> ## Content Index > Fetch the complete content index at: https://www.sotkn.com/llms.txt > Use this file to discover other available public pages before exploring further. # What Is an NVIDIA Cloud Partner? Why Offtake Waits on the API - URL: https://www.sotkn.com/explained/nvidia-cloud-partner-explained/ - Published: 2026-10-28T01:15:17.000Z - Updated: 2026-10-28T01:16:25.000Z - Description: An NVIDIA Cloud Partner is a specialist GPU cloud in NVIDIA's partner network. Offtake needs the API 12 weeks before the cards. Software readiness, not the badge, is the stuck step. - Author: Sean - Tags: #desk-explained, Mechanism EXPLAINED · Mechanism · Last reviewed: 28 Oct 2026 · Next review: after NVIDIA reports the quarter ended October 2026 An NVIDIA Cloud Partner (NCP) is a member of the NVIDIA Partner Network whose primary business is selling hosted GPU cloud and managed services on NVIDIA products. It is not a GPU. It is NVIDIA’s designation for the specialist clouds between the chip and the hour — and the paperwork NVIDIA uses when it buys those hours back. **In short.** NVIDIA’s partner page defines an NCP as an AI cloud provider delivering infrastructure purpose-built for modern AI workloads at production scale, inside the NVIDIA Partner Network. A June 2024 NVIDIA blog set the reference architecture those partners are supposed to build. Requirements for AI Clouds v2.3, dated 25 June 2026, then says what an NCP must demonstrate **to be considered for offtake**: API readiness and transport at least **12 weeks** before GPU delivery; development capacity, high-performance storage and the data-movement system API-integrated **8 weeks** before. Our arithmetic: **84 days** and **56 days**, a **28-day** gap. The stuck step is that software clock plus live GPU allocation — not the badge on a partner page. Why an investor cares: a [neocloud](https://www.sotkn.com/explained/neocloud-explained/) is the business. NCP is NVIDIA’s gate on that business. [GPU as a service](https://www.sotkn.com/explained/gpu-as-a-service-explained/) is the hour. [Take-or-pay](https://www.sotkn.com/explained/take-or-pay-gpu-contracts-explained/) is the customer paper. This page is the NVIDIA paper that decides whether those hours can be sold back to NVIDIA, and on what stack they can be sold to everyone else. ## What is an NVIDIA Cloud Partner? NVIDIA’s Partner Network page puts NCP next to OEMs, distributors, data-centre partners and systems integrators. The NCP line is specific: partners whose **primary business model** is offering hosted software and services, in a cloud or managed-services model, to end-user customers, leveraging NVIDIA products. The dedicated Cloud Partners page repeats it: NCPs are AI cloud providers in the NPN delivering infrastructure purpose-built for modern AI workloads at production scale. As of this review, that page’s partner grid names **Crusoe, GMI Cloud, Lambda, Lightning AI, Nebius, SoftBank and Yotta**. That is NVIDIA’s marketing list on 28 Oct 2026, not a census of every cloud that owns NVIDIA GPUs. A name on the grid is not offtake. Three related NVIDIA labels sit next to NCP and are not the same thing: the partner type itself, the October 2024 **Reference Platform NCP** specialisation for large clusters on the tested reference architecture, and **Exemplar Cloud**, a benchmark badge. NVIDIA’s May 2025 technical blog said Exemplar started with the NCP ecosystem; the live page now lists specialist clouds **and** hyperscalers — AWS, Microsoft Azure and Oracle Cloud Infrastructure sit on the same table as CoreWeave, Crusoe, Lambda and Nebius. ## How does it work? Two clocks run. One is the hardware blueprint. The other is the offtake checklist. The blueprint is the NCP reference architecture NVIDIA published on 26 June 2024\. It is a full-stack list, not a logo: GPU servers from NVIDIA and its manufacturing partners (the 2024 post names Hopper and Blackwell); storage tested for DGX SuperPOD and DGX Cloud; Quantum-2 InfiniBand and Spectrum-X Ethernet for east-west traffic; BlueField-3 DPUs for north-south traffic; and NVIDIA AI Enterprise software. The cash consequence is that the partner’s bill of materials is NVIDIA’s — GPUs, networking, DPUs and software — before a single hour is sold. The offtake checklist is Requirements for AI Clouds. Version 2.1 landed 26 February 2026; v2.2 on 10 April 2026; **v2.3 on 25 June 2026**. NVIDIA calls it the primary requirements set for any NCP providing GPU and AI compute services needed to run DGX Cloud, a public reference for what a large GPU consumer requires, and a demand that those services be generally available — not a one-off build for NVIDIA alone. The dated clocks in v2.3 are the ones that bind. - **T-minus 12 weeks (84 days).