NVIDIA GB200 Supply Chain: The Global Ecosystem Explained

The NVIDIA GB200 (Grace/Blackwell) high-end AI accelerator represents one of the most complex and demanding hardware products ever assembled. Delivering terascale-and-beyond compute for next-generation AI workloads, a single GB200-based system (e.g. the GB200 NVL72 rack) integrates tens of GPUs, dozens of CPUs, and specialized interconnect, all with power draw measured in hundreds of kilowatts. As a result, building a GB200 system requires an extremely large, global supply chain encompassing dozens to hundreds of distinct suppliers. These include major silicon foundries, memory makers, advanced packaging houses, substrate fabricators, connector and cable manufacturers, board assemblers, coolant and power-electronics vendors, and numerous other component suppliers. Our analysis shows that over 50 unique subcomponent categories must be sourced for a GB200 rack (^[1]), with multiple potential vendors in each category. Major players span from TSMC (for GPU fabrication and CoWoS packaging) and SK Hynix/Micron (for HBM3e memory) to companies like Amphenol (for NVLink connectors), Foxconn (as a system integrator), and specialized OSAT firms (Amkor, ASE, Powertech, etc.). Moreover, recent reports highlight that even Tier-1 materials (e.g. substrates and fan-out laminates) are under strain, forcing Nvidia and its partners to adopt new packaging standards (such as panel-level fan-out) and to develop U.S. packaging capacity (^[2]) (^[3]).

This report provides a comprehensive examination of the GB200 supply chain, covering the chip architecture, fabrication and packaging process, memory and substrate suppliers, board and system integration, cooling and power subsystems, networking and interconnect components, OEM and hyperscale integration, and prevailing bottlenecks. We draw on industry analyses, technical disclosures, and market reports to quantity and name many of the companies involved. For example, a detailed SemiAnalysis breakdown explicitly identifies “over 50 different subcomponents of the GB200 rack” with distinct supply channels (^[1]). We also analyze expert cost models (e.g. Epoch AI’s BOM model) and official datasheets to illustrate how major cost centers (such as HBM memory and packaging) dominate the bill of materials (^[4]) (^[5]). Real-world case studies of OEMs and cloud providers (e.g. Foxconn’s server division, ZT Systems, AWS UltraServers) highlight the roles of system integrators and hyperscale customers in sourcing and assembling GB200 systems (^[6]) (^[7]) (^[8]). Throughout, we cite numerous sources on supply-chain constraints (digitimes, Wccftech, TheRegister, etc.) to show how CoWoS packaging and HBM supply remain limiting factors, and how the industry is responding.

In sum, the GB200 supply chain is vast and multinational. Beyond NVIDIA’s own IP and TSMC’s fabs, it includes Taiwanese substrate firms (Ibiden, Unimicron, Nan Ya PCB), U.S. advanced-packaging partners (Amkor), memory suppliers (SK Hynix, Micron), global electronic component companies (Murata, Texas Instruments, Infineon, etc.), and many ODM/OEMs (Foxconn/Hon Hai, Quanta, Supermicro, Dell, HPE, etc.). Each rack-scale system effectively constitutes an “AI supercomputer” whose construction depends on coordination among hundreds of vendors. We systematically document these layers below, providing tables of key components vs suppliers, data on production volumes and bottlenecks, and expert commentary from industry sources. The result is an in-depth report on the massive ecosystem of suppliers and vendors behind NVIDIA’s GB200.

