JuiceFS, a fast growing file storage for massive volume of data

At the 68th IT Press Tour, Joe Zhou, Developer Relations Engineer at JuiceFS developer Juicedata, presented an update on the company's vision for cloud-native storage and the rapidly evolving role of object storage in modern data infrastructure. Rather than focusing solely on JuiceFS itself, Zhou framed the discussion around a broader industry transition in which object storage has become the foundational storage layer for AI, analytics, databases, and cloud-native applications. His central argument was that while object storage has become the dominant persistence layer because of its economics and scalability, it remains too primitive for most enterprise workloads. As a result, an entire generation of software—including distributed databases, vector stores, streaming platforms, and file systems—is emerging to provide richer interfaces while continuing to use immutable object storage as the underlying medium.

According to Zhou, object storage has evolved from being a simple archival technology into the de facto backend of modern cloud infrastructure. He pointed to a growing list of platforms that are now fundamentally built on object storage rather than block or file storage. Examples include AI databases such as LanceDB, Chroma, and Milvus; cloud databases including Neon, which was recently acquired by Databricks; distributed SQL platforms such as TiDB; streaming systems like WarpStream; analytics engines; and newer cloud-native storage systems such as turbopuffer, used by companies including Anthropic and Notion. JuiceFS itself belongs to this category of technologies that leverage object storage while presenting applications with more familiar interfaces. The breadth of these examples illustrated that object storage is no longer confined to backup or archival use cases but has become the persistence layer underpinning modern data platforms.

The appeal of object storage, Zhou explained, is based on several structural advantages that are difficult for traditional storage architectures to match. Object stores expose an extremely simple API built around operations such as PUT, GET, and Compare-and-Swap. This simplicity enables hyperscale cloud providers to deliver extraordinary scalability while maintaining operational efficiency. Unlike conventional file systems, object stores employ a flat namespace rather than hierarchical directories, allowing virtually unlimited scaling without the metadata bottlenecks associated with traditional storage architectures. Public cloud object storage services also routinely advertise eleven nines (99.999999999%) of data durability and support multi-region availability for high resilience. Combined with features such as immutability and exceptionally low storage costs—typically around two cents per gigabyte per month in major cloud regions—object storage has become the most economical and reliable long-term storage platform available.

Despite these strengths, Zhou argued that object storage is fundamentally unsuitable for many application workloads when used directly. Most enterprise software expects richer file system semantics than object stores provide. Objects cannot be modified in place, meaning even minor updates often require entire files to be rewritten. Directory hierarchies do not actually exist, instead being simulated through indexed prefixes that become increasingly expensive to manage as environments scale. Batch metadata operations such as renaming large directory trees are slow and costly because they involve manipulating enormous numbers of object keys. Object stores also exhibit higher latency than conventional file systems, cannot execute applications directly, and perform poorly when managing structured datasets composed of many related files. These limitations create friction for AI pipelines, software development environments, analytics platforms, and enterprise applications originally designed around POSIX file systems.

According to Zhou, these shortcomings explain why many modern cloud-native systems are effectively rebuilding traditional interfaces on top of object storage. Instead of abandoning POSIX or relational database interfaces, companies increasingly expose familiar APIs while using object storage solely as the persistence layer. JuiceFS provides POSIX compatibility, Neon delivers PostgreSQL semantics over object storage, and numerous modern databases perform similar abstraction for their respective workloads. The trend reflects an industry consensus that object storage offers compelling economics but requires additional software layers before it becomes practical for mainstream computing.

A significant portion of the presentation examined Amazon Web Services' recently introduced S3 Files service, released in April 2026. Zhou described the product as "a decent approach" that validates JuiceFS's overall architectural direction but argued that AWS's implementation remains constrained by important design decisions. S3 Files enables customers to mount an S3 bucket as a POSIX-compatible NFS file system by placing Amazon Elastic File System (EFS) in front of S3. Within this architecture, EFS acts as both the metadata layer and high-performance cache while S3 remains the authoritative copy of all data.

The system employs a strict one-to-one relationship between files and objects. Small files, particularly those below the default threshold of 128 kilobytes, remain optimized for low-latency access through the EFS layer. Write operations are initially committed to EFS before being synchronized back to S3 after approximately sixty seconds. Although this approach improves responsiveness compared with accessing S3 directly, Zhou argued that it introduces additional complexity and several significant limitations.

One of the most important concerns involves write amplification. Because each file corresponds directly to a single object, modifying even a tiny portion of a large file requires substantial data movement. For example, appending only a few bytes to a two-gigabyte video file requires retrieving the complete object, merging the new data, and rewriting the entire object through Amazon's multi-stage append workflow. This process increases latency while consuming additional network bandwidth and storage operations.

Metadata operations also become increasingly expensive at scale. Zhou illustrated this using a simple rename command. Renaming a directory containing one million files in a conventional POSIX file system typically requires only metadata updates. Under S3 Files, however, such an operation eventually triggers background rewrites of every corresponding object because each object's key incorporates the file path. Consequently, operations that appear trivial to applications may generate substantial backend activity, increasing operational costs and execution time.

Additional trade-offs include batching delays introduced by asynchronous synchronization between EFS and S3, conflict resolution policies that always treat S3 as the authoritative source, and an architecture limited exclusively to Amazon Web Services. Customers cannot extend the solution across multiple public clouds or integrate alternative object storage platforms. Pricing also becomes more complicated because organizations pay separately for EFS capacity, EFS read and write operations, synchronization processes, and underlying S3 storage and requests. According to Zhou's comparison, S3 Files successfully delivers POSIX compatibility and directory hierarchies but continues to struggle with in-place updates, efficient metadata operations, application execution, and workloads involving structured datasets.

JuiceFS approaches these challenges through a fundamentally different architecture centered on strict separation of data and metadata. Rather than mapping one file to one object, JuiceFS divides every file into immutable four-megabyte chunks. Metadata—including directory structure, permissions, and file mappings—is maintained independently within a dedicated metadata engine. Because only modified chunks require rewriting, appending data to a large file typically updates only the final chunk rather than recreating the entire object. Similarly, renaming a directory becomes a lightweight metadata transaction instead of triggering extensive object rewrites.

