NVIDIA AI Factory Ecosystem
From TSMC, HBM and optical networking to power and cooling, cloud and software, out to the partners NVIDIA has invested in directly and named on the GTC stage: how a single AI factory threads the entire value chain
At GTC, NVIDIA used its DSX AI Factory reference design to make the point to the market: standing up an AI super-datacenter takes far more than buying GPUs. It also needs leading-edge process nodes and CoWoS packaging, HBM, networking and optical interconnect, server racks, digital-twin design, power and liquid cooling, the physical facility, cloud compute and software, with no link in the chain dispensable. NVIDIA itself holds the thickest-value 'compute + networking + CUDA software' full-stack core; moving outward along the chain, many second-tier leaders (cold plates, optical modules, power, ODMs, neoclouds) benefit heavily from a single link and carry even more upside leverage. This topic puts NVIDIA at the center and lays the chain out layer by layer: upstream silicon foundation, then core platform, interconnect, systems, design, physical infrastructure (power / cooling / facility), and cloud and software. It also flags the ecosystem partners NVIDIA has strategically invested in or holds (13F positions plus private placements) and named publicly at GTC. The framing strictly separates 'directly named supplier vs. general beneficiary,' 'closed stake vs. announced intent only,' and 'public vs. private.' It is a factual map and positioning read, with no return forecasts of any kind (YMYL).
Wafer Foundry & Advanced Packaging
The physical lifeline for the AI factory's compute silicon. NVIDIA runs no fabs of its own; every datacenter GPU lands on TSMC's leading-edge process plus CoWoS / SoIC advanced packaging. Advanced-packaging capacity is the one physical bottleneck on GPU shipments, and NVIDIA has locked up the bulk of TSMC's CoWoS capacity. Back-end assembly and test then runs through ASE, Amkor and peers.
Sole leading-edge wafer foundry plus CoWoS-S/L 2.5D and SoIC 3D advanced-packaging supplier for every one of NVIDIA's datacenter GPUs.
World's largest OSAT, handling TSMC's CoWoS back-end plus its own VIPack advanced packaging and system-level test; participates in NVIDIA accelerator packaging and test.
Largest US-based OSAT and TSMC's turnkey partner for advanced packaging (CoWoS / InFO) and test in the US (Arizona).
High-Bandwidth Memory (HBM)
The core source of the AI factory's memory bandwidth. Every training and inference GPU stacks multiple HBM dies, the highest value-density, highest-margin and tightest-supply category in memory today. HBM allocation for NVIDIA's flagship platforms is heavily concentrated among three suppliers and commands a premium by generation (HBM3E → HBM4).
NVIDIA's lead HBM supplier: front-runner on HBM3E (including GB300's 12-layer) and HBM4 (lead supplier for Rubin).
Second HBM source: qualified by NVIDIA onto the Rubin HBM4 supplier list, adding a second source.
Third HBM source and the only US-based HBM IDM: HBM3E for H200 / Blackwell Ultra, on the Rubin HBM4 supplier list.
NVIDIA Accelerated Computing Platform (Ecosystem Core)
The center of the whole topic and the source of demand. NVIDIA delivers GPU + CPU + DPU + networking + software as an integrated rack-scale system and defines the DSX AI Factory reference design. It is both the value anchor of the chain and the force pulling demand through every downstream link.
Networking & High-Speed Interconnect
The back-end network that wires tens of thousands of GPUs into one 'supercomputer.' NVIDIA uses its own NVLink (scale-up) plus Spectrum-X Ethernet / Quantum InfiniBand (scale-out), and opens it to third parties through NVLink Fusion; the Ethernet camp pushes open standards as an InfiniBand alternative. Players in this layer therefore tend to both partner and compete with NVIDIA.
Custom AI silicon (XPU) plus electro-optical interconnect: optical DSP / SerDes / datacenter interconnect plus Ethernet switching, partnered with NVIDIA NVLink Fusion.
Dual engine in AI networking: merchant Ethernet switch / router silicon (Tomahawk / Jericho) plus custom XPUs (AI accelerator ASICs) co-developed with hyperscale customers.
Leader in high-end Ethernet switching for the AI datacenter; the Etherlink platform wires thousands of GPUs into a lossless, low-latency back-end network.
Specialist in in-rack AI connectivity: Aries PCIe / CXL retimers + Scorpio fabric switching + smart cables, solving signal integrity across GPU-CPU-accelerator interconnect.
Leader in active electrical cables (AEC): low-cost copper for short-reach in-rack scale-up, pushing back the crossover point where copper gives way to optics.
Enterprise / neocloud AI networking: in-house Silicon One switching plus the N9100, the first partner switch built on NVIDIA's Spectrum-X silicon.
Leading white-box switch ODM, building 800G / 1.6T AI fabric switches for hyperscale customers and supporting GPU clusters at the ten-thousand-card scale.
Positioned in AI back-end / datacenter networking: closed the Juniper acquisition in 2025-07, folding in AI-native datacenter networking (Mist AI, high-speed Ethernet switching); its systems business also appears in the 'servers' layer.
Optical Modules & Interconnect (incl. CPO)
The 'nerves' of GPU clusters scaling out across nodes. The 800G → 1.6T high-speed optical modules and their upstream EML / silicon-photonics laser chips and optical engines are among the densest shipment links in NVIDIA's ecosystem; NVIDIA also outsources much of its in-house optical-module assembly to specialist contractors. CPO (co-packaged optics) enters commercial ramp from 2026 with Quantum-X / Spectrum-X Photonics, but pluggable 800G / 1.6T modules still dominate near-term.
World shipment leader in high-speed optical modules: first tier in 800G / 1.6T datacom modules, first to volume silicon photonics, lead supplier for long-reach optical interconnect across NVIDIA's ecosystem.
Leading high-end optical / optoelectronic contractor, doing precision optical and optoelectronic assembly and packaging for NVIDIA, Cisco, Lumentum and others; the core contractor for NVIDIA's in-house optical modules.
Leader in InP laser / EML chips and optical modules: self-supplied end to end from InP wafers and EML chips up to 800G / 1.6T modules.
First tier in high-speed optical modules, with 800G / 1.6T datacom modules locked to large overseas customers and 1.6T qualified.
Vertically integrated optical-component and module house: 1.6T transceivers shipping, with command of all three technology paths (EML, silicon photonics, VCSEL).
Optical-component / optical-engine platform: one-stop from passive to active optical components, early to market on 800G / 1.6T optical engines, sitting upstream of optical modules.
Leading domestic optical-communications-component vendor: vertically integrated from optical chips to components to modules, with in-house optical chips and volume-delivery capability on silicon-photonics 1.6T.
Leader in domestic high-speed laser chips: DFB / EML laser chips and high-power CW silicon-photonics light sources, at the very top of the optical-module chain.
Server Systems & Rack Integration (ODM / OEM)
The vendors that assemble NVIDIA GPU boards into DGX / HGX / MGX / GB200 NVL72 full racks. Board specs are locked by NVIDIA and OEMs cannot change them; the money is in system-integration and liquid-cooled full-rack manufacturing service fees, with margins diluted by the cost pass-through of the compute silicon. GB200 / GB300 full-rack orders are heavily concentrated among Taiwan ODMs. This layer lists only the representative vendors with larger full-rack / board share; NVIDIA's official Blackwell system partners also include ASRock Rack (under ASRock) and others not carded individually (ASUS, GIGABYTE and Pegatron are carded below).
World's leading AI server contractor and the #1 ODM by GB200 / GB300 full-rack shipments, first to complete full-rack volume delivery.
Foxconn group's A-share-listed core contract assembly platform for AI servers and GB200 / GB300 NVL72 full racks; largest-share vendor for NVL72 board assembly.
One of NVIDIA's main qualified ODMs and a core contractor for GB200 / GB300 NVL72 full racks.
Leading US OEM, among the first to ship GB200 / GB300 rack-scale solutions (PowerEdge XE9712 / XE9780).
One of NVIDIA's qualified ODMs, a GB200 / GB300 NVL72 full-rack contractor, and the parent of Wiwynn.
Wistron's ODM arm focused on white-box AI servers and racks for hyperscalers, supplying North American cloud giants directly.
