| How we know | What it shows | What it does not establish |
|---|---|---|
| Disclosed | The company’s own filing, release, product page or investor record describes an AI-server or AI-rack role | Customers it does not name, or AI revenue it does not report |
| Partner-disclosed | A named partner or platform owner’s document names the company, such as NVIDIA’s co-design credit to Lite-On | Scale beyond that document’s product or program |
| Positioning | A dedicated AI-server product or reference design with no named customer | That shipping racks use the part |
| Teardown-verified | An independent teardown names the part in a unit it opened | Volumes, pricing, exclusivity or future sourcing |
| Analyst-attributed | A named analyst’s inference, such as SemiAnalysis on the NVLink backplane | Anything on its own; it cannot set an evidence level |
| Reported, unconfirmed | Press citing unnamed sources, or paywalled research seen only in summaries | Anything; it never sets an evidence level or a part link |
Start with four roles
An AI rack involves platform owners that design the chips and the reference system, builders that manufacture and sell complete racks, component makers that supply parts, and upstream enablers that supply materials or foundry capacity. NVIDIA and AMD are platform owners; Hon Hai, Quanta, Wistron, Wiwynn, Dell and Supermicro are builders; Luxshare, Delta and AVC make components; Ajinomoto’s build-up film is an upstream input.
stock3d labels each connection by specificity: AI-rack evidence, Data-center products, Adjacent technology and Upstream enabler. For a builder, AI-rack evidence means its own AI servers or racks, not a component role. Each listing also shows how the evidence is known, as in the table above. The labels describe specificity, not investment quality.
Read partner lists for what they are
NVIDIA’s October 2024 post on its GB200 NVL72 contribution to the Open Compute Project names more than 40 infrastructure providers building on Blackwell, from Amphenol, AVC and Chenbro to Luxshare, SUNON and TE Connectivity. It does not say which part each supplies, to which customer or in what volume. NVIDIA’s May and October 2025 posts do the same for a future 800 V DC architecture, and its March 2025 photonics release names module makers and ecosystem partners.
stock3d pairs each list with the company’s own disclosure and maps a part only where that disclosure describes the product. Some partner statements are narrower and stronger: NVIDIA says the GB300 NVL72 power supply’s energy storage was co-designed with Lite-On, which ties one company to one current design.
Reference: NVIDIA: GB200 NVL72 designs contributed to OCP (Oct 15, 2024).
Check who the customers are
Many suppliers sell most of their AI products to a few unnamed buyers. Astera Labs’ largest end customer was more than 70% of its 2025 revenue; Credo’s ten largest customers were about 90% of its fiscal 2026 revenue; Lumentum’s two largest end customers were 26.6% and 15.0% of FY2026 revenue; and SK hynix’s two largest were 14.8% and 12.4% of its Q1 2026 revenue. Fabrinet names one: NVIDIA was 16.3% of its FY2026 revenue, down from 27.6% a year earlier.
Builders report in different units. Dell gives AI-server orders, revenue and backlog, Hon Hai and Foxconn Industrial Internet give shipment growth rates, and Wistron gives an AI-infrastructure share of revenue. None breaks out NVL72-class racks by customer, so rack counts and market shares in the press are treated as reported and never combined.
Reference: Astera Labs: Form 10-K for 2025.
Map parents, subsidiaries and acquisitions
Several listed companies hold other listed companies in the same rack. Hon Hai holds 84.16% of Foxconn Industrial Internet on a reciprocal basis, Wistron holds 35.46% of Wiwynn, TSMC holds 34.84% of Global Unichip, and AVC consolidates Fositek with a 16.87% holding. Treating a parent and its listed subsidiary as two independent suppliers can count one exposure twice.
Other units sit inside larger groups or are private. SPIL belongs to ASE Technology, Kaori Thermal to Kaori Heat Treatment, Boyd Thermal to Eaton since March 2026, CoolIT to Ecolab since July 2026, and ZT Systems’ manufacturing business to Sanmina since October 2025. Molex, Stäubli, Rittal, Cooler Master and LiquidStack are not publicly listed, and stock3d never substitutes a listed owner, investor or partner for them.
Reference: Wiwynn: 2025 Annual Report.
Identify the listed line you would own
Check the issuer, exchange and ticker together. TEL is TE Connectivity on the NYSE, not Tokyo Electron; NVT is nVent, not NVIDIA; Delta Electronics in Taiwan is separate from the Thai-listed Delta Electronics (Thailand); and Samsung Electronics, Samsung Electro-Mechanics and Samsung SDI are three separately listed companies.
Many rack suppliers trade only in Taipei, Shanghai, Shenzhen, Tokyo or Seoul. stock3d treats a US line as verified only when it appears in Nasdaq’s symbol directories or as an active depositary receipt in Citi’s directory. SK hynix added Nasdaq ADSs (SKHY) on July 10, 2026, ten ADSs per share; ASE trades in New York as the ADR ASX; and Ibiden and Delta have OTC ADRs.
Reference: SK hynix: Nasdaq prospectus (Form 424B4).
Use the model as a research starting point
Choose a sector, inspect how its parts work, and open a company’s evidence record. Compare the product role, the evidence basis, the business mix and the caveat, and follow the source before moving into deeper financial research on Yahoo Finance or Fynca.
The rack is an interactive 3D model of an illustrative NVL72-class composite, not a bill of materials, and its company links are curated, not exhaustive. One illustrated part, the casters and leveling feet, has no verified listed supplier. That gap stays visible; adding loosely related names would not make a connection more accurate.