Introducing MinIO ExaPOD: The Reference Architecture for Exascale AI

AI has shifted the baseline for infrastructure, from managing petabytes to operating seamlessly at exabyte scale. Agentic systems, long-context language models, simulation pipelines, and large-scale observability now demand storage architectures purpose-built for this data reality. 

Today we’re thrilled to announce ExaPOD,  MinIO’s validated reference architecture for the exascale era: a one-exabyte, highly dense and efficient building block that scales linearly in capacity, performance, and operational simplicity. It gives enterprises a predictable, sustainable foundation for AI at scale.

TL;DR

• 1 EiB usable building block.
• Extreme power efficiency: Average 900 W of power per PiB (usable capacity including cooling)
• Extreme space efficiency: All-inclusive 36 PiB usable capacity per rack
• Up to 19.2 TB/s aggregate throughput @ 1 EiB (640 servers, 122.88 TB drives).
• Indicative all-inclusive (CapEx + OpEx + AIStor Licensing) cost: $4.55–$4.60 / TiB‑month (usable) at 1 EiB scale.
• 1:1 Clos leaf‑spine network, full bisection bandwidth (one‑hop east‑west).
• Built on Supermicro servers, Intel Xeon 6781P, Solidigm D5‑P5336 NVMe. 

Who is ExaPOD for?

ExaPOD is designed for organizations operating at the frontiers of data scale, where performance, resilience, efficiency, and governance must coexist at exabyte levels.

• Hyperscalers & neo‑clouds launching agent platforms and multi‑tenant RAG services.
• Enterprises running LLMops, digital twins/simulations, and observability at billions of objects/day.
• Regulated industries that need immutability, legal hold, and long‑term retention, without giving up performance.

Why ExaPOD now?

Data has evolved from a byproduct of computation to the fuel of intelligence itself. Across hyperscale data centers and real-time systems, the rate of creation now outpaces our ability to comprehend it, petabytes have exploded into exabytes within a single design cycle. Autonomous systems interpret the physical world in microseconds; medical networks diagnose in real time; financial engines anticipate anomalies before they occur. Each interaction, each inference, decision, or feedback loop, spawns new data that refines the next iteration. The result is a self-amplifying ecosystem where information begets intelligence and intelligence, in turn, accelerates information.

ExaPOD establishes exabytes as the new operational baseline. Purpose-built for modern workloads, ExaPOD sustains the volume, variety, velocity, and persistence demanded by large-scale data environments:

• Volume: Continuous ingest, training, inference, and simulation generate billions of objects and petabytes daily.
• Variety: Handles video, 3D telemetry, embeddings, logs, and metadata, from kilobytes to multi-terabytes.
• Velocity: Maintains terabit-scale throughput for real-time pipelines without loss.
• Persistence: Enforces immutability for retention, reproducibility, and compliance.

In short, real-time, AI-driven operations have made exascale data infrastructure a necessity, not an aspiration. ExaPOD gives enterprises the architecture to keep pace, removing bottlenecks, controlling cost, and ensuring data remains the competitive advantage it was meant to be.

From DataPOD to ExaPOD

Back in 2024, MinIO’s DataPOD proved the 100 PiB class: object‑first, software‑defined, and operations‑simple. Today, ExaPOD extends the same philosophy, S3‑native, reliability‑first, into a 1 EiB+ architecture powered by MinIO AIStor software and hardware from our partners Intel, Solidigm, and Supermicro. No embedded S3 API gateways. No scale-limiting metadata DBs. Just a fast, durable substrate that keeps agents, LLMs, and simulations fed.

Architecture at a glance

Four Pillars

• Linear capacity scale. Expand horizontally beyond 1 EiB without re‑architecture.
• Linear performance scale. High‑density SMC compute with Intel Xeon 6th Gen Processor + Solidigm 122.88 TB QLC NVMe + 400 GbE delivers predictable low‑latency throughput across ingest, training, checkpointing, and inference..
• Network excellence. 1:1 Clos fabric (non‑blocking) ensures full bisection bandwidth with one‑hop leaf‑spine paths.
• Hyperscale unit economics. Software‑defined on COTS hardware lowers CapEx/OpEx while keeping ops simple.

The 1 EiB reference build

Cluster snapshot

• Racks: 32 (48U)
• Servers: 640 (20 per rack)
• Per‑rack raw capacity: 57.6 PiB
• Per‑rack usable: 36 PiB 
• Erasure coding: Stripe Size 8 (5 data, 3 parity), customizable

Performance & Density: Choosing Your Lever

With extremely power and space-efficient 122.88 TB drives populating 640 servers, ExaPOD delivers up to 19.2 TB/s aggregate throughput at 1 EiB usable. If your priority is performance density, stepping to 61.44 TB drives roughly doubles aggregate throughput (≈ 38.4 TB/s) at similar usable capacity, but requires ~1,280 servers and additional racks/power. Choose based on supply chain, pricing, and whether density or throughput tops your list.

Networking 

A 1:1 Clos leaf‑spine fabric keeps the design simple and predictable: no oversubscription, full bisection bandwidth, and one‑hop paths between any two servers. Result: stable tail latencies under concurrency and fewer surprises as you scale.

TCO: What To Expect

Modeled economics for the 1 EiB reference build land at $4.55–$4.60 / TiB‑month (usable) including CapEx, OpEx, and AIStor licensing. Your numbers will vary by hardware timing, power/cooling and lease rates. Smaller footprints tend to carry a CAPEX premium; multi‑EiB rollouts often improve unit costs.

The reference configuration uses a 5-data, 3-parity erasure coding scheme that strikes the optimal balance between performance and durability. The configuration is fully customizable where customers can adopt wider stripe size (depending on the workload) to minimize server count and further reduce total cost of ownership.

Note on economics

The cost figure reflects MinIO’s modeled economics for a 1 EiB usable deployment based on current engagements. Actual pricing will vary with hardware procurement timing, vendor negotiations, drive availability (e.g., 61.44 TB vs 122.88 TB), regional power rates, cooling efficiency, and facility lease costs. Support tiers and services also affect totals. Treat $4.55–$4.60 / TiB usable per month as indicative, not a guarantee.

Wrapping Up: Your Path to Exascale Data

As data continues its meteoric rise, the infrastructure supporting it must evolve just as swiftly. ExaPOD stands as MinIO's definitive response to the exascale shift, offering a validated, integrated design that's resilient, high-performing, and economically sound. Whether you're transitioning from our DataPOD or embarking on a new journey, ExaPOD provides a repeatable blueprint that scales effortlessly, ensuring your organization stays ahead in a world where large-scale centers and real-time processing generate exabytes in the blink of an eye.

Eager to explore more? Get started:

• Read the white paper for the full architecture and BOM.
• Talk to a MinIO architect to tune drive size, racks‑per‑increment (¼, ½, ¾, 1 EiB), and to model TCO for your power and lease rates.
• Reach out to us for a personalized demo. 

We'd love to hear from you. Connect with us at https://www.min.io/contact-us. Together, let's architect the data future!