Quick Take
Quick Take: A 2026 GPU purchase clears six NVIDIA gates: ODM, MGX, allocation, NCP, NPN, and KYC. Build certification is not a quota. A China order-path badge is not a restricted-SKU license after November 2023 controls. Fit is an RFQ lock. Unfit is a partner directory or a legal memo.
What Are NVIDIA ODM, MGX, Allocation, NCP, NPN, and KYC Gates?
Quick Answer: The six controls run in order, and no single badge covers them. Manufacturing permission is the ODM gate. A modular reference design is the MGX gate. Scarce Blackwell supply is the allocation gate. Selling compute as a service is the NCP type. Registered, Preferred, and Elite are NPN tiers. Ownership, site, and end use are the KYC screen.
NVIDIA does not hand a GPU to a generic distributor and stop. A card moves through authorized build, a certified design, a scarce-supply decision, a cloud or partner sell-through path, and an end-customer screen. A channel gate is a control that must clear before the purchase can move to the next party. Certification to build is not the same thing as a quota, and a partner tier is not the same thing as a cloud-partner type.
This page fits procurement, compliance, and channel managers who are locking an RFQ: who may integrate a system, who may resell compute, and what papers the purchase order must carry. It is unfit as a complete partner directory, unfit as a legal memo, and unfit as instructions for moving restricted GPUs around a license. The China four-layer order path is on this page. A Southeast Asia partner roster is still deferred. This overview does not invent that second map.
Read the six names as questions, not as logos on one certificate:
- ODM — who may manufacture boards or systems under NVIDIA authorization.
- MGX — who may ship a system on the modular reference design. MGX is not allocation.
- Allocation — who receives scarce Blackwell-class GPUs first in a tight cycle.
- NCP — NVIDIA Cloud Partner, a type that sells compute as a service.
- NPN — NVIDIA Partner Network, a broader membership with Registered, Preferred, and Elite grades.
- KYC — the end-customer screen: ownership, physical site, and stated end use.
Partner names later in this article are source examples. Each one is pending confirmation against the official NVIDIA partner locator before anyone treats it as a directory entry. Reported dates from the financial press and from export-control coverage are labeled as reported. They are not a TAOHAN legal opinion.
Gate Table: What Each NVIDIA Control Blocks on a B300 or H200 RFQ
Quick Answer: Treat each row as its own yes or no. A factory authorization does not certify the chassis design. A certified chassis design does not reserve GPUs for your quarter. Cloud compute sales are a partner type. Membership grades are a network tier. The last row screens destination, and example names stay pending confirmation.
Use this table when you open a 2026 B300 or H200 RFQ. The buyer takeaway is what the gate refuses to grant, not a marketing description of the program. Counts of partners are not measured here. Where a source story uses a round number, the cell says pending confirmation.
| Gate | What NVIDIA controls | Buyer takeaway |
|---|---|---|
| 1. ODM | Who may build boards or systems under authorization | No authorization means that party cannot touch the hardware manufacturing path |
| 2. MGX | Who may use the standardized modular reference design | A compliant design may ship. That is not a GPU quota |
| 3. Allocation | Who gets scarce Blackwell-class GPUs first | Hyperscaler, NCP, and sovereign-AI priority shows up in tight cycles. Any short name list is pending confirmation |
| 4. NCP | Cloud type inside the partner program: compute as a service | A small published set. Type is not a membership tier |
| 5. NPN | Broader network: Registered, Preferred, Elite | Thousands of partners globally. Tier is not NCP |
| 6. KYC | End-customer screen on every deal | Allocation sets quantity. KYC sets destination |
Two limits belong on the same page as the table. First, commercial list price is not a channel net, and this article publishes no prices at all. Second, a row that names a company is an illustration from source notes, not a headcount you can paste into a bid. If your RFQ needs a live locator hit, ask for the entity, the region, and the locator screenshot dated the day of the quote.
NCP vs NPN: Why a Preferred or Elite Tier Is Not a Cloud Partner
Quick Answer: Membership and cloud status answer different questions. Distributors, server builders, integrators, and some clouds sit in a three-grade network. A cloud that sells GPU compute sits in a separate type. A Preferred grade is not that type. An Elite grade is not that type. Published cloud-type counts conflict in source notes, so leave the headcount unfixed.
