{"slug": "quobyte-adds-ai-capabilities-and-supercharges-sales", "title": "Quobyte adds AI capabilities and supercharges sales", "summary": "Quobyte is enhancing its parallel file system to support AI inferencing and agentic context loading for GPUs, addressing bottlenecks in loading large context data. The company is also expanding its sales team to market these capabilities, which include GPU Converged storage and KV caching, to customers.", "body_md": "file\n\n# Quobyte adds AI capabilities and supercharges sales\n\nParallel file system supplier Quobyte is strengthening its capabilities to provide AI Inferencing and agentic context to GPUs and building up its sales side to present them to customers.\n\nCEO and co-founder Bjorn Kolbek said: “A year ago it was all about training and refining or fine tuning models. Suddenly it has completely moved away [towards] inference, inference, inference. Whenever you walk into a customer prospect everyone's talking about inferencing.”\n\n“Because of course you train a models] once and then you run it at scale. So what's happening is that context loading is what everyone is now concerned about. …It's called multi-turn agentic AI. What it basically means is you have your conversation with the LLM, and you ask questions, and it's an ongoing conversation. That's fairly small if it's you just asking questions. But then you ask it to load documents … that need to be stored and it's growing over time. And whenever you come back, a month later or a day later, that information needs to come back as well.”\n\n“First, it's in the GPU's memory, …and in the host memory. Then it might go to local storage, but in the end, it needs to be swapped out to persistent storage. And then, when you [comeback and] ask that question, it needs to go into a GPU as fast as possible.”\n\nThis historical context can be substantial; “We're talking 100 gigabytes. 100 gigabytes or more. There are so many things where context needs to be loaded fast and it's big. And now suddenly, if you look at the architecture of the GPU nodes, you have the nod, you have GPUs in it, you have eight network cards for the GPU back-end network. Then you have 100 gig front-end for everything else in the storage. So there is a bottleneck.”\n\nConsequently Qoobyte is getting involved in the KV Caching area: “Yes. The way we approached it; there are papers published about how you can load the KV context from two nodes. We have talked about [GPU Converged](https://www.blocksandfiles.com/ai-ml/2025/12/16/quobyte-brings-gpu-converged-storage-to-ai-clusters/1716790 ). We've now rolled this out into production. And this comes in very handy because the storage system is running on the DGX or GPU nodes. If my context is distributed across them, because it needs to be redundant, then, if I need to load it, we can use all the backend nodes. So let's say eight times 200 gigabits to pull your context from the other nodes. So the bottleneck is gone.”\n\nQuobyte initially envisioned GPU Converged primarily as a tool to save cost and power, as no one can get hardware at the moment. Kolbeck said: ”Now it turns out to be an accelerator for this context because we load over this massive back-end network. … There is a massive tier of ultra high-performance storage. It's also highly reliable.”\n\nWhat software is used to store the context data?\n\nKolbeck said: “We're just a storage system. The customer uses whatever software they use to store it. It can be database, it can be Parquet files. That’s, in the end, their decision. I know with RAG it was really all about vector databases, but this is really more of a big chunk of data in main memory and then the parts that you need are paged into the HBM.\n\nIs Quobyte writing directly to HBM?\n\nKolbeck: “We're writing to host memory.” The GPU server’s CPU? “DRAM. Exactly. And then, because you don't want to write into the HBM. You want to stage it into the host memory only when the context is there [and] you want to use the GPU. In the end, the GPU must be busy at all times. To keep the GPUs busy, you load the context into memory and only then start allocating the GPU.”\n\nNvidia has its STX and DOCA scheme for KV Cache tiering. Quobyte’s system is, we understand, similar to this. Indeed, Quobyte’s SW can run on the ARM processors in the BlueField DPUs.\n\nQuobyte is not just supporting Nvidia GPUs, and is working with AMD and Intel accelerators. Kolbeck said: “One of our big advantages is that we do not need a kernel driver because that's problematic with some of the alternatives. We need to be able to use any kernel. … Not having a kernel driver is a major advantage in that field.” He likes the point that AMD has shared memory hosts and also mentioned that comiong lower-power processors dedicated to inferencing look very interesting.\n\nQuobyte is strengthening its sales activities. It has just hired Adrian Burch as its EMEA Regional Sales Director. He was previously CRO for UK Energy for a year and, prior to that, EMEA Sales Director for WEKA for 3 years. Burch reports to Andrew Perry, Quobyte’s VP of Sales, who came on board in March. At the time we wrote Perry’s “focus will be on expanding Quobyte’s presence in key verticals such as Neoclouds, Financial Services, Media & Entertainment, Life Sciences, and Research—sectors where Quobyte’s ability to deliver linear scalability and 100 percent uptime is a critical competitive advantage.”\n\nBurch tells us: “Andrew's only been in his second quarter with the business and he's already dramatically turned it around. Andrew's brought in people he knows and trusts, which, fortunately, I'm one of those.”\n\n##### Bootnote\n\nExample papers about loading KB+V cache context from two or more nodes can be found [here ](https://arxiv.org/html/2510.09665v2)and [here](https://discos.sogang.ac.kr/file/2025/intl_conf/CLOUD_2025_H_Lee.pdf).\n\nQuobyte’s GPU-converged storage brings data in its parallel file system and object storage software closer to GPUs, and scales as GPU servers are added. The idea is to simplify and speed data storage for GPU servers by using a GPU server’s existing drives and clustering them, converging them, into a shared pool.", "url": "https://wpnews.pro/news/quobyte-adds-ai-capabilities-and-supercharges-sales", "canonical_source": "https://www.blocksandfiles.com/file/2026/07/13/quobyte-adds-ai-capabilities-and-supercharges-sales/5270293", "published_at": "2026-07-13 10:38:00+00:00", "updated_at": "2026-07-13 10:46:59.915372+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-infrastructure", "ai-products", "ai-tools", "ai-research"], "entities": ["Quobyte", "Bjorn Kolbek", "Nvidia", "AMD", "Intel"], "alternates": {"html": "https://wpnews.pro/news/quobyte-adds-ai-capabilities-and-supercharges-sales", "markdown": "https://wpnews.pro/news/quobyte-adds-ai-capabilities-and-supercharges-sales.md", "text": "https://wpnews.pro/news/quobyte-adds-ai-capabilities-and-supercharges-sales.txt", "jsonld": "https://wpnews.pro/news/quobyte-adds-ai-capabilities-and-supercharges-sales.jsonld"}}