US enterprise OEM offering AI servers and liquid-cooled full racks built on NVIDIA HGX / GB200 (including the Cray supercomputing line).
Hong Kong-listed global PC / server OEM offering NVIDIA-platform AI servers and Neptune warm-water direct-liquid-cooled racks.
US OEM / ODM known for modular racks and liquid cooling, offering GB200 / GB300 NVL72 and HGX B300 full-rack solutions.
Veteran server ODM, a second-tier player in NVIDIA AI server / rack contracting.
Official NVIDIA Blackwell system partner offering AI servers and full racks built on HGX / MGX / GB200 NVL72.
Official NVIDIA Blackwell system partner and a leading HGX / GB200 NVL72 AI server vendor (the GIGAPOD full-rack solution).
Official NVIDIA Blackwell system partner offering AI server / board and full-rack contracting.
AI Factory Design & Digital Twin (EDA + Omniverse DSX Blueprint)
NVIDIA moves the 'design' of the AI factory into Omniverse as a digital twin: the DSX blueprint announced at GTC in 2026-03 uses a digital twin to optimize power, cooling, layout and capacity before ground is broken. This layer covers the design / simulation partners named publicly at GTC; among them, the two EDA leaders also design NVIDIA's own silicon. Note: Schneider (ETAP power-distribution simulation), Vertiv, Eaton and Trane are also DSX SimReady asset partners, carded in the 'power' and 'cooling' layers; DSX blueprint and design-build partners further include EPC engineering giant Bechtel (private), Schneider's ETAP power-distribution simulation, AI building-optimization firm Phaidra (private), datacenter operator Switch (private), sovereign cloud Nscale (in the 'cloud' layer) and others not carded individually (general contractor Jacobs is carded below).
One of the top three EDA houses, in a dual role: EDA tools that design and verify NVIDIA's own silicon, plus flagship Omniverse DSX digital-twin partner (datacenter-scale thermal / fluid simulation).
One of the two EDA leaders, both an EDA / IP tools supplier that designs NVIDIA silicon and a strategic NVIDIA holding; has closed the Ansys acquisition, adding multiphysics simulation.
Omniverse DSX blueprint design partner: Xcelerator / Simcenter connects to Omniverse for the AI factory's power-cooling-automation framework and SimReady cooling-system assets.
Omniverse DSX blueprint design partner: connects the DSX reference design into the CATIA / 3DEXPERIENCE MBSE platform to build a 'virtual twin' of the AI factory and speed first production.
Omniverse DSX blueprint design partner: connects the blueprint into Windchill PLM, linking engineering / product-design data with high-fidelity real-time simulation.
Omniverse DSX blueprint design partner: integrates the Omniverse library with the DSX blueprint as the 'digital thread' running through the AI factory's entire build lifecycle.
Omniverse DSX blueprint engineering partner: EPC and facility design for AI factories / datacenters, taking the DSX digital-twin blueprint into physical construction.
Power Delivery & Electrical Systems
As AI factory power density jumps, power delivery runs end to end from 'chip-level power' out to 'rack / facility / grid.' NVIDIA has publicly launched an 800V HVDC (high-voltage DC) architecture alliance, naming three classes of direct partners: electrical systems, power semiconductors and power components. The companies below are representative listed names; the official 800V HVDC partner list also includes ADI, STMicro, Onsemi, Renesas and others not carded individually (ABB and Hitachi, via subsidiary Hitachi Energy, are carded below). Generation / backup / grid-tie links (GE Vernova gas turbines, Caterpillar standby generation, Mitsubishi Electric heavy-electric distribution) sit as general beneficiaries, and municipal / utility power companies also see pull from AI factory load.
Direct NVIDIA partner in datacenter power plus thermal-management systems, co-developing the GB200 NVL72 full-rack power-and-cooling reference architecture with NVIDIA.
Direct NVIDIA partner on 800V HVDC datacenter electrical systems, with a matching 800VDC reference architecture; acquired Boyd Thermal in 2026 to add liquid cooling.
Direct NVIDIA partner on 800V HVDC electrical systems, co-developing gigawatt-scale AI factory power / liquid-cooling reference designs and certified CDUs; ETAP is also a DSX power-distribution simulation partner.
Direct power-semiconductor partner in NVIDIA's 800V HVDC and MGX AI factory ecosystem, with share in GB200 board-level power (PMIC / MOSFET).
Direct NVIDIA partner on 800V HVDC power components and a lead supplier of AI server power (PSU / power shelf), once the sole supplier of the GB200 integrated busbar controller.
Power-silicon vendor supplying GPU board-level power (VRM / PMIC) to NVIDIA directly, on the 800V HVDC silicon supply alliance (listed as MPS).
GaN / SiC power-semiconductor vendor named by NVIDIA to co-develop the 800V HVDC architecture, supplying the Kyber rack-scale system (powering Rubin Ultra and others).
Direct NVIDIA partner on 800V HVDC power components (the only A-share name on the list), deeply involved in Blackwell-architecture datacenter power hardware design.
Gas-turbine and grid-equipment supplier providing upstream generation and grid-tie for AI datacenters (not an NVIDIA-named rack-power partner).
Power / analog semiconductor partner named officially in NVIDIA's 800V HVDC datacenter power architecture, supplying high-voltage DC-DC, gate drivers, power management and isolation devices.
Power-semiconductor partner named officially in NVIDIA's 800V HVDC architecture, primarily supplying SiC / Si power devices and gate drivers.
Power partner named officially in NVIDIA's 800V HVDC architecture, supplying AI server power (PSU), power shelves and HVDC power modules.
High-density power-module vendor, focused on GPU board-level vertical / lateral power delivery (VPD / LPD); technically aligned with AI compute power delivery but not in NVIDIA's official 800V alliance.
Electrical-systems partner named in NVIDIA's 800V HVDC ecosystem, supplying datacenter medium-voltage distribution, transformers and electrification, covering AI factory power distribution from grid-tie to rack.
Supplier of datacenter standby generator sets (diesel / gas) and on-site power systems, providing backup for AI factories during grid outages.
Partner named in NVIDIA's 800V HVDC ecosystem (via subsidiary Hitachi Energy), supplying HVDC transmission, transformers and grid-connection equipment.
Heavy-electric / distribution partner named in NVIDIA's ecosystem, supplying datacenter UPS, distribution systems, transformers and power devices.
Cooling & Liquid Cooling
The GB200 NVL72 is a liquid-cooled full rack, with cooling shifting wholesale from air to direct liquid cooling (DLC). Per-rack cooling value scales up with power density, making this a textbook link where 'second-tier leaders carry more upside.' Cold plates, CDUs and quick-disconnects each have their own specialist leaders; several have cleared NVIDIA Recommended Vendor List (RVL) qualification, building a switching barrier. Note: Vertiv's CDU and liquid cooling sit in the 'power' layer; Boyd has folded into Eaton and CoolIT has been acquired by Ecolab.
Lead supplier of GB200 / GB300 cold plates, holding NVIDIA Recommended Vendor List (RVL) status.
Leader in datacenter liquid-cooling systems (CDU / XDU coolant distribution units), co-defining the full-rack cooling reference architecture with NVIDIA (also in the 'power' layer).
Liquid-cooling products qualified for NVIDIA Recommended Vendor List (RVL) status, in the GB200 / GB300 cooling supply chain.
Built-to-spec liquid-cooling solution provider working with NVIDIA, supplying row / rack-level CDUs and coolant distribution manifolds.
Leader in domestic datacenter thermal management, offering a full liquid-cooling chain of cold plates + CDUs + quick-disconnects, with supply into NVIDIA-ecosystem datacenters.
Supplier of Airedale by Modine-branded CDUs (400kW–2MW), serving high-density AI compute liquid cooling and hybrid cooling.
Rooted in power-electronics thermal management, extended into AI datacenter liquid cooling; reported domestically as a GB300 liquid-cooling module supplier.
Omniverse DSX blueprint design partner, using the blueprint to optimize thermal management for gigawatt-scale AI factories and providing SimReady cooling-system assets.
One of the lead suppliers of NVIDIA GB200 / GB300 liquid-cooling systems (cold plates / water-cooling modules), a cold-plate source alongside Auras.
