The software-defined data platform quietly feeding the world's largest GPU clusters - one starving model at a time.
In a rented data hall outside Reno, a row of NVIDIA H100s is humming politely, drawing power, billing the customer, and waiting. Waiting for data. The model is half through an epoch and the checkpoint will not write fast enough. Down the hall, another rack pulls weights off a file system that wasn't designed for any of this. The bottleneck is not the silicon. It never is. The bottleneck is the boring grey box that holds the bytes.
This is the world WEKA exists in, and the world WEKA built itself for. A company that calls itself, with a slightly straight face, a storage company. In the age of frontier models, that is roughly like calling NASA a logistics firm.
For thirty years, enterprise storage was an exercise in tidy hierarchies. Fast tier, slow tier, archive. Block, file, object. NAS for the office, SAN for the database, tape for the auditors. It worked beautifully for a world where the most demanding workload was a payroll run.
Then GPUs got hungry. Then they got hungrier. Then large language models came along and started training on petabytes of text in parallel across thousands of accelerators that share nothing - except the file system underneath them. Suddenly the tidy hierarchy was a traffic jam.
WEKA's three co-founders had seen this story before, just in slower motion. Liran Zvibel, Maor Ben-Dayan and Omri Palmon had built XIV, an enterprise storage startup IBM bought in 2007 for around $350 million. They knew what the inside of a storage stack looks like when you sand off the marketing. And in 2013 they bet, loudly, that the next decade would not need another array. It would need a parallel, software-defined data layer that treated flash, GPUs and clouds as one fabric.
Worth saying out loud: the trio essentially put the XIV band back together. WEKA is what happens when storage veterans get a second act and a much bigger stage.
The bet was specific and unfashionable: write a file system from scratch, in user space, in a language fast enough to compete with hardware. Make it parallel by default. Make it run anywhere - bare metal, cloud, hybrid, edge. Make it talk POSIX, S3, NFS, SMB and GPU Direct without breaking a sweat. Then refuse to ship until it actually worked.
That last part is why WEKA spent most of the 2010s in a kind of productive stealth. The Register was already calling it a "stealthy startup swimming against the tide" in 2016. The general-purpose AI moment had not arrived. The people who needed WEKA the most did not yet know they would need it.
Then they did.
The WEKA Data Platform is, technically, a parallel file system with object semantics, ridiculous throughput, and an opinion about where your data should live. In practice, it is the thing your AI team installs when their training jobs stall and nobody can figure out why.
In 2025 WEKA gave it a new name and a new shape: NeuralMesh. The pitch is that storage should behave less like a warehouse and more like a network - a mesh of intelligent nodes that scale linearly, self-heal, and feed GPUs without bottlenecking on any single component. NeuralMesh Axon takes the same idea to exascale. WEKApod appliances put it in a box for customers who don't want to assemble it.
The core software. Parallel, POSIX-compliant, cloud-portable, mildly opinionated.
AI-native mesh storage built for training, inference, and agentic AI context memory.
Exascale variant for frontier model builders and the largest AI factories on earth.
Pre-validated turnkey hardware, co-engineered with NVIDIA DGX SuperPOD reference designs.
Liran Zvibel, Maor Ben-Dayan and Omri Palmon - all XIV alumni - found WEKA. The plan: a parallel file system written from scratch for a world that does not exist yet.
Years of engineering. Norwest leads a seed round. Mellanox, Qualcomm and others join. The product slowly, stubbornly takes shape.
WEKA ships. Early adopters are life sciences, financial services, and a handful of HPC labs who can't get any other file system to keep up.
Series C closes, then Series D for $135M. NVIDIA certifications stack up. WEKA's customer logos quietly turn into a who's who of model builders.
Series E - $140M at a $1.6B valuation, led by Valor Equity Partners with NVIDIA and others. Doubles previous valuation. Oversubscribed entirely by existing investors.
Rebrand to NeuralMesh. Axon ships for exascale. WEKA announces a BlueField-4 architecture at NVIDIA GTC DC promising 100x tokens per watt.
NeuralMesh integrated with NVIDIA STX to serve as persistent context memory for agentic AI - extending the role of storage from byte store to model state.
It is easy to claim performance. It is harder to claim the customer list WEKA quietly accumulated. Stability AI. Cohere. Contextual AI. Hugging Face. ElevenLabs. CoreWeave. Nebius. Together AI. SLAC and the Vera C. Rubin Observatory. Eleven of the Fortune 50. A non-trivial number of government agencies that prefer not to be named in marketing copy.
Source: company announcements, Crunchbase, TechCrunch. Bars scaled to Series E.
WEKA's partnerships read like a map of the AI hardware industry. NVIDIA, deeply - co-engineered for BlueField DPUs and certified for DGX SuperPOD, with NVIDIA also on the cap table. AWS, Azure, Google Cloud, Oracle Cloud Infrastructure - because nobody is building one model in one place anymore. HPE, Supermicro, Lenovo and Cisco for the customers who still want a box they can touch.
BlueField-3 and BlueField-4 integration, DGX SuperPOD certification, NVIDIA AI Enterprise compatibility. And an investor.
Native deployments across every major cloud. WEKA travels with the data, not against it.
Reference architectures and WEKApod appliances for customers who want it pre-assembled.
The official line is more polished, but underneath the marketing language WEKA's mission is straightforward: keep accelerators busy. Remove the storage bottleneck between compute and data. Make the data layer something teams stop thinking about, because it has stopped being the slowest thing in the building.
There is also a quieter argument woven through the company's recent messaging. With NeuralMesh on BlueField-4, WEKA claims roughly 100x tokens per watt versus traditional architectures. That number is doing more work than it looks like it is doing. It reframes AI from a performance race - where the question is how fast - into an efficiency race, where the question is how much value per joule.
In a world where AI energy demand is becoming a political story, that reframing matters. WEKA is selling speed, but increasingly it is selling restraint.
The next era of AI - the one labelled, slightly clumsily, "agentic" - depends on a thing that current architectures struggle with: persistent, low-latency, shared context. Agents that plan over hours have to remember what they did. Models that collaborate need to share state. Reasoning workloads, the kind that drove WEKA to retitle its product around "the age of reasoning", are not just compute-bound. They are memory-bound, then context-bound, then storage-bound, in that order.
WEKA's bet, refreshed for 2026, is that the data platform underneath all of this becomes the central nervous system - not just storage, but the substrate where models live, checkpoint, reason, and remember. Its recent integration with NVIDIA STX, positioning NeuralMesh as context memory for agentic AI, points exactly there.
If the 2010s were about training, the 2020s are about inference, and the 2030s are about memory - WEKA is the rare infrastructure vendor already shipping for the third act.
The H100s are still humming. The customer is still being billed. But the file system is no longer the slow part. Checkpoints land in seconds. The inference cluster next door pulls weights through the same fabric without complaint. The training job hits its epoch boundary on time, which used to be a notable event and is now boring.
There are no banners. Nobody is filming a launch video. The team running the cluster has, frankly, stopped thinking about storage at all, which is the most expensive compliment you can pay a storage company. The grey box in the back of the rack is doing exactly what it was supposed to do, and the people who built it - three former XIV engineers, $465M in patient capital, eleven years of writing code that other people would never see - are already working on whatever bottleneck shows up next.
WEKA is not the loudest company in AI. It might end up being one of the most load-bearing. The bored GPU is bored no longer. Whether that is enough to justify a $1.6B valuation is a question for a different essay - but the customer list, for now, is voting yes.