The Community Edition of JuiceFS is released under the Apache 2.0 open-source license and supports multiple external metadata databases, including Redis, TiKV, MySQL, PostgreSQL, and FoundationDB. Users can combine these metadata engines with virtually any mainstream object storage platform, including Amazon S3, Google Cloud Storage, Microsoft Azure Blob Storage, Alibaba Cloud OSS, Tencent Cloud COS, Ceph, and MinIO. Applications access the storage through several interfaces, including POSIX via FUSE, Java and Python software development kits, a Kubernetes CSI driver for container environments, and an S3 Gateway providing compatibility with existing object-based applications.

Recent Community Edition enhancements include configurable storage-class tiering. Administrators can now automatically place directories or individual files into different cloud storage classes such as Amazon S3 Standard-Infrequent Access, Intelligent-Tiering, or Glacier Instant Retrieval. This feature allows organizations to optimize storage costs according to workload requirements without altering application behavior.

The Enterprise Edition expands considerably on the open-source platform. Instead of relying on external databases, it introduces a proprietary distributed metadata engine based on the Raft consensus protocol. This metadata layer scales horizontally while maintaining three-copy redundancy for fault tolerance. Enterprise deployments also gain access to a distributed cache architecture shared across thousands of clients, enabling improved performance for large distributed AI clusters and analytics environments.

Additional Enterprise Edition capabilities include native cross-region replication and multi-cloud mirroring. Organizations may choose cache-only mirrors, where data remains stored centrally while caches are positioned closer to compute resources, or full mirrors that replicate complete datasets across multiple cloud regions or providers. These capabilities support hybrid cloud architectures while reducing latency for globally distributed workloads.

Scalability represents another major Enterprise Edition enhancement. JuiceFS recently increased the supported limit from one hundred billion to five hundred billion files within a single volume. Zhou noted that one existing customer deployment already stores approximately 1.47 pebibytes of data and more than 404 billion inodes, demonstrating that the architecture has moved beyond theoretical scalability into real-world production environments.

Several customer examples illustrated how these capabilities are being used in practice. MiniMax, one of China's leading artificial intelligence laboratories, operates JuiceFS in a hybrid cloud architecture where GPU clusters remain inside the company's own data center while object storage resides elsewhere. Cache-only mirrors position frequently accessed data closer to the GPUs to minimize latency, and the organization is evaluating full replication across multiple locations. JuiceFS argues that this architecture provides the flexibility to balance infrastructure costs against AI training performance without requiring duplicate storage management.

Beyond MiniMax, Zhou highlighted a growing list of customers spanning AI, cloud infrastructure, robotics, internet services, and software development. Named adopters include HeyGen, GMI, PixVerse, Momenta, Horizon Robotics, Xiaomi, Lovart, NAVER, Trip.com, fal, D-Robotics, Cerebrium, Fly.io, and Jerry. These deployments demonstrate the platform's applicability across generative AI, autonomous driving, cloud-native applications, and enterprise software development.

Finally, Zhou discussed JuiceFS's commercial model. Enterprise Edition pricing is based solely on the amount of source-region storage capacity managed by the platform rather than the number of connected clients or compute nodes. This allows organizations to expand large GPU clusters without incurring additional JuiceFS licensing costs. Equally important, JuiceFS does not impose its own data transfer charges because clients communicate directly with the underlying object storage rather than routing traffic through JuiceFS-managed infrastructure. Customers therefore pay only the standard data transfer fees charged by their chosen cloud provider.

Overall, the presentation positioned JuiceFS as part of a broader shift in enterprise storage architecture. Rather than replacing object storage, the company argues that the future lies in enhancing it with higher-level interfaces that preserve cloud economics while eliminating operational limitations. By separating metadata from data, chunking files into immutable blocks, and supporting multiple clouds and storage providers, JuiceFS aims to provide the performance and flexibility of a distributed POSIX file system while retaining the scalability, durability, and low cost that have made object storage the foundation of modern AI, analytics, and cloud-native infrastructure.

Read More

IO River behind the Soccer World Cup broadcast with a modern CDN approach

IO River used its presentation at the 68th IT Press Tour to introduce a new approach to edge infrastructure management that challenges the traditional content delivery network (CDN) model. Founded in 2022 by CEO Edward Tsinovoi and CTO Michael Hakimi, both former Akamai engineering leaders responsible for edge computing initiatives, the company argues that the architecture underpinning today's Internet delivery infrastructure has failed to keep pace with the growing complexity of cloud-native applications and the explosive rise of generative AI. Rather than building yet another CDN, IO River has developed what it describes as a "Virtual Edge" platform that virtualizes multiple CDN and edge providers into a unified operational layer. Its goal is to provide enterprises with the resilience, flexibility, and performance optimization previously achievable only by the world's largest Internet companies.

The founders' background plays a central role in the company's strategy. During their time at Akamai, Tsinovoi and Hakimi helped develop large-scale edge computing technologies and gained firsthand experience operating one of the Internet's largest distributed infrastructures. According to the company, this experience highlighted an important structural weakness within the industry: despite decades of technological progress, most organizations continue to depend on a single global CDN provider to deliver applications, websites, APIs, and streaming content. While this model worked reasonably well during the early decades of the web, IO River argues that it has become increasingly fragile in an era characterized by massive AI workloads, globally distributed users, and constantly changing traffic patterns.

Since its formation, IO River has attracted financial backing from several prominent investors. The company is supported by S Capital, formerly the Israeli branch of Sequoia Capital, together with Venture Guides, New Era, and Pags Group. In early 2026, it completed a $20 million funding round that increased total investment to approximately $25 million. Beyond financial support, the company has assembled an advisory board consisting of experienced executives from the networking, storage, cloud, and cybersecurity industries. Advisors include Ronni Zehavi, founder of Cotendo and HiBob; Aryeh Mergi, previously associated with XtremIO and M-Systems; Ash Kulkarni, Chief Executive Officer of Elastic and former Akamai executive; Marty Kagan, founder of Cedexis and Hydrolix; Ofir Ehrlich, founder of CloudEndure and Eon; and Pavel Gurvich, formerly of Guardicore. Collectively, these advisors bring experience from companies that have significantly influenced cloud infrastructure and enterprise software.

