The Compute Hub
The Compute Hub is BenchGecko's AI infrastructure situation room. It tracks the complete supply chain from sand to API across five layers.
Five Supply Chain Layers
Layer 1: Foundries
Semiconductor fabrication plants that manufacture AI chips.
- TSMC -- manufactures 90%+ of leading-edge AI chips (N3, N5 nodes)
- Utilization rates, queue depth, and process node transitions
- FCI (Foundry Concentration Index) tracking single-source dependency risk
Layer 2: Chips
AI accelerators designed for training and inference.
| Chip | Maker | Use Case |
|---|---|---|
| H100/H200 | NVIDIA | Training and inference (current workhorse) |
| B200/GB200 (Blackwell) | NVIDIA | Next-gen training |
| MI300X | AMD | High-memory inference |
| TPU v5p | Google Cloud training and inference | |
| LPU | Groq | Ultra-fast inference (220 tok/s/$) |
| WSE-3 | Cerebras | Wafer-scale training (180 tok/s/$) |
The Chip Efficiency Race on the BenchGecko homepage ranks accelerators by tokens per second per dollar.
Layer 3: Memory
HBM (High Bandwidth Memory) is the bottleneck of AI infrastructure.
- HBM3E -- current generation, severe scarcity
- Lead times stretching past 52 weeks
- Suppliers: SK Hynix, Samsung, Micron
- MDI (Memory Demand Index) tracking supply tension
Layer 4: Systems
Rack-level and pod-level AI compute systems.
- NVIDIA DGX and HGX configurations
- Google TPU pods
- Custom hyperscaler builds
- Power density per rack (going from 30kW to 120kW+)
Layer 5: Energy
Power supply for AI datacenters is becoming a binding constraint.
- 10 datacenter regions ranked by power readiness
- 8 nuclear deals tracked (Microsoft/Constellation, Amazon/Talen, Google/Kairos, etc.)
- Grid strain analysis by region
- Renewable energy commitments vs actual delivery
Composite Indices
| Index | Current | Trend | What It Measures |
|---|---|---|---|
| AICDI | 61/100 | High | Overall compute demand pressure |
| FCI | 62 | +1 | Foundry concentration risk |
| MDI | 84 | +4 | Memory supply tension |
| HBM Tension | 48 | +3 | HBM lead time and availability |
| Capex/Rev Gap | 3.4x | +0.2x | Hyperscaler spend vs AI revenue |
Cross-Layer Insights
The Compute Hub generates cross-layer signals. For example: "Bubble Index sits at 278%, Gecko Pulse reads 20/100 (healthy). Capex keeps outpacing revenue." These insights connect infrastructure supply with economic fundamentals.
Related Pages
- AI Economy Dashboard -- capex and company financials
- Model Pricing -- how compute costs flow to API prices
- Gecko Pulse -- composite health score
- Methodology
- BenchGecko Homepage