AUROXEON
Digital Design
Surface, architecture from the grid to the gate.
Surface runs on a national estate, in an enterprise rack, on an edge module and - on the roadmap - inside dedicated silicon. It is the same architecture at every scale. This page is the descent through all of them.
Scroll to descend
Enterprise AI inherited the cost structure of the internet.
The current stack was built to serve everyone at once. That assumption set the compute, the energy and the infrastructure dependency - not the difficulty of the work being done. Auroxeon's own framing is AI built for your business, not the internet.
Enterprise, sovereign and physical systems have a different requirement set: controlled infrastructure, persistent organisational knowledge, auditability, local execution, predictable cost. Different requirements can demand a different architecture.
This is not a claim that the current generation of models is deficient. It is a claim about fit.
Compute concentration
Capability is coupled to access to a constrained accelerator supply chain.
Infrastructure control
Regulated and sovereign operators need the execution boundary inside their own estate.
Physical systems
Machines, sensors and industrial edge cannot depend on a round trip to a data centre.
Surface is a field that keeps its shape.
Work enters as a probe. The system routes it across a persistent state surface, executes under control, and leaves the route behind. Execution is deterministic and replayable; the surface is permanently changed by the work it has done - which is what learning without retraining means here. In silicon, foundation knowledge and decision rules are built into the hardware, with a small memory layer that learns.
Evidence is the brand.
Every figure carries a workload, a baseline, the hardware it ran on, a date and a maturity state. Where methodology has not been published, the figure says so rather than borrowing confidence it has not earned.
Benchmark 01 - deployed density at equal power
One large model, or 1,675 Surface nodes.
Same power envelope, one mark per node
Baseline: one 200B-parameter LLM. Surface node draw stated as ~15 W per node against 300–700 W per GPU, with 1 vCPU per node.
Benchmark 02 - repeated workload
The second time is not the first time.
A conventional stack pays full price for the same problem every time. Surface retains the route, so equivalent work costs less on every subsequent execution. This is the economic core of the architecture, and the figure that most deserves independent validation.
Sovereignty is a boundary you can draw.
Not a policy, not a region setting. A real edge around data, compute, execution, knowledge and eventually silicon - governance identical in every environment, and deployable in a data centre, in cloud, on-site or air-gapped. Auroxeon is pursuing this in region: Tees Valley sovereign AI, with SIRC on Net Zero AI.
The whole company, by maturity.
One register. Filter it, open any row. Nothing on this site is stated without a position here, and a roadmap item can never be dressed as a validated one - the states are drawn at different densities, not different colours, so the distinction survives greyscale, projection and colour blindness.
States shown are placeholders for the prototype. Auroxeon assigns the real position of every row.
Run Surface on your infrastructure. Prove the economics first.
Not beta access. A bounded technical evaluation against your own baseline, on your own hardware, producing a Surface Evaluation Report you own. Enterprise evaluation is open; indicative duration 8–12 weeks. Named sectors: Life Sciences, Defence, Legal, Finance.
The thesis is the descent.
Software into infrastructure into silicon. Each step down the compute stack widens the addressable position and deepens the defensibility, because the architecture and the hardware are designed against each other rather than one being fitted to the other. The product architecture already spans it: Geode node clusters in the hall, SurfaceLEGACY as the bridge from existing AI estates.
Public materials are deliberately limited to what can be evidenced. Diligence materials are released against a qualified request.
Entry
Enterprise runtime
SurfaceCORE on customer infrastructure. Evaluation to production. Server, cloud, on-site.
In pilot
Expansion
Edge and physical systems
SurfaceEDGE into industrial estates, machines and sensors. Multi-processor blades.
Roadmap
Position
Dedicated silicon
Hardware and architecture co-designed. Rules and knowledge in the hardware.
Roadmap
Bring us a workload and a baseline.
Enterprise evaluation is open. Analyst briefings and investor diligence run as separate paths, with separate material. Auroxeon Ltd, Tees Valley · Auroxeon US LLC, Austin.