AI teams
Plan dedicated infrastructure around recurring workloads, with the configuration and operating scope agreed for your needs.
Managed clusters / Owned by you
Your GPU cluster.
Our operating desk.
Own the infrastructure your workloads depend on. Compute Function brings hardware, site selection and deployment together, then operates and manages the cluster.
Start with your workload, a bill of materials, or hardware you already own.
One cluster. Clear roles.Operating model
- Ownership
- You
- Deployment
- Compute Function
- Operation & management
- Compute Function
Who it is for
Teams that want ownership with an operator behind it.
Enterprises & research
Build a cluster around your access, location and workload requirements, with a defined management relationship.
Infrastructure owners
Bring a new build or existing equipment to a desk that can coordinate deployment and take on cluster operation.
Venture funds & family offices
A compute operating partner for the capital behind it.
For VCs: support the portfolio’s compute needs.
Bring a portfolio company’s requirement to one desk for capacity sourcing, owned-cluster planning and operation. Compare its deployment options and operating budget before deciding how much capital to commit.
- Scope capacity around each company’s workload, runway and readiness.
- Compare renting existing capacity with owning a managed deployment.
- Keep hardware, site, management and any spot use visible in the plan.
For portfolio companies using the cluster internally, assess compute cost and avoided spend separately from third-party rental income.
For family offices: own infrastructure with a named operator.
Evaluate direct ownership of a defined GPU deployment, with Compute Function operating the cluster. Underwrite the equipment, site, customer demand, recurring costs and exit assumptions as one asset-level plan.
- Identify the equipment, ownership structure and operating responsibilities.
- Distinguish contracted demand from utilization that still needs to be won.
- Stress-test rental rates, power requirements, costs and residual value.
The deployment guide
Know what happens before the next decision.
Each stage has something you bring, work we take on, and an output you can review. Open a stage to see what it involves.
Hardware lead times, site readiness and your approval process determine the schedule. Milestones are built around your deployment.
Start with a deployment brief ↗01
Discovery
Start with the workload, not a purchase order.
You do not need a finished bill of materials. The first job is to turn what you want to run into a deployment brief.
- You bring
- Workload, expected usage, budget range, preferred region and target date. Share existing hardware or a bill of materials if you have one.
- We take on
- Scope GPUs, memory, networking, storage and access requirements. Identify the constraints that affect ownership, location and operating cost.
- You receive
- A deployment brief: proposed configuration, site requirements, cost categories, target milestones and the questions still to resolve.
Your decisionConfirm the brief reflects the workload and budget you want us to plan around.
02
Sourcing & diligence
Compare hardware and sites together.
A server quote is useful only if the site can power, cool and connect that configuration on the required schedule.
- You bring
- Your priorities: location, lead time, fabric, density, access and which requirements are flexible.
- We take on
- Match the bill of materials to vendors. Review site power, cooling, network and readiness. Put quotes in the same fields and flag dependencies or missing items.
- You receive
- A shortlist with the equipment scope, site options, quoted lead times and outstanding checks. Ready infrastructure is kept separate from planned capacity.
Your decisionChoose the configuration and site combination you want to take forward.
03
Commercial plan
Approve the budget and the operating responsibilities.
You should be able to see what is being bought, what recurs each month, and who is responsible for every part of the deployment.
- You bring
- Your selected option, purchasing entity, approval process and any access or service requirements the operation must meet.
- We take on
- Separate equipment, site, deployment and management costs. Align the delivery milestones, ownership structure, acceptance tests and operating scope.
- You receive
- An itemized commercial plan, deployment schedule and management scope, with the billing parties and approval points identified.
Your decisionApprove the scope and sign the applicable agreements before orders and deployment commitments are made.
04
Deployment & acceptance
Bring the cluster up and prove it is ready.
Delivery is a checkpoint. A usable cluster also needs the right configuration, connectivity and acceptance results.
- You bring
- The people who will receive access, your access policies and a representative workload or acceptance requirements.
- We take on
- Coordinate delivery, installation and bring-up. Check GPU visibility and node health, fabric connectivity, storage access and the workload checks agreed in the plan.
- You receive
- An inventory and configuration record, acceptance results, access handover and a list of any issues still open.
Your decisionReview acceptance against the agreed criteria before the deployment moves into the ongoing operating phase.
05
Ongoing management
Keep the infrastructure in an operating rhythm.
Compute Function takes the cluster operating role. Your team keeps control of the workloads and access decisions that belong to it.
- You bring
- Workload priorities, authorized contacts, maintenance windows and approval rules for changes that affect your environment.