The rapid rise of large-scale AI has intensified demand for ever-more-powerful compute engines. NVIDIA, as a market leader in AI accelerators, has continually pushed GPU performance with successive architectures (e.g. Volta, Pascal, Ampere, Hopper) and now Blackwell (the architecture behind “Grace Blackwell” GB200). The GB200 design couples NVIDIA’s new Blackwell Tensor Core GPUs (for AI compute) with the Grace CPU into an integrated “superchip” for data center deployment (^[9]) (^[10]). In a first demonstration, NVIDIA announced the GB200 NVL72 rack in early 2024, an entire cabinet containing 36 GB200 compute cards (totaling 72 GPUs and 36 CPUs) interconnected by NVLink into a single massive system (^[11]) (^[9]). The NVL72 achieves unprecedented scale: 720 PFLOPS of FP8/FP6 Tensor Core performance in sparse mode (360 PFLOPS dense) and over 13 TB of HBM3e memory in one liquid-cooled rack (^[12]) (^[10]). Such scale comes at a cost: NVIDIA's device specification lists a maximum load of 1.2 kW per GB200 GPU, and the full rack draws on the order of 100–150 kW of power (^[6]). Building these systems demands not only large amounts of semiconductor silicon and memory, but also specialized cooling, power delivery, and interconnect infrastructure on an industrial scale.

Silicon and Architecture. The core of the GB200 is a multi-chip package containing two Blackwell GPUs and one Grace CPU linked by 900 GB/s NVLink C2C interconnect. This “Grace Blackwell Superchip” package (called the GB200 module) is in turn assembled onto a server board. Each compute tray in an NVL72 rack holds two of these modules (2 Grace CPUs + 4 GPUs). In total, the NVL72 rack houses 36 such modules (72 GPUs and 36 CPUs) in 18 1U compute trays, connected through NVLink switch trays to form one GPU domain. NVIDIA specifies 372 GB of HBM3E per Grace Blackwell Superchip and 13.4 TB per NVL72; its current product page does not give a separate per-GPU memory allocation.

Supply Chain Complexity. The GB200 is effectively not a single processor but a rack-scale supercomputer in a box. Its supply chain encompasses every layer of the electronics ecosystem from raw silicon to housings. On one end, NVIDIA provides the chip design and some proprietary ASICs (the GPUs and NVSwitch chips), but outsources wafer fabrication to TSMC. On the other, hyperscale server OEMs (Foxconn, Supermicro, Dell, etc.) assemble the boards, power systems, and racks for end customers. In between are dozens of specialty vendors: memory manufacturers (SK Hynix, Micron), packaging houses (Amkor, ASE, Powertech), substrate companies (Unimicron, Ibiden, Nanya, etc.), board fabricators, passive component manufacturers (Murata, TDK, Yageo, etc.), connector and cable makers (Amphenol, TE Connectivity, US Conec, etc.), cooling companies (CoolIT, Enago, etc.), and many more. Stakeholders range from multinational chip foundries to local laminated-glass substrate suppliers. Summing all tiers, the number of distinct suppliers is easily in the hundreds. One industry analysis explicitly notes that assembling a GB200 rack involves “over 50 different subcomponents” with multiple vendors for each (^[1]).

Historical Context. To appreciate the scale, it is instructive to consider GPU evolution. Early NVIDIA GPUs in the GTX/Tesla series had a single monolithic silicon block with off-chip GDDR/small HBM stacks. By the time of Ampere/Hopper (2020–2022), NVIDIA was already using 7nm GPUs with HBM2e memory (e.g. H100 with 80 GB of HBM). Each new generation added more memory (for example, H100 offered 80 GB of HBM, while NVIDIA specifies 372 GB of HBM3E per Grace Blackwell Superchip) and higher-performance packaging (CoWoS). The shift to GB200 has amplified this trend: it uses the most advanced packaging (CoWoS-L with multi-die stacking, now moving to panel-level FO) and unprecedented memory. As a result, suppliers who catered to H100 production (HBM makers, substrate fabs, OSAT labs) now face even far greater demand. Well-known bottlenecks from the Hopper era – limited HBM supply and CoWoS packaging capacity – have continued to limit Blackwell rollout (^[2]) (^[13]), driving NVIDIA to explore alternative solutions (e.g. panel-level fan-out and onshoring of packaging facilities).

