Decentralized science
Molecular folding and climate simulation on millions of idle Apple chips — no supercomputer time bought.
Decentralized · Private · Apple-native
A supercomputer that
grows in the dark.
Orchard coordinates the idle Neural Engines and on-device Foundation Models of millions of iPhones, iPads, and Macs into a cooperative swarm — solving massive problems while 100% of raw user data stays on the device.
- Marginal compute cost
- $0
- Raw data off-device
- Never
- Runs when
- Plugged in · Wi-Fi · idle
Four pillars
-
01
Local node execution
Every device is a node, running native on-device inference via the Foundation Models framework on the Neural Engine. No token fees. Activates only when charging, on Wi-Fi, and idle — typically overnight.
-
02
Local micro-swarms
Your iPhone, iPad, and Mac on one LAN shard a high-parameter model across their unified memories — MLX-style peer-to-peer clustering runs models no single device could hold.
-
03
Global agentic workflows
A Task Router fragments a massive problem into thousands of micro-prompts. Edge agents return structured output; millions of results merge into one accurate map.
-
04
Cryptographic privacy
Federated learning with Secure Aggregation and Differential Privacy. Only mathematically masked gradient updates ever leave a device — individual data is impractical to reconstruct.
What a swarm this size can grow
A commons knowledge graph
Local agents collaboratively read, summarize, and index public web data into a collectively owned semantic database.
Collaborative creativity
Cinematic AI video rendered frame-by-frame, diffusion tasks distributed across millions of GPUs overnight.
We don't wish the physics away
- Bandwidth is the bottleneck. Consumer Wi-Fi is no NVLink. Orchard targets asynchronous, batch-processed work — never real-time conversational swarm inference.
- iOS limits background compute. Heavy work runs in the charging/idle sleep state, chunked into resumable units that respect thermal and battery policy.
- Nodes are untrusted. Redundant assignment and consensus voting guard every result before it counts.