Inference Storage Planning for a Thousand-GPU AI Datacenter
The core of thousand-GPU datacenter storage planning is a three-tier architecture plus KV tiering: the measured one-FX100-per-8-nodes ratio, a fully reproducible Python TCO model, and gated joint tests to control build risk.
The storage design of the thousand-GPU plan
Mingxin's thousand-GPU plan (128 nodes × 8 × MI308X) includes a fully reproducible TCO model: Clos network BOM, three-tier storage (local / all-flash pool / capacity tier) with KV tiering, power and PUE, depreciation modeled line by line, and a 10,000-run Monte Carlo sensitivity analysis (dc1k_plan).
The storage ratio anchor comes from measurement, not guesswork: a single 100 GbE port sustains 8-node KV read-back (R2/R3 bandwidth data), so FX arrays deploy at 1:8; network and component anchor prices all cite public sources.
Gated construction: stop-loss if gates are missed
G1 delivery acceptance (materials inspected, array self-test with 24/24 drives recognized) → G2 single-node baseline (R1 basis reproduced ±10%) → G3 main joint-test gate (TTFT reduction ≥25%, throughput inside the measured +29–40% band, 8-node shared read-back approaching the NIC ceiling) → G4 stability (72-hour stress + fault injection transparent). Each gate stops the project if missed — customers do not pay for unverified promises.
The complete costing model is reproducible Python after NDA — a datacenter is a heavy-asset decision, so we make verifiability part of the plan itself.
FAQ
Can an already-built datacenter be retrofitted?
Yes: KV tiering is an incremental architecture that leaves compute and network mostly untouched. Start with a single-array pilot (the ten-week G1–G4 process) and scale after verification.
Can we run the TCO model ourselves?
Yes: Python source is provided after NDA — change parameters and recompute every conclusion, including parameter sets unfavorable to Mingxin. That is part of the verifiable commitment.
What share of investment should storage take?
The joint-test basis is within ~10% of GPU investment; the exact share depends on the workload profile (inference share, context length, switch frequency). The calculator produces a range first.
Data sources (verifiable)
Related reading
- GPU Rental Platforms: Raising Per-GPU Output with Storage
- Checkpoint and Data-Read Acceleration for Training Clusters
- Storage Acceleration for AMD MI308X Inference Clusters
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.
This site presents business-cooperation information and constitutes neither an investment offer nor any promise of returns. Measured data come from signed / official test reports (see the Evidence Library); vendor specs, public sources and estimates are labeled as such.