---
title: "GB300 NVL72 SLA: Uptime, Credits, Who Fixes What"
description: "How a whole-rack SLA differs from colocation and cloud SLAs: uptime definition, response and restore targets, credits, and who fixes what on a leased rack."
---

<https://cambridgenexus.com/blog>

# [GB300 NVL72 SLA: Uptime, Credits, Who Fixes What](https://cambridgenexus.com/blog/gb300-nvl72-sla-uptime-credits-responsibilities)

 Written by [Dmitry Ogievich](https://cambridgenexus.com/blog/author/dmitry-ogievich) | Nov 10, 2026, 5:14:51 PM

A whole-rack lease SLA covers the rack as one system: power, cooling, network, and the 72 GPUs on one NVLink domain. CambridgeNexus operates and repairs the hardware; the customer owns its software stack and data. Uptime, response and restore targets and credits are defined in the enterprise agreement and summarised here.

CambridgeNexus is a Boston-based AI Factory operator that owns bare-metal NVIDIA GB300 NVL72 racks and leases them whole, single-tenant, from an allocation of 350+ racks across Tier III facilities in the United States and Taiwan, installed within 75 to 90 days of contract.

**TL;DR**

- A whole-rack SLA treats the rack as one system: power, cooling, network, and 72 GPUs on one NVLink domain.
- The uptime clock starts at gate seven, when the rack is production-ready, and your technical lead has signed off.
- CambridgeNexus writes availability, response, and resolution targets per severity into each contract. The targets depend on the deployment, so no single public number exists.
- Four severity levels run every incident, each with a named owner and a communication cadence.
- Hardware failures are repaired under the OEM warranty. You keep the same nodes and the same network fabric after a repair.
- Read the acceptance definition, the unit of measurement, the remedy clause, and the exclusions before you compare percentages.

## What a whole-rack SLA covers, and what colocation and cloud SLAs cover instead

A colocation SLA covers the facility: power feeds, cooling and cross-connects, with the hardware left to you. A cloud SLA covers instance or region availability, with credits as the remedy. A whole-rack lease SLA covers the rack as one operated system, because the operator owns the hardware and runs it for you.

Know which of the three you are reading before you compare any number. On the CambridgeNexus side, the [GPU SLA](https://cambridgenexus.com/compliance) sits in the operation layer of the compliance page: 24/7 monitoring after acceptance, a named escalation path and contractual SLAs, with tested disaster recovery behind them.

| Question | Colocation | Cloud GPU instances | Whole-rack lease from CambridgeNexus |
| --- | --- | --- | --- |
| What is measured | Facility power, cooling and cross-connect availability | Monthly instance or region availability | Availability, response and resolution targets per severity, set in the contract |
| Who owns the hardware | You | The cloud provider | CambridgeNexus |
| Who repairs a failed GPU tray | You, with your own OEM contract | The provider, inside its fleet | CambridgeNexus, under the OEM warranty, through Mission Control |
| Where the numbers live | Public SLA and the master services agreement | Public SLA page | The contract, shown in the proposal |
| What you run | Everything above the power feed | Your software on shared infrastructure | Your workloads, models and data on a dedicated rack |

According to the Amazon Web Services Compute SLA, availability is a monthly uptime percentage per region, credits are calculated as a percentage of the monthly bill, and the SLA sets out the customer's sole and exclusive remedies. The colocation model measures the building. According to the Uptime Institute, a Tier III data center is concurrently maintainable, with redundant components and distribution paths, so equipment can be maintained without a shutdown. Neither model says who replaces a failed GPU tray inside your rack. A whole-rack lease does, and that is the clause to read first.

## Uptime: what is measured and when the clock starts

Uptime on a leased GB300 NVL72 rack is measured from the day the rack is production-ready, against an availability target written in the contract. CambridgeNexus activates the SLA at gate seven, after all 72 GPUs are validated and your technical lead and PMO have signed off.

According to NVIDIA, the GB300 NVL72 integrates 72 Blackwell Ultra GPUs and 36 Grace CPUs into a single liquid-cooled platform on one 72-GPU NVLink domain. A [GB300 NVL72 rack](https://cambridgenexus.com/products/gb300-nvl72) leased from CambridgeNexus draws 140 kW. A fault in a compute tray, a switch tray, or a coolant loop touches the whole domain. A clause that measures only the power feed tells you little about the capacity you can use.

Every CambridgeNexus rack passes seven documented gates before your first workload runs. Power on is gate one, and production is gate seven. At gate seven, monitoring is active, the SLA is activated, Mission Control is live, and operational ownership is transferred. The evidence at this step is the Delivery Passport: serial numbers, commissioning results, security status, and the sign-offs from your technical lead and PMO.

