Storage news ticker - 22 September

extract

DISK Storage news ticker - 22 September Evolving enterprise data observability supplier Acceldata launched its xFactory, a private AI software factory to build, test and deploy AI agents, applications and analytics across hybrid enterprise environments. Using xFactory, enterprises can create working, tested and governed agents, applications and analytics from plain business requests, built on data that stays where it lives. … Acceldata says streaming architectures are changing. Open table formats are becoming foundational to modern lakehouses. Data increasingly spans on-premises and cloud environments. At the same time, reliability, governance, and operational simplicity need to advance alongside new capabilities. The latest Acceldata Open Data Platform (ODP) releases move both release lines forward. ODP 3.3.6.5-1 advances the 3.3.6 line with Apache Kafka 4.3.0, Apache Hive 4.1.0, Apache Iceberg V3, expanded Ranger governance, Hue-Trino integration, and improvements to platform operations and upgrades. ODP 3.2.3.7-2 and ODP 3.2.3.7-3 bring significant modernization to established environments with Hive 4.0.1, Iceberg 1.6.1 across multiple engines, NiFi Registry high availability, more granular Ranger controls, easier Mpack upgrades, improved administration, and stronger data-integrity safeguards. … Taiwan’s ADATA announced the latest release of its TRUSTA AI Scaler Extended Memory integrated HW/SW system at the USA’s AI Infra Summit 2026. It integrates GPU memory, DRAM, and SSD resources to optimize memory and storage utilization for different AI workloads, enabling larger AI models to run on systems with limited GPU memory. For AI inference and fine-tuning, this approach can reduce deployment costs by more than 50 percent compared with GPU-only architectures, offering enterprises a more accessible path to scalable on-premises AI. The free and open source AI Scaler Toolkit supports mainstream model families including Llama, Qwen, Mistral, GPT-OSS, DeepSeek, Phi, and Gemma, while integrating with AI Agents such as OpenClaw, NemoClaw, and Hermes Agentic. This broad compatibility gives enterprises and developers greater flexibility to optimize resources across diverse use cases and Agentic AI workflows. ADATA has not said it’s an implementation of Nvidia’s KV cache scheme. We understand that it’s aimed at n-prem inference and fine-tuning, not at Neocloud-scale, rack level KV cache type workloads. … AWS's datacenters were damaged by Iranian attack drones in Bahrain and the United Arab Emirates, in the Middle East (UAE) (me-central-1) Region, with its three availability zones. In a service health post, AWS said it was “unable to restore access to the resources and data hosted exclusively in the mec1-az2 Availability Zone. We continue to work on recovering regional resources, as well as zonal resources hosted in the other affected Availability Zones (mec1-az1 and mec1-az3). … Since the disruption began in March, most customers have been able to re-establish their operations in other Regions by restoring backups or copying data that remained accessible. AWS Support remains available to help customers who need assistance moving their applications to alternate Regions.” The disruption began in March so it has taken AWS 6 months to issue this note; quite an embarrassment. It admits that “as we work to restore these facilities, the ongoing conflict in the region means that the broader operating environment in the Middle East remains unpredictable. We recommend that customers with workloads running in the Middle East consider taking action now to backup data and potentially migrate your workloads to alternate AWS Regions. We recommend customers exercise their disaster recovery plans, recover from remote backups stored in other regions, and update their applications to direct traffic away from the affected regions. For customers requiring guidance on alternate regions, we recommend considering AWS Regions in the United States, Europe, or Asia Pacific, as appropriate for your latency and data residency requirements.” … Object storage player Cloudian tells us it has added Samsung's PM1753 family of enterprise SSDs to the roster of supported SSDs for HyperStore deployments. The PM1753 is a PCIe 5.0 x4 enterprise SSD available in capacities from 1.92 TB to 30.72 TB and in 2.5-inch, E1.S, and E3.S form factors. Samsung rates the family at up to 14,500 MB/s sequential read and 10,000 MB/s sequential write, up to 3,300K random