Everpure (NYSE: P) claims its FlashBlade systems can now deliver up to 20x faster Time to First Token (TTFT) for LLM inference by pre-staging context directly into GPU memory, part of a bundle of data management capabilities the company says will be available this October.

The announcement, made in London on September 30, 2026, extends the "Data Primacy" vision Everpure introduced at its Pure//Accelerate conference in June. That principle holds that data, not applications, must drive enterprise architecture in the AI era. The new capabilities target three bottlenecks the company says are stalling enterprise AI at scale: fragmented context, unpredictable inference costs, and slow deployments.

"Enterprise AI is hitting a wall not because the models are lacking, but because data is not ready for real-time, autonomous agents," said Prakash Darji, General Manager, Data & Digital Experience at Everpure. "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."

Agents get governed access to live data

The updates center on Everpure Data Intelligence, which discovers, classifies, and contextualizes enterprise information at its source — across the Everpure Platform, public clouds, SaaS applications, and third-party storage. Three additions stand out.

First, native MCP integration. Everpure implements the open Model Context Protocol so AI agents and security tools can query live data catalogs in natural language. Agents can locate relevant data and determine its sensitivity class as an input to AI, agent workflows, and analytics — without custom API work.

Second, turn-key deployment. The capabilities run through the existing Pure1 console, which Everpure says eliminates separate management servers and lengthy professional services engagements, shortening time-to-value.

Third, privacy-first file intelligence. The system shows administrators who can access each file share and how stale that share is, without ever reading file contents. Teams can close exposure and reclaim capacity before opening shares to AI agents.

Performance at the source

On the performance side, the headline feature is PureKVA, a Key-Value Accelerator that pre-stages context into GPU memory on FlashBlade. Beyond the claimed 20x TTFT improvement, Everpure says it supports enterprise multi-tenancy with zero dataset relocation, eliminates GPU idle time, increases token throughput, and cuts response lag for real-time applications. The pitch: inference performance without moving data off its system of record.

A second feature, Always-On DeepReduce compression, scans storage blocks continuously across FlashBlade to find sub-block similarities that traditional deduplication misses, even on already-compressed content. Usable capacity expands automatically, with no impact on write performance and no manual scheduling, which the company says shrinks hardware footprint and cross-cloud spending.

The third piece addresses cost directly. An Intelligent Token Optimization reference architecture built on open-weight models lets enterprises run inference on their own infrastructure, cutting API token usage billed by external providers and making AI spend more predictable.

Security as a byproduct

Everpure argues the same data context that makes information safe for AI consumption also strengthens cyber resilience. Because the platform classifies which data is sensitive and who touches it, that context determines how data gets protected and what gets recovered first after an incident.

Individually, the capabilities address specific gaps in simplicity, security, performance, and cost. Together, Everpure positions them as a single, continuously updated foundation for running AI in production — one designed to keep pace as agentic workflows scale. The company has not published pricing for the October release; competitive pressure from rival storage vendors bundling similar AI data services will test whether the performance claims hold in customer deployments.