The central thesis presented by IO River is that today's CDN market remains fundamentally organized around assumptions established during the 1990s. Most organizations select a single CDN provider and entrust that vendor with all edge delivery responsibilities. Although this simplifies procurement and operations, it also creates a single point of failure. The company argues that this model no longer reflects the operational realities of modern digital businesses, particularly those supporting AI-driven applications, globally distributed APIs, or mission-critical online services.

To support this argument, IO River highlighted the frequency of outages affecting major edge providers. According to the company, every major CDN experiences a significant global outage approximately every one to three years, while smaller regional incidents occur on a monthly basis. Even service level agreements promising 99.9 percent availability still permit roughly six to ten hours of downtime annually without financial penalties. IO River cited several high-profile examples from recent years, including multiple Cloudflare incidents lasting between four and twenty-five hours during 2025 and early 2026, AWS CloudFront disruptions lasting seven and fifteen hours, a six-hour Akamai outage in 2025, a seven-hour Google Cloud incident in 2024, and a Microsoft Front Door outage in 2023. Regardless of the precise duration of individual events, the company's broader point was that even the largest cloud infrastructure providers remain vulnerable to significant service interruptions.

From the customer's perspective, the consequences of these outages can be severe. IO River argued that a single twelve-hour outage affecting a major online business could generate customer losses exceeding ten billion dollars through lost transactions, reputational damage, operational disruption, and reduced productivity. In many cases, these financial impacts far exceed the total lifetime revenue earned by the infrastructure vendor itself. As a result, the traditional model of relying on contractual service level agreements offers limited practical protection when outages occur.

Rather than creating another competing CDN network, IO River positions itself as a virtualization layer above existing edge infrastructure. The company emphasizes that it is neither a conventional CDN, nor simply a global load balancer, nor a traditional multi-CDN switching platform. Instead, it describes its architecture as a "Virtual Edge" that separates application services from the underlying infrastructure providers. This abstraction enables organizations to treat multiple CDNs as interchangeable infrastructure resources while managing them through a single operational platform.

The architecture consists of three distinct layers. The first layer focuses on intelligent traffic steering. Rather than directing all requests toward a single CDN provider, IO River dynamically distributes traffic across more than fifteen premium global and regional edge providers, including Akamai, Cloudflare, Fastly, AWS CloudFront, Google Cloud CDN, Azure CDN, and others. Traffic routing decisions are driven by artificial intelligence models that continuously evaluate provider reliability, network performance, latency, and operating costs. Instead of intercepting traffic inline, the platform performs routing through DNS and CNAME manipulation. This architectural choice provides an important resilience benefit: if IO River itself experiences a failure, DNS continues directing traffic according to the most recent routing configuration, allowing applications to remain operational rather than introducing another infrastructure dependency.

The second architectural layer provides unified management and operational abstraction. One of the longstanding challenges of multi-CDN deployments is that every provider exposes different configuration languages, APIs, scripting models, and management tools. For example, Fastly relies on its VCL language while Akamai uses EdgeWorkers and proprietary configuration mechanisms. Maintaining equivalent application behavior across multiple providers therefore requires substantial engineering effort. IO River addresses this complexity by presenting administrators with a single management interface that automatically translates configurations into provider-specific implementations. Organizations can therefore define caching policies, routing behavior, security settings, and operational rules once rather than maintaining separate implementations for every CDN platform.

The third layer virtualizes higher-level application services traditionally bundled within CDN offerings. Rather than relying exclusively on each CDN vendor's proprietary implementations, IO River enables customers to deploy application services independently of the underlying infrastructure provider. These services include web application firewalls, bot management, API protection, edge compute functions, and serverless execution environments. Importantly, IO River does not attempt to develop all these capabilities internally. Instead, it partners with specialized technology vendors, including Check Point, Google, and others, integrating their products into the Virtual Edge platform. The company argues that this partnership model contrasts with incumbent CDN vendors, which historically attempted to develop every capability in-house and consequently often deliver "good enough" implementations rather than best-of-breed functionality.

A major operational advantage claimed by IO River is its ability to detect infrastructure degradation before providers publicly acknowledge service problems. Because the platform continuously monitors performance across multiple CDN vendors simultaneously, it can identify abnormal latency, packet loss, or service degradation in real time. When problems emerge, traffic is automatically redirected toward healthier providers. According to the company, this often occurs before vendors such as Cloudflare or AWS CloudFront officially recognize or announce outages. Consequently, customer applications may continue operating normally while users relying on a single provider experience widespread disruption.

The company also addressed one of the historical barriers to multi-CDN adoption: pricing. Traditionally, distributing traffic across multiple providers reduced purchasing volumes with each vendor, weakening customers' negotiating leverage and increasing per-gigabyte costs. IO River argued that this disadvantage has largely disappeared. Through aggregated purchasing, optimized traffic allocation, and closer collaboration with infrastructure providers, organizations can now maintain favorable commercial terms while simultaneously benefiting from increased redundancy and performance optimization.

Although still a relatively young company, IO River presented encouraging signs of commercial traction. After approximately one year of active sales activity, the platform has acquired around fifty enterprise customers with no reported customer attrition. Initial adoption has been strongest within industries where uptime directly affects revenue, including over-the-top streaming services, online media publishers, gaming companies, and educational technology providers. Representative customers include Minute Media, Nexon, and Plarium. More recently, the platform has expanded into additional sectors including e-commerce, travel, hospitality, and software-as-a-service. One highlighted customer within the hospitality sector is Accor Hotels, illustrating the platform's appeal beyond traditional digital-native businesses.

The company's pricing model reflects its position as an orchestration platform rather than a pure infrastructure provider. Customers pay a tiered platform subscription based on deployment scale, with optional charges if traffic is purchased through IO River rather than directly from CDN providers. Additional usage-based pricing applies to application services such as web application firewall processing or request-based security features. This flexible model allows organizations either to retain existing infrastructure contracts or to consolidate procurement through IO River depending on their operational preferences.