read IOPS and 650K random write IOPS, with 1 DWPD endurance over five years. The drives include power loss protection, AES-XTS 256-bit encryption, and OCP compliance. Interestingly, despite today’s elevated cost of flash, it's seeing increasing demand for flash as more AI workloads move to object storage. These workloads are typically performance sensitive, making flash a natural fit. To mitigate the cost, some customers are also interested in hybrid… flash front end with an HDD layer behind it. … Cohesity, released its fifth annual Cohesity Global Cyber Resilience Report, finding that for 78 percent of organizations, the primary goal of cyber recovery is restoring systems rather than maintaining business operations. Restoring systems alone, however, does not guarantee a business can resume normal operations. Recovery depends on more, including whether restored environments can be verified as clean and safe, applications and dependencies are functioning properly, and employees have reliable access to systems and data. The research found that organizations may struggle to resume normal operations even after systems are restored. Among those that experienced a material cyberattack in the past 12 months, 60 percent encountered moderate or significant delays because they lacked confidence that restored data and systems were clean and safe to use. Sixty percent also reported identity or access issues after systems were restored. Download the full Cohesity Global Cyber Resilience Report here. … Boston-based startup DataCebo released SDV Enterprise 2.0, software that lets an organization build a generative relational model of its own databases, inside its own environment, which captures the relationships, constraints and business context embedded in an organization’s relational data and provides a reusable foundation for software, analytics and AI applications without repeatedly moving or exposing the underlying production data. SDV (Synthetic Data Vault) 2.0 builds one generative relational model that captures the schemas, relationships, statistical patterns, constraints and business rules across an enterprise database. Using only a representative subset of data, models can typically be trained in minutes to an hour, often on a standard CPU, across schemas of any relational depth. Training takes place inside infrastructure the organization controls, without sending production data to an external model provider. SDV Enterprise 2.0 connects directly to Oracle, SQL Server, BigQuery, Spanner, and AlloyDB and runs within the customer’s environment. SDV 2.0 is available now through self-service, consumption-based pricing, beginning at $500 per month for unlimited tables. Organizations can get started here. … As enterprises deploy AI agents into production, they need trusted data, context, the business controls, and the flexibility to work across the platforms, models and tools already in use. To meet these demands, data mover Fivetran + dbt Labs is advancing its vision for Open Data Infrastructure: a vendor-neutral, interoperable architecture that lets organisations independently choose and evolve their storage, compute, data movement, transformation and visualisation technologies at every layer. This enables AI systems to work across platforms using data and context the enterprise owns, not a single vendor. dbt v2 is a full Rust rewrite of the dbt engine built for the scale that teams run at today and for how agents write SQL. It parses a 10,000-model project up to 10x faster than v1 and gives teams and their agents accurate real-time feedback, surfacing errors, column checks, and lineage before anything runs. With this release, the two engine era of Core and Fusion ends. Now, dbt is one engine with two versions: dbt Core v1, the python implementation, is dbt v1. Fusion, the Rust implementation, has become dbt v2. Both versions remain Apache 2.0-licensed and security-supported. More info here. … Storage industry veteran Mark Cree has set himself up as an executive advisor at Upside Insights. His CV has familiar names on it: Scale Computing, Hammerspace, AWS, InfiniteIO, StorSpeed, Cisco, NuSpeed, and Neo Networks. … Micron has developed a 512 GB DDR5 RDIMM enabling up to 12TB of DDR5 DRAM in a single 24-slot dual-socket server. It delivers speeds up to 9,200 MT/s and reduces operating power by more than 60 percent compared with four 128GB RDIMMs. For memory-bound workloads such as Spark SVM-based data analytics, it can deliver up to 1.4x higher performance compared to 256GB DDR5 configurations. AMD and Intel are both actively validating the module across next-generation server platforms. Micron expects the 512GB RDIMMs to be in volume production sometime in the second half of 2027. … Toshiba has a new Canvio portable 2.5-inch disk drive with capacities of up to 4TB. Its supplied with a USB Type-A to Micro-B cable. USB 3.2 Gen 1 technology delivers transfer speeds of up to 5Gbit/s while maintaining backwards compatibility with USB 2.0 devices. USB-powered operation also makes portable storage simple without the need for an external power supply, and the HDD is pre-formatted for Microsoft Windows so users can get started immediately. Mac systems are supported following reformatting. The drive us encased in grip-enhancing ribbed detailing and the top cover is manufactured using 65 percent recycled material. The new Canvio Ready with 1TB, 2TB and 4TB capacities will be available starting Q4. More info - - here. … Subconscious (Cambridge, MA; founded 2025) builds agent-oriented inference: co-designed models (TIM / “Marathon” variants of Qwen, GLM, etc.) plus a custom runtime. OrangeLine is that runtime—described as a fork/optimization of systems like SGLang, tuned for agent traces rather than one-shot chat. It compresses and manages context on the GPU while a run is in progress, instead of relying on the usual “summarize and throw away” compaction that happens outside the model. Long agent jobs fill the window with tool calls and results that stop mattering after a few turns. OrangeLine: - Scores messages as the trace proceeds and compresses low-relevance spans in place on the GPU. - Prunes the KV cache mid-run and frees memory so the same hardware can keep going. - Uses Subconscious Cache: when a span is pruned, information is not only dropped from the prefix. Suffix KV states are kept, so tokens after the pruned region do not have to be fully re-encoded. A normal prefix cache only reuses what sits before the cut; OrangeLine reuses both sides. Founder Jack O’Brien said: “Our biggest draw is helping our customers achieve longer-running sessions at lower costs. We compress the context window at runtime, so a long agent run stops carrying dead weight, and the model stays fast deep into the job. It drops into an existing cluster in place of vLLM or SGLang, holds 5M+ tokens, and gets 2.3x more concurrent work out of the same GPUs. Customers can self-host if they want, with nothing leaving their cloud.” More information - https://www.subconscious.dev/ - here. … VAST Data says Beamr Imaging (BMR) joined the VAST Cosmos Community as a Technology Partner to deliver new AI-powered video workflows for broadcasters and content owners. The collaboration with VAST Data combines the VAST AI Operating System, NVIDIA Video Super Resolution and Beamr’s content-adaptive video technology. The joint workflow, first demonstrated at IBC 2026 in Amsterdam, upscales and AI-enhances video archives and then compresses them with Beamr’s CABR technology, which can reduce bitrate by up to 50 percent versus standard encoding. The pipeline can migrate content to HEVC or AV1, runs end-to-end on NVIDIA RTX PRO GPUs, and is deployable in the cloud or on-premises, aiming to modernize and monetize existing media libraries while controlling storage and delivery costs. … Australian Neocloud provider Sharon AI has expanded a strategic infrastructure partnership with VAST Data, with a VAST AI OS installation scaling to 600 petabytes forming a data backbone capable of supporting approximately 100,000 GPUs. … VAST Data has hired Jean-Thomas Acquaviva as AI & HPC Technical Director. Before he spent a little more than 12 years at DDN, almost a year at KoDe software, CEA for more than six years, Intel for 2 periods and at the University of Versailles Saint Quentin. … VSORA is a French, fabless semiconductor, startup making new inference chips for AI with an architecture that allows them to reduce the steps and cycles between processing and memory. It says its design delivers lower energy costs, higher processing performance, reduced latency, improved throughput, and better resilience. It’s a chiplet-based design offering 288GB of HBME3E memory and 3,200 Tflops. The chips can be built into servers (26 Pflops, 2.3 TB HBM3E) and racks (2-5 Pflops, 18.4 TB HBM3E). A developer uses the VSORA SDK to generate an app. VSORA’s processors are already in manufacturing with TSMC, and are expected to be commercially available by the end of this year for customers.