** The NCP must demonstrate API readiness and transport establishment at least 12 weeks ahead of GPU delivery. - **T-minus 8 weeks (56 days).** The NCP must provide development capacity — ancillary CPU nodes only — and high-performance storage with the API integrated, and must meet the Data Movement Systems requirements, 8 weeks before GPU and cluster delivery. - **SLAs to be considered for offtake.** Managed Kubernetes control-plane target: financially-backed **99.95%+** uptime for production. Our arithmetic: that annual target leaves about **4.38 hours** of downtime a year. Home-directory storage: over 99.99% availability for unplanned incidents, exclusive of scheduled maintenance. Telemetry must arrive within **120 seconds**. SOC 2 Type 1 or better is required, covering security, availability and confidentiality. - **Exemplar workload performance (BM01).** Run NVIDIA’s public DGXC benchmarking release on every Scalable Unit handed off — NVIDIA’s example is a **512-GPU** cluster — and achieve within **5%** of an NVIDIA-provided target. The test is a performance number. The unit of hand-off is the SU. Capacity cannot be “handed” to DGX Cloud by phone, Slack or email: v2.3 requires a programmatic Resource Index. For NVL72, the API must support NVLink domain-aware allocation. Break-fix must pull a node without breaking the rest of the tenancy, and must reconfigure NVLink when it does. That stack has to be live **before** the cards land. ## How is it different from a neocloud, a hyperscaler, or Exemplar Cloud? Four labels, four jobs. | Label | What it is | Who writes the paper | What it does not mean | | -------------------------- | --------------------------------------------------------------- | -------------------------------------- | ----------------------------------------------------------------- | | **Neocloud** | A specialist GPU cloud that bills accelerator hours | The operator | NVIDIA has blessed the stack, or will buy the idle hours | | **NVIDIA Cloud Partner** | An NPN partner type for hosted NVIDIA GPU cloud | NVIDIA’s partner programme | Offtake, Exemplar status, or a complete list of GPU clouds | | **Reference Platform NCP** | A specialisation for large clusters on the NCP reference design | NVIDIA, invitation-only specialisation | Every NCP has it | | **Exemplar Cloud** | A benchmark badge on named GPU types | NVIDIA’s Exemplar programme | The operator is an NCP — AWS, Azure and OCI sit on the same table | A [neocloud](https://www.sotkn.com/explained/neocloud-explained/) can be an NCP. The live grid above is that overlap. A neocloud can also sell hours without the designation. A [hyperscaler](https://www.sotkn.com/explained/hyperscaler-explained/) sells GPU hours at much larger scale and is not an NCP on NVIDIA’s partner-type list; it can still sit on the Exemplar Cloud table. NVIDIA as **customer** is a third path. CoreWeave’s 8-K dated 15 September 2025 reports a new order form under a 10 April 2023 master services agreement, initial value **$6.3 billion**, under which NVIDIA is obligated to purchase residual unsold cloud capacity through **13 April 2032**, subject to termination and to “satisfaction of delivery and availability of service requirements.” NVIDIA also supplies CoreWeave with GPUs and is a stockholder. Our arithmetic on those two dates: **2,408 days** from 9 September 2025 to 13 April 2032\. That contract is NVIDIA offtake. It is not the definition of NCP. The phrase that matches v2.3 is the condition: delivery and availability of service requirements. The badge does not write the invoice. The live stack does. ## Why does it bind? It binds twice: on NVIDIA’s allocation, and on the partner’s software clock. GPUs are still the scarce object. An NCP’s product is [GPU as a service](https://www.sotkn.com/explained/gpu-as-a-service-explained/). Without an allocation, the partner has a hall and a logo. The reference architecture is how that allocation is supposed to be pulled through as one bill of materials, not as a pile of cards in a generic cage. The second bind is the one v2.3 dates. Offtake consideration requires API readiness **84 days** before GPU delivery, and storage plus the data-movement path API-integrated **56 days** before. A partner that energises a hall on GPU day minus 30 has already missed both clocks. The cards can sit in the cage. NVIDIA’s checklist will not treat them as delivered capacity until the Resource Index, Kubernetes control plane, telemetry pipe and storage mesh exist as generally available services. Our read: a neocloud can show contracted megawatts and still not have offtake-grade hours. The hall is not the stuck step. The stack that has to be live 12 weeks earlier is. [GPU utilisation](https://www.sotkn.com/explained/gpu-utilization-explained/) still decides whether a sold hour earns. NCP offtake is the prior question: whether NVIDIA will take the residual hour at all. ## Who gets paid — and where is the stuck step? Four groups of cash. **NVIDIA** is paid first, when the NCP buys the bill of materials in the reference architecture — GPUs, InfiniBand or Spectrum-X, BlueField, and AI Enterprise. NVIDIA can be paid again as a customer, when it offtakes residual hours, the way the CoreWeave 8-K describes. Equity in that case is not the hour. **The NCP** is paid when it sells GPU hours to its own customers — on-demand, reserved, or take-or-pay — and when NVIDIA offtakes what those customers did not take. The partner’s [TCV](https://www.sotkn.com/explained/total-contract-value-explained/) and [remaining performance obligation](https://www.sotkn.com/explained/remaining-performance-obligation-explained/) are the customer book. Residual offtake, where it exists, is a backstop on utilisation, not a substitute for a signed customer. The 8-K is the primary for CoreWeave’s numbers. **The lenders and landlords** are paid on the clocks our [GPU financing](https://www.sotkn.com/explained/gpu-financing-explained/) and [data-center financing](https://www.sotkn.com/explained/data-center-financing-explained/) pages track. An NCP designation does not draw a loan. A live, offtake-eligible cluster does. **The end customer** pays for a GPU hour on a named stack. Exemplar Cloud is how NVIDIA wants that customer to compare performance per TCO. It is not how the partner gets the cards. The stuck step is **live, offtake-eligible capacity**: GPU allocation **plus** the API, storage, telemetry and Kubernetes stack that v2.3 dates to 12 weeks and 8 weeks before delivery. A partner page, an Exemplar cell on a GPU-type table, or a signed hall lease does not move that clock. Our read: NVIDIA gets paid for the stack, then for the residual hour, and keeps the right to say when that hour counts as delivered. The NCP gets paid for the hour only after both gates open. ## What would change the view? - **NVIDIA publishes a dated NCP roster and which partners are offtake-eligible.** The marketing grid would then stop being a partial list. - **Requirements for AI Clouds drops the 12-week API clock, or NVIDIA offtakes halls that miss it.** The software bind would thin out. - **Exemplar Cloud and NCP are merged into one public badge.** The four-label split above would collapse. - **NVIDIA stops buying residual unsold capacity from specialist clouds.** The CoreWeave 8-K path would then be a one-off. ## Common misconceptions - **People assume NCP means the cloud will get the next GPU allocation.** Actually NCP is a partner type. Allocation is a separate NVIDIA decision. The marketing grid is not a delivery queue. - **People assume Exemplar Cloud is the NCP list.** Actually NVIDIA’s Exemplar Cloud table includes AWS, Azure and OCI. Exemplar is a benchmark badge. - **People assume a residual offtake deal is the definition of NCP.** Actually CoreWeave’s $6.3 billion order form is NVIDIA as customer, conditional on delivery and availability. v2.3 is the public checklist those words point at. - **People assume the hall is the gate.** Actually v2.3 dates API readiness 84 days before GPU delivery. A powered cage without that API is not offtake-grade capacity. ## FAQ **What is an NVIDIA Cloud Partner?** An NVIDIA Cloud Partner is an NVIDIA Partner Network member whose primary business is selling hosted GPU cloud and managed services on NVIDIA products. It is a partner type, not a GPU serial number and not a guarantee of offtake. **How is an NCP different from a neocloud or a hyperscaler?** A neocloud is the business: a specialist GPU cloud. NCP is NVIDIA's partner designation for clouds that sell hosted services on its stack. A hyperscaler can sell GPU hours without being an NCP. NVIDIA's own Exemplar Cloud table lists AWS, Azure and OCI alongside specialist clouds. **What does NVIDIA require before offtake?** NVIDIA's Requirements for AI Clouds v2.3 says NCPs must show they can meet the SLAs to be considered for offtake. API readiness and transport must be demonstrated at least 12 weeks before GPU delivery; development capacity, high-performance storage and the data-movement system must be API-integrated 8 weeks before. **Is Exemplar Cloud the same as NCP?** No. NCP is the partner type. Exemplar Cloud is a separate performance-validation programme. NVIDIA's Exemplar Cloud page lists hyperscalers as well as specialist clouds. Reference Platform NCP is a further NPN specialization for partners built on the NCP reference design. **Is this a list of stocks to buy or sell?** No. This page explains how the NVIDIA Cloud Partner designation works and where the money goes. It does not rank companies or give buy or sell calls. Not investment advice. Do your own research. ## Related terms Related: [neocloud](https://www.sotkn.com/explained/neocloud-explained/); [GPU as a service](https://www.sotkn.com/explained/gpu-as-a-service-explained/); [take-or-pay GPU contracts](https://www.sotkn.com/explained/take-or-pay-gpu-contracts-explained/); [GPU utilisation](https://www.sotkn.com/explained/gpu-utilization-explained/); [total contract value](https://www.sotkn.com/explained/total-contract-value-explained/); [remaining performance obligation](https://www.sotkn.com/explained/remaining-performance-obligation-explained/); [neocloud vs hyperscaler](https://www.sotkn.com/explained/neocloud-vs-hyperscaler/). ## Markets and stack doors Layer: [Neoclouds](https://www.sotkn.com/stack/neoclouds/). Also [Hyperscalers](https://www.sotkn.com/stack/hyperscalers/) and [Semiconductors](https://www.sotkn.com/stack/semiconductors/). Markets: [US](https://www.sotkn.com/markets/us/). Company models: [What CoreWeave does](https://www.sotkn.com/stack/neoclouds/what-does-coreweave-do/), [What Nebius does](https://www.sotkn.com/stack/neoclouds/what-does-nebius-do/), [What Lambda does](https://www.sotkn.com/stack/neoclouds/what-does-lambda-do/), [What Crusoe does](https://www.sotkn.com/stack/neoclouds/what-does-crusoe-do/). [Get the Wire](https://www.sotkn.com/#/portal/signup) — short daily notes when cloud capacity and capex move. **How we check this** Figures come from NVIDIA’s Cloud Partners and Partner Network pages, the 26 June 2024 NCP reference-architecture blog (October 2024 Reference Platform NCP update), the 18 May 2025 Exemplar Clouds technical blog, NVIDIA’s live Exemplar Cloud page, Requirements for AI Clouds v2.3 (25 June 2026), and CoreWeave’s 15 September 2025 8-K. The 84-day, 56-day and 28-day clocks, the 4.38-hour annual downtime budget on a 99.95% SLA, and the 2,408-day span on the CoreWeave dates are our arithmetic from those primaries. Partner names on the NCP grid are as that page stood on this review. Where we interpret, it is labelled “our read”. Not investment advice. Do your own research. **Last reviewed:** 28 Oct 2026 · **Next review:** after NVIDIA reports the quarter ended October 2026 Spot an error? [Tell us](mailto:sean@thekopinotes.com). [Editorial standards →](https://www.sotkn.com/editorial-standards/) ## Sources 1. [NVIDIA — NVIDIA Cloud Partners](https://www.nvidia.com/en-us/data-center/gpu-cloud-computing/partners/?ref=sotkn.com) 2. [NVIDIA — NVIDIA Partner Network (NPN)](https://www.nvidia.com/en-us/about-nvidia/partners/?ref=sotkn.com) 3. [NVIDIA Blog — NVIDIA Unveils Reference Architecture for AI Cloud Providers](https://blogs.nvidia.com/blog/ai-cloud-providers-reference-architecture/?ref=sotkn.com) (26 Jun 2024; Reference Platform NCP update Oct 2024) 4. [NVIDIA DSX — Requirements for AI Clouds v2.3](https://docs.nvidia.com/dsx/ncp/nvidia-requirements-for-ai-clouds/home?ref=sotkn.com) (25 Jun 2026) 5. [NVIDIA Technical Blog — Announcing NVIDIA Exemplar Clouds](https://developer.nvidia.com/blog/announcing-nvidia-exemplar-clouds-for-benchmarking-ai-cloud-infrastructure/?ref=sotkn.com) (18 May 2025) 6. [NVIDIA — Exemplar Cloud](https://www.nvidia.com/en-us/data-center/ai-cloud-performance/?ref=sotkn.com) 7. [CoreWeave Form 8-K](https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm?ref=sotkn.com) (15 Sep 2025; $6.3bn order form, residual capacity through 13 Apr 2032) [Sean](https://www.sotkn.com/author/sean/) Writes Second Order by TKN — a plain-English map of AI infrastructure and semiconductors. [Author page →](https://www.sotkn.com/author/sean/) · [Editorial standards →](https://www.sotkn.com/editorial-standards/)