Everpure announced a set of platform capabilities aimed at getting enterprise AI out of pilots and into production, all of them due in October. The headline is PureKVA (Key-Value Accelerator) for FlashBlade, which pre-stages inference context directly into GPU memory and, by Everpure’s count, delivers up to 20x faster Time to First Token. Alongside it are an always-on version of the DeepReduce compression engine across FlashBlade systems, native Model Context Protocol (MCP) support in Everpure Data Intelligence so agents can query live data catalogs in natural language, a file-share exposure tool, and a reference architecture built on open-weight models. Everpure frames the whole set as an extension of its Data Primacy vision, the idea that data, not applications, should drive enterprise architecture in the AI era.
PureKVA and Always-On DeepReduce on FlashBlade
PureKVA works on the same problem we covered in our KV 캐시 오프로드를 플래시에 적용 work: long-context inference spends GPU time rebuilding context it has already computed. Everpure’s approach has FlashBlade stage that context straight into GPU memory, with no dataset relocation out of primary storage, which the company says eliminates GPU idle time, raises token throughput, and cuts response lag for real-time applications while supporting enterprise multi-tenancy. The up-to-20x TTFT figure is Everpure’s claim and comes with no published test configuration, so treat it as a ceiling until the company shares the setup behind it.
DeepReduce itself is not new; Everpure introduced Purity DeepReduce on FlashBlade in February, a similarity-based reduction that runs off the write path, citing a median of about 2:1 on top of data that was already compressed. What’s new is the always-on mode: the engine now scans storage blocks continuously across FlashBlade systems to find sub-block similarities that conventional deduplication misses, including inside pre-compressed content, and usable capacity expands automatically with no manual scheduling and, per Everpure, no impact on write performance.
MCP, File Intelligence, and an Open-Weight Reference Architecture
에버퓨어 데이터 인텔리전스, the classification layer that came out of the 1touch acquisition, now implements MCP so AI agents and security tools can query its live data catalogs in natural language, find relevant data, and read its sensitivity class before using it as an input to AI, agent workflows, or analytics. The catalog spans the Everpure Platform, public clouds, SaaS applications, and third-party storage. Deployment runs through the existing Pure1 console, with no separate management servers and, Everpure says, no professional services engagement required.
Privacy-First File Intelligence shows who can access each file share and how stale it is without reading file content, so teams can close exposure and reclaim capacity before opening shares to agents. Everpure ties the same classification work to cyber resilience: knowing which data is sensitive and who touches it is what determines how it gets protected and what gets recovered first. Rounding out the set is an Intelligent Token Optimization Reference Architecture that uses open-weight models to give enterprises more control over their data and more predictable AI spend by cutting API token consumption from external providers. Everpure did not name the models or GPUs in the reference design.
“Enterprise AI is hitting a wall not because the models are lacking, but because data is not ready for real-time, autonomous agents. We are eliminating that friction. By making enterprise data continuously governed, automated, and instantly accessible, we’re giving organizations the foundation to move AI out of the lab and into production with the necessary confidence,” said Prakash Darji, General Manager, Data & Digital Experience at Everpure.
The capabilities announced today will be available this October. They follow June’s Data Stream launch and land a month after Everpure topped both axes of Gartner’s enterprise storage Magic Quadrant 또.
아마존