Production agent fleets
The first full implementation of the architecture, aimed at resident-session capacity for agentic inference. Its flagship metric is how many live agents a rack holds inside a wake objective, rather than peak token throughput.
Used for: customer-facing assistants, coding agents, research and analysis agents, and workflow automation, at fleet scale.
Architecture and microarchitecture definedRollout-heavy post-training
The same orchestrator and memory-I/O blocks, with dense compute and an optional high-bandwidth memory interface in place of the flash tier. Hardware session fork and bit-exact determinism apply directly to rollout trees.
Used for: reinforcement learning after pre-training, preference optimisation, and agentic self-play.
Numerics studiedAssistants that remember across days
A single system-on-chip carrying a reduced lifecycle state set, with on-package memory and non-volatile storage, so a session survives a power cycle without recomputation.
Used for: phone and home assistants, robots resuming interrupted tasks, and vehicles keeping driver context across ignition cycles.
ConceptSession semantics in someone else's silicon
A subset of the instruction and interface architecture, delivered as an IP block rather than a part. Compliance is enforced in the compiler from the first day, so software written for the largest part runs on the smallest one.
Used for: handset and device SoC vendors adding persistent assistant memory, and accelerator vendors wanting session semantics on a shared compiler target.
DefinitionLong-form generative video
Dense compute alongside a dedicated encode and variational-autoencoder block, over the same memory tiers, so scene and character state persists across a sequence rather than being regenerated per shot.
Used for: generative video platforms and studios. Scope is demand-gated and deliberately unscheduled.
Concept · demand-gatedSovereign and national-scale serving
The dense compute block tiled across a wafer, with edge dies carrying the orchestrator and memory interfaces. Listed for completeness of the architecture's reach rather than as a near-term plan.
Used for: national laboratories and sovereign AI programmes serving a very large single model.
Structurally deferredDesignations are internal programme names. No product is available for sale, and nothing here is a specification, datasheet or offer to supply. Diagrams are schematic representations of the design, not layouts.
What an operator would get
Objectives the architecture is designed against: targets for a system in development, not specifications of an available product.
Capacity measured in agents
Resident sessions per dollar inside a wake objective, not tokens emitted in a burst.
Idle that costs almost nothing
A quiet session falls to cheaper memory automatically, compresses on the way, and returns before anyone notices it left.
No recompute on return
Restoring context should not mean re-running prefill. State comes back as state.
Encrypted the moment it sleeps
State is encrypted under a per-session key as a side effect of leaving the active tier.
Deterministic and auditable
Lifecycle transitions are journaled in a hash chain and can be replayed and verified.
Sharing without copying
Prompts, schemas and documents held once and computed against in place.