NPN refers to a membership tier. NCP refers to a partner type. Collapsing those words is the most common RFQ error on a GPU buy, because a high grade sounds like permission to sell cloud compute. It is not. Preferred is not NCP. Elite is not NCP.
| Axis | NPN | NCP |
|---|---|---|
| What it is | Membership tier: Registered, Preferred, or Elite | Partner type: cloud |
| Who | Distributors, server OEMs, system integrators, and some clouds | Clouds that sell GPU compute |
| Breadth | A global network measured in the thousands | A short published list. Count pending confirmation. Do not freeze “tens” or “about 80” as fact |
| China locator claim in source | Many NPN names plus a national-distributor layer | Conflicting NCP claims: Alibaba Cloud only in one source; Xiechuang Data and Runze Technology in a later map. Pending locator confirmation |
Source notes, still pending confirmation on the locator, separate China-market claims that buyers mix together:
- One source lists Alibaba Cloud as the only China Cloud Partner (NCP).
- A later map lists Xiechuang Data and Runze Technology as China NCP examples.
- L&T is NPN Preferred on an AI cloud and DGX track. That grade is not NCP.
- Hongxin Electronics is the follow-up’s dual Elite (Compute and Networking) example, not Digital China.
- Digital China and Unisplendour sit on the national distributor layer, not as NCP.
The same legal name can be an NCP in one region and only a reseller of a local NCP in another. Identity is the contracting entity plus the region, not the brand string on the slide. When two regions share a logo, put two rows on the RFQ and ask which row holds the cloud-type status.
Language inside NCP materials often mentions AI Factory, Reference Platform, or SuperPOD style offers. Those phrases describe how compute is packaged. They do not convert an NPN reseller into an NCP, and they do not prove allocation for your quarter.
Which China Channel Tiers Can Place an NVIDIA GPU Order in 2026?
Quick Answer: Source notes split China GPU order paths into four layers: national distributor, NPN Elite, NPN Preferred, and NCP. After November 2023 export controls, those badges do not reopen a compliant H100, H200, or B-series buy. Names below are examples pending official locator confirmation, not a partner directory.
A later China-market note fills the roster this overview had deferred. It is still not a partner directory. It is a four-layer order-path map: who may raise a purchase order toward NVIDIA in the China market, at what qualification, and what that badge does not buy after export controls. Southeast Asia names remain out of scope.
| Layer | Qualification | Source examples (pending confirmation) |
|---|---|---|
| 1 | National primary distributor | Unisplendour; Digital China |
| 2 | NPN Elite | Hongxin Electronics (claimed sole dual Elite: Compute + Networking); Advanced Datacomm; China Electronics Port |
| 3 | NPN Preferred | L&T, also described as a mainland DGX / AI cloud partner. That track is still Preferred, not NCP |
| 4 | NCP cloud partner | Xiechuang Data; Runze Technology |
Two source claims about China NCP names now sit side by side. An earlier note listed Alibaba Cloud as the only China Cloud Partner on the official locator. The follow-up map lists Xiechuang Data and Runze Technology. Do not collapse the two lists into one. Confirm the live locator for the contracting entity and the region on the day of the quote.
The follow-up also corrects a label buyers mix up. Hongxin Electronics is the name tied to dual Elite (Compute and Networking). Digital China and Unisplendour sit on the national distributor layer. Unisplendour is also described as distributor plus Elite, which is a different claim from dual Elite Compute plus Networking. Public reporting links Hongxin’s Elite GPU sell-through to an acquired NVIDIA Elite seller. Treat that chain as pending confirmation.
After U.S. export controls in November 2023, source notes say the compliant order path for H100, H200, and B-series through these China qualifications is largely closed. The remaining use of the badge is non-restricted SKUs, channel relationship, and KYC credibility. It is not a license waiver. It is not a reason to skip destination screening.
- A China qualification is scarce. After export control it is not a restricted-SKU license.
- National distributors move volume. Elite partners sell technical integration. Warehouse scale is not allocation.
- Ask what share of the quote is actual GPU compute versus general ICT. A large distributor can have a thin GPU mix.
