Strategic PartnersIssue 024 records
Everpure
Second Everpure edition, four new sources. Everpure published its first PureKVA result with a disclosed method on October 30: on one Gemma-4 26B endpoint serving 40 simulated agents, returning-turn latency without a cache tier jumped from about 0.1 to 45 seconds once GPU memory filled near 22 agents, while Pure KVA held 0.11 to 0.30 seconds; hardware is undisclosed and each configuration ran once. Everpure's October 23 public sector post gives state, local and education leaders a 90-day AI readiness plan, but its named SLED customer examples are explicitly not AI outcomes. Independent heise analysis (October 20) stresses that catalog metadata access does not grant data-read rights and that Everpure competes with entrenched catalogs, and Everpure's CTO concedes inference and disclosure governance is unsolved industry-wide. No new named SLED AI deployment with measured results was found; applicability remains interpretive.
- Evidence records
- 4
- Cross-source patterns
- 1
- Evidence classes
- 3 vendor claim1 independent reporting
- Outcomes
- 2 emerging1 effective1 mixed
- Source freshness
- 3 new this fortnight1 recent
- Research completed
- 2026-10-31
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Agent access governance, not storage performance, is the gating question, and Everpure concedes open limits
Heise notes Everpure's catalog metadata access does not authorize agents to read data and that buyers must weigh overlap with existing catalogs. Everpure's public sector framework says agency AI stalls on ownership, skills and trusted data. Everpure's CTO concedes governing what models infer and disclose is unsolved, including at Everpure. None of the three reports SLED deployment results.
Operating questionBefore connecting agents, who separately approves catalog visibility, data-read access and output review for each workflow, and how are those decisions logged?
Supporting evidenceheise online / iX (Harald Weiss)Everpure (Austin Boone)Everpure (Par Botes)
Evidence ledger
Full record · every source keeps its link and limitations
Everpure, formerly Pure Storage, shifts from storage to data intelligence
heise reports Everpure's repositioning from storage to enterprise data management around Data Intelligence (built on the 1touch acquisition announced in February 2026): discovery and classification across SMB/NFS shares, S3 and, per Everpure, nearly 300 connectors to third-party storage, SaaS and mainframes, without copying data to a central repository. It describes File Intelligence (permissions and staleness of shares), MCP access for agents and security tools, and Pure1-based deployment, and states that metadata access does not by itself authorize agents to read the underlying data. It frames the move as competing with Salesforce Data 360, Dell's AI Data Platform, Informatica, IBM and Collibra.
Why it matters, evidence and limitations
- Why it matters
- Useful for state agencies, universities and districts that already run data catalogs or governance tools: the article's central question is whether Everpure complements, replaces or duplicates them. No SLED customer or deployment is mentioned.
- Evidence and measured results
- Independent reporting based on a press event and vendor statements; no testing, customer interviews or measured results. The Gartner figures cited (up to 80% agent accuracy and 60% cost improvements from semantic context) are relayed third-party claims, not Everpure results.
- Limitations and uncertainty
- Published October 20, before the stream's last run, but not previously archived; included because it is the most recent independent analysis of the October-availability release and adds the competitive and authorization framing. 'Eventually' deployed through Pure1 suggests availability may still be phased.
Readiness Is Not Curiosity: Why Your Agency's AI Program Is Waiting on People, Not Models
Everpure's public sector team argues that agency AI programs stall on ownership, workforce skills and untrusted data rather than models. It proposes four layers built together (mission, people, operating model, data foundation), a four-step fluency ladder (aware, fluent, practitioner, builder), treating shadow AI as a sign the sanctioned path is too hard, and a 90-day plan with exit criteria (signed baseline, named owner, workflow tested against real examples, stop/improve/scale decision). It cites Spring Branch ISD, Louisiana's Office of Technology Services and the Mississippi Department of Revenue as non-AI capacity examples, and promotes Everpure Data Stream.
