Inside Channel Manager

Scality entrusts the contextual memory storage of AI agents, without performance guarantees. Avery Calloway - 10/09/2026 - 6 minutes of reading Scality Unveils AI Inference Factory at Scality Day in Paris Scality has announced its latest innovation, the AI Inference Factory, during Scality Day in Paris - a software stack designed for on-premises inference that offloads the contextual memory of AI models to storage. This move aims to address the limitations of GPU memory for AI agents as the complexity of context expands during operations. Although Scality has launched this product with a fixed budget model in mind, it has not yet guaranteed performance outcomes. AI models and reasoning agents require context retention throughout a session, which can grow increasingly cumbersome. Scality's announcement, made on October 8, aims to tackle these demands, promoting the use of AI Inference Factory on enterprise infrastructure instead of relying on cloud providers. This software stack integrates validated open-weight models, an inference service, a control layer, and Scality's own Scality AI Data Infrastructure (ADI). The target market includes businesses, government entities, and new cloud service providers focused on AI. Scality offers this stack through both licensed and fully managed service options, delivering a complete solution that it validates, deploys, and maintains. Shared KV cache for agents' context. To produce each token, the AI model utilizes a Key-Value (KV) cache, storing previous computations related to the context. This technical memory mechanism enables AI agents and large language models to respond swiftly and cost-effectively during complex tasks or conversations. Instead of revisiting and recalculating entire conversation histories for every word generated, the system retains prior attention states directly within the graphics card memory (VRAM). According to Scality, reasoning models and agent loads, particularly during long sessions, are the most impacted. The company has opted to move the KV cache from GPU servers to ADI, marketed as a shared cache with multi-petabyte capacity. Furthermore, the architecture separates the two phases of inference: a group of GPUs first reads the request and updates the cache (prefill), while a second group then accesses this cache to generate tokens (decode). "With ADI as a shared cache, any decode GPU can retrieve any context without size limitations," explains Giorgio Regni, Scality's Chief Technology Officer. It's important to note that this KV cache does not overlap with an agent's memory from one session to another. Instead, it maintains the context of the ongoing session, including instructions, read documents, tool results, and reasoning already produced. Retention of historical data, preferences, and organizational knowledge requires different mechanisms, such as retrieval-augmented generation or long-term memory, which are not part of this announcement. This distinction is crucial for agent functionality, allowing an agent that waits for tool feedback or human validation to resume tasks without the GPU recalculating all accumulated context. Hence, this development addresses the costs associated with extended and interrupted sessions, not the learning capabilities of agents. ADI operates on a scale of multiple petabytes, using NVMe or HDD disks connected to GPUs via RDMA. Data is distributed across flash memory, disks, and optionally tape. Laboratory performance insights from Scality. The performance numbers shared stem from Scality's internal tests. The Gemma-3 27B model loads in 1.9 seconds - nearly ten times faster than loading from a local NVMe drive. Restoring a context of 14,000 tokens from ADI takes 166 milliseconds before producing the first token, which is only 83 milliseconds longer than loading from HBM. Accessing cached data is 14 times faster than recalculating it for the same context, going up to 72 times faster for 439,000 tokens. The cache exceeds GPU memory capabilities by over 80 times and can restore 1,000 sessions without recalculation. According to the company, latency between generating tokens remains steady, with context restored before the first token is emitted. However, these tests are based on a modest setup: a server with eight GPUs, five storage servers, and four RDMA links operating at 100 Gbit/s. Scality has not released any service-level commitments. Its control layer effectively distributes requests according to service-level agreement (SLA) targets, yet these are ultimately determined by the client. The company cites two external studies. DistServe, presented at OSDI 2024, reported a 7.4 times increase in requests served at similar latencies thanks to the separation of phases, while Moonshot AI noted a 75% increase in requests processed with its Mooncake shared cache. Fixed budget per token amidst unpublished pricing. Scality's second key argument revolves around costs. In cloud environments, providers charge per token generated, with bills escalating alongside usage. Additionally, a provider may alter a model's version, risking disruption to dependent applications. By deploying on-premise, enterprises can maintain control over model versions and pay for infrastructure rather than consumption. The stack supports various models, including Mistral, Gemma, gpt-oss, Qwen, Kimi, GLM, and DeepSeek, compatible with Dell, HPE, Lenovo, and Supermicro servers. This variety addresses common needs within the market. Yet, Scality has not provided pricing details or client references. On-premise deployment incurs costs for GPU purchase, energy, and operational maintenance. In the cloud, token prices are competitive. For example, as of October 8, Anthropic charges $0.10 per million tokens for its Claude Haiku 5.5, and Google Cloud has set spending caps for computational loads. On-premise solutions excel in scenarios of massive, consistent usage, while cloud services are advantageous when usage is low or variable. Thus, cost savings will likely vary by case until Scality reveals its pricing. IDC Vice President Nataliya Yezhkova points out the market dynamics, stating that on-site inference is now a viable option "as long as the infrastructure can efficiently preserve and serve model states at scale." Scality isn't pioneering this field alone; Red Hat has also entered the arena with its AI Inference Server since May 2025, supporting a distribution of the vLLM engine, alongside the distributed inference project llm-d. A response to architectural challenges yet to be proven. The distinction between prefill and decode, as measured by DistServe, and the reuse of the KV cache are established practices that Scality hasn't invented. Fundamentally, the approach makes sense. The differentiating factor lies in how it handles inference state storage, as opposed to the engine work emphasized by Red Hat. In this paradigm shift, agent context transitions from a computational burden to a stored data format, allowing GPU memory constraints to transform into storage capacity inquiries that IT departments are capable of managing. Still, Scality's evaluations are based on a single test bed using just two models, Gemma-3 27B and GLM-5.2. It remains to be seen whether this framework can hold up under production loads. An IT department considering this stack will require testing with its own agents, specific session volumes, and documented performance guarantees. Ultimately, the fulfillment of the promise of a fixed budget will depend on how well these requirements are met.

CTERA Networks bridging the gap between cloud storage and local storage. Liran Eshel, CEO, Co-Founder The explosive growth in unstructured data, the increased need to access files anywhere and from any device, and the desire to collaborate frequently across the distributed enterprise are all driving companies to use scalable, costefficient cloud storage technology to address these needs. Palo Alto based CTERA Networks caters to this need of both enterprises and service providers and in the process has created a buzz amongst the tech circles. The company enables its customers to use the cloud storage infrastructure of their choice whether private, public or hybrid, while ensuring end-to-end security and overcoming performance and scalability challenges. 'Single Pane Of Glass' Management Founded in 2008 by Liran Eshel, CEO, and Zohar Kaufman, VP R&D, the company provides a cloud storage platform that enables enterprises and service providers to create, manage and deliver cloud storage services, whether internal or external. The company's unique platform includes built-in functionality for file sync and share, hybrid on-premises/cloud backup,managed branch storage and mobile collaboration which is unified under 'single pane of glass' management. "Our platform extends storage infrastructure to make it "cloud enabled", and integrates with the majority of object-based and other scalable storage systems, as well as cloud service providers. CTERA is a technology vendor - we do not provide cloud services, but rather empower service providers and enterprises to do so," adds Eshel. It is the only company to provide a platform that addresses multiple cloud storage "killer apps" and supports both hardware and software-based deployment scenarios to optimize performance across remote office and branches,as well as mobile workers. Providing Game Changing Services CTERA's core development team originated from Check Point Software, so security is very much in the company's DNA and is embedded into all aspects of the company's solutions, which feature built-in source-based encryption and other security measures. The company's cloud storage platform is comprised of cloudside management software, cloud storage gateway appliances and ndpoint software agents for PCs and mobile. Customers can choose any mix of software/hardware to serve their storage, backup and file sharing needs in an optimal way. Active Directory integration, including support for the most complex AD forests, makes it easy for enterprises to deploy CTERA's solution, while global source-based de-duplication ensures optimal erformance over the WAN. CTERA integrates with storage infrastructure from EMC, IBM, HDS, OpenStack, Caringo, DDN, Scality and leversafe, as well as with public cloud providers such as AWS and Rackspace. Winning the Trust of Clients Due to its game changing technology, in a very short period of time, the company has won the hearts of many firms which include some of the Fortune 100 enterprises across multiple industry sectors including financial services,life sciences and energy. Major service providers also use CTERA's platform to drive their businessgrade cloud storage services.They include, among others, Telefonica, CenturyLink, Orange (France Telecom), Telecom Italia and Swisscom. With customers in production with 10,000s of appliances deployed and 100,000s of active users, CTERA's technology is proven to scale to very large deployments. CTERA plans to continue its good run and expand its platform both to support additional usecases as well as to extend further the value of cloud storage for security, compliance and data mining purposes in the future. Management Liran Eshel, CEO, Co-Founder and Zohar Kaufman, VP R&D, Co-Founder Description A provider of a cloud storage platform for enterprises and service providers, enabling the delivery of a suite of storage applications over the infrastructure of their choice