# Reserved Compute

> Request reserved GPU capacity supplied by Nodus. Plan your configuration and term, review market estimates, and get a proposal from Nodus.

Source: https://www.nodus-compute.ai/docs/guides/marketplace/
Build revision: 53bfc045e800bf7bb44fa18623013c7aa9c9971e

**Beta:** this feature may change.

Reserved Compute is **Beta**: dedicated GPU capacity supplied by Nodus for your configuration and reservation window. Nodus manually reviews each request and prepares a proposal covering availability, price, access, and delivery. Your agreement is with Nodus.

Start with the [reserved provider index](https://www.nodus-compute.ai/reserved-compute/#estimated-pricing). Choose RTX, H-series or B-series, then a model, node count and contract length. One node is eight GPUs. The index compares providers’ published reservation rates for the **exact selected term**, in USD per GPU-hour. A year and 12 months match the same commitments; a three-year price never enters a one-year estimate.

View the average, range and provider count. Compare contract lengths using their matching rates and provider counts. The 30-day budget is a planning figure at your selected contract rate, not a payment schedule or one-month offer. The full-term estimate and range cover the entire reservation. Every matching provider has equal weight. Regional prices are averaged within each provider first. At least two providers are required; otherwise the page shows insufficient data and keeps quote requests available. Payment schedules, locations, included resources and lead times differ; some published capacity is built to order or may be out of stock. Prices are checked every five minutes while the page is open. Expired or unavailable observations are removed. Only public provider reservation prices enter the index; no internal Nodus prices or assumed discounts are used.

Full-term estimates use matching reserved rates and your requested GPU-hours. They do not confirm stock, cluster pricing or a Nodus offer. Nodus confirms the actual contract against your requirements. The public index can model 1–260 weeks, 1–60 months or 1–5 years; the console request form accepts durations up to 365 days. Planning uses seven-day weeks, 30-day remaining months and 365-day full years. Unsupported terms or models show **Request a quote** instead of an estimated price.

**Request this capacity** opens an email draft to <admin@nodus-compute.ai> with your configuration, sources and estimate. Add your company, preferred location, start date, application and interconnect requirements, then send it to request an offer. Nothing is submitted until you send the email. **Copy request brief** lets you use another email client. For a larger deployment, [book a conversation](https://www.nodus-compute.ai/book/).

For a tracked request, sign in at [Request reserved capacity](https://console.nodus-compute.ai/reserved-compute/request). Choose a GPU model, **total GPU quantity**, duration, location and start timing. You can request a single GPU; the final offer confirms how GPUs are arranged across physical servers. Choose dedicated servers with host SSH access or a connected cluster. These choices request a configuration, not confirmed stock.

Use **Other / help me choose** when you need guidance. Put your compute needs, budget, storage, networking, software or other requirements into the single optional **Anything else?** box. Pricing is requested by default; you do not need to estimate a GPU-hour budget. Company and email are prefilled when available. Check your contact name and follow-up email, then select **Review request** and **Submit request**. **Back** keeps your entries. Durations use days, weeks or 30-day months, up to 365 days.

**As soon as available** and **Flexible / planning ahead** are start preferences. **Specific date** asks for your preferred start date. Nodus confirms delivery dates, configuration and commercial terms before an agreement. Submission does not reserve inventory or authorize payment.

After submitting, **Request received** shows your reference and contact email. The request enters Nodus’s admin inbox and an internal email alert is queued. Nodus follows up using your contact email. A call is optional: **Book a call** opens the booking page; submission never books an appointment automatically. Find your saved brief and progress under **Your requests**. Confirm your contact email there to allow supplier sourcing to start. Nodus may use Google research and automated first enquiries to find suitable capacity. Only your GPU configuration, quantity, location and timing are shared; your identity, budget and private notes remain with Nodus.

The console sidebar shows **Reserved compute** only when the selected project has a current or upcoming server assignment. Offline and suspended servers remain visible. A sourcing request alone does not add this section. The [request form](https://console.nodus-compute.ai/reserved-compute/request) remains accessible before capacity is assigned. After fulfillment, open **Reserved compute → server → Connect** for SSH access.

|Status|What it means|
|-|-|
|Received|Your requirements are saved for Nodus to review.|
|Sourcing|Nodus is checking suitable capacity.|
|Offer sent|Nodus has sent an offer through the agreed contact channel.|
|Closed · Won / Lost|Nodus has recorded the outcome of the conversation.|

These statuses track the conversation; they do not reserve hardware or confirm payment or delivery. Nodus discusses pricing, final agreements and delivery with you directly. Your contact details and private notes remain with your organization and Nodus operators.

You can edit or withdraw your request. Editing returns progress to **Received** until Nodus reviews the new requirements. When updating an older request, confirm the GPU model, node count, memory per GPU and GPU-hour budget; unknown configuration units are never automatically treated as GPU nodes. For on-demand compute, use **Compute console** in the header.

## Plan a reserved-compute request

Nodus supplies dedicated GPU capacity under an agreed reservation. Your request describes the configuration and window your project needs. To run code with managed logs, outputs, spending limits, and recovery instead, use [Nodus Compute Jobs](https://www.nodus-compute.ai/docs/guides/jobs/).

|Requirement|What to specify|
|-|-|
|Hardware|Accelerator type or constraints, device count, and required memory|
|Application|Training, fine-tuning, inference serving, or another compute task|
|Networking and location|Required interconnect, network access, and location constraints|
|Reservation window|Start date and duration, including any scheduling flexibility|
|Budget|An optional target price or commercial requirements; you can ask Nodus to quote|

### Is reserved capacity available immediately?

Submitting a request does not allocate hardware. Nodus reviews the requirements and follows up when suitable capacity is available. Review the configuration, quantity, price, access, and delivery terms in the proposal. Nodus confirms the final agreement, funding and delivery directly with you.

### Should I reserve capacity or run a job?

Choose [Reserved Compute](https://www.nodus-compute.ai/reserved-compute/) when you need capacity for a planned window. Choose [Jobs](https://www.nodus-compute.ai/docs/guides/jobs/) when you want Nodus to run a command and return logs and outputs. If you need to call a hosted model instead of arranging hardware, start with [Inference (Beta)](https://www.nodus-compute.ai/docs/guides/inference/).

For model sizing, use the [70B VRAM guide](https://www.nodus-compute.ai/blog/vram-needed-for-70b-model/) to distinguish inference, fine-tuning, and full training memory requirements before choosing a GPU quantity.

## Worked planning example: eight accelerators for two weeks

This is a hypothetical requirements brief, not an available offer or a customer result. Suppose a planned run needs eight accelerators for a continuous 14-day training window:

|Detail|Example requirement|
|-|-|
|Device memory|At least 80 GiB per accelerator|
|Quantity|Eight accelerators available together|
|Window|14 consecutive days, with exact UTC start and end times|
|Network|State the required interconnect and whether the job spans hosts|
|Storage|Specify dataset size, checkpoint storage, and persistence after the window|
|Acceptance|Confirm the configuration, access method, delivery time, and agreed validation before the run|

Eight devices × 24 hours × 14 days is **2,688 device-hours** of planned capacity. That arithmetic is a sizing aid, not an invoice: the proposal determines the billable unit, reservation term, price, and other charges. Eight devices with 80 GiB each do not automatically behave like one device with 640 GiB of memory; your training framework and interconnect must support the required distribution.

Submit the brief through [Request capacity](https://console.nodus-compute.ai/reserved-compute/request). If dates or hardware are flexible, include acceptable alternatives. Delivery follows the agreed contract; submitting a request alone does not guarantee availability or begin billing.

## API resources

A **MarketplaceEntry** records your request, including its configuration, dates, budget and review status. Its optional `spec.sourcing` holds the private GPU sourcing brief and accepts only `unit: GPU`. Older API requests without this brief remain supported. These details guide manual sourcing, including storage, topology, software and security requirements. The API stores total GPU quantity in `quantity` with `sourcing.unit: GPU`. Optional `gpusPerServer` records a confirmed layout; new console requests leave it unspecified. `rateUSDPerUnitHour: ""` means pricing is requested, and `memoryGiB: 0` means unspecified memory for demand. Supply still requires positive memory and a positive rate. Existing detailed briefs and explicit budgets remain readable.

`sourcing.startTiming` records `ASAP`, `SpecificDate` or `Flexible`; `sourcing.desiredHours` records the preferred duration. ASAP and flexible requests require confirmed specific dates and a reviewed positive budget before a new legacy contract proposal can be created. Their stored interval is not a guaranteed delivery window. Existing proposals keep their frozen dates and independent confirmation checks. After signing in with `nodus login`, update the dates and requirements in this example and submit it with `nodus apply -f request.yaml`:

request.yaml

```yaml
# Non-binding request; Nodus confirms server layout, pricing and access before fulfillment.
apiVersion: nodus.dev/v1
kind: MarketplaceEntry
metadata:
  name: reserved-training
spec:
  intent: Demand
  title: 1 RTX PRO 6000 for model development
  company: Example Lab
  contactEmail: capacity@example.com
  hardware: RTX PRO 6000 Blackwell
  quantity: 1
  memoryGiB: 96
  region: United States
  network: No preference
  startsAt: "2027-01-17T00:00:00Z"
  endsAt: "2027-02-16T00:00:00Z"
  minimumHours: 720
  rateUSDPerUnitHour: "" # Request a quote; this cannot authorize a contract.
  notes: Model development with Ubuntu, host SSH access and local NVMe storage.
  sourcing:
    topology: Servers
    unit: GPU
    startTiming: SpecificDate
    desiredHours: 720
    contactName: Research lead
  state: Open
```

`status.intake` records Nodus’s customer-visible manual progress for the matching request revision. Only an authorized Nodus operator can update it. **MarketplaceContract** records remain available for existing integrations: each stores a versioned reservation proposal and both organizations’ confirmations. A confirmation records agreement on the proposal; it does not charge, provision capacity or execute a legal document. The request workspace handles new offers and agreements manually.

Reference coverage includes thirteen RTX, H and B models across published terms. The request brief includes the sources and calculation. Monthly dedicated rentals do not establish annual contract pricing; smaller reference configurations do not guarantee an eight-GPU host.