WEKA has launched its third-generation storage appliance, the WEKApod 3, headlined by a configuration that claims 1.1 exabytes of effective capacity in a single 56U rack. The milestone is real—but it is not a measure of how much NAND flash sits inside the chassis. Instead, it reflects what happens when 441.5 petabytes of raw NVMe capacity are combined with always-on data reduction in the company’s new NeuralMesh 6 software stack. For enterprise IT teams moving from AI proof-of-concepts into production, unraveling that distinction is the first step toward an honest capacity plan.

The New Appliance Family: Three Configurations, One Software Heart

WEKApod 3 arrives in three variants. The Prime Max configuration—the one behind the exabyte headline—uses 2U, two-node chassis, each packed with up to 70 Micron 245.76 TB 6600 ION QLC NVMe SSDs. Fully populated with roughly 1,800 drives across 56 rack units, the raw flash capacity lands at 441.5 PB. WEKA then applies NeuralMesh 6’s data reduction to reach the advertised 1.1 EB of logical capacity. The company’s own materials clarify that this assumes a 3:1 effective-to-raw ratio in a fully loaded rack.

Two other configurations are available. WEKApod Nitro targets performance-heavy workloads, trading some density for higher throughput per rack unit. WEKApod Prime sits between the two as a general-purpose balance of capacity and speed. All run NeuralMesh 6, which will ship in the second half of 2026. Existing software customers get the upgrade at no additional cost. Appliances can be ordered now, with deliveries starting in fall 2026, according to TechRadar’s initial report.

What the 1.1 EB Number Actually Means for Your Rack

When a storage vendor says “exabyte,” most people imagine a warehouse of drives. WEKA’s claim is different: the 1.1 EB is effective, application-visible capacity. The system can logically host that much data—models, datasets, checkpoints, logs—provided the data compresses and deduplicates well enough. If your data doesn’t reduce, you are buying 441.5 PB of raw flash, which is still a formidable number but less than half the headline figure.

This is not deceptive marketing; it is standard enterprise storage practice, amplified to a symbolic scale. Deduplication, compression, thin provisioning, and metadata efficiency have been used for decades to advertise usable capacity. What’s new is applying these techniques inside a single rack densely packed with high-capacity QLC drives and wrapping the whole thing in a contractual performance guarantee. According to NAND Research, WEKA offers a commercial commitment that covers both the reduction ratio and the performance impact of the reduction engine—an unusual move that turns a software variable into a contractual yardstick.

Why Data Reduction Works—and When It Won’t

NeuralMesh 6 runs data reduction by default, using fingerprinting, similarity hashing, inline deduplication, and compression. The processing happens outside the primary write path to keep latency low. WEKA claims up to 6x savings on AI training data, where repeated model checkpoints, cloned development environments, and nearly identical document collections create abundant opportunities to squash duplicates.

But every environment is different. Encrypted files, pre-compressed media, compressed archives, highly random research outputs, and certain database formats may deliver minimal savings. The 1.1 EB number is a planning scenario, not a guarantee. A rigorous evaluation demands testing with your actual data mix—model artifacts, object storage, images, logs, and system files—under sustained workloads. Don’t skip a proof of concept that measures:

  • Your true reduction ratio after protection overhead and metadata
  • Ingest and random-write performance during peak periods
  • Rebuild behavior with these large SSDs
  • Endurance and garbage collection effects on QLC media

Micron’s 245.76 TB 6600 ION drives are built on G9 QLC NAND. They excel at capacity economics, but they are not infinite endurance devices. The product brief lists reference specs—up to 13.7 GB/s sequential reads and 3 GB/s writes—but also warns that lifetime and power consumption vary sharply by workload. Plan your endurance budgets carefully, especially for write-heavy pipeline stages.

Performance: The Numbers That Need Decoding

WEKA also announces rack-level performance: up to 10.2 TB/s throughput and 210 million IOPS. These are extraordinary figures, but they require the same scrutiny as the capacity claim. First, the top-line stats likely belong to the Nitro configuration, not the capacity-optimized Prime Max. Second, throughput in TB/s doesn’t directly equate to faster token generation, lower time-to-first-token, or higher concurrent users for your inference workload. The metrics that matter for production AI are increasingly about efficiency:

  • Tokens served per rack
  • Tokens per GPU
  • P99 storage latency
  • S3 and POSIX concurrency under load
  • GPU idle time waiting for data
  • Watts per usable terabyte

If a higher throughput number comes from a chassis filled with different drives and a different balance of compute to storage, the raw number alone won’t predict your real-world experience. Ask for configuration-specific data sheets and insist on seeing tested results with workloads that resemble your own.

Unified File and Object Access: A Bridge for Windows-Centric Environments

One of NeuralMesh 6’s more practical features is native unified access: an object written via S3 can be read via POSIX, and a file written via NFS or SMB appears as a native S3 object without a separate copy or gateway. For enterprises that still run Windows file servers, SQL Server databases, and traditional NAS alongside modern container platforms and AI toolchains, this can dissolve operational friction. Data preparation, training, and inference stages often use different protocols; a single data set accessible by all can slash copy time and storage duplication.

That said, interoperability isn’t magic. Test your Windows clients against the S3 and SMB implementations. Verify access control lists, Active Directory integration, backup software compatibility, and existing ransomware recovery workflows before conformance becomes a production assumption. The appliance’s new multi-tenancy features—Composable Clusters for physical resource allocation and a Virtualized RDMA Data Fabric for network isolation—are similarly critical for shared environments. If multiple departments or customers will use the same rack, you’ll need clear performance boundaries and encryption segregation.

The Density Gamble: Power, Cooling, and Failure Domains

Packing 1,800 SSDs into a single rack reduces floor space and cabling, but it also concentrates risk. Each Micron 6600 ION drive can draw up to 30W under load. Even at a fraction of that, the rack-level power and cooling requirements are substantial before adding CPUs, NICs, memory, and network switches. Micron argues its QLC SSDs cut power per terabyte compared to HDDs, but a dense all-flash rack requires careful thermal engineering. Don’t treat “dense” as “thermally simple.”

Failure domains also grow. A single node failure affects dozens of drives. A rack-level maintenance event touches an enormous logical capacity. The cable-based, backplane-free interconnect design that WEKA describes may reduce the blast radius of a shared backplane failure, which is a genuine engineering plus. But you still need to understand the protection scheme’s usable capacity impact, rebuild times for very large SSDs, and firmware upgrade procedures across the fleet. Service access in the tightly packed 2U chassis also warrants a close look—can you hot-swap a drive without disturbing neighboring nodes?

Direct Supply Chain: Faster Delivery, Tighter Lock-In

By sourcing hardware directly instead of relying on third-party OEMs, WEKA can offer more predictable pricing and delivery, as TechTarget reported. That’s a real advantage when data center construction is slowing and grid connection queues stretch for years. But it also means you commit to one supplier’s hardware roadmap, support model, and component longevity. If your organization prizes disaggregated, pick-the-best-component architectures, this integration may chafe. If rapid deployment and a single throat to choke during outage calls are priorities, the appliance model might be exactly what you need.

What to Watch Next

WEKApod 3 sets a new bar for storage density, but it’s not the only player chasing the exabyte-rack dream. Competitors will likely respond with their own software-defined appliances that lean heavily on data reduction. The trend points toward a future where effective capacity, not raw terabyte count, becomes the primary purchasing metric—but only if vendors provide transparent, testable guarantees. For buyers, the lesson is clear: the big number gets your attention, but the small print in the reduction contract and the real-world POC results determine whether 1.1 EB is your capacity or just a conversation starter.