Datacenters & Facilities (Colocation / Compute Real Estate)
The physical landing site for the AI factory. Interconnection-oriented, large-scale datacenter REITs supply rack space, power and interconnect; NVIDIA takes its reference architecture into these campuses through DGX-Ready colocation certification, and the DSX digital-twin blueprint is deployed live in flagship campuses among them. This layer ranks by colocation / compute-real-estate carrying capacity and depth of NVIDIA partnership, not NVIDIA equity stakes or global IDC market share. NVIDIA's DSX-ecosystem datacenter / colocation partners also include the listed NextDC (Australia, NXT), NTT (Japan, 9432), Sify (India, SIFY), and the private AirTrunk (under Blackstone), DayOne, Digital Edge, Princeton Digital Group, STT GDC India and other regional providers; this layer lists only representative interconnection-oriented large-scale REITs, not carded individually.
Large global datacenter REIT and NVIDIA DGX-Ready provider, with campuses hosting NVIDIA's AI Factory Research Center plus the Omniverse DSX blueprint.
Global interconnection-oriented datacenter REIT and NVIDIA DGX-Ready colocation partner, deploying DGX + liquid cooling across IBX campuses.
Leading third-party datacenter operator in China (the 'China Equinix'), with AI compute colocation benefiting from domestic AI demand and NVIDIA GPU sales into China.
Neutral third-party IDC in China, with AI datacenter colocation benefiting from domestic AI demand and NVIDIA GPU sales into China.
Cloud Services & AI Factory Operators (Hyperscaler + Neocloud)
The operators that actually 'run' the AI factory, renting out compute by usage or long-term contract. Hyperscalers are NVIDIA's priority-allocation customers for flagship GPUs (and mostly develop their own silicon in parallel); neoclouds (pure-GPU clouds) are a new species deeply bound to NVIDIA's capital and capacity, with CoreWeave and Nebius serving as both operators and NVIDIA holdings. The core/key ranking in this layer reflects exposure to NVIDIA GPUs / AI factories and capital ties, not global public-cloud market share (so Oracle / CoreWeave leading does not mean their cloud share exceeds AWS / GCP). DSX-ecosystem AI cloud partners also include the listed Indosat (Indonesia, ISAT), GMO Internet (Japan, 9449), NAVER Cloud (Korea, parent 035420), YTL (Malaysia, YTLPOWR) and other regional telco / conglomerate clouds, plus the private firmus, Sharon AI, GMI Cloud, ResetData, Yotta and other regional neoclouds whose NVIDIA exposure is diluted by core business or not directly investable; this layer lists only the representatives with stronger exposure / capital ties, not carded individually.
Hyperscaler (Azure), a Blackwell launch-tier large customer for NVIDIA, deploying GB200 / GB300 NVL72 clusters (primarily for OpenAI workloads).
OCI placed a very large purchase of NVIDIA superchips, buying roughly 400,000 GB200 for the Stargate (Abilene) project and renting compute to OpenAI.
Largest pure-GPU cloud (neocloud), an NVIDIA holding plus anchor customer, the 'at scale' launch platform for new cards.
Hyperscaler (AWS), offering NVIDIA GB200 NVL72 instances plus the H100 / P5 series.
Hyperscaler (Google Cloud), the first to offer both A4 (B200) and A4X (GB200 NVL72) instances.
AI infrastructure neocloud, with a direct NVIDIA stake and priority GPU allocation.
A Bitcoin miner turned AI cloud: contracted to supply NVIDIA GB300 compute to Microsoft (about $9.7 billion over 5 years) and reached a 5GW DSX-aligned AI infrastructure partnership with NVIDIA in 2026-05.
AI datacenter real-estate developer, leasing roughly 400MW long-term (North Dakota) to NVIDIA-held CoreWeave.
Neocloud / AI compute infrastructure, an NVIDIA investment, involved in building Stargate Abilene.
GPU cloud (neocloud), an NVIDIA investment.
European sovereign AI datacenter neocloud, an NVIDIA investment and a DSX blueprint deployment partner.