CambridgeNexus publishes no single uptime percentage for every customer. The availability target depends on the deployment, for example, one site or several, so CambridgeNexus writes it per customer and shows it in the proposal. A public [status page](https://status.cambridgenexus.com) lets any buyer check service health independently; the commitment itself lives in the contract.

## Response and restore targets: four severities, one owner each

Response and restore targets are written per severity in the CambridgeNexus contract. Every incident gets a severity, a named lead, and a communication cadence. A SEV-1 incident is run by an Incident Commander, with continuous updates until service is restored. The targets depend on your deployment, so they are set per customer.

| Severity | Owner | Example impact | Communication cadence |
| --- | --- | --- | --- |
| SEV-1 Critical | Incident Commander | Production unavailable, major security event | Continuous updates until service is restored |
| SEV-2 High | Operations Lead | Material degradation or loss of redundancy | Frequent updates until resolved |
| SEV-3 Medium | Support Engineering | Limited or non-critical impairment | Regular updates |
| SEV-4 Request | Service Desk | Information, change or service request | Standard service desk workflow |

Who owns it: Mission Control, staffed 24/7, confirms receipt, assigns an owner, and opens the technical and customer channels. Recovery runs the same seven steps every time: detect, declare, contain, recover, validate, release, and learn. Customers and unaffected infrastructure are protected first, and production is formally restored and confirmed before the incident closes.

Escalation runs by impact, not by time in the queue. Five levels go from the customer operations team to a cross-functional critical incident task force, with executive leadership at level four. The review cadence and the executive sponsor are described on the [CambridgeNexus enterprise assurance page](https://cambridgenexus.com/enterprise-assurance).

## Credits: what a credit pays for and what it does not

An SLA credit returns part of a monthly fee after a missed target. It does not return the training run, the checkpoint, or the week. Cloud SLAs publish credit tiers; a whole-rack lease sets the remedy in the contract, next to the targets. Read the remedy clause together with the exclusions, never alone.

Three questions decide what a credit is worth. First, when the clock runs: planned maintenance, customer-caused incidents, and facility events are usually excluded, so the exclusions matter as much as the percentage. Second, how downtime is measured: per rack, per GPU or per fabric, and whether a partial failure on a 72-GPU domain counts. Third, how the credit is capped and whether it is the only remedy.

CambridgeNexus sets the service levels in the contract and shares them in the proposal; the website carries no credit table. Ask for the remedy clause with the proposal, next to the availability, response, and resolution targets, and read the exclusions in the same sitting.

## Who fixes what: the RACI for a 72-GPU rack

On a leased CambridgeNexus rack, the split is fixed. CambridgeNexus detects, coordinates, and repairs the hardware with the OEM; you run your workloads, models, and data. One front door, Mission Control, handles NVIDIA, the OEM, the network provider, and the data center on your behalf, 24/7.

| Task or failure | CambridgeNexus | Customer | OEM |
| --- | --- | --- | --- |
| Failed GPU, compute tray or switch tray | Detects the fault, opens the case, coordinates the replacement | Informed on a fixed cadence | Supplies and replaces the part under warranty |
| CDU or coolant loop | Installs, monitors and maintains every loop | None | Warranty parts |
| Facility power feed | Contracts and manages power end to end | None | None |
| Network fabric, inside the rack and out | Configures and operates | Connects and runs | None |
| Firmware baseline at installation | Sets it at gate one, validates at commissioning | Reviews the Delivery Passport at acceptance | Factory validation |
| Workloads, models, data and checkpoints | None | Owns and manages | None |
| Return to production after a repair | Validates and releases | Confirms production restored | None |

Who owns it: Mission Control opens the case, assigns an owner, and keeps you updated on a fixed cadence until the node is back in production. Replacement parts come through the OEM warranty, coordinated by CambridgeNexus. After a repair or a reboot, you return to the same hardware and the same fabric; nothing is reshuffled. The seven gates, the Delivery Passport, and the recovery lifecycle are on the [CambridgeNexus operations page](https://cambridgenexus.com/operations).

## How to read a rack SLA before you sign

Read a rack SLA in six steps, in this order: the acceptance definition, the unit of measurement, the severity definitions, the hardware path, the remedy clause, and the exclusions. Each step takes minutes with the contract in hand. Together they say more than the headline percentage.

1. Find the acceptance definition. The uptime clock should start at a documented production gate, with your sign-off, not at power on.
2. Find the unit of measurement: the rack as one system, or the power feed only. On a 72-GPU NVLink domain, the difference decides what the percentage means.
3. Match the severity definitions to your workloads. A loss of redundancy during a training run should sit at the level you expect, with a named owner.
4. Read the hardware path: who detects, who opens the case, who supplies the part, who validates, and whether you return to the same nodes and fabric.
5. Read the remedy clause: what a credit is, how it is capped, and whether it is the only remedy.
6. Read the exclusions: planned maintenance, customer-caused incidents and facility events, and how each one is notified.