Why it matters, evidence and limitations
- Why it matters
- Directly SLED-focused: written for state, local and education leaders and names three SLED customers. The framework itself is vendor-neutral and maps to public records, accessibility, procurement and audit requirements the post lists.
- Evidence and measured results
- Vendor opinion and framework with no measured AI outcomes. The customer examples are explicitly not AI stories (backup recovery, database refresh, auditability) and their figures (16+ hours to near-instant refresh, 90 minutes to 20 seconds reporting) are vendor-relayed. The IDC survey cited was commissioned by Everpure.
- Limitations and uncertainty
- Marketing content ending in a Data Stream sign-up offer. The claim of a year of AI program work with SLED customers is not substantiated with named AI outcomes. Customer metrics are not AI results and should not be presented as such.
The Agent Snowball Effect: How Pure KVA Keeps Long-Running Agents Fast
Everpure publishes a load test of Pure KVA 1.2 (a KV-cache backend for vLLM using a 32 GiB host-memory tier and NFSv4.1 storage) serving 40 simulated agents on one Gemma-4 26B endpoint. Without KVA, median time to first token on returning turns stayed near 0.1 seconds until about 22 agents, then rose to 17 seconds within 30 seconds and about 45 seconds by the end; with KVA it stayed between 0.11 and 0.30 seconds. On an idle server, NFS restores were about 10 times faster than rebuilds at the agents' context lengths and 32 times faster at 118K tokens. Everpure estimates matching KVA without a second tier would need about 1.8 times the GPU memory.
Why it matters, evidence and limitations
- Why it matters
- Relevant to university research computing and any state or campus team planning on-premises agent inference on limited GPUs; it shows pilots with few agents can hide a sharp latency cliff. No SLED deployment is cited.
- Evidence and measured results
- Vendor-run benchmark with a published appendix (model, harness, think-time distribution, arrival schedule, corpus size, tiers). Each configuration ran once with one seed; the GPU type and count, server hardware and NFS backend system are not stated in the post. No independent replication.
- Limitations and uncertainty
- Responds to earlier criticism (StorageReview, covered in the October 29 edition) that the up-to-20x TTFT claim lacked a published configuration, but the hardware is still undisclosed and results are single runs. The 1.8x GPU memory figure is a vendor estimate, not a measured comparison against a second-tier alternative such as open-source cache offload.
AI Models Get the Glory. Your Data Decides the Outcome
Everpure's Par Botes argues that production AI is constrained by data context rather than compute: redundant re-embedding and re-retrieval show up as GPU or token pressure, and buying capacity capitalizes the waste. He says governance speeds teams when access rules are built into data management. He concedes that cost per useful token is hard to measure, that tooling to separate productive from redundant work is immature, and that governing what a model may infer and disclose from permitted documents is unsolved industry-wide, including at Everpure; he recommends logging what agents reach for.
Why it matters, evidence and limitations
- Why it matters
- Applies to agencies and universities justifying GPU or token spend: the post suggests auditing redundant data work before expanding capacity. The admitted gap on inference and disclosure control matters for FERPA, HIPAA and public records contexts. No SLED example is given.
- Evidence and measured results
- Opinion from a vendor executive with no data, measurements or case studies. The value lies in the explicit limitations it concedes, which are stated by the vendor about its own products.
- Limitations and uncertainty
- Published October 15, before the last run, but not previously archived; included now because it is the vendor's own statement of limits for the agent-access features covered in this stream. Contains no evidence for its compute-waste claims.
How to read this edition
Source findings, measured results and limitations come from the cited publications. Patterns, operating questions, role takeaways and implementation considerations are Lighthouse Advisory interpretation, stated as questions to validate locally rather than guaranteed outcomes. Vendor and operator claims are labeled as claims. Full research method.
- Vendor claim
- A supplier-provided assertion that has not been upgraded to independent evidence.
- Independent reporting
- Independent reporting with attributable sources but without a formal evaluation design.