Mingxin

The Storage Substrate of a MaaS Platform: Turning Context Caching into a Billable Capability

Direct answer

A MaaS platform's three storage pains — fast model on/offboarding, tenant session retention, and context-cache pricing — map to measured anchors: loading 6.2–9.3×, switching utilization 46.7%→62.8%, KV read-back 12.6×.

How MaaS differs from single-service inference

A MaaS platform serves many models and tenants at once: models on/offboard frequently (new releases, version updates, elastic scaling), tenant sessions must persist across requests, and leading API vendors already publish context-cache hit discounts as a pricing line — a self-hosted MaaS without the same capability starts with an inferior cost structure.

All three pains land on the storage tier: weight-distribution speed (R9 anchor: a 70B model takes 23 minutes over NFS), switching idle time (R1: 46.7% utilization at 20 switches/hour), and KV persistence (R2: re-compute TTFT 149.5 s).

One all-flash pool, three returns

Weight pool: one copy shared cluster-wide — DeepSeek-32B/70B serving loads 6.2×/9.3× (R9), 8-GPU concurrent loads 16% faster wall-clock (R1) — new-model onboarding and elastic scale-out in minutes.

Switch acceleration: effective utilization 46.7% → 62.8% at the 20-switches/hour level (R1) — directly more sellable inventory for a GPU-hour platform.

Tenant KV pool: isolated by tenant namespace; session recovery reads back at 11.85 s vs 149.5 s recompute (R2). Context retention becomes a value-added pricing item — a business model the major API vendors have already validated (see their public price pages).

FAQ

Can KV leak across tenants?

No: KV is doubly isolated by tenant namespace plus prefix hash; identical content across tenants is still not shared (security takes precedence over dedup gains).

How should context caching be priced?

Reference the public market: major API vendors bill cache-hit input at significant discounts (see their price pages). A self-hosted platform's discount headroom comes from the measured compute savings (+29–40% throughput, R2/R3).

What starting scale makes sense?

The one-FX100-per-8-nodes ratio is enough to start (the ten-week G1–G4 joint-test process); after verification, scale linearly at 1:8.

Data sources (verifiable)

R1FX100 Comprehensive LLM Inference & Training Benchmark (8× AMD MI308X)2026-07-03
Download report PDF ↓
R2FX100 KV-Cache Benchmark (480B, TP8 long-context, signed)2026-07-05
Download report PDF ↓
R9Mingxin FX100-HBMM vs NFS Baseline on Huawei Ascend 910B2026-05-30
Contact us for access →

Related reading

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.