Software & Developer Ecosystem (NVIDIA Platform Partners)
NVIDIA's own software (CUDA / AI Enterprise / NIM / Omniverse) is the core moat (see the core layer). This layer covers listed software companies in deep partnership on NVIDIA's platform or announced co-development at GTC, embedding NVIDIA accelerated computing, Nemotron models and the Agent Toolkit into their own platforms to ship enterprise AI agents. The focus here is listed enterprise AI application software; the lower-level infrastructure software named in NVIDIA's DSX 'AI factory software' dimension (high-performance storage VAST Data / WEKA / DDN, cluster and container orchestration Mirantis / Rafay / Spectro Cloud / Red Hat (part of IBM), AI building-optimization Phaidra and others) is mostly private unicorns or already acquired, with no standalone investable name, so it is not listed separately.
Builds autonomous AI agents on NVIDIA's platform and co-developed the open-source reasoning model Apriel-Nemotron with NVIDIA.
Co-building an 'operations AI' stack with NVIDIA, integrating NVIDIA accelerated computing / Nemotron / cuOpt into the Ontology.
GTC public: released a Secure-by-Design AI blueprint, embedding Falcon protection into NVIDIA's AI agent architecture.
GTC public: named an enterprise-software partner in NVIDIA's Agent Toolkit, advancing enterprise AI agents.
GTC public: partnering with NVIDIA to deliver a full-stack AI platform, with Cortex connecting to NVIDIA models / NeMo Retriever.
NVIDIA Strategic Investments · Listed Holdings & Stakes
NVIDIA does more than sell silicon; it also binds the ecosystem with capital. As of 2026Q1, public 13F holdings stood at about $18.4 billion across 7 names: Intel (about 51.6%, the largest), CoreWeave (about 19.9%), Synopsys (about 10.4%), Coherent (about 10.1%, new), Nokia (about 7.3%), Nebius (about 0.7%) and Generate Biomedicines (about 0.1%, new), of which CoreWeave / Synopsys / Coherent / Nebius are carded in their respective links above. There are also strategic arrangements announced publicly in 2026 but not yet in the 13F: Marvell (closed an about $2 billion Series A preferred private placement on 2026-03-31), Coherent (closed an about $2 billion common-stock private placement on 2026-03-02, now new in the 13F), Lumentum (about $2 billion investment intent), Corning (about $500 million warrant subscription plus rights to exercise up to 15 million shares, not common stock) and IREN (warrants for up to about $2.1 billion), with framing detailed in each link. Exited prior holdings (Arm, Recursion, WeRide, Applied Digital, SoundHound and others) are no longer held. The names below are pure holdings / stakes with no other link to belong to.
NVIDIA's largest single listed holding; the strategic partnership lets Intel x86 CPU / foundry work alongside NVIDIA's datacenter and PC products.
Telecom / networking equipment; NVIDIA's strategic stake is an entry into AI-RAN and datacenter networking.
Optical fiber / specialty glass; NVIDIA uses a warrant structure to strategically bind optical / fiber manufacturing capacity.
AI protein / drug-design biotech; a new small holding extending NVIDIA's AI thesis into scientific discovery.
NVIDIA Strategic Investments · Private Ecosystem (Models · Robotics · Compute)
A bigger slice of NVIDIA's capital map sits in private companies: frontier large models, humanoid robotics / embodied intelligence, and compute infrastructure. Retail investors cannot buy these directly, so this page is a display map only (OpenAI / Anthropic have on-site reports to click into). The compute names Crusoe / Lambda / Nscale are listed in the 'cloud' layer; there are also Cohere, Mistral, Perplexity, Scale AI, Reflection AI, Together AI and other model / data / inference platforms NVIDIA has backed across rounds. The framing strictly separates 'closed participation vs. announced ceiling only.'
One of NVIDIA's largest single AI strategic investments and one of its largest GPU customers.
Maker of the Claude models; NVIDIA strategic investment plus compute partnership (committed to purchase up to 1GW of Grace Blackwell / Vera Rubin).
Maker of the Grok models / Musk's AI venture; NVIDIA has joined its funding rounds.
Humanoid robotics / embodied intelligence; an NVIDIA strategic investment (physical AI positioning).
Europe's leading open-source large-model maker; NVIDIA has invested across rounds.
UK end-to-end self-learning autonomous driving; an NVIDIA strategic investment (physical AI / automotive).