Meet the Cube.

The Cube is Edgescale’s Physical AI appliance and flagship product. It is a fully managed, industrial-grade appliance that runs mission-critical AI directly at the point of work. It is not a server to configure. It arrives pre-validated, connects to the local network, and runs – bringing the power of AI to the physical world.

The Cube includes server-class compute and storage, a hardened operating platform built on Red Hat, the Data Manifold operational data layer, the Sovereign AI runtime, the Edgeport application and agent hosting layer, and embedded operators that maintain and govern the environment autonomously. Each Cube carries a 99.9% uptime SLA and handles roughly 100 cameras, 300,000 tags, 25,000 task-hours, and about 20 concurrent agents at a single site.

The Cube installs on the local network, physically close to the machines and sensors it reads from and writes to. That can be a factory floor, a substation, a pump station, a warehouse, a hospital wing, or any remote site with no reliable connectivity.

Placing compute at the source is what lets the architecture jump the usual bottlenecks: limited WAN bandwidth, air gaps, and legacy firewalls. Each Cube also maintains an independent out-of-band cellular connection for management traffic, so remote support never depends on the customer’s production network. Mounting options include rack, wall, rail, surface, and portable.

Through the Data Manifold, an autonomous operational data layer with 300+ native connectors for legacy and modern systems: OPC UA, MQTT, Modbus, SQL, REST, RTSP, file shares, and direct device integrations. It pulls from PLCs, SCADA, historians, ERP, MES, CMMS, and camera feeds without custom pipeline code.

Data integration is where most industrial AI projects stall. Traditional integration runs months of engineering and $200,000 to $500,000 before anything of value is produced. The Data Manifold takes a zero-ETL approach: purpose-built agents build, test, and maintain the integrations, then adapt them as source systems change. It also acts as an anti-corruption layer, so stable operational systems stay protected while AI applications iterate.

Yes. The Data Manifold runs integrated brokers and databases locally, including MQTT and Kafka on a broker-based architecture, plus time-series and vector stores. Data is ingested, buffered, and structured on arrival, with 12TB of local NVMe available for caching and processing.

The result is a Unified Namespace: a single real-time, coherent data layer that agents and applications can query at the site, rather than a set of point-to-point integrations. Only distilled results move upstream, which is how sites cut WAN bandwidth by 90% to 99% while keeping raw video and sensor data local.

The Cube’s Sovereign AI runtime supports 7B to 70B quantized models, mixture-of-experts models, and multimodal models. Supported families include GPT-OSS 20B, Llama 3.1 8B and 70B, Llama 3.3, Mixtral 8x7B, Gemma 3 12B and 24B, E5-Large for embeddings, Whisper for speech, PaliGemma 2 for image, and YOLO.

Inference runs locally on vLLM with an OpenAI-compatible API, with tool calling, guardrails, fine-tuning, and benchmarking supported. Atom runs models of 1B to 3B quantized and below. Prism runs 70B to 400B and larger.

Yes, through Edgeport, the application and agent hosting layer. Leading third-party software runs natively, including Palantir Edge, Confluent Platform, and Ignition Edge. Custom applications deploy as standard Docker and OCI containers, and agents run on common frameworks including LangChain, LlamaIndex, and AutoGen.

Edgeport is also how the same application reaches an entire footprint. Build once, deploy across the fleet, manage centrally, with no per-site engineering and no local IT.

Hybrid Inference is a semantic routing layer that reads each prompt before it goes anywhere. Sensitive content routes to local private models running on the appliance. Safe workloads route to frontier cloud models. Teams get access to the best models available without putting proprietary operational data into a public cloud.

Red Hat research on this approach shows intelligent routing cuts unnecessary calls to large public models by up to 40% with no loss in quality, and vLLM Semantic Router benchmarks show accuracy improving by more than 10% with latency and token usage nearly halved. That is the difference between AI as a compliance risk and AI as a competitive advantage.

Yes. Offline is a normal operating state, not a failure mode. The Cube continues running local workloads and inference during a partition, governing itself from last-known-good verified artifacts, with bounded degradation and integrity-verified checkpoint recovery when connectivity returns.

For sites with intermittent connectivity, updates sync when a link is available. For strict air-gapped sites, updates move through secure, staged procedures. This matters more than it sounds: many operators will not consider an operational use case at all unless it survives a complete internet or cloud outage.

Through the Software Distribution Fabric. Each appliance creates a cloud-like environment at its location, so you define the application once, deploy it everywhere it needs to run, and manage the fleet from a central plane. There is no per-site engineering and no local IT requirement.

Custom on-premise engineering site by site is what turns a minutes-in-the-cloud deployment into an 18-month field program, and it leaves every site configured a little differently. Documented rollouts using the drop-ship model have compressed from 18 months to 6 weeks, with updates applied uniformly across the footprint.

In copilot mode, operators work in natural language on the screens they already use: tablets, ruggedized handhelds, push-to-talk devices, HMI displays, and AR/VR headsets. They ask questions, get answers grounded in their own site data, and document work as they go. They do not install, configure, or maintain anything. The human makes the decision; the AI does the research, surfaces context, and reduces the cognitive load.

Autonomous mode is a separate pattern, used where systems operate at machine scale without a human in the decision loop. People direct; systems execute; governance sits between the AI and any physical action. Either way, the point is amplifying the operators who already know the plant, not replacing them.

The appliance is online in about 15 minutes. Full production status, including integration with local systems and data flowing to applications and agents, typically takes days to weeks depending on the number of source systems.

For context, one automotive just-in-time supplier reached production in roughly 12 weeks with no custom development, then scaled from 2 Cubes to 205 planned Cubes across their plant expansions. A comparable custom build takes 18 to 24 months to reach the same point.

Edgescale AI does. The Cube is a fully managed appliance with embedded operators – autonomous agents on the device that handle the DevOps, SRE, and solution engineering work of keeping the environment healthy. Software and model updates are delivered continuously with zero downtime, and each Cube is remotely managed over its own out-of-band cellular connection.

Nothing on this list lands on the operations team. That is deliberate. The most common objection we hear is that a new system will turn into another IT project the plant has to babysit, so the Cube is built to require nothing from the site.

A single Cube carries a 99.9% uptime SLA. Deployed as a coordinated cluster, the platform reaches 99.99% or better. Cubes operate standalone or in clusters, and continue running local workloads through a WAN or cloud outage.

Current-generation Cube MSRP is approximately $45,000 per year per Cube, on a typical term of three or five years, payable annually or all upfront. Professional services run approximately $100,000 per configuration and cover pilot, installation, configuration, and training. The deployment model is self-hosted, with a hosted option available.

The cost structure is the point. Cloud AI inference is metered and compounds with every use case added, every agent run, and every query. A Cube is a fixed annual cost regardless of workload, which is what makes multi-use-case expansion financially predictable.

Edgescale AI supports three commercial models that map to how critical infrastructure actually buys:

• Leased – a fixed annual fee for the complete solution, booked as OpEx
• Capitalized – purchased outright, depreciable over its useful life, rate-base eligible for utilities, and carried on the balance sheet
• Hybrid – capitalized hardware plus a software subscription, similar to a standard maintenance contract

The capitalized path is what lets a CFO treat AI capability as infrastructure rather than as an unpredictable services line item. Owning sovereign inference scales with value. Renting metered inference scales with exposure.

Edgescale AI uses an embedded engineer model. We place a technical lead alongside your implementation team, take ownership of the physical infrastructure, connectivity, and integration, and let your team focus on business logic and the application layer.

Engagements follow a Land, Expand, Scale motion. Land: a single Cube proves value at one facility within weeks. Expand: the same configuration replicates horizontally across the footprint through Edgeport, with no per-site engineering. Scale: capacity and new use cases stack on the same infrastructure, with Prism handling enterprise-scale inference.

Some teams can. The question is what it costs. A do-it-yourself on-premise AI stack requires ML engineering talent, takes 18 to 24 months to reach production, and runs approximately $12 million for an SI-led platform build per generation, with cost and schedule variance around 50%. It also leaves your team owning integration, certification, lifecycle management, and security architecture across the full stack, permanently.

The Cube lowers total solution cost by 60% to 80% versus sourcing hardware, software, orchestration, and engineering separately, and it pulls use case ROI forward by 12 to 24 months. Complex multi-vendor coordination is the work Edgescale AI takes on so customers do not have to.

These are reference ranges from documented deployments:

• Payback: 2 to 3 months for agent and co-pilot deployments
• Avoided integration cost: $200,000 to $500,000 for data integration alone
• Cloud cost reduction: 75% to 90% versus an equivalent cloud architecture
• Bandwidth reduction: 90% to 99% less WAN traffic by processing locally and forwarding insights
• Quality and yield: 5% to 20% improvement
• Operator productivity: 10% to 30% improvement
• Unplanned downtime: 25% to 30% reduction (World Economic Forum)
• Safety incidents: 10% to 20% reduction
• Cost per frame: approximately $0.001 processed locally versus approximately $0.03 in the cloud

Value concentrates fastest in uptime. On a high-throughput line, a single avoided stoppage can pay for the deployment.

By architecture, not by policy. Every Cube is single-tenant and deployed on premise. Operational data and models stay inside the facility, processed locally, and nothing leaves unless you explicitly configure federation for a specific workflow such as central model tuning.

This is a design principle rather than a configurable option, and the risk math supports it: 80% to 95% of breaches involve data in transit or in external cloud systems. Data and operational know-how are core intellectual property. Keeping them on site keeps them yours.

Governance sits between the AI and any physical action, enforced at the infrastructure level through a system we call The Engine. AI may propose an action, but only The Engine admits it, checking the proposed action against formally verified invariants and recording a signed, immutable record of the decision. Execution is the mechanical consequence of an admitted decision.

There is no code path from a model output to an actuator that bypasses the admission gate. This is the structural answer to the December 2025 joint guidance from CISA, the NSA, and seven allied agencies that language models almost certainly should not make safety decisions for OT environments.

Security is a structural property of the platform, verified before runtime and immutable at runtime:

• Hardware-rooted trust – TPM 2.0 attestation, measured boot, and dm-verity, with hardware-bound identity and no password-based device authentication anywhere on the appliance. A Cube that fails attestation does not enter service.
• Immutable operating platform – secure boot, FIPS-ready cryptography, encryption at rest and in transit, zero-trust network access, and least-privilege access throughout.
• Immutable signed evidence – every admission decision is a content-addressed, signed artifact independent of any logging subsystem, so clearing logs cannot retroactively change what was admitted.
• Islanding by default – the appliance keeps governing itself through network partition, with bounded degradation and verified recovery.
• Brain-off-board safety for robotics – the robot keeps its real-time control loop while the Cube provides higher-order cognition, so degradation of the Cube never interrupts active control.

Where the architecture diverges from a standard, we say so. Third-party certification against frameworks such as IEC 61508 and IEC 62443 is independent, and a continuous adversarial audit surfaces and triages findings rather than claiming a clean sheet.

Regulation across critical infrastructure now converges on one technical demand profile: passive, local, real-time monitoring at distributed physical sites, with audit-grade evidence retention, in environments where cloud connectivity is unavailable, insufficient, or prohibited. The Cube satisfies that profile by construction. Frameworks map by sector:

Electric utilities and bulk electric system operators – NERC CIP-015-1 internal network security monitoring, with enforcement beginning October 2028, plus CIP-003-9 and CIP-012-2
Oil and gas – TSA Pipeline Security Directives and API 1164; EPA methane monitoring under OOOOb and OOOOc
Manufacturing – IEC 62443, plus CIRCIA reporting obligations
Defense industrial base – CMMC 2.0 and ITAR
Ports and maritime – USCG MTS cybersecurity rule
Mining – MSHA monitoring requirements
Life sciences – FDA 21 CFR Part 11 and GxP
EU operations – NIS2 and the EU Cyber Resilience Act

Compliance here is a mandate with deadlines and penalties, not a feature. The Cube is built to meet it as architectural ground truth rather than as a set of configuration options.

All firmware and software updates are signed and verified before they are applied. The immutable operating platform, secure boot, and measured boot prevent unauthorized changes, which is also the primary defense against supply chain attacks.

Updates are delivered continuously with zero downtime and no truck roll. Sites with intermittent connectivity sync when a link is available. Strict air-gapped sites use secure, staged update procedures. The underlying concern is real: obsolete systems carry latent risk, and major incidents traced to unpatched vulnerabilities routinely cost tens of millions.

Edgescale AI provides the logical security of the infrastructure as a managed service: operating platform patching, network security, intrusion detection, attestation, and lifecycle management.

The customer is responsible for the physical security of the appliance at their facility, though the hardware includes tamper-resistant features and hardware-bound identity. The customer also governs the AI models, the data flows, and the invariants that define which actions are admissible in their environment.