회사 소식 Storage news ticker - 22 September

Evolving enterprise data observability vendor Acceldata launched xFactory, a private AI software factory for building, testing and deploying AI agents, applications and analytics across hybrid environments. With xFactory, enterprises can create validated, governed agents from plain business requests while data stays in place.

Acceldata notes streaming architectures are shifting, open table formats underpin modern lakehouses, and data spans on-prem and cloud. Reliability, governance and operational simplicity must advance alongside new capabilities in its updated Open Data Platform (ODP).
ODP 3.3.6.5-1 brings Apache Kafka 4.3.0, Hive 4.1.0, Iceberg V3, expanded Ranger governance, Hue-Trino integration and platform upgrades.

ODP 3.2.3.7-2 and 3.2.3.7-3 modernize legacy deployments with Hive 4.0.1, multi-engine Iceberg 1.6.1, NiFi Registry HA, finer Ranger controls, simpler Mpack upgrades and stronger data integrity.

Taiwan’s ADATA unveiled TRUSTA AI Scaler Extended Memory HW/SW system at AI Infra Summit 2026. It combines GPU memory, DRAM and SSDs to optimize resources for AI workloads, enabling larger models on GPU-memory constrained servers. For inference and fine-tuning, it cuts deployment costs over 50% vs GPU-only architectures for scalable on-prem AI.

The open-source AI Scaler Toolkit supports Llama, Qwen, Mistral, GPT-OSS, DeepSeek, Phi, Gemma and integrates with OpenClaw, NemoClaw, Hermes Agentic. ADATA states this is not an Nvidia KV cache implementation, targeting on-prem inference and fine-tuning rather than large neocloud rack-scale KV cache workloads.

AWS data centers in Bahrain and UAE within its me-central-1 region suffered damage from Iranian attack drones across three AZs. AWS cannot restore resources exclusive to mec1-az2 and continues recovering assets in mec1-az1 and mec1-az3. Since the March outage, most customers restored operations in alternate regions via backups. Six months later, AWS warned regional conflict creates uncertain operating conditions, advising users to back up data and migrate workloads to US, Europe or Asia Pacific AWS regions per latency and data residency needs.

Object storage vendor Cloudian added Samsung PM1753 enterprise SSDs to HyperStore’s supported hardware. This PCIe 5.0 x4 SSD ranges 1.92TB–30.72TB in multiple form factors, delivering high sequential and random IO performance, 1 DWPD endurance, power-loss protection and AES-XTS 256-bit encryption. Flash demand rises as performance-sensitive AI workloads shift to object storage; many customers adopt flash frontends paired with HDD backends to control costs.

Cohesity released its fifth annual Global Cyber Resilience Report. 78% of organizations see cyber recovery’s main goal as restoring systems instead of sustaining business operations. Yet system restoration alone cannot guarantee normal operations. Among firms hit by major cyberattacks in the prior year, 60% faced delays from doubts over clean recovered systems, and another 60% encountered identity and access issues post-recovery.

Boston startup DataCebo launched SDV Enterprise 2.0, software to build generative relational database models within customer environments. It captures relational schemas, constraints and business rules to generate reusable synthetic data for analytics and AI without exposing production data. Training runs on customer infrastructure using a small data subset, often on standard CPUs. It connects to Oracle, SQL Server, BigQuery, Spanner and AlloyDB. Pricing starts at $500 monthly for unlimited tables.

As enterprises deploy AI agents into production, they need trusted data, governance and cross-platform flexibility. Fivetran and dbt Labs advance Open Data Infrastructure, a neutral, interoperable architecture freeing users from vendor lock-in. dbt v2 is a full Rust rewrite of the engine, parsing 10,000-model projects up to 10x faster than v1 and delivering real-time feedback for agent-authored SQL. The old Core/Fusion split ends; dbt Core v1 (Python) and dbt v2 (Rust) both remain Apache 2.0 licensed.

Storage veteran Mark Cree joined Upside Insights as executive advisor, with past roles at Scale Computing, Hammerspace, AWS, Cisco and others.

Micron’s new 512GB DDR5 RDIMM enables up to 12TB DDR5 DRAM in a 24-slot dual-socket server, running at 9,200 MT/s and cutting power more than 60% vs four 128GB RDIMMs. It delivers up to 1.4x higher performance for memory-heavy analytics. AMD and Intel validate the module for next-gen servers; volume production is scheduled H2 2027.

Toshiba rolled out Canvio Ready portable 2.5-inch HDD up to 4TB. USB 3.2 Gen1 with USB bus power, pre-formatted for Windows, reformat required for Mac. Its ribbed casing improves grip, and the top cover uses 65% recycled material. 1TB/2TB/4TB models launch Q4.

2025 Cambridge startup Subconscious builds agent-focused inference: co-designed Qwen/GLM variants and OrangeLine runtime, an optimized SGLang fork built for agent traces instead of one-shot chat. OrangeLine dynamically scores and compresses context on GPU, prunes KV cache mid-run and retains suffix KV states to avoid full re-encoding. It supports over 5M tokens and delivers 2.3x higher concurrent workloads on the same GPUs; customers can self-host with no data leaving their cloud. Founder Jack O’Brien says it eliminates redundant context for long-running agent sessions.

VAST Data announced Beamr Imaging as a Cosmos Community technology partner for AI media workflows. Combining VAST AI OS, NVIDIA Video Super Resolution and Beamr’s CABR compression, the pipeline upscales archives and reduces bitrate by up to 50% compared to standard encoding, supporting HEVC/AV1 on NVIDIA RTX PRO GPUs in cloud or on-prem.

Australian neocloud provider Sharon AI expanded its partnership with VAST Data. A 600PB VAST AI OS deployment acts as data backbone supporting roughly 100,000 GPUs.

VAST Data hired Jean-Thomas Acquaviva as AI & HPC Technical Director, with extensive prior experience at DDN, CEA, Intel and Versailles Saint Quentin University.

French fabless semiconductor startup VSORA develops AI inference chips with chiplet architecture cutting memory-processing cycles. Its design delivers better power efficiency, latency and throughput with 288GB HBM3E and 3,200 Tflops. Servers can hit 26 Pflops and racks reach 2–5 Pflops. VSORA chips are now in TSMC manufacturing and will ship commercially by year-end.

Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!