Credit-line, revenue, gross-margin, and quota-share figures that appear in channel commentary are not frozen on this page. If a seller quotes those numbers, mark them pending confirmation and keep them off the purchase order.
ODM vs OEM vs MGX, and Is Supermicro AMD-Only?
Quick Answer: Board-build permission, system-brand assembly, and a modular reference design are three different permissions. Supermicro belongs in the NVIDIA system-builder and modular-design set. It is not an AMD-only house, and that fact does not place it outside the design program. Source notes name seven board builders and more than ten design-certified names. Both lists await locator confirmation.
ODM authorization is an answer to “who may manufacture.” OEM, in this channel, is an answer to “who may ship a branded system that uses NVIDIA parts and designs.” MGX is an answer to “whose chassis follows the modular reference.” None of the three answers “who has this quarter’s GPUs,” and none of them answers “whose end customer cleared KYC.”
Supermicro is an NVIDIA OEM and an MGX-class system builder. It is not an AMD-only company, and it is not evidence that a builder sits outside NVIDIA’s design program. A source error that calls Supermicro AMD-only should be corrected before it reaches a purchase order. AMD platforms can exist in the same vendor’s catalog. That catalog fact does not erase NVIDIA OEM or MGX-class status.
The source lists seven authorized ODM names. Do not add an eighth. Treat the list as pending confirmation:
- Hon Hai / Foxconn
- Quanta
- Inventec
- Wistron
- Foxconn Industrial Internet
- Wiwynn
- Pegatron
MGX and OEM examples named alongside that claim include Supermicro, ASUS, Dell, HPE, Lenovo, and Gigabyte. The source says the certified set is 10 or more. The names above are the ones written down here. They are not a closed roster, and they are pending confirmation. Missing names are not a finding that those vendors failed a gate. They are simply not in this overview.
On an RFQ, write three columns even when one vendor appears in more than one. Column one is manufacturing authorization. Column two is MGX design certification for the exact SKU. Column three is allocation for the quarter. A yes in column two with a blank in column three is a design, not a delivery date.
How Do Allocation and KYC Split Quantity from Destination?
Quick Answer: Getting cards and being allowed to ship them are separate decisions. The first decision is who receives scarce parts in a tight cycle. The second decision reviews parent ownership, a real installation site, and declared end use. System sellers of B300-class machines still face the second decision. Design approval is not a waiver.
Allocation decides whether you can get scarce GPUs. KYC decides whether you can sell or ship them. The screen has three checks: ownership through to the ultimate parent, a physical site where the cards are real, and a stated end use. That third check is the anti-diversion screen. Reported diversion cases are exactly what KYC is meant to stop. This page does not describe routes, hubs, or ways around the screen.
An OEM that sells B300-class systems still faces KYC. MGX only means a compliant design may be built. It does not mean the builder has quota, and it does not mean the purchase order skips the destination screen. In tight cycles, source narrative says smaller neo-clouds feel the shortage first. That is a sequencing story, not a promise that a hyperscaler name on a slide equals your delivery.
Illustrative allocation examples, not an exhaustive list and pending confirmation, include Microsoft, Oracle, and UAE G42 as a sovereign-AI style buyer. Do not turn those three names into a closed “five or six” allow-list. If a seller quotes a short priority list, ask which quarter it covers and whether your legal entity is on it.
- Quantity question: is allocation confirmed for this SKU and this quarter?
- Destination question: does KYC clear parent, site, and end use for this purchase order?
- Design question, still separate: is the system an MGX-certified design?
- Type question, still separate: is the seller an NCP in this region, or only an NPN grade?
What Did the Reported 2026 BIS Rule and July Allow-List Change?
Quick Answer: Label every date below as press-reported, not as an NVIDIA bulletin. Rules from 7 October 2022 capped China-sale performance. A reported 31 May 2026 test looks at ultimate parent location. A reported 14 July 2026 allow-list removed more than half of Asian buyers. On-site reviews were then reported in Singapore, Malaysia, and Japan.
