AI Datacenter Construction
Complete build-out plans from 128-GPU joint tests to thousand-GPU datacenters: Clos network BOM, three-tier storage with KV tiering, power/PUE, and a fully reproducible Python TCO model.
Core Capabilities and Measured Basis
- Thousand-GPU plan (128 nodes × 8 × MI308X) with a complete TCO model: network, storage, power and depreciation modeled line by line, 10,000-run Monte Carlo sensitivity analysis (dc1k_plan, reproducible)
- Gate-based joint-test methodology: G1 delivery acceptance → G2 single-node baseline → G3 main joint-test gate (TTFT reduction ≥25%, throughput within the measured +29–40% band) → G4 72-hour stability; stop-loss if gates are missed
- Network and component anchor prices are publicly sourced: TH5 switches, 400G NICs / optics all carry cited public sources
- Production storage ratio: one FX100 serves 8 GPU nodes (100 GbE sustains 8-node KV read-back, measured anchor)
Supporting Evidence
LMCache parallel-read patch, multi-GPU KV tiering scale-out, concurrent model loading, model-switch effective TPS, training checkpoint concurrent writes (Qwen2.5-32B/7B).
Download report PDF ↓Qwen3-Coder-480B-FP8, single 8-GPU instance (TP=8 standard production topology), long-context cold recovery, three-way comparison across concurrency levels.
Download report PDF ↓Dual 4-GPU instances, nine load levels, full metrics (TTFT p50/p90/p99, TPOT, throughput, disk bandwidth); fs:// shared-pool cross-instance hot sharing verified; two independent runs within 5% deviation.
Download report PDF ↓The Engagement Path
Joint test first, decisions second: gate-based acceptance with built-in stop-loss. The full costing model is provided as reproducible Python after NDA — customers can rerun it with their own parameters. Every key figure on this site carries a report ID and is open to third-party verification.