Reference Architecture

MinIO ExaPOD is a software-defined, modular reference architecture purpose-built for exabyte-scale AI infrastructure. It scales DataPOD’s 100 PiB model to 1 EiB units, combining high-density NVMe flash with 400GbE fabrics for consistent low-latency throughput.

Estimate TCO
Key Innovations
Stackable 1 EiB blocks for exabyte-scale AI
  • ■Hyperscale Modularity — Stackable 1 EiB blocks enable zettabyte infrastructures with minimal overhead
  • ■Sustainability — 36 PiB usable per 48U rack, ~700–900 W / PiB (including cooling)
  • ■Performance — Up to 19.2 TB/s at 1 EiB, linear scaling, low TTFB for AI workloads
  • ■Network Excellence — 1:1 CLOS leaf-spine, zero oversubscription, full bisection bandwidth
  • ■Hyperscale Economics — Software-defined COTS hardware, reduced TCO vs proprietary systems
Why Exabyte-Scale Now
AI workloads demand infrastructure beyond petabyte designs
  • Volume — Billions of objects, petabytes/day from ingest, training, inference, simulation
  • Variety — Video, 3D telemetry, embeddings, logs, metadata; KB to TB object sizes
  • Velocity — Real-time pipelines need terabit-scale throughput without loss
  • Persistence — Compliance and reproducibility require immutable retention of high-cardinality data
Industry Drivers
Workloads driving exabyte-class storage
  • AI Agents — LangGraph-style agents produce logs, prompt chains, embeddings; thousands of concurrent agents generate hundreds of billions of objects
  • Multi-Modal & Simulation — Video streams, 3D telemetry, digital twins; single runs can produce multi-PiB/day
  • Long-Context LLMs & RAG — 100K+ token context, fine-grained reads; storage throughput becomes the GPU bottleneck
  • Observability & Compliance — Fine-grained telemetry, governance logs; high cardinality, immutable retention to exabyte scale
ExaPOD Design Pillars
Production-ready 1 EiB reference architecture
  • Linear capacity scale — horizontal expansion beyond 1 EiB
  • Linear performance scale — NVMe + 400GbE, predictable latency
  • Network excellence — 1:1 subscription CLOS, no oversubscription
  • Hyperscale unit economics — commodity hardware, lower $/PB
Vendor Foundations
Supermicro, Intel, Solidigm, 400GbE — validated stack
Supermicro SYS-212H-TN

2U, 24 U.2 NVMe bays; liquid cooling, PCIe Gen5 validated

Intel Xeon 6781P

80 cores, 136 PCIe Gen5 lanes — saturates NVMe + dual 400GbE without oversubscription

Solidigm D5-P5336

122.88 TB QLC NVMe per drive; 61.44 TB option trades density for throughput (more servers)

F5 (optional)

Load balancing, SSL/TLS offload, WAF for S3 at exabyte scale

Hardware Reference Design — 1 EiB
Cluster, rack, and server specifications

Cluster

32 racks × 48U
640 storage servers (20 per rack)
32 ToR + 32 OOB switches
10 spine switches (1:1 non-blocking)
Erasure code: 8 drives (5 data, 3 parity)
~12% headroom over 1 EiB usable

Single Rack (48U)

20 × SYS-212H-TN servers
57.6 PiB raw → 36 PiB usable
1 × 64-port 400GbE ToR switch
1 × OOB management switch

Storage Server

1 × Xeon 6781P, 256 GB DDR5-6400
24 × 122.88 TB Solidigm D5-P5336
2.88 PiB raw → 1.82 PiB usable
NVIDIA ConnectX-8 / BlueField 4 DPU
At a Glance
Throughput
19.2 TB/s
Usable
1 EiB
Power
~900 W / PiB
TCO
$4.55–4.60/TiB/mo
TCO Model
Customize capacity, drive sizes, and $/TB for your deployment
Open TCO Model