IO River's go-to-market strategy varies geographically. Within the United States, the company primarily sells directly to enterprise customers. In Europe, however, it relies more heavily on systems integrators and regional partners. Examples include GNN in Germany and Equativ in France. Interestingly, the company reported increasing collaboration with CDN providers themselves. Vendors such as Fastly now occasionally introduce IO River into customer opportunities because the platform's visibility into multiple providers helps demonstrate each CDN's unique strengths rather than simply comparing pricing. This evolution suggests that some infrastructure vendors increasingly view multi-provider orchestration as complementary rather than purely competitive.

Operationally, IO River maintains an approximately even split between engineering and commercial functions. Research and development activities are concentrated in Tel Aviv, leveraging Israel's strong cybersecurity and networking engineering ecosystem, while sales, marketing, and customer-facing operations are primarily based in the United States. The company has also filed multiple U.S. patents covering aspects of its Virtual Edge architecture, reflecting efforts to protect its technological innovations as the platform matures.

Financially, the recently completed funding round provides an estimated two-year operating runway, allowing the company to continue expanding both product capabilities and market presence. Looking forward, IO River views generative AI as one of the primary drivers reshaping edge computing requirements. AI applications generate highly dynamic workloads, unpredictable traffic patterns, and increasingly distributed processing demands that challenge traditional single-provider architectures. By abstracting multiple infrastructure providers into a unified operating platform, IO River believes organizations will be better positioned to optimize performance, resilience, cost, and geographic distribution as AI workloads continue to grow.

Overall, IO River presented a vision in which edge infrastructure evolves similarly to cloud computing, where organizations routinely operate across multiple providers instead of depending on a single vendor. Rather than replacing existing CDN providers, the company seeks to orchestrate them, allowing customers to leverage each provider's strengths while minimizing the operational complexity traditionally associated with multi-CDN deployments. Through AI-driven traffic steering, unified configuration management, integrated application services, and resilient architecture that avoids introducing additional single points of failure, IO River aims to make enterprise-grade edge resilience accessible to organizations that previously lacked the resources to engineer sophisticated multi-provider infrastructures. In the company's view, the Virtual Edge becomes the "easy button" for the multi-cloud, multi-edge era, giving businesses of all sizes the operational flexibility, reliability, and vendor independence that until now has largely been reserved for Internet giants such as PayPal and other hyperscale online platforms.

Read More

Paradigm4 promotes flexFS, a real alternative to classic high performance file storage

Paradigm4 used its presentation at the 68th IT Press Tour to introduce flexFS, a cloud-native parallel file system designed to bridge the long-standing gap between traditional POSIX file systems and modern object storage. The presentation featured CTO and flexFS inventor Gary Planthaber alongside business development executive Dave Clock, Chief Revenue Officer Andy Cosgrove, and technical sales partner David Freund. Although Paradigm4 is best known for its origins in life sciences analytics, the company argued that flexFS has evolved into a broadly applicable storage platform capable of addressing performance and cost challenges across AI, analytics, machine learning, and enterprise computing.

The story behind flexFS begins with Paradigm4's own infrastructure requirements. While developing large-scale genomic analytics applications, the company needed a high-performance POSIX-compatible file system capable of delivering tens or even hundreds of gigabytes per second in public cloud environments. Existing solutions failed to meet both performance and cost objectives. Paradigm4 evaluated numerous open-source and commercial offerings, including JuiceFS, ObjectiveFS, S3FS, Goofys, S3 Backer, Amazon EFS, Amazon FSx for Lustre, Lustre, DDN, and Weka. According to the company, open-source products generally lacked either full POSIX compliance or sufficient throughput, while enterprise parallel file systems provided the required performance but at price points unsuitable for genomics research organizations operating under constrained budgets.

Unable to identify an acceptable alternative, Paradigm4 developed flexFS internally. Initially built exclusively to support the company's own analytics platform, the file system gradually matured into a standalone product after it became clear that many industries faced the same challenge. Today flexFS has reached version 1.9 and is offered in both commercial and Community Edition forms. The free Community Edition supports up to 5 TB of storage using the customer's own object storage bucket, lowering the barrier to adoption while allowing developers and organizations to evaluate the technology before committing to larger deployments.

Paradigm4 argues that modern AI infrastructure suffers from a fundamental architectural mismatch. Most enterprise applications, AI frameworks, and analytics tools continue to rely on POSIX file semantics, while cloud providers increasingly encourage customers to use inexpensive object storage services such as Amazon S3, Microsoft Azure Blob Storage, Google Cloud Storage, and Oracle Cloud Infrastructure Object Storage. Although object storage provides excellent scalability, durability, and economics, it lacks the low-latency file semantics expected by most software. Organizations therefore compensate by deploying expensive network-attached file systems or maintaining duplicated datasets across multiple storage tiers. According to Paradigm4, these compromises increase infrastructure costs, slow data pipelines, reduce GPU utilization, and ultimately limit AI productivity.

Rather than adapting traditional on-premises parallel file systems for cloud deployment, Paradigm4 describes flexFS as an "object-native parallel filesystem." This distinction reflects a fundamentally different architectural approach. Instead of storing complete files on block devices, flexFS divides every file into multiple chunks. Each chunk is assigned its own object identifier and written directly into the underlying cloud object store. By leveraging the hyperscaler's object infrastructure, flexFS automatically benefits from massive parallelism, scalability, and durability without requiring specialized storage hardware.

One of the major challenges with object storage is metadata latency. Listing directories, locating files, or retrieving metadata can become significantly slower than with conventional file systems. To address this limitation, flexFS employs its own persistent low-latency metadata server. This component maintains the namespace, file attributes, and object mappings independently of the cloud provider, allowing applications to experience near-traditional file system responsiveness while still storing all data inside object storage.

