
Infrastructure
Eight strata,from the powerfeed to the model.
Qode Core operates every layer in between — because a guarantee at the top of a stack only holds if the same operator holds the bottom.
The stack
Eight strata.One operator.
Read top to bottom and it is a buying order. Read bottom to top and it is a build order.
- Scope
- Compute · Cloud · Data · Platform · Operations · Security
- Deployment
- Colocation · Dedicated · Managed
- Jurisdiction
- United Arab Emirates
Strata 07 and 08 — partners and applied research — shape what reaches the other six. They are not something you provision, so they are not specified here.
- 08InnovationContext
- 07PartnersContext
- 06Security
- 05Intelligent Operations
- 04AI Platforms
- 03Data Foundation
- 02AI Cloud
- 01GPU Compute
Strata 01 – 03
Compute, cloud,data.
The floor everything stands on, what makes it consumable, and where the training set lives.

01 · GPU Compute
The floor everything else stands on.
Accelerated compute for training, fine-tuning and inference — customer-owned or Qode Core-provisioned.
- Workloads
- Training · fine-tuning · inference
- Ownership
- Customer-owned or Qode Core-provisioned
- Accelerator class
- Sized per engagement
Full specification
Accelerated compute for training, fine-tuning and inference, alongside general-purpose capacity for the data and platform layers that surround it. This is the only layer in the stack where the physical object and the product are the same thing — which is why it is the layer customers most often want to own outright, and why colocation exists.
Consumes: Power envelope, cooling capacity and structured cabling from the facility.
Provides: Schedulable accelerated and general-purpose capacity to the cloud layer.
- Workloads
- Training · fine-tuning · inference
- Ownership
- Customer-owned or Qode Core-provisioned
- Accelerator class
- Sized per engagement
- Node baseline
- NVIDIA driver & container toolkit, pinned
- Workload runtime
- gVisor-sandboxed, digest-pinned images
- Confidential compute
- Where the GPU supports it

02 · AI Cloud
What makes dedicated hardware consumable.
Virtualisation, orchestration, networking and storage over the compute floor, with the control plane kept in-jurisdiction.
- Tenancy
- Private · dedicated · hybrid
- Control plane
- In-jurisdiction
- Ingress
- Terminated in-facility, authenticated
Full specification
Virtualisation, orchestration, networking and storage services layered over the compute floor. Without this layer a rack is an asset on a balance sheet; with it, the same rack is capacity a team can provision against. The control plane — the thing that decides what runs where — stays inside the jurisdiction, because a control plane operated from elsewhere is an administrative path out of it.
Consumes: Accelerated and general-purpose capacity from the compute layer.
Provides: Provisioned environments, networks and volumes to the data and platform layers.
- Tenancy
- Private · dedicated · hybrid
- Control plane
- In-jurisdiction
- Ingress
- Terminated in-facility, authenticated
- Egress policy
- Private (RFC1918) & IMDS denied by default
- Image supply
- Signed, from approved registries only
- Footprint
- Sized per engagement

03 · Data Foundation
Where the training set actually lives.
Storage, pipelines, lineage and governance. Residency is a property of how storage is provisioned, not a policy applied afterwards.
- Residency
- In-country by default
- Governance
- Lineage · access · retention
- Cross-border replication
- Not configured
Full specification
Storage, pipelines, lineage and governance for the data that trains and serves models. Most sovereignty conversations are really conversations about this layer: a model is a derivative of its training data, and the question of which legal regime governs that data is settled here rather than at the application tier. Residency is a property of how storage is provisioned, not a policy applied afterwards.
Consumes: Volumes, networks and object storage from the cloud layer.
Provides: Governed datasets, lineage records and feature access to the platform layer.
- Residency
- In-country by default
- Governance
- Lineage · access · retention
- Cross-border replication
- Not configured
- Lineage
- Read, write and movement, with authority
- Retention
- Set by your policy, enforced by ours
- Capacity
- Sized per engagement
Strata 04 – 06
Platform, operations,security.
The runtime your models sit on, the service that keeps it running, and the controls that make the guarantee enforceable.

