AIStor

CoreWeave's LOTA benchmarks are transparent and well run, but the warm-cache headline numbers measure local NVMe inside the GPU nodes, not the object store behind them. Using the same open-source Warp tool with no cache at all, MinIO AIStor delivered roughly 33.5 GiB/s of durable, erasure-coded GET throughput per storage node over plain TCP, making the case for a faster store rather than another layer in front of it.

Join our live workshop (Nov 12 or 18) and learn to build, secure, and monitor an AIStor Table Share for querying on-premises data from Databricks.

Every coding assistant, enterprise chatbot, and agent loop resends the same tool schemas, system instructions, and policy text on every turn, and the GPU rebuilds all of it into attention state before it can emit a single output token. Prompt caching computes that prefix once and reuses the KV state, and MemKV takes it from a per-process optimization to a shared NVMe-backed tier that survives routing across replicas, HBM eviction, worker restarts, and concurrency.

Cache hits are not the outcome. Part 3 turns the architecture from Parts 1 and 2 into an evaluation framework: capture an honest tail-latency baseline first, treat MinIO's published 53-second-to-703-millisecond TTFT result as a proof point to reproduce rather than a business-case input, and follow the measurement chain from repeated context through prefix reuse and avoided recompute to unit economics, ending in a buyer checklist that judges a shared context tier on P99 TTFT and jitter rather than throughput or capacity.

A prefix cache that only helps one process is useful, but requests move across replicas, HBM fills, sessions spill, and workers restart, so reuse that lives inside a single worker is not a fleet architecture. Part 2 works through what the serving stack needs once KV state leaves local GPU memory: a tier that is larger than HBM, fast enough that restore beats recompute, shared across workers, and reachable through the runtime's own KV transfer path, which is memory behavior at cluster scope rather than storage.

Agentic AI applications resend the same project rules, tool schemas, and document context on every turn, so an expensive GPU fleet spends much of its time rebuilding a prefix it has already processed. Part 1 of three reframes prompt caching as an operating-margin lever rather than a model feature, maps prompt caching, prefix caching, KV cache, and KV cache offload to the business questions each one answers, and argues that reusable context needs a memory path rather than ordinary enterprise storage.

Worker A gathers evidence, publishes an accepted handoff, starts another edit, and is deleted mid-sentence with the unfinished tail left visible. Worker B starts in a fresh runtime with no session state, verifies the last accepted boundary in the same authorized AIStor Memory Workspace, discards the unchecked tail, and continues the work rather than restarting it.

Personal AI agents become useful as they learn you, but that familiarity should not require one agent accumulating your entire life. AIStor Memory gives each agent a bounded relationship with its own learned history, active workspace, and credential scope, all under your control.

Every agent begins with the experience your organization has already earned.

Moving your Iceberg catalog to MinIO AIStor no longer means re-registering tables by hand. The mc table migrate command reads metadata from your existing catalog, whether that's Nessie, Polaris, Hive Metastore, AWS Glue, or a SQL-backed catalog, and registers your tables directly in AIStor's built-in Iceberg REST catalog. Your Parquet files stay where they are.

MinIO joins the Databricks Software-Defined Storage Ecosystem. Live, zero-copy access to on-premises data.

MinIO AIStor joins ClickHouse House Mates as a validated, trusted storage partner for ClickHouse Private deployments. AIStor gives ClickHouse the high-performance, fully S3-compatible object storage layer it needs to scale without limits.

Until recently, data that stayed on-premises was data that Databricks couldn't reach. If your analytics and AI workloads ran in Databricks, and your most valuable data lived on-prem, you had two options: build and maintain a replication pipeline to copy data into the cloud, or accept that certain datasets simply wouldn't participate in your cloud analytics.Both options carry real costs. But a third option now exists: Databricks querying on-premises data directly, with no copies and no pipelines, through the open Delta Sharing protocol embedded natively in MinIO AIStor.

Databricks needs your on-prem data. AIStor Table Sharing delivers it live — no replication, no pipelines.

Modern AI architecture rests on a comfortable assumption: when AI slows down, the fix is more compute or a better model. Bigger GPUs. Denser clusters. New architectures. That assumption is now costing organizations real money.

Storage, not models or compute, is now AI's biggest bottleneck. The fix is an architectural one, not a bigger GPU.

106 million vectors indexed 12x faster. AIStor with NVIDIA cuVS and GPUDirect RDMA rewrites the benchmark.

Legacy storage talks to the AI factory. AIStor lives inside it — running natively on NVIDIA BlueField-4 Vera

Delta Sharing + AIStor enables zero-copy access to on-prem data from Databricks without duplication

Table Sharing connects Databricks directly to on-prem AIStor data via Delta Sharing without copying/syncing

Technical comparison detailing 13,000+ commits separating AIStor from unmaintained OSS

Announcing AIStor Free, Enterprise Lite, and Enterprise subscription options

Why "Awardable" matters for gov agencies—MinIO AIStor validated in CDAO Tradewinds for federal AI infrastructure

Former Intel Capital exec Ran Kurup joins MinIO as CCDO—lead strategic growth in AI infrastructure era