The platform also offers an optional Proxy Group that functions as a write-back cache similar to a content delivery network (CDN). Unlike traditional caching approaches that require entire files to be cached, flexFS supports fractional caching. For example, administrators can configure the system to cache only the first hundred blocks of every file while allowing larger data ranges to stream directly from object storage. This capability enables organizations to optimize cache utilization for workloads that primarily access file headers or metadata while avoiding unnecessary consumption of expensive local SSD capacity.

Paradigm4 emphasized deployment flexibility as another key advantage. flexFS supports single-region public cloud deployments, multi-region architectures, multi-cloud configurations, hybrid cloud environments, fully on-premises installations, and converged deployments where storage services run directly on compute nodes. During the presentation, the company highlighted joint work with Oracle demonstrating performance on Oracle Cloud Infrastructure approaching that of locally attached NVMe storage. This suggests that object-backed storage need not impose the performance penalties traditionally associated with cloud storage.

Beyond core storage functionality, flexFS incorporates several operational features intended to simplify enterprise administration. Duplicate files are identified using hard links supported by checksum verification and byte-for-byte validation to ensure data integrity. The system includes an optimized file search utility driven directly by the metadata server, allowing administrators to perform large-scale directory searches more efficiently than standard POSIX file system operations. Software updates are designed to be non-disruptive, with metadata server pauses lasting less than one second while client mounts automatically reconnect through FUSE session handoff. For containerized environments, flexFS provides a Kubernetes Container Storage Interface (CSI) driver together with Helm charts to simplify deployment into Kubernetes clusters.

Security and resilience are also important components of the platform. Paradigm4 stated that flexFS has achieved ISO 27001 certification, demonstrating adherence to internationally recognized information security management standards. Data durability is rated at eleven nines (99.999999999%), leveraging the inherent resilience of hyperscale cloud object storage. Because flexFS presents a standard POSIX interface, the company positions it as a drop-in replacement for managed cloud file services including Amazon EFS, Amazon FSx for Lustre, Oracle Cloud Infrastructure File Storage, Google Cloud Filestore, and Microsoft Azure Files, allowing customers to migrate workloads without application modifications.

The company illustrated the platform's economic benefits through a detailed customer case study involving one of the world's five largest pharmaceutical companies. Covering a period from September 2022 through March 2026, the deployment grew to approximately 1.14 petabytes containing more than 160 million files. During those 43 months, actual infrastructure costs using flexFS combined with Amazon S3 totaled approximately $2.53 million.

Paradigm4 compared this real-world expenditure against a modeled alternative architecture based on conventional AWS managed storage services. The comparison assumed that storage requirements would be distributed across 25 percent Amazon FSx for Lustre Persistent SSD, 40 percent Amazon EFS Standard Regional, 10 percent Amazon EBS gp3 block storage, and 25 percent Amazon S3 Standard object storage. Under this scenario, total infrastructure spending would have reached approximately $5.65 million over the same period.

The resulting savings exceeded $3.13 million, representing a 55 percent reduction in storage costs. During calendar year 2025 alone, savings reached approximately $1.44 million, or 59 percent. By March 2026, the monthly operating cost using flexFS had fallen to roughly $110,000 compared with an estimated $274,000 for the conventional AWS architecture. The analysis also identified approximately $332,000 in wasted spending associated with over-provisioned Lustre capacity that flexFS eliminated.

Cost efficiency continued improving as deployment scale increased. In 2022, when the environment stored approximately 25 terabytes, effective costs averaged around $90 per terabyte per month. By early 2026, after scaling to 1.14 petabytes, costs had declined to roughly $66 per terabyte per month. Competing managed services remained largely flat throughout the same period, with Amazon EFS estimated at approximately $307 per terabyte per month and Amazon FSx for Lustre around $174 per terabyte per month. Paradigm4 used these figures to argue that object-native storage architectures become increasingly advantageous as data volumes grow.

While genomics remains an important market, Paradigm4 presented several new use cases demonstrating flexFS's applicability across modern AI and analytics workloads. One emerging application is data lakehouse acceleration. In benchmark testing using the TPC-H workload at scale factor 100, Apache Spark combined with Gluten completed execution in just 176 seconds using cached flexFS compared with 1,191 seconds when operating directly on Amazon S3. This represented a 6.8-fold performance improvement, illustrating how intelligent caching and optimized metadata management can significantly reduce analytics processing times without abandoning object storage economics.

Another growing application involves modernization of coupled-architecture database systems. Paradigm4 suggested that massively parallel processing (MPP) data warehouses, graph databases, and vector databases can all benefit from flexFS without requiring application code changes. The company estimates that organizations could reduce total cost of ownership by as much as 60 percent through storage consolidation and improved infrastructure utilization.

Artificial intelligence and machine learning represent another strategic growth area. flexFS supports widely used AI frameworks including PyTorch, TensorFlow, and JAX, providing high-performance shared storage for training datasets, model checkpoints, and distributed learning environments. Because many AI training jobs repeatedly access the same datasets, the platform's caching architecture helps improve throughput while minimizing expensive transfers from object storage. Faster checkpoint operations also reduce training interruptions and improve recovery following hardware failures.

Paradigm4 also introduced the concept of agentic AI workspaces as an emerging workload category. Autonomous AI agents frequently create temporary working files, process very large documents, and require rapid point-in-time recovery when errors occur. flexFS provides POSIX-compatible scratch spaces while supporting efficient byte-range access into large PDF files and other datasets. The platform also enables point-in-time recovery, allowing organizations to restore files accidentally deleted or corrupted by autonomous AI agents. As enterprises increasingly deploy agentic AI systems capable of modifying data independently, these recovery capabilities could become increasingly valuable.

Throughout the presentation, Paradigm4 emphasized that flexFS is not simply another cloud file service but rather a foundational storage architecture intended to reshape how organizations build AI infrastructure. The company believes that the separation between inexpensive object storage and traditional file systems creates unnecessary complexity that affects nearly every modern workload. By combining cloud object economics with POSIX compatibility, flexFS aims to eliminate that architectural compromise while improving performance and reducing costs simultaneously.

