
Capacity Partnership
Turn idle GPUsinto revenue.
A managed AI-inference platform on your idle GPU nodes, built on HZ AI. You supply the hardware; we bring the buyer, the software and the billing.
- Programme
- Datacentre capacity partnership
- For
- Facility and GPU fleet operators
- Structure
- Revenue share on a contracted floor
- Platform
- HZ AI orchestration
- First tranche
- ≈1 MW of idle fleet, right-sized
- Jurisdiction
- United Arab Emirates
The problem
Your GPUs are darkmore often thanyou would like.
Idle capital and standing power costs, with no straightforward way to turn either into revenue.

Idle between committed workloads
Hardware depreciates in the gaps between reserved jobs, used or not.

Power you pay for but do not draw
Minimum-take and demand charges make unused megawatts a standing cost.

Self-monetising excess is hard
A serving stack, metering, billing and buyers — and your infrastructure exposed to reach them.
Read the detail
Idle between committed workloads. Capital-intensive hardware sits unproductive in the gaps between reserved jobs, depreciating whether it is used or not.
Power you pay for but do not draw. Interconnect and power allocations carry minimum-take and demand-charge pressure. Unused megawatts are a standing cost.
Self-monetising excess is hard. Doing it yourself means a serving stack, metering, billing and buyer relationships — and exposing your infrastructure to reach them.
Idle capital and standing power costs, with no straightforward way to turn either into revenue. That is the gap this programme is pointed at.
What the programme is
A managed platformthat fills the gaps.
Our lightweight agent sells your excess GPU capacity to committed buyers. We run the stack; you keep the hardware and reclaim it when you need it.
Your workloads keep priority. Only the gaps are monetised, and you can reclaim them on your terms.
Illustrative. Your committed share, reclaim terms and revenue split are set in the contract, not by this diagram.
How this works
We install a lightweight agent on your excess GPU nodes and sell their capacity to committed buyers. Qode Core runs the full stack — provisioning, privacy, metering, billing — and you keep control of the hardware, reclaiming it whenever you need it.
How the deal works
You supply.We operate.The buyer commits.
Two contracts: ours with you, ours with the buyer. The buyer transacts only with Qode Core.
The supply
You
- Run our agent on idle GPU nodes
- Earn revenue on hardware that would sit dark
- Provider details hidden from buyer-facing APIs
- Reclaim capacity when you need it back
The platform
Qode Core
- Provisioning, security, metering, billing
- The connective layer between supply and demand
- We carry the software and the buyer relationship
The demand
The buyer
- A committed anchor buyer of model endpoints
- Buys per token, with throughput and latency commitments
- Owns nothing and consumes the API
The two contracts
Two separate contracts result: Qode Core to you, covering supply and revenue share; Qode Core to the buyer, covering the demand channel and its service levels. The buyer transacts only with Qode Core.
Data and privacy boundary
The serving path staysinside your facility.
Prompts and responses are processed on your hardware and do not leave it. Only metering data — token counts and metadata — crosses to our control plane.
- Caddyingress
- Gatewayauth · meter
- tlsproxymTLS
- WorkloadvLLM · GPU
Outbound-only mTLS · token counts and metadata only · never prompt or response content
Outside your facility
- Backend
- Metering & billing
- Reconciler
- Ask endpoint
- No inbound except Caddy
- Content never leaves your facility
- Infrastructure hidden from buyer APIs

What lands in your facility
A light,outbound-only footprint.
Your nodes call out to our control plane over mTLS. The control plane never dials into your network.
- What we deploy
- Our agent plus a serving plane, managed and updated by us.
- Network posture
- Outbound-only mTLS and HTTPS. No inbound path except buyer traffic.
- Egress posture
- No private (RFC1918) or metadata (IMDS) egress; approved registries only.
- Host prerequisites
- Driver, container toolkit, Docker, firewall and NTP, pinned by our playbook.
Read the detail
What we deploy. The pilot deploys our agent plus a serving plane on your nodes. Both are managed and updated by us.
Network posture. Outbound-only mTLS and HTTPS to our control plane. There is no inbound path except the one buyer traffic needs.
Egress posture. Buyer workloads deny private (RFC1918) and cloud-metadata (IMDS) egress. Images pull only from approved registries.
Host prerequisites. NVIDIA driver and container toolkit, Docker, firewall and NTP — installed and pinned by our provisioning playbook before enrolment.
Agent security
What our agent can do,and what it cannot.
The agent is a constrained tenant on your hardware. Here is exactly what it is allowed to touch.
Runs under tight isolation
- Runs buyer workloads under gVisor, a syscall-level sandbox, not a bare container
- Runs under a dedicated service account we create, scoped to running and isolating those workloads
- Runs only images our control plane signs and digest-pins before dispatch
- Reaches our control plane over mTLS, outbound only
- Meters usage and drains on your reclaim notice
Blocked by default
- Reach your internal network — private (RFC1918) egress is denied
- Read cloud or instance metadata (IMDS)
- Pull images from outside approved registries
- Open inbound ports or bind publicly without your opt-in
- Bypass private-egress deny under the enterprise profile
The honest boundary
Your staff hold physical and root access, so isolation cannot shield buyer data from a hostile host administrator. The contract covers host-level access controls, audit logging and breach notification — it is not a promise that we can lock you out of your own machines. Confidential-compute GPUs are an optional stronger runtime where the hardware supports them.

