunities:
Structural Supply Oversupply: Post-mining GPU remnants + AI startup failure wave + low utilization rates of “East Data West Computing” nodes.
Domestic Chip Fragmentation: Ecosystems like Ascend, Cambricon, and Biren urgently require efficient aggregators.
Information Asymmetry Arbitrage: Domestic GPU public cloud pricing transparency is very low, creating high arbitrage potential. 九、做市商创业路径建议 / IX. Strategic Roadmap for Compute Market Makers Phase 1（6个月）：聚合层 / Aggregation Layer
$18 B H100 Spot折扣范围 / H100 Spot Discount Range 20% - 90% off 做市商典型毛利率 / Typical Market Maker Gross Margin 15% - 50% AWS H100 vs CoreWeave 价格差距 / AWS vs. CoreWeave Price Gap
$0.02 / GPU hr Vast.ai 在线GPU物理规模 / Vast.ai Online GPU Scale 20,000+ GPUs CoreWeave 最新估值 / CoreWeave Valuation $19 Billion 全球闲置GPU占总GPU比例 / Global Idle GPU Ratio 40% - 60% 中国本土GPU做市商玩家数量 / Chinese GPU Market Maker Players < 10 家 (companies) 做市商可服务市场规模 (SAM) / Serviceable Addressable Market $5 - 10 B 1. 2. 3. • • • 🦞 Preemptible VM Compute Market Maker Full-Dimensional Deep Analysis
🦞 Preemptible VM Compute Market Maker Full-Dimensional Deep Analysis