COMPUTE MARKET MAKER SYSTEM — CROSS-INDUSTRY CASE CONJECTURES

Case 1: AI Training Clusters — Elastic Hedging (North America / AWS vs. Private Infra)

ROLES Supply-side Data Center A: A private H100 × 4,000 cluster in Silicon Valley Supply-side Data Center B: AWS us-east-1 Reserved Instance pool Demand-side: Anthropic / OpenAI / LLM startups

PAIN POINTS Training workloads are cyclical: the pre-training phase demands 24/7 full-load capacity, whereas the fine-tuning phase requires only 30%. Compute nodes cannot be shut down mid-way (checkpoints are highly expensive), but keeping them idle incurs massive capital costs.

MARKET MAKER SOLUTION The market maker aggregates both Data Center A and AWS to provide real-time quotes: Data Center A’s compute is quoted at $3.20/H100-h, AWS spot at $3.80; the market maker captures a $0.40 spread while quoting $3.50 to fine-tuning clients. When Data Center A suddenly secures a major pre-training contract and needs to reclaim its hardware, the market maker instantly routes the active fine-tuning workload to AWS spot with zero client disruption. Data Center A reclaims full capacity — much like how a high-frequency trading market maker absorbs your liquidity and redistributes it . ▪ ▪ ▪ ▪ ▪ ▪

Case 2: Movie Rendering — Seasonal Surges (Hollywood / Bollywood / China Movie Metropolis)

ROLES Data Center: Vancouver rendering farm (60% annual load, 95% during year-end peak season) Demand-side A: A post-production house in Los Angeles (Thanksgiving window, requires an extra 10,000 GPU-h within 2 weeks) Demand-side B: An animation studio in Mumbai (Diwali project, but budget is only half of Vancouver’s standard rates)

PAIN POINTS Scaling up for peak season presents a dilemma: buying hardware is uneconomical, while temporary cloud rentals are too expensive. Conversely, in the off-peak season, infrastructure sits idle; data centers want to sell but lack immediate demand channels.

MARKET MAKER SOLUTION The market maker captures cross-regional spreads : buying Vancouver’s off-peak capacity at $2.00/h (stockpiling 8,000 hours); selling to Mumbai at peak for $2.80/h (earning a 0.80 spread); and serving Los Angeles’s urgent demand at $3.50/h (premium service, delivered within 15 minutes). Operating like global energy traders who buy European LNG in summer and sell to Asia in winter, the market maker capitalizes on a combination of cross-regional arbitrage + time arbitrage + option value for immediate premium fulfillment . ▪ ▪ ▪ ▪ ▪ ▪

Case 3: Autonomous Driving Simulation — Sudden Compute Peaks (Munich / Shanghai / Detroit)

ROLES Automaker: A major German Tier 1 automotive supplier Scenario: Prior to launching a new OTA smart driving update, 72 hours of uninterrupted massive simulation testing are required

PAIN POINTS Their private 24,000 GPU simulation cluster suffices for daily baseline routines. However, OTA deployment cycles are irregular — suddenly requiring double the capacity (50,000 GPUs) for a tight 3-day window before plunging back to the baseline. Capital expenditure for seasonal peaks is wasteful, and standard public cloud bursting is exorbitantly expensive.

MARKET MAKER SOLUTION The market maker manages a global compute reserve pool , continuously absorbing fragmented capacity worldwide (5,000 GPUs from South Korean internet cafes off-hours, 2,000 GPUs from pivoted Icelandic mining farms, 3,000 GPUs from London financial institutions after-hours), packaging them into “Surge Bundles”: “72h Urgent Expansion Bundle” 5,000 GPU — $4.20/h “48h Standard Expansion Bundle” 3,000 GPU — $3.60/h “24h Lightweight Bundle” 1,000 GPU — $2.90/h The market maker effectively sells capacity options — the automaker pays an upfront premium to secure urgent expansion rights over the next 3 months, settling at the strike price when exercised. This closely mirrors how airlines purchase aircraft utilization rights ahead of season. ▪ ▪ ▪ ▪ ▪ ▪ ▪ ▪ ▪ ▪

Case 4: Edge Computing / CDN — Last- Mile Liquidity (Southeast Asia / Latin America)

ROLES Demand-side: TikTok / Shopee / Regional live streaming platforms Supply-side: Dispersed edge nodes across cities in the Philippines and Indonesia

PAIN POINTS 8-11 PM in Indonesia marks the peak window for live streaming traffic, which plummets to near zero by 2 AM. Furthermore, peak and off-peak hours vary geographically — Jakarta and Surabaya’s traffic peaks are offset by exactly 1 hour.

MARKET MAKER SOLUTION The market maker executes microsecond-level cross-regional scheduling , streaming live node prices every second. At Jakarta’s peak ($0.15/core- h) matching Surabaya’s trough ($0.04/core-h), the market maker reroutes Surabaya’s idle compute to Jakarta via optical backbones, capturing the $0.11 spread. This functions precisely like real-time wholesale electricity markets combined with PJM transmission congestion pricing . Node prices naturally diverge based on congestion density; the market maker acts as the critical “inter-regional transmission infrastructure + clearing house.” ▪ ▪ ▪ ▪

Case 5: Synthetic Biology / Drug Discovery — The Long Tail of Scientific Research (Boston / Cambridge / Zhangjiang)

ROLES Demand-side: Small biotech startups / academic research laboratories Supply-side: High-performance computing clusters of enterprise pharma giants (70% idle during weekends and nights)

PAIN POINTS Enterprise pharma clusters follow standard 5x9 business hours, meaning they are completely vacant on weekends. Conversely, small biotechs cannot afford local HPC setups, and public cloud spot instances are too unstable for complex workflows. Moreover, journal submission deadlines and grant cycles are highly randomized.

MARKET MAKER SOLUTION The market maker targets off-hours arbitrage : contracting idle pharmaceutical clusters on weekends at $0.80/h, then repackaging the capacity into a “Scientific Weekend Bundle” sold to small labs at $2.00/h. AlphaFold 2 or molecular dynamics models run overnight, returning clean results by Monday morning. The market maker effectively becomes a temporal arbitrageur of compute , financializing enterprise “dark hours” into foundational startup productivity. ▪ ▪ ▪ ▪

Summary: Four Spreads of the Compute Market Maker

Core Thesis: The definitive day compute transitions from “purchasing static hardware” to “sourcing real- time liquidity” marks the exact dawn and ultimate justification for the existence of compute market makers. Currently, global GPU idle rates sit between 30% and 50% — the sheer depth of this underlying pool is exponentially larger than the aggregate liquidity found across most cryptocurrency markets.

Type

Analogy 案例 / Case Study

Time Arbitrage

Futures / Options

Case 3: Automaker Peak Surge Bundles

Regional Arbitrage

Cross-market Spread

Case 2: Vancouver → Mumbai | Case 4: Jakarta → Surabaya

Fragment Aggregation

Absorbing bulk orders, retail splitting

Case 5: Pharma Weekends → Research Bundles

Risk Hedging

Options / Swaps

Case 1: Elastic switching between pre-training & fine- tuning