Finally, Paradigm4 concluded by raising a broader industry question. Just as the concepts of the Data Lakehouse and Coupled-Architecture Database Management Systems have become recognized categories within enterprise analytics architectures, the company believes there may be room for a new category called the "File Lakehouse." This proposed concept would describe storage platforms that combine the scalability and economics of object storage with the performance and application compatibility of high-performance parallel file systems. During the IT Press Tour, Paradigm4 actively sought feedback from industry analysts and journalists on whether this terminology should be formalized as part of future AI, machine learning, and analytics reference architectures. Whether or not the industry ultimately adopts the label, the presentation clearly positioned flexFS as a technology designed to unify cloud object storage and enterprise file systems into a single architecture capable of supporting next-generation AI workloads at significantly lower cost.

Read More

PoINT Software and Systems confirms its leadership in data management

Almost 4 years after we met in Paris PoINT Software & Systems we had the privilege to talk again with Thomas Thalmann, CEO, in Sofia, Bulgaria, for the 67th edition of The IT Press Tour.

PoINT is a privately held German software vendor founded in 1994, with roots in storage and archiving dating back to 1985 through work with Philips and Digital Equipment Corporation. Certified as "Software Made in Europe" and recipient of the Storage Newsletter Cloud Storage Award 2026, the company's core mission centers on helping organizations manage data growth efficiently, reduce costs, and build cyber-resilient storage infrastructures.

The company frames its market relevance around five intersecting categories of pressure facing organizations today: explosive growth in unstructured data and migration complexity on the technical side; rising storage and energy prices on the economic side; data sovereignty concerns on the political side; compliance, archiving obligations, and cybercrime risk on the legal side; and CO2 footprints and e-waste on the ecological side. PoINT's response to all five centers on intelligent data tiering, placing the right data in the right place at the right time, with a strong emphasis on tape as a cost-efficient medium that consumes no energy when inactive and provides natural air-gapping against ransomware.

The company offers three main software products. The PoINT Storage Manager, launched in 2007, handles file tiering and archiving by moving inactive files from primary NAS systems to secondary storage including tape, optical, object stores, or public cloud, using policy-based rules while maintaining transparent access for end users. It counts over 200 installations worldwide, with a notable deployment at Daimler spanning multiple locations with WORM, versioning, encryption, and multi-tenancy. The PoINT Archival Gateway delivers S3-to-tape functionality, exposing an Amazon S3-compatible REST API while writing data directly to tape without intermediate disk layers, dramatically reducing costs compared to all-disk or public cloud approaches. Available in Compact and Enterprise editions, the Enterprise configuration scales to 32 interface nodes, 12 tape libraries, 384 drives, and 153.6 GB/s native throughput, with geo-distribution, automatic failover, and erasure coding across two sites. It is also packaged as the ORION S3, a turnkey system developed with BDT offering up to 392PB of native capacity. The PoINT Data Replicator handles backup and replication of object and file data to S3-compatible systems, supporting S3-to-S3 and File-to-S3 modes for use cases including cloud repatriation, legacy NAS migration, and continuous backup via Kafka and SQS change tracking.

Notable customers include Sixt, Daimler, Amgen, PostFinance, and EMBL-EBI, which deployed the gateway to archive Kubernetes workloads via S3 and achieve read/write throughput exceeding 1PB per week. Technology partners include HPE, NetApp, Fujitsu, Dell EMC, Cloudian, and Spectra, with resellers including SVA, Cristie, and Computacenter.

Read More

Leil continues to innovate to address large scale storage challenges

Leil joined The IT Press Tour for the second time following the 1st session in April 2024 for the 55th edition when we unveiled the company to the world.

Leil is an Estonian startup founded in 2022 and headquartered in Tallinn, built by engineers with deep expertise in parallel file systems and distributed storage. Currently seed-funded, the company's core mission is to bridge the growing gap between the economic potential of high-capacity hard disk drives and the legacy software architectures originally designed for flash and SSD storage. In short, Leil builds software that makes HDDs perform the way they were physically designed to, something no mainstream storage platform currently does.

The company frames its market opportunity around what it calls the "SMR Paradox." Shingled Magnetic Recording drives offer significantly more capacity per disk, and hyperscalers like Google, Meta, and AWS have already adopted SMR at 100% across their infrastructure using custom-built software. However, the remaining 90% of the enterprise market has achieved zero SMR adoption, simply because no accessible, enterprise-grade software exists to manage these drives properly. Legacy architectures treat modern high-capacity HDDs like slow SSDs, generating small random I/O patterns that waste 30 to 60% of potential capacity economics, require months of tuning per petabyte added, and demand PhD-level specialist staff to operate.

Leil's answer is a two-layer product stack. Leil FS is an open-source parallel file system strictly optimized for high-capacity HDDs, serving as the community adoption engine and baseline for innovation. Leil OS is the commercial enterprise distribution built on top, offering hyperscale-grade efficiency, a management UI, seamless deployment, and 24/7 SLA support. Both are underpinned by the proprietary SMRT Engine, the company's core intellectual property. Key capabilities include a 25% usable capacity gain over generic software-defined storage on identical hardware, tape-level cost per TB from €0.99/TB/month without the retrieval penalty associated with tape, and deployment in 10 minutes via standard repository commands versus 6 to 12 months for traditional SMR integration projects. On performance, Leil OS serializes writes into sequential streams, claiming to unlock 99.7% of theoretical maximum HDD throughput, while its implementation of SNIA Command Duration Limits prevents tail latency spikes critical for AI training workloads. For resilience, a Head Depopulation technology allows Leil to retire only the failing platter surface of a drive rather than triggering a full rebuild, achieving zero-downtime recovery.

Target use cases span AI and HPC warm-tier storage, active archives, enterprise backup qualified with Veeam and Acronis, media post-production for 4K and 8K workflows, on-premises Kubernetes, and CCTV storage. Real deployments include a national broadcaster using Leil OS for multi-petabyte video-on-demand storage, supercomputing centers running national archive projects, and autonomous driving research programs staging telemetry datasets for ML pipelines. The company goes to market through a 100% channel model with white-label and OEM options, and technology alliance partnerships with WD, Seagate, Nvidia, Intel, and AMD. Because Leil OS is built atop the open-source Leil FS under GPL-3.0, customers always retain a guaranteed exit path with no vendor lock-in.