Current State. As of early 2026, Blackwell (GB200) chips are in full mass production, and NVIDIA has begun the transition to the GB300 NVL72 (Blackwell Ultra) platform, announced at GTC 2025 and ramping through the first half of 2026 with 288 GB of HBM3e per GPU and higher FP4 throughput than GB200. TSMC’s Arizona fab has produced the first U.S.-made NVIDIA Blackwell wafer (heralding onshoring of AI chipmaking) (^[14]) (^[15]), although full 2-chip/CPU packages still rely on Taiwanese OSAT lines (^[3]) (^[16]). NVIDIA has begun shipping integrated NVL72 racks (e.g. DGX systems) in small volume and major cloud players (AWS, Microsoft, Google, Meta) are deploying Blackwell nodes. Analysts expect surging demand: TrendForce projects Blackwell (B200/GB200) to represent over 80% of NVIDIA’s high-end GPU shipments by 2025 (^[17]), and forecast shipments to ramp from a few hundred thousand in 2024 to on the order of 1.5–2 million units in 2025 (^[18]) (^[17]). Key OEMs (Foxconn, Supermicro, Dell, HPE, GIGABYTE, Inspur, etc.) have all announced or are preparing Blackwell-based servers. This market enthusiasm has led to a “domino effect” in the supply chain: Taiwanese server manufacturing giants (Hon Hai/Foxconn, Accton, Chenbro, Auras) are poised to reap growth by supplying major subsystems (servers, network switches, chassis, liquid-cooling systems) for the Blackwell wave (^[19]) (^[20]).

Structure of this Report. We explore the GB200 supply chain in depth. The next section details the GB200 hardware architecture (Chips, NVSwitch, NVL72 rack). We then examine chip fabrication and packaging, highlighting the roles of TSMC and OSAT partners. Following that, sections cover memory and substrate suppliers (HBM3e, coWoS substrates), system integration and board assembly (PCBs, OEMs), power and cooling subsystems, and interconnect and networking components. Throughout, we integrate data on production volumes, cost breakdowns, and supply constraints. We include tables summarizing major component categories vs example suppliers, and an estimated BOM cost breakdown for a Blackwell B200 GPU. Case studies (e.g. Foxconn’s DGX gear, ZT Systems’ ACX200, AWS UltraServers) illustrate real-world examples of how the ecosystem is put together. Finally, we discuss the implications for the industry, including ongoing supply-chain bottlenecks, new packaging innovations, and the future of AI hardware manufacturing.

The GPU architecture at the heart of GB200 is NVIDIA’s Blackwell tensor-core GPU, paired with the Grace CPU in a combined module. Each GB200 module (also called a Grace Blackwell Superchip) consists of two Blackwell GPUs and one Grace CPU, interconnected by NVIDIA’s NVLink-C2C interconnect. NVIDIA specifies 372 GB of HBM3E per Grace Blackwell Superchip and 13.4 TB per NVL72; its current product page does not state a separate per-GPU allocation.

NVIDIA-scale system. In practice, NVIDIA deploys GB200 in very large multi-GPU servers. The flagship NVL72 system uses 36 GB200 modules (72 GPUs, 36 CPUs) in a single rack. NVIDIA’s own datasheet describes NVL72 as a “‘single massive 72-GPU rack’ that can overcome communication bottlenecks” using NVLink and dedicated liquid cooling (^[10]). The architecture not only stitches GPUs on a board but also links entire boards via NVLink switch fabrics: NVL72 uses nine 1RU NVLink switch trays per rack to form one unified NVLink domain (^[21]) (^[7]). Each NVL72 server (often called a DGX or similar) contains two such boards in 1U, and 18 servers (36 modules) fill the rack (^[6]).

Key performance figures highlight the scale: each NVL72 rack delivers 720 petaflops of FP8/FP6 Tensor Core performance in sparse mode (360 petaflops dense) and carries 13.4 TB of HBM3e memory (^[8]). Amazon, for example, explicitly cites “up to 72 NVIDIA Blackwell GPUs interconnected using fifth-generation NVLink — all functioning as a single compute unit” for its EC2 P6e-GB200 UltraServer offering (^[8]). AWS P6-B200 instances use eight NVIDIA Blackwell B200 GPUs and offer up to 1,440 GB of HBM3e GPU memory; P6e-GB200 is AWS's separate GB200 family. In all cases, these systems require advanced infrastructure: direct-to-chip liquid cooling loops, extremely high-voltage power delivery, and bespoke networking.