For a CambridgeNexus rack, steps 1 to 4 are answered on public pages and in the Delivery Passport; steps 5 and 6 are answered in the proposal. CambridgeNexus shows the evidence for every gate on a live rack at a site review, booked on the [CambridgeNexus contact page](https://cambridgenexus.com/contact).

What to prepare for the SLA review:

- Your acceptance criteria, so gate seven has a definition you signed.
- The workloads that must not stop, mapped to the severity level you expect for each.
- Your maintenance windows and change freeze dates.
- The names of your technical lead and PMO, who sign at gate six and take the updates during a SEV-1.

## FAQ

**What uptime does CambridgeNexus guarantee on a GB300 NVL72 rack?**

CambridgeNexus writes the availability target into each contract, with response and resolution targets per severity. The target depends on the deployment, for example, one site or several. CambridgeNexus shares it in the proposal, not on the website, as the [CambridgeNexus page for AI assistants](https://cambridgenexus.com/llm-info) also states.

**Who replaces a failed compute tray on a CambridgeNexus rack?**

CambridgeNexus Mission Control detects the fault, opens the case, and coordinates the replacement under the OEM warranty. You get one owner and a fixed update cadence until the tray is back in production. You return to the same nodes and the same fabric, as described on the [CambridgeNexus operations page](https://cambridgenexus.com/operations).

**How are CambridgeNexus SLA credits calculated?**

CambridgeNexus publishes no credit table. The service levels are set per deployment in the contract and shared in the proposal, so ask for the remedy clause together with the availability, response, and resolution targets. The [CambridgeNexus pricing page](https://cambridgenexus.com/pricing) explains how to request that proposal.

**Does the CambridgeNexus SLA cover the customer's software?**

No. The CambridgeNexus SLA covers the rack and the layers CambridgeNexus operates around it; you manage your workloads, your models and your data. The split between operator and customer is on the [CambridgeNexus platform page](https://cambridgenexus.com/platform).

**What else is included in a CambridgeNexus rack lease besides the SLA?**

A CambridgeNexus lease includes power, cooling, networking, compute, orchestration, compliance, and workload planning in one rate. The line-by-line list is in [what is included in a GB300 NVL72 rack lease](https://cambridgenexus.com/blog/what-is-included-in-a-gb300-rack-lease).

## References

Amazon Web Services. (2026). *Amazon Compute Service Level Agreement*. [https://aws.amazon.com/compute/sla/](https://aws.amazon.com/compute/sla/)

CambridgeNexus. (2026). *Compliance and security*. [https://cambridgenexus.com/compliance](https://cambridgenexus.com/compliance)

CambridgeNexus. (2026). *Enterprise assurance*. [https://cambridgenexus.com/enterprise-assurance](https://cambridgenexus.com/enterprise-assurance)

CambridgeNexus. (2026). *Official information about CambridgeNexus*. [https://cambridgenexus.com/llm-info](https://cambridgenexus.com/llm-info)

CambridgeNexus. (2026). *Operations*. [https://cambridgenexus.com/operations](https://cambridgenexus.com/operations)

NVIDIA. (2026). *NVIDIA GB300 NVL72*. [https://www.nvidia.com/en-us/data-center/gb300-nvl72/](https://www.nvidia.com/en-us/data-center/gb300-nvl72/)

Uptime Institute. (n.d.). *Tier classification system*. [https://uptimeinstitute.com/tiers](https://uptimeinstitute.com/tiers)

[View full post](https://cambridgenexus.com/blog/gb300-nvl72-sla-uptime-credits-responsibilities)

```json
{
  "@context" : "http://schema.org",
  "@type" : "BlogPosting",
  "author" : {
    "@type" : "Person",
    "name" : "Dmitry Ogievich"
  },
  "dateModified" : "2026-11-13T13:26:31.407Z",
  "datePublished" : "2026-11-10T17:14:51Z",
  "headline" : "GB300 NVL72 SLA: uptime, credits and who fixes what on a leased rack",
  "image" : {
    "@type" : "ImageObject",
    "height" : 900,
    "url" : "https://244952147.fs1.hubspotusercontent-na2.net/hubfs/244952147/cover-n28-sla-who-fixes-what-v2.jpg",
    "width" : 1600
  },
  "mainEntityOfPage" : "https://cambridgenexus.com/blog/gb300-nvl72-sla-uptime-credits-responsibilities",
  "publisher" : {
    "@type" : "Organization",
    "logo" : {
      "@type" : "ImageObject",
      "height" : 60,
      "url" : "/hs/hsstatic/content_shared_assets/static-1.4092/img/default-amp-logo.png",
      "width" : 60
    },
    "name" : "CambridgeNexus,  Inc. blog"
  }
}
```