Platform
Ask the database. Run the model. Never open the files.
RainDB and ZChat are the two things you run on Umbra today. Same idea: the work happens while the data stays locked.
The stack
Four layers
Umbra silicon
Encrypted logic at production throughput.
Umbra runtime
Deterministic execution on ciphertext.
RainDB · ZChat · SDK
Databases, models and custom workloads on one runtime.
Your application
Business logic that never sees plaintext.
RainDB
Fully FHE-encrypted database
Queries execute against ciphertext. Results return encrypted. The engine holds no plaintext row at any point.
Deploy beside Umbra silicon: on a node or next to existing infrastructure: for analytics where no single role may read raw records.
No plaintext row. At any point. In any process.
ZChat
Private encrypted AI
Private inference on ciphertext end to end. Prompts and outputs stay encrypted through execution: no plaintext GPU boundary.
Reference configuration: 6 × U100 + 2 × NVIDIA GPUs
Umbra SDK
Build once. Carry forward.
Compile encrypted workloads against the Umbra runtime. C and Python bindings ship first. Integration done on Umbra U100 today runs unchanged on Umbra V-series1.
Umbra U100 ships at 40M ops/s per card. Same runtime, same SDK, same integration work across generations.
SDK overview
Integration guide, API reference and workload templates. Sent to your work address.