Summarizing the components of the full GB200 system (e.g., one NVL72 rack) illustrates the breadth of suppliers needed:

• Compute Modules: 36 GB200 modules per rack, each containing 2 Blackwell GPUs + 1 Grace CPU (^[9]) (^[6]). (NVIDIA designs GPUs; fabrication by TSMC.)
• Memory: NVIDIA specifies 372 GB HBM3E per Grace Blackwell Superchip and 13.4 TB HBM3E per GB200 NVL72 rack (^[22]).
• NVLink Backplane and Cable System: 72 GPU connections plus switch fabrics. (Connectors/cables from Amphenol, Samtec, US Conec, etc. (^[23]) (^[24]).)
• PCIe Interconnect: PCIe expansion supports external networking adapters. NVIDIA identifies Quantum-X800 InfiniBand, Spectrum-X800 Ethernet, and BlueField-3 DPUs as networking technologies for Blackwell-scale deployments; the components used depend on the system configuration (^[22]).
• Switch ASICs: Nine 1RU NVLink switch trays per rack, with two NVIDIA NVSwitch ASICs in each tray (18 ASICs total) (^[25]).
• Power Delivery: Rack busbars (48 VDC) driving 12V VRMs on blades (^[26]). (Busbar and high-current connectors from companies like Alpha & Omega Semiconductor, TE Connectivity (RapidLock), etc.)
• Cooling System: Custom cold-plate assemblies, pumps, quick-disconnects, and fitting lines. NVIDIA specifies a liquid-cooled GB200 NVL72 design but does not identify component-level cooling suppliers on its product page (^[22]).
• Chassis and Rack: 18×1U server chassis, rack, top-of-rack switch space, etc. (Chassis by Foxconn, Supermicro, Chenbro, etc.; rack/PDU frame by Vertiv, Schneider.)
• Auxiliary Control: Baseboard management controllers (ASPEED AST2500 BMCs, sensors) for each server.

The system-level integration itself was historically the domain of NVIDIA and its ODM partners. ZT Systems, for example, announced an ACX200 rack integrating GB200 modules into a liquid-cooled server with NVLink switch trays (^[7]). Similarly, Foxconn’s server division claimed to be first in line to ship GB200 systems by late 2024 (^[27]) (^[28]). These OEMs manage procurement from many part vendors and do the final assembly. A continuously updated, sourced map of which firms design GB200-class racks in-house (and where their new US plants are) is maintained in GPU Smith's NVIDIA server vendor index and US AI-server manufacturing tracker.

At the core of the GB200 supply chain is the semiconductor manufacturing process. NVIDIA designs the Blackwell GPU and Grace CPU, but outsources fabrication to TSMC. This begins at the wafer level: the first NVIDIA Blackwell wafer was produced at TSMC’s Arizona fab in October 2025 (^[15]) (^[14]). This milestone “signals that Blackwell has reached volume production” in the onshore fab (^[15]). However, crucially, the wafer gross die fabrication is only the first step. Because Blackwell uses a multi-chip package, each wafer’s individual dies (two GPUs + one CPU) must be assembled together with HBM memory in advanced packaging processes. Currently, NVIDIA relies on TSMC’s CoWoS (Chip-on-Wafer-on-Substrate) advanced packaging, which to date exists only in TSMC’s Taiwanese sites. As Ming-Chi Kuo and others note, the Arizona-fab wafer still “needs to be shipped to Taiwan for TSMC’s CoWoS advanced packaging, only then [is] production of the Blackwell chip [complete]” (^[3]).