- We take on
- Run the agreed monitoring and provisioning scope, triage infrastructure faults, coordinate vendor or on-site remediation, and manage planned maintenance and approved configuration changes.
- You receive
- Visibility into cluster health, incidents, maintenance and unresolved actions. Reporting cadence and support coverage are set when management is scoped.
Your decisionApprove changes outside the agreed operating scope and set priorities when a maintenance action affects workloads.
06
Capacity review
Plan the next move around actual usage.
The operating plan should evolve as the workload grows, the hardware changes or part of the cluster becomes available.
- You bring
- Updated demand, upcoming workloads, expansion plans and any blocks or hours you can release.
- We take on
- Review the configuration and capacity plan. Scope additions or changes, and assess suitable unused windows for Liquid Compute spot placements.
- You receive
- A next-step plan for expansion, reconfiguration or optional spot availability, including constraints and reclaim terms.
Your decisionApprove any new spend or release of capacity. Spot placements depend on demand and technical fit.
After go-live
What operating the cluster actually involves.
Compute Function takes the infrastructure operating role. Together we define the coverage, escalation contacts, service levels and change approvals for the environment.
- Provisioning & access
- Configure the infrastructure and access boundaries in the operating scope. Your team authorizes users and owns its application and data access policies.
- Health & monitoring
- Define the GPU, node, storage and network signals to watch, the alert thresholds and the route for escalating an issue.
- Faults & remediation
- Triage infrastructure issues, record the affected equipment and coordinate repairs, replacement parts or warranty actions with the relevant providers.
- Maintenance & changes
- Plan firmware, software and configuration work around agreed maintenance windows, with the approvals and rollback approach set before changes.
- Operating visibility
- Keep the inventory, configuration, incident history and outstanding actions available for review. Agree reporting frequency and points of contact.
- Capacity planning
- Review usage and upcoming needs. Plan additions or configuration changes, and identify any capacity that can be offered to spot demand.
Your team controls
The workload roadmap, application code and data, user authorization, budget and approvals for changes outside the operating scope.
We agree before handover
Support hours, incident escalation, response commitments, maintenance windows, reporting cadence and responsibility for replacement costs.
Understand the quote
See the costs before you commit.
A deployment has an equipment budget and an operating budget. The proposal separates the categories, identifies who bills each item, and shows which costs recur.
- Equipment
- GPU servers, networking, optics, cabling and storage.Purchase cost; warranties and replacements identified separately.
- Deployment
- Shipping, installation, integration, configuration and acceptance work.Project costs and milestone payments shown in the deployment quote.
- The site
- Rack space, power, cooling, connectivity and any site services.Recurring commitments, usage charges and site-specific setup costs.
- Management
- The operating scope, monitoring, maintenance coordination and support.Management fees, support coverage and out-of-scope work identified in the proposal.
Compute Function fees / Indicative USD
Budget for the operator as well as the hardware.
Planning ranges for discussion. Final fees are quoted against configuration, site, cluster size and operating scope. Hardware, site, freight and third-party charges sit alongside these fees.
- Deployment & bring-up
- $25k–50k / small deploymentOne-time planning assumption. Scope sourcing coordination, installation, configuration and acceptance work before an order.
- Ongoing management
- $500–1,500 / server / monthIndicative range for an eight-GPU node. Monitoring, maintenance coordination, support coverage and operating responsibilities determine the quote.
- Desk-sourced capacity placement
- 2.5% of collected rental revenueIllustrative placement-fee assumption for demand sourced through the desk. Confirm the applicable rate and billing basis before a placement.
The model below fixes fees at $25k deployment, $1k per server per month for management, and a 2.5% placement fee. These are indicative Compute Function budgeting assumptions, not a binding fee offer.
Illustrative ownership economics / USD
Underwrite the cash flows, not just the silicon.
Explore a fully cash-funded B300 deployment rented to external buyers. Paid utilization is fixed at 100%, with no revenue delay. Adjust colo and electricity prices to match your site. Compute Function fees stay fixed in this model. The IRR is an annualized, unlevered, pre-tax asset return after modeled fees and expenses.
Fixed fees & operating assumptions
- CF deployment
- $25,000
- CF management / server / month
- $1,000
- CF placement / rental revenue
- 2.5%
- Air-cooled PUE
- 1.3
- Networking / server
- $60,000
- Freight / server
- $2,500
- Annual repair reserve / hardware cost
- 1%
- Selling costs / exit value
- 5%
Colo is charged on IT capacity. Electricity is separate: IT kW × PUE × 730 hours/month × your electricity price.
Initial cash outlay
$2,555,000
Annualized asset IRR
20.9%
Monthly cash after expenses¹
$80,337
Cash returned / cash invested
1.39×
Illustrative, not a target or promised return. Negative results are shown. “No IRR” means the cash flows do not produce a conventional IRR. ¹From the first operating month; excludes terminal sale proceeds.
- Monthly gross rental receipts
- $105,120
- CF placement fee / revenue month
- $2,628
- CF management / month
- $4,000
- Colo capacity / month
- $11,700
- Electricity including PUE / month
- $4,555
- Total site cost / month
- $16,255
- Repair reserve / month
- $1,900
- Net hardware proceeds at exit
- $649,800
- Total CF fees across hold
- $263,608
- Operating break-even / GPU-hour²
- $0.97
²At 100% paid utilization; covers ongoing costs and placement fees, not recovery of the initial investment.
IRR sensitivity
Compare rental rates at 100% paid utilization. All other inputs stay as selected.
| $/GPU-hr | Asset IRR | Monthly cash¹ |
|---|---|---|
| $3.50 | 3.4% | $57,561 |
| $4.50 | 20.9% | $80,337 |
| $5.50 | 39.5% | $103,113 |
Cash-flow timing and calculation
All equipment, infrastructure, freight and deployment fees are paid at month 0. Rental receipts, placement fees, colo, electricity, management and repair reserves all start in month 1. There is no revenue delay. The final month adds the hardware exit value, less selling costs.
- Month 0
- −$2,555,000
- Year 1
- $964,042
- Year 2
- $964,042
- Year 3 (includes exit)
- $1,613,842
IRR solves the monthly cash flows and compounds the monthly rate over 12 months. The cash multiple is all positive net cash flows divided by all cash outflows.
Scope: 8 GPUs per server, 8,760 hours/year, constant rate and 100% paid utilization from the first operating month. Placement fees apply to all modeled rental receipts, assuming all demand is desk-sourced. Exit value applies to hardware only. Excludes financing, tax, depreciation benefits, fund-level fees, additional shared storage or software licenses, additional downtime, revenue-rate decay and unbudgeted expansion. Internal-use savings are not rental cash receipts.
Hardware reference: Managed GPUs publishes B300 server estimates of $550k–650k, networking around $50k–70k/server for an illustrative server. The B300 hardware preset is $570,000 per server. Networking uses a benchmark midpoint, with a separate freight allowance. Colo defaults to $195/IT kW/month plus electricity at $0.08/kWh, both editable, with fixed PUE 1.3; these are planning assumptions, not a vendor quote. NVIDIA’s B300 system specifications describe an eight-GPU system; this model is not a DGX price quote. The fixed 15 kW/server power budget follows the full-load maximum in Supermicro’s air-cooled B300 datasheet; actual draw depends on configuration and workload. Shared networking and storage power beyond this server budget are excluded. References checked October 17, 2026.
When your cluster has room
Give unused capacity a route to spot demand.
Liquid Compute brings available blocks or hours to inference, reseller and burst workloads through the desk. Define when the capacity is available and the notice you need to reclaim it.
Explore Liquid Compute ↗Your workload comes first
Spot placements need to fit your available windows, isolation requirements and operating commitments.
Placements depend on demand and technical fit. Utilization and revenue are not guaranteed.
Before you build
Who owns the equipment?
You own the cluster. The equipment, purchasing structure, site arrangements and responsibilities are documented in the agreements for your deployment.
Who runs it?
Compute Function operates and manages the cluster. The management agreement defines the scope of provisioning, monitoring, maintenance, support, access and service levels.
Can we bring hardware we already own?
Yes—bring the configuration, location and current operating arrangements. The desk reviews technical fit and the management scope before proposing a transition.
How long does deployment take?
Hardware availability, site readiness, power, cooling and integration determine the schedule. The desk confirms milestones for your specific deployment.
What will it cost?
Hardware, networking and storage, colocation, deployment work and ongoing management are scoped during quoting. You review the cost breakdown and commercial terms before committing.
Can unused capacity go to spot buyers?
Liquid Compute gives available capacity a route to inference, reseller and burst demand through the desk. Available windows, technical fit and reclaim terms are agreed per placement. Demand and revenue are not guaranteed.
Start with the plan
Tell us what you want to build.
Bring your workload, budget and target region. If you already have equipment or a bill of materials, include the details you know.
- First, we clarify the brief. We follow up on the workload, budget, region and any requirements that affect the configuration.
- Then, we identify the next step. That might be configuration work, a site review, vendor quotes or a review of hardware you already own.
- You review before committing. The scope, cost categories and deployment milestones come together before an order or management commitment.
Need rented capacity instead? Request compute.