Technical whitepaper
Deadwood platform overview
A concise reference for engineers and procurement: how training touches your warehouse, how usage is metered, and how settlement aligns with production traffic.
Abstract
Deadwood is compute infrastructure for supervised preference learning and production inference. Teams integrate a single API, train models against governed data sources—including Snowflake warehouses without bulk export—and route predictions through versioned deployments.
This note summarizes control planes, metering, and settlement assumptions buyers evaluate alongside Premium and Enterprise contracts.
Warehouse-native training
Premium connects your Snowflake account under credentials you issue. Training jobs submit authorized queries; large datasets remain in your cloud boundary. Deadwood charges for orchestration, training units, and API usage—Snowflake bills query compute separately unless Enterprise terms include a dedicated cluster envelope.
Metering and SLAs
API calls, training runs, and model swaps are counted per workspace. Dashboards expose usage, tier limits, and overage math. Accelerator and Autonomous SKUs (where offered) publish uptime targets; incident response follows subscribed notification channels.
Settlement and treasury
Premium maps usage to real-time Avalanche postings for auditable balances and creator payouts where marketplace features apply. Local and Enterprise variants may batch or alternate chains per contract—implementation detail is pinned in order forms.
Security posture
Transport uses TLS; Premium advertises SOC 2 Type II and GDPR-aligned processing with encryption at rest for Deadwood-managed artifacts. Customer VPC peering, SSO, and residency are Enterprise negotiation items.
Next steps
For tier limits, Snowflake billing split, and trial mechanics, see the FAQ on the Snowflake Premium page. For custom data residency, SLAs, or warehouse topology, route through sales with your architecture diagram.