Read More

StorPool jumps into KVM-based HCI

Fourth session with StorPool Storage with The IT Press Tour in their city, Sofia, Bulgaria, following several articles I wrote as I unveiled the company to the world in 2014.

StorPool is a Bulgarian software-defined storage company founded in 2011, entirely self-funded, profitable, and growing, with roughly 60 employees across Bulgaria and the USA. Serving over one million end users globally across 30 countries on five continents, the company positions itself as the leader in modern block-based software-defined storage, with a mission to create a better world through better data storage and management.

At its technical core, StorPool delivers an ultra-fast, highly reliable, and linearly scalable block storage platform with latency below 0.1ms, up to 100 million IOPS, five-nines availability, and scalability ranging from 10TB to 50+ petabytes with no workload interruption. The platform includes built-in backup and disaster recovery tools and integrates natively with major KVM-ecosystem platforms including OpenStack, CloudStack, Proxmox, OpenNebula, and Kubernetes.

The company organizes its go-to-market around four major industry trends. The first is the VMware exodus triggered by Broadcom's acquisition, which StorPool addresses with a drop-in vSAN replacement, StorPool One, a fully managed KVM platform replacing the entire VMware Cloud Foundation stack at 64% lower five-year TCO, and an Oracle Virtualization bundle delivering 71% savings versus VMware over five years. The second is European data sovereignty, where StorPool responds as a fully European, non-US-owned company participating in the EuroStack initiative. The third is AI infrastructure, where the platform powers GPU-as-a-Service and inference workloads for customers including Redmond.ai and Cloudalize. The fourth is hardware cost pressure, where StorPool's HCI mode consumes only 10–15% CPU and RAM overhead, consolidates over 20 physical components down to 7, and can run approximately 3,000 virtual machines on just 10 servers.

Customer outcomes speak to the platform's broader economic impact: a 15% margin increase for CloudSigma, 60% higher per-rack VM density for Namecheap, a reduction from 50 to just 5 storage staff at Dustin, and elimination of downtime at Atos. End-user workloads running on StorPool-powered infrastructure include those of NASA, ESA, CERN, Siemens, and Deutsche Börse Group. StorPool frames this not merely as storage optimization but as a ripple effect improving the economics of the entire data center, accelerating the broader shift from siloed, manually operated IT toward API-driven, automated, always-on infrastructure.

Read More

NGX Storage unveils ExaScale, its full disaggregated NVMe storage model

NGX Storage joined the 67th edition of The IT Press Tour last week for the second time. We introduced NGX to the world in December 2022 when we met the team in Lisbon, Portugal, for the 47th tour.

NGX is a European enterprise storage vendor founded in 2015 and headquartered in Ankara, Türkiye, within a university-linked technology park. Branding itself as "Made in Europe," the company operates R&D centers in India and commercial operations across four continents. Its founding premise is straightforward: data storage should be powerful, flexible, and manageable across every protocol from a single platform.

The company frames its value proposition around a clear industry pain point. Enterprise storage has become fragmented, expensive, and operationally overwhelming, with organizations juggling separate NAS, SAN, and object storage platforms, each with its own tools and infrastructure costs. The AI era compounds this complexity, as training and inference workloads demand fast access to enormous datasets, microsecond NVMe latency, and extreme throughput at scale. NGX's argument is that these converging pressures require a fundamentally new storage architecture built for extreme performance and data scale from the ground up.

The company offers a unified portfolio organized into five product lines. The NGX-H is a dual-controller hybrid system scaling to 38PB, supporting Fibre Channel, iSCSI, NFS, SMB and S3. The NGX-AFA mirrors that architecture but runs exclusively on NVMe SSDs for latency-sensitive workloads up to 34PB. The NGX ExaScale is a scale-out NVMe block storage platform using NVMe-oF with RDMA/TCP, designed for AI and HPC workloads and scaling to hundreds of petabytes in future releases. The NGX HyperIO is the scale-out object storage platform, built on high-density nodes with erasure coding, geo-dispersed protection, self-healing, and multi-site replication, suited for analytics, backup, and large-scale archival. Finally, a Scale-Out NAS capability built on top of the AFA and Hybrid platforms supports AI workloads, HPC, and data lake architectures at exabyte scale. Across all products, the platform includes inline compression and deduplication, thin provisioning, full RAID levels, and three-way and four-way mirroring.

NGX also offers a MetroScale Cluster capability for active-active datacenter deployments, delivering zero RTO and zero RPO, validated for SAP, Oracle, Red Hat, and Microsoft Windows environments. Its customer base is 99% enterprise, spanning verticals including finance, healthcare, defense, media, oil and gas, and education, with use cases covering virtualization, VDI, HPC, AI/ML, backup, and business continuity. Technology partners include Intel, Nvidia, Veeam, VMware, Kioxia, and Western Digital. The company is actively expanding into Poland, Spain, Malaysia, South Korea, and the UAE, and is currently developing the third generation of its unified storage platform, reinforcing its positioning as a full-lifecycle storage vendor with no hardware vendor lock-in.

Read More

Caeves promotes a modern file tiering solutions fueled by AI

Caeves Technology joined the IT Press Tour last week in Sofia, Bulgaria.

The company is a New Jersey-based startup founded in 2024 by the team behind Talon Storage Solutions, an entity co-founded in 2012 that developed edge caching technology before being acquired by NetApp in March 2020. After five years scaling NetApp's cloud data services, the founding team launched Caeves, operating in stealth mode before releasing its flagship product, Caeves Intelligent Deep Storage, first in private preview on Microsoft Azure in August 2025, then in general availability worldwide in February 2026. The leadership team includes Shirish H. Phatak as CEO and CTO, Jaap van Duijvenbode as VP of Product and Customer Experience, and Andrew Mullen as SVP of Sales and Alliances.

