WEKA announced today the third generation of its WEKApod AI storage appliances, pairing custom-engineered hardware with the simultaneously launched NeuralMesh 6 software to create what the company calls the densest AI storage and memory system on the market. The lineup spans WEKApod Nitro, WEKApod Prime and WEKApod Prime Max, and the headline claim is a big one: a single rack delivering 1.1 exabytes of effective capacity, the first single-rack system to break the exabyte barrier.

The numbers underneath that claim are equally aggressive. The system builds its effective capacity on 441.5 petabytes of raw storage with NeuralMesh data reduction doing the rest, while throughput reaches 10.2 terabytes per second per rack and IOPS hit 210 million per rack. WEKA says the platform leads every publicly available alternative in per-rack-unit density, by 267 percent in effective capacity and 114 percent in throughput over the next-best system in each category.

“AI infrastructure built for traditional workloads cannot power the inference era,” said Liran Zvibel, co-founder and CEO at WEKA, arguing that rack space, energy, supply chain and operational density now determine whether AI deployments produce margin or destroy it.

That framing reflects a market where physical constraints have become the story. The release cites a decline in US datacenter construction in 2025, grid connection queues stretching four to seven years in major markets, and a Morgan Stanley projection of a 49 gigawatt US power shortfall through 2028. In that environment, storage that consumes rack space and power inefficiently competes directly with GPUs for the same scarce resources, which is why WEKA is positioning tokens per rack and tokens per watt as the metrics that matter for production inference.

The engineering choices follow the thesis. WEKApod 3 runs on a PCIe Gen 6 internal fabric with a cable-based, backplane-free drive interconnect that isolates failures to a single drive, and NVIDIA ConnectX SuperNIC networking enables Spectrum-X Ethernet connectivity to keep the storage-to-GPU data path clear. A software-managed thermal architecture allows operation at 35 degrees Celsius ambient, degrading gracefully under stress rather than shutting down. Multiple patents are pending across the chassis, interconnect, thermal and serviceability designs. At the capacity extreme, WEKApod Prime Max packs 70 NVMe drives into a two-rack-unit, two-node chassis, and with Micron’s 245.76 terabyte 6600 ION drives can put 15.8 petabytes in a 2U footprint.

There is also a procurement angle. WEKA controls component sourcing directly rather than inheriting OEM channel dependencies, a pointed differentiator at a moment when NAND markets remain constrained and lead times have stretched beyond normal planning cycles.

All three appliances are orderable today through WEKA’s distributor and VAR network, sold for the first time as configurable SKUs from under one petabyte to beyond 100 petabytes, with delivery beginning in Fall 2026 and NeuralMesh 6 shipping on the first systems. For a company trusted by 30 percent of the Fortune 50, the message to AI clouds and enterprises is direct: the inference era will be won on density, and WEKA just set the scorecard.