You stay in control
You never lose accessto your own compute.
The programme runs only on nodes you mark as excess. Your committed workloads always come first.
01
Serve
Excess nodes serve buyer traffic and earn.
02
Notice
You give the reclaim notice you set.
03
Drain
In-flight work drains gracefully. Drop-free drain is the pilot target.
04
Reclaimed
Nodes return to you.
This does not reduce the compute you can use.
Your capacity comes first
The programme runs only on nodes you mark as excess. Your committed workloads always come first. We monetise the idle gaps, and you set the terms for taking capacity back.
You define what counts as excess. Nothing else is ever enrolled or touched.
Committed workloads keep first call on your GPUs. We fill only the idle gaps.
Need the headroom back? Reclaim enrolled nodes on the notice you set. Your own capacity is never blocked.
It turns nodes that would sit idle into revenue, and leaves your own utilisation untouched.
The economics
A floor you can count on.Upside when thedemand is there.
The buyer commits a minimum spend — your floor. Above it, you share every token served, with no idle risk on your side.
Illustrative. The floor level and the revenue split are set in your contract. Upside moves with real buyer demand, not with a projection.
Contracted floor — identical in all three
A worked example
What a month ofidle GPUs is worth.
Earnings per megawatt of facility capacity per month, by GPU class and by how much of the month sits idle and gets sold.
| GPU class | 10%sold | 25%sold | 50%sold | 75%sold |
|---|---|---|---|---|
| H100$1.85 / GPU-hr ≈588 GPUs / MW | $40K | $99K | $199K | $298K |
| H200$2.50 / GPU-hr ≈556 GPUs / MW | $51K | $127K | $253K | $380K |
| B200$4.25 / GPU-hr ≈417 GPUs / MW | $65K | $162K | $323K | $485K |
| B300$5.75 / GPU-hr ≈400 GPUs / MW | $84K | $210K | $420K | $630K |
- 10%
- $40K
- 25%
- $99K
- 50%
- $199K
- 75%
- $298K
- 10%
- $51K
- 25%
- $127K
- 50%
- $253K
- 75%
- $380K
- 10%
- $65K
- 25%
- $162K
- 50%
- $323K
- 75%
- $485K
- 10%
- $84K
- 25%
- $210K
- 50%
- $420K
- 75%
- $630K
Formula — $ per MW per month = GPUs per MW (all-in facility kW) × 730 hr × idle share sold × rate × Qode Core fee.
Assumption — Illustrative, and it assumes matched buyer demand. Rate and floor anchors are July 2026. These are not a forecast of your revenue; they are the arithmetic, shown so you can run it against your own fleet.
Appendix — “keep us at X%” means three different things
When a datacentre says “keep us busy”, it can mean one of three very different things — and only one of them is something we contract to.
Commercial
Committed revenue against the capacity you hand over. Predictable money, whatever demand does.
Contracted — take-or-pay
Technical
GPU busy time and share of throughput in use. An operations metric that moves with model mix and traffic.
Reported, not an SLA
Power
Facility draw against contracted megawatts. You already pay for it, and only real demand draws watts.
Your contract, your risk
We commit to commercial utilisation, through the revenue floor. Technical and power utilisation are dashboards, not service levels.
Getting started
A low-risk pilot.
From first call to serving traffic on a roughly 1 MW tranche of your idle fleet — right-sized to your comfort, with a clean exit at every step.
01
Lock the facts
A go/no-go gate up front. We confirm the scoping details together before anything is installed.
02
Provision
The agreed tranche, via our playbook: driver, toolkit, Docker and firewall, then enrol.
03
Stand up serving
Controller and ingress deployed inside your network, ready to serve buyer traffic.
Exit criteria
Measured service levels, a passed reclaim drill, and a buyer volume commitment.
What we need to scope it
Inventory. GPU class and count, RAM, NIC, OS, driver versions, power and cooling headroom.
Network. Topology, public egress bandwidth, available public IPs for ingress.
Terms. Your definition of excess, reclaim notice, operator model, governance posture.
Already a technology partner? The partner dossier

Let's scope a pilot
Turn your idle fleetinto revenue.
A roughly 1 MW pilot, right-sized to your comfort, then we scale on your terms.