The company tackles two major structural problems in enterprise data management. The first is economic: according to Gartner, 30% of enterprise storage budgets are spent on cold or redundant data that delivers zero active business insight, while data volumes double every two years with no sign of slowing. The second is AI readiness: 85% of unstructured data is created once and never touched again by analytics or artificial intelligence tools. When data is archived to cold tiers or legacy systems, it becomes completely invisible to Microsoft 365 Copilot, Azure AI Search, and all modern AI tools, directly undermining the ROI of enterprise AI investments. As Jensen Huang noted at Nvidia GTC in March 2026, unstructured data remains largely impossible to query, search, or index at scale without a smarter infrastructure layer.

Caeves Intelligent Deep Storage is a cloud-native solution built exclusively on Microsoft Azure, combining intelligent tiering, multi-protocol access supporting SMB and NFS, and native integration with the Microsoft 365 ecosystem. The platform deploys in under 30 minutes entirely within the customer's own Azure tenant, with no data ever leaving the customer's environment. It provides automatic tiering from Hot to Cool to Archive landing on Azure Object Storage, reducing storage costs by up to 70% without any loss of access or performance. A key differentiator is the Caeves Copilot Connector, which indexes historical archives directly into Microsoft Graph, making them accessible through Microsoft 365 Search and Microsoft Copilot with no custom RAG pipeline or additional infrastructure required. Unlike most competitors, Caeves stores data in native Azure Blob format, meaning customers retain full access at all times with no proprietary encoding, no extraction fees, and no lock-in, at a cost ranging from $0.01 to $0.03 per GB per month.

The pricing model is entirely capacity-based and available through the Microsoft Marketplace, offering a free tier up to 5TB for pilots and testing, with rates decreasing from $0.03/GB/month for small teams down to $0.01/GB/month beyond 1PB. For organizations managing more than 200TB, the cost of Caeves is typically recovered within the first month of tiering savings alone.

Looking ahead, a minor release is planned for September 2026, followed by a major version in Q1 2027 featuring an Enterprise Management Plane with ROI dashboard, policy management, compliance and governance tools, and an MCP Server enabling integration with Claude, Gemini, LangChain, Azure OpenAI, and Microsoft Foundry. The long-term vision is to position Caeves as the intelligence and context layer for enterprise data estates within the Azure ecosystem, evolving from a storage optimizer into a full data intelligence platform with autonomous tiering operations and deep integration with Copilot Studio.

Read More

67th Edition of The IT Press Tour in Sofia, Bulgaria

The IT Press Tour, a media event launched in June 2010, announced participating companies for the 67th edition organized March 31st and April 1st in Sofia, Bulgaria.

During this edition, the press group will meet 6 hot and innovative companies:

• Caeves Technology, a young data management player leveraging Azure for cold file data,
• Leil Storage, promoter of a modern spin down storage approach coupled with ab advanced parallel file system,
• NGX Storage, an alternative solution for scale-out storage,
• PoINT Software & Systems, an established data archiving player,
• StorPool, a reference in SDS,
• and Storware, an innovator in backup.

I invite you to follow us on Twitter with #ITPT and @ITPressTour, my twitter handle and @CDP_FST and journalists' respective handle.

Read More

Lustre is built to last according to The Lustre Collective

The Lustre Collective aka TLC joined the recent IT Press Tour in Silicon Valley last week and it was a key session to understand the mission and direction of the new entity as it was founded just before SC25 last November.

TLC is a newly formed company created to ensure the long-term innovation, stability, and relevance of the Lustre parallel file system, one of the most widely deployed storage technologies in HPC, enterprise AI, and large-scale data infrastructure. Launched publicly at Supercomputing 2025, TLC is founded by long-time Lustre leaders and original developers who have collectively driven Lustre’s architecture, evolution, and community releases for more than two decades. Their goal is to provide independent, expert stewardship focused solely on Lustre’s future.

Lustre itself has a 25-year history and remains the dominant parallel filesystem for demanding workloads, powering a majority of the world’s top supercomputers and large AI systems. According to data highlighted in the presentation, Lustre is used by over 60% of the Top 100 HPC systems and underpins exascale machines and large commercial AI deployments, including systems operated by NVIDIA, national laboratories, and hyperscalers. Its longevity is attributed to its open-source, vendor-neutral GPLv2 license, symmetric bandwidth, linear scalability, POSIX compliance, and proven reliability at extreme scale. 

TLC was formed in response to structural gaps in the Lustre ecosystem. While many vendors and cloud providers actively contribute to Lustre, development priorities are often shaped by individual commercial interests. TLC positions itself as a neutral, independent organization that works across vendors, hyperscalers, research institutions, and enterprises to identify and address shared long-term needs of the Lustre community. Unlike venture-backed startups, TLC is not pursuing acquisition or IPO strategies; instead, it operates more like a permanent engineering collective, reinvesting revenue directly into Lustre development and expertise. 

Technically, Lustre continues to evolve to meet modern AI and cloud demands. The platform delivers industry-leading performance, supporting tens of terabytes per second of throughput, hundreds of millions of IOPS, tens of thousands of clients, and hundreds of thousands of GPUs. As illustrated in the architecture diagrams, Lustre provides fully parallel data and metadata paths, flexible use of HDD, QLC/TLC NVMe, and client-side NVMe caching, multi-rail RDMA networking, and protocol re-export via NFS, SMB, and S3 gateways. Security features include strong authentication, encryption, and fine-grained multi-tenant isolation. 

TLC’s roadmap focus includes accelerating Lustre’s transition toward greater resilience, usability, and cloud readiness. Near-term development areas include erasure-coded files, undelete/trash functionality, fault-tolerant management services, client-side compression, GPU peer-to-peer RDMA, and improved recovery mechanisms. Longer-term priorities include metadata redundancy, metadata writeback caching, enhanced multi-tenancy, easier quality-of-service controls, and modernized tooling and monitoring. 

The Lustre Collective monetizes through services rather than licensing, offering consulting, production support, feature development, performance tuning, training, and deployment assistance. Overall, TLC positions itself as a trusted partner for enterprises, hyperscalers, appliance vendors, and research institutions, working to ensure that Lustre remains the definitive data foundation for exascale HPC, enterprise AI, and large-scale distributed computing for decades to come.

Read More