Do not present the timeline below as an official NVIDIA bulletin. No primary NVIDIA notice is cited for these rows. Frame them as a press-reported compliance response, pending confirmation. The phrase “more than half of Asian customers” is a news narrative. It is not a measured SKU count, and it is not a TAOHAN census of the channel.
| When (as reported) | What was reported | Gate it tightens |
|---|---|---|
| From 7 October 2022 | Successive BIS advanced-computing controls. China-sale performance caps moved from A100-class parts toward A800 and H800 class, then toward further China-specific SKUs | Performance ceiling at the source |
| November 2023 | Source notes: the compliant H100, H200, and B-series order path through China qualifications is largely closed. Badges remain for non-restricted SKUs and relationship proof | Allocation and destination at the China channel edge |
| 31 May 2026 | BIS rule, as reported: a license looks at ultimate parent location. Worldwide subsidiaries of China- or Macau-headquartered parents need a separate license path | KYC and destination, upstream of the reseller |
| 14 July 2026 | NVIDIA buyer allow-list, as reported. Financial Times first, Reuters follow. More than 50 percent of Asian customers removed. Neo-clouds described as hardest hit | Destination, at the channel edge |
| Mid and late July 2026 | On-site checks reported in Singapore, Malaysia, and Japan, covering contracts and end users | Destination, at execution |
The practical RFQ consequence is narrow. If the ultimate parent sits in a jurisdiction the reported BIS rule treats as controlled, mark the export license as pending confirmation. Do not assume an NPN grade, an MGX certificate, or an NCP logo waives that license. Do not add operational color about how cards have moved. One compliance sentence is enough: the screen exists so the stated end user is the real end user.
Figures you can quote without pretending they are a directory:
- 6 GPU channel gates on one purchase, as framed on this page.
- 7 ODM names in the source list, pending confirmation, with no eighth name added.
- 10 or more MGX or OEM certifications claimed by the source, pending confirmation.
- 3 NPN grades: Registered, Preferred, Elite.
- More than 50 percent of Asian customers removed, as reported on 14 July 2026, not a SKU census.
How to Lock an NVIDIA GPU RFQ Against Six Channel Gates
Quick Answer: Run six checks before the purchase order. Record the buying company and its region. Put system build, cloud capacity, and resale on separate rows. Ask design certification and quarterly supply as two questions. File parent, site, and end-use papers. If export licensing may apply, mark it unconfirmed. Price any backup path on its own row.
The steps below are a buying checklist. They are not a method for avoiding a license, washing origin, or routing around KYC. If a seller offers a workaround, stop and restate the legal buyer.
- Step 1. Name the legal buyer entity and region. NCP status is the entity plus region, not the logo.
- Step 2. Split the commercial line. Separate an ODM or OEM system, cloud NCP capacity, and NPN resale. If the quote is a China-market order path, name the layer: national distributor, NPN Elite, NPN Preferred, or NCP. Do not mix those paths on one row, and do not treat the layer as a restricted-SKU license.
- Step 3. Ask whether the SKU is an MGX-certified design and, separately, whether allocation is confirmed for that quarter.
- Step 4. Attach a KYC pack covering ultimate parent, site, and end use. Do not promise third-country delivery as a workaround.
- Step 5. If the parent is in a BIS-controlled jurisdiction, mark the license as pending confirmation. Do not assume a partner tier waives it.
- Step 6. RFQ any alternate path, such as a different OEM, cloud, or quarter, as a separate line.
A clean brief also states the GPU family in plain language, for example B300 or H200, the support term, and the site country that will appear on the KYC pack. It does not hide the parent behind a trading name. If SKU Lookup does not carry the NVIDIA system bill of materials, send the part list with the brief rather than inventing a catalog hit.
Fit. You are comparing a system OEM, a cloud capacity offer, and a reseller quote, and you need to know which gate each quote actually clears.
Unfit. You want a ranked partner directory, a legal sign-off, or a playbook for moving restricted GPUs. Use the official locator and counsel for those jobs. This article will not supply them.
Real-World Deployment Scenarios
Quick Answer: Map the four scenarios below before freezing CapEx. Title-bound lots need serial-to-PO papers. Futures slots need a named entitlement, not a screenshot. Audit packs need invoice chain and serials. A China-path badge needs a layer name and does not reopen H100, H200, or B-series after November 2023 controls.
Scenario 1: Cross-Border Parallel Import Lot (Title Bound)
Workload Profile: Buyer receives HGX nodes via free-port logistics with multi-hop packing lists.
Architectural Fit: Demand ODM/NCP/KYC evidence, serial-to-PO mapping, and country-of-origin paperwork before wire. Missing title documents = diversion risk, not a paperwork nicety.
Scenario 2: Futures Allocation Transfer (Contract Bound)
Workload Profile: Allocation slots sold mid-queue with deposit already paid.
Architectural Fit: Verify the named entitlement on NVIDIA/ODM systems matches the payee entity—allocation screenshots alone are not title.
Scenario 3: Enterprise Compliance Audit (Governance Bound)
Workload Profile: Internal audit asks how GPU assets entered the estate.
Architectural Fit: Keep a dated evidence pack (invoice chain, SN list, support entitlement) so finance and security share one narrative.
Scenario 4: China-Market Partner Quote (Qualification Bound)
Workload Profile: Seller cites a national distributor, NPN Elite, Preferred, or NCP badge on a China-path quote.
Architectural Fit: Write the layer on the RFQ row. After November 2023 controls, treat H100, H200, and B-series as license-pending. Confirm the locator for the legal entity. Do not read the badge as allocation.
Frequently Asked Questions
Quick Answer: Buyers usually ask seven things. Cloud type is not a network grade. Design certification does not reserve supply. Supermicro is an NVIDIA system builder, not AMD-only. China order paths sit in four layers, and NCP names still conflict across sources. The destination screen covers parent, site, and use. The July allow-list remains a press claim.
NCP is a cloud partner type that sells GPU compute as a service. NPN is a membership tier: Registered, Preferred, or Elite. Preferred is not NCP, and Elite is not NCP. The same brand string can be an NCP in one region and only a reseller of a local NCP in another, so identity is the legal entity plus region. Source material describes a short published NCP list; the count is pending confirmation.
No. MGX certification means a builder may ship a system on NVIDIA's modular reference design. Allocation is a separate gate that decides who receives scarce Blackwell-class GPUs in a tight quarter. Ask both questions on the RFQ. A certified design does not skip KYC.
Supermicro is an NVIDIA OEM and MGX-class system builder, not an AMD-only company, and not proof that it sits outside NVIDIA's design program. Source examples place it with other MGX or OEM names, pending confirmation on the official partner locator. That status still does not grant allocation.
Sources disagree on China NCP names. One lists Alibaba Cloud only; a later map lists Xiechuang Data and Runze Technology. L&T is NPN Preferred, not NCP. Digital China is a national distributor, not dual Elite. Hongxin is the dual Elite Compute plus Networking example. Confirm the official locator. This page is not a partner directory.
Source notes describe four China order-path layers: Unisplendour and Digital China as national distributors; Hongxin, Advanced Datacomm, and China Electronics Port as NPN Elite examples; L&T as Preferred; Xiechuang Data and Runze Technology as NCP examples. An earlier source named Alibaba Cloud as the only China NCP. Confirm the locator. After November 2023 controls, none of these badges is a compliant H100, H200, or B-series license. This page is not a directory.
KYC checks ownership through to the ultimate parent, whether the physical site and the cards are real, and the stated end use. Allocation decides quantity. KYC decides destination. An OEM selling B300-class systems still faces that screen. Do not offer a third-country delivery path as a workaround.
Press reports, pending confirmation, say a buyer allow-list dated 14 July 2026 removed more than half of Asian customers, with smaller neo-clouds hit first. That percentage is a news narrative, not a measured SKU count, and it is not an official NVIDIA bulletin on this page. Confirm the live allow-list on the RFQ.
References
Quick Answer: Start with NVIDIA's public partner overview, the MGX product page, and the partner locator before you freeze any name. Use the Bureau of Industry and Security rules index for the export frame, and the data-center platform page for product context. Timeline rows on this page stay labeled pending confirmation until a primary notice is in hand.
Key Takeaway
Key Takeaway: Do not freeze a GPU purchase order until six answers are written down: builder permission, design certification, quarterly supply, cloud type versus network grade, and a destination screen on parent, site, and end use. TAOHAN supplies NVIDIA GPU systems with competitive pricing after that screen.