Indeed, all of TSMC’s current AI-chips packaging facilities are in Taiwan (^[29]), so NVIDIA must still send wafers to OSAT partners there. A journalist from The Register emphasizes that “Nvidia remains reliant on Taiwanese packaging plants to turn those wafers into its most powerful GPUs” (^[16]). In practice, the GPU dies and memory stacks are aligned on a large organic substrate (or interposer) using CoWoS-L (2.5D fan-out) methodology (^[30]) (^[31]). The substrate itself—essentially a high-density PCB interconnect—is supplied by specialized manufacturers (see next section). Multiple bonding, underfill, and RDL processes integrate the components into a single package.

This advanced packaging is extremely complex and capacity-limited. CoWoS lines must handle very large-die GPUs with fine pitch interconnects. According to supply-chain analyses, the packaging yield rate is a critical concern: failed wafers waste both the expensive logic dies and HBM stacks (^[32]). NVIDIA’s recent delays have been largely attributed to CoWoS bottlenecks. Foxconn and Digitimes both reported that “advanced packaging tech used to stitch together the compute dies and HBM3e modules” presented challenges, exacerbated by “CoWoS capacity [that] remains extremely limited” through 2025 (^[13]) (^[2]). Thus, packaging capacity remains an important supply-chain consideration; the relevant discussion appears in the Discussion: Implications and Future Directions section below.

Major partners: The key players in packaging include TSMC (packaging operations) and leading OSAT houses:

• Amkor Technology (U.S.) – Signed an agreement with TSMC to provide “turnkey advanced packaging and test services” in a new U.S. OSAT fab (slated by 2027–28) (^[33]) (^[29]). Amkor is expected to handle US packaging of future NVIDIA chips, including Blackwell and beyond.
• ASE Group (Taiwan) – A semiconductor assembly and test provider. PTI’s corporate group page does not identify ASE as a PTI affiliate, and this article does not attribute a specific GB200 packaging role to ASE without a direct announcement.
• Powertech Technology (PTI) (Taiwan) – Another major substrate/packaging supplier; recent reports mention Powertech’s role in next-gen panel-level fan-out (PFLO) for NVIDIA BLACKWELL (^[34]).
• Samsung Electronics (South Korea) – While Samsung is a leader in HBM DRAM, it has not been reported as a supplier of HBM3e for NVIDIA. (NVIDIA’s HBM3E comes primarily from SK Hynix and Micron.)
• Intel/Moog – (Intel’s EMIB or Foveros alternatives; not currently used by NVIDIA.)

NVIDIA also contributes specialized chiplets: the NVIDIA NVSwitch ASIC used in racks is itself an NVIDIA-designed chip, but likely fabricated/co-packaged with their GPUs; it uses the same CoWoS infrastructure.

In summary, the chip production supply chain includes TSMC wafer fabrication and advanced packaging. Public sources cited here do not establish individual GB200 assembly, test, or panel-level packaging supplier relationships; companies should not be presented as GB200 suppliers without direct confirmation. Historical shipment forecasts are discussed in the quantitative-estimates section as forecasts, not confirmed production totals.

A major part of the GB200 BOM is HBM3E memory. NVIDIA specifies 372 GB of HBM3E per Grace Blackwell Superchip and 13.4 TB per NVL72 rack; the current product page does not provide a separate stack count or per-GPU allocation. HBM3E is manufactured by only a few suppliers:

• SK Hynix (South Korea): Leading HBM vendor; has 16–24 GB HBM3e products. SK Hynix is expected to supply a significant volume for NVIDIA’s AI GPUs (as it did for Hopper).
• Micron Technology (USA): Offers 24 GB HBM3e “Gen2” stacks and has explicitly identified NVIDIA as a key customer. Micron’s CEO stated that NVIDIA is “one of its primary customers” for HBM3e (^[35]). In fact, Micron’s first HBM3e device (8-Hi 24 GB) is already qualified for NVIDIA’s H200 series, and launch in early 2024 (^[36]) (^[35]).
• Samsung Electronics (South Korea): Also a major HBM player; Samsung has HBM2/HBM2e/3 lines but has not been publicly confirmed to supply HBM3e to NVIDIA.

