Blimp - AI Data Engine for Inference and Training

Blimp is a low cost smart cache and AI query optimizer for your analytics, AI, and ML pipelines on Apache Iceberg and Amazon S3. It is the data engine for inference and training. It lowers your inference token spend with a fresh context layer, and lowers your training cost by keeping GPUs at full utilization. Simply point your existing engines at one S3/Iceberg endpoint, no data migration needed.

• No migration. Point your engine at Blimp and you're done, same SQL, same tables, same credentials, one S3/Iceberg endpoint. Blimp runs alongside Spark, Trino, Databricks, and Snowflake, and works with any Iceberg catalog. Cached data is served interchangeably over the S3 API or an NFS mount.
• Lower inference token spend with a fresh context layer. An AI harness matches your query patterns, then authors and materializes Iceberg materialized views for you. No rewrites, no schema changes, no MV management. Models and agents read a compact, per entity materialized view kept current via change data capture (CDC) instead of stuffing prompts with stale RAG. Fewer input tokens, better signal. The first run builds the view; every similar query after is served subsecond from cache.
• Lower training cost, full GPU utilization. A 100 Gbps NVMe cache feeds your GPUs from fast S3 and NFS endpoints, eliminating data starvation so accelerators stay saturated instead of idling on I/O. Measured on live clusters: up to 10 GB/s GET, 4 GB/s NFS write, submillisecond time to first byte, and >90% accelerator utilization on MLPerf with built-in fio, warp, and MLPerf benchmarks to verify it on your own cluster.
• Fresh features for inference. Blimp's AI pipeline refreshes your feature datasets and online/KV store (Redis, DynamoDB, feature store) continuously off the CDC stream, so your model's per request lookups return current values with no scheduled jobs to maintain for fraud detection, product recommendation, and customer service.
• Elastic, fast to adopt. Two cluster options: the cost-efficient 2/1 (28 vCPUs, 128 TB) and the high-throughput 8/1 (100 vCPUs, 512 TB), both infinitely scalable. Clusters spin up in 10 to 20 minutes and tear down in 5 to 10; select instance size, start/stop, and resize storage dynamically from the UI or programmatically. From the first click: a demo in minutes, a pilot within an hour, and a production sidecar within a day.
• Hack proof by architecture, fully ACID. Every write commits as an atomic data entity, no partial writes, no torn reads, no orphaned files. Every identity is anchored on the blockchain, so keys cannot be spoofed and identities cannot be forged, and every message between components is cryptographically signed.