04 · AI Platforms
The runtime your models sit on.
Managed environments to train, fine-tune, serve and govern models — the layer Intellake, Oryn, Kyro and Insight 360 are built on.
- Qode Core products
- Intellake · Oryn · Kyro · Insight 360
- Deployment
- In-tenant, inside the perimeter
- Serving runtime
- vLLM, behind a metering gateway
Full specification
Managed environments for training, fine-tuning, serving and governing models — and the layer Qode Core's own products are built on. Intellake, Oryn, Kyro and Insight 360 are not add-ons bolted to someone else's platform; they are what this stratum looks like when it is used in anger, which is the only reason we can speak about it with any authority.
Consumes: Governed datasets from the data layer and schedulable capacity from compute.
Provides: Model endpoints, training environments and governance records to applications.
- Qode Core products
- Intellake · Oryn · Kyro · Insight 360
- Deployment
- In-tenant, inside the perimeter
- Serving runtime
- vLLM, behind a metering gateway
- Governance tier
- Tynapse trust layer — verify before execute
- Model catalogue
- Agreed per engagement
- Open-weight models
- Run in-tenant, weights never leave

05 · Intelligent Operations
Infrastructure is a service you run, not a product you ship.
Continuous telemetry across compute, storage, network, power and cooling, with programmatic response to known failure modes.
- Operating model
- Monitor · Optimise · Automate · Scale
- Framework
- People · Process · Technology · Resilience
- Telemetry
- Compute, storage, network, power, cooling
Full specification
Continuous telemetry across compute, storage, network, power and cooling, with programmatic response to known failure modes before a human is paged — and a defined escalation path for the modes that are not known yet. This stratum is the one a customer experiences most and specifies least, and it is the reason the same hardware performs differently under two different operators.
Consumes: Telemetry from every stratum beneath it.
Provides: Availability, capacity headroom and an auditable record of every change.
- Operating model
- Monitor · Optimise · Automate · Scale
- Framework
- People · Process · Technology · Resilience
- Telemetry
- Compute, storage, network, power, cooling
- Automated response
- Known failure modes, before a page
- Coverage & escalation
- Agreed per engagement
- Availability commitment
- Contracted, not advertised

06 · Security
What makes the guarantee enforceable rather than contractual.
Physical, network and tenancy isolation, identity and access, and audit lineage. Air-gapped deployment is a supported model.
- Isolation
- Physical · network · tenancy
- Air-gapped deployment
- Supported model
- Tenancy isolation
- Syscall-level sandbox, not a bare container
Full specification
Physical, network and tenancy isolation; identity and access management; audit lineage. Every sovereignty claim this company makes reduces to a control in this stratum — and a control that cannot be demonstrated to an auditor is a sentence in a contract, which is a different and much weaker thing. Air-gapped deployment lives here too: not a hardened path, but the absence of one.
Consumes: Facility access control and network boundaries from the layers beneath.
Provides: An enforceable perimeter and an audit record to every stratum above.
- Isolation
- Physical · network · tenancy
- Air-gapped deployment
- Supported model
- Tenancy isolation
- Syscall-level sandbox, not a bare container
- Service accounts
- Dedicated, scoped to the workload
- Control-plane reach
- Outbound-only mTLS. Never inbound.
- Key custody
- Held in jurisdiction

Capacity
No GPU counton this page.
Accelerator count, mix, coverage hours and availability are sized to a workload and written into a contract. The site's power and supported platforms are on the facility brief.
Why there is no figure here
The site's power and the GPU platforms it supports are published. What a deployment actually gets — accelerator count, mix, coverage hours and availability — is sized to a workload and written into a contract. A figure published here would belong to someone else's deployment, and quoting it at you would be the first inaccurate thing we said. What is specified on this page is structural: it is true of every engagement.
Deployment models
Three ways in.One line that moves.
The infrastructure does not change between models. The boundary between your responsibility and ours does — so here it is, drawn.

AI Colocation
Your hardware. Our facility.

Dedicated
Our hardware. Reserved for you.

Managed
Your workload. We run the rest.
| Responsibility | AI ColocationYour hardware. Our facility. | DedicatedOur hardware. Reserved for you. | ManagedYour workload. We run the rest. |
|---|---|---|---|
| Hardware specification | Customer | Shared | Qode Core |
| Hardware procurement | Customer | Qode Core | Qode Core |
| Facility, power & cooling | Qode Core | Qode Core | Qode Core |
| Physical access control | Qode Core | Qode Core | Qode Core |
| Network & connectivity | Shared | Qode Core | Qode Core |
| Hypervisor & OS patching | Customer | Shared | Qode Core |
| Monitoring & alerting | Customer | Shared | Qode Core |
| Incident response | Customer | Shared | Qode Core |
| Capacity planning | Customer | Shared | Qode Core |
| Workload, models & code | Customer | Customer | Customer |
| Data governance policy | Customer | Customer | Customer |
AI Colocation
Your hardware. Our facility.
- Qode Core holds
- Facility, power & cooling · Physical access control
- Shared
- Network & connectivity
- You hold
- Hardware specification · Hardware procurement · Hypervisor & OS patching · Monitoring & alerting · Incident response · Capacity planning · Workload, models & code · Data governance policy
Dedicated
Our hardware. Reserved for you.
- Qode Core holds
- Hardware procurement · Facility, power & cooling · Physical access control · Network & connectivity
- Shared
- Hardware specification · Hypervisor & OS patching · Monitoring & alerting · Incident response · Capacity planning
- You hold
- Workload, models & code · Data governance policy
Managed
Your workload. We run the rest.
- Qode Core holds
- Hardware specification · Hardware procurement · Facility, power & cooling · Physical access control · Network & connectivity · Hypervisor & OS patching · Monitoring & alerting · Incident response · Capacity planning
- You hold
- Workload, models & code · Data governance policy
The last two rows never move. In every model, the workload and the policy governing its data remain yours.
Contractual detail
Note the last two rows. They do not move. Across every deployment model the workload and the policy governing its data remain yours — what you are buying is the stack underneath them, not a share of them.
Contractual detail, commercial terms and the precise split within each shared responsibility are agreed per engagement — which is a fact about how infrastructure is bought, not a gap in this page.
The facility
The building,and how to check it.
Two 1,200 sqm buildings, a substation with 5 MW available today, a Tier III design at launch and a Tier IV target by month six.
- Buildings
- 2 × 1,200 sqm
- Power
- 5 MW available today
- Expansion
- 10–15 MW, subject to utility approval
- Resilience
- Tier III design · Tier IV target, month 6
- Cooling
- AC + chillers in place · DLC at fit-out
RenderSite verification
Eight things toconfirm on the floor.
A page cannot prove a figure. Walk the floor and check these — for us, or for anyone else you are considering.
Power envelope per rack
Against the accelerator density you actually intend to deploy, not the hall average.
Cooling design
Air, rear-door or direct-to-chip — and the delta between design load and current draw.
Floor loading
The constraint that decides whether a high-density row is possible at all.
Tier and its certificate
Ask to see the certificate, not the claim. They are different documents.
Carrier diversity
How many carriers, and whether they enter the building through the same duct.
Physical access control
Who can reach the cage, how that is recorded, and how long the record is kept.
Operator reach
Which staff hold administrative access, under which jurisdiction they are employed.
Exit
What leaving looks like: notice, data return, media destruction, and who certifies it.
We will answer all eight on a site visit, in person, against the specific deployment you are scoping — including the certificates, as each one is issued.

Specify a deployment
Tell us the workload.We will tell you the stack.
Model size, data classification, throughput target, regulator. Those four answers determine which strata you provision and which deployment model carries them.