The competitive situation: As of 2023–2024, Micron was first to ship HBM3e, but SK Hynix and Samsung were already preparing their products. There have been concerns about HBM supply for AI: one report notes that in 2024, Micron’s entire HBM3e capacity was sold to customers (mostly NVIDIA) and even fully booked into 2025 (^[37]) (^[38]). Thus, the HBM DRAM supply chain is very constrained, reflecting the handful of players in a niche market. NVIDIA must coordinate orders directly with SK Hynix and Micron (and perhaps Samsung) to secure the billions of dollars of memory needed.

Beyond chips, a key material is the substrate or interposer onto which the dies are mounted. NVIDIA uses organic substrates in CoWoS packaging. High-end substrates (with many layers of copper) are produced primarily by Taiwanese firms. For example, Yole Intelligence notes that Unimicron, Ibiden, Nan Ya PCB and similar companies dominate the IC substrate market (93% share) (^[39]). These companies fabricate the glass-reinforced laminates and build-up layers needed for chip modules. In years prior, substrate shortages have become choke points; Taiwan-based substrate makers have ramped capacity for AI chips. A Taiwan news report specifically warns that NVIDIA’s GB200 is facing delays because “key materials used for chip substrates are in short supply due to insufficient production capacity” (^[40]). The cited Digitimes report is behind a paywall; readers should treat its reported substrate and CoWoS constraints as a secondary report rather than a confirmed NVIDIA disclosure. In response, both substrate makers and TSMC have announced expansions: for instance, in late 2025 TSMC invested heavily in new CoWoS lines and substrate fabs (Erudite Asia on 15 Oct 2025).

Example: Independent cost models may estimate memory and packaging costs, but their assumptions should be labelled as model assumptions rather than GB200 specifications. NVIDIA publishes 372 GB of HBM3E per Grace Blackwell Superchip and 13.4 TB per NVL72, not the 192 GB-per-GPU premise used in the prior estimate.

Below is a summary table of major GB200 components and representative suppliers:

Component Category	Function / Description	Key Suppliers / Vendors
GPU Compute Dies	NVIDIA Blackwell Tensor-Core GPUs	NVIDIA (design); custom-built TSMC 4NP process
CPU Die	NVIDIA Grace 72-core Arm CPU	NVIDIA (design)
HBM3E memory	372 GB per Grace Blackwell Superchip; 13.4 TB per NVL72 rack	Supplier relationships should be identified only where publicly confirmed
Advanced Packaging (CoWoS)	Multi-chip integration	TSMC's CoWoS advanced-packaging service; supplier relationships for individual GB200 packages are not publicly confirmed here
IC Substrates / Interposers	Organic/glass build-up substrates for CoWoS/FO	Public sources cited here do not confirm GB200 supplier relationships; SUMCO’s published product lineup is silicon wafers, not PCBs or IC substrates
PCBs / Server Boards	Motherboard hosting GB200 modules	Foxconn (Hon Hai), Quanta, Wistron, Inventec, etc.
Connectors & Cables (NVLink)	High-speed NVLink interconnect (GPU↔switch, QSFPs, etc.)	Amphenol (Paladin), US Conec (DensiLink), TE Connectivity, Samtec (SkewClear, FireFly)
NVSwitch ASICs	NVIDIA’s switch chips for multi-GPU fabrics	NVIDIA (design); TSMC (manufacturing)
Power Delivery (VRMs)	Voltage regulation modules on-board	NVIDIA’s public hardware guide does not identify component-level VRM suppliers
Busbars & Harness	Rack-level power bus bar and server power cabling	NVIDIA documents power shelves and a bus bar but does not identify component-level suppliers
Cooling Systems	Direct-to-chip liquid cold plates, pumps, quick-disconnect fittings	NVIDIA documents liquid cooling for GB200 NVL72; component-level supplier relationships are not identified here ([22])
Chassis & Rack Cabinets	Server enclosures, racks, power distribution units	Foxconn, Supermicro, Chenbro, Auras (liquid cooling modules), Vertiv, Schneider
Network Adapters (NICs)	Ethernet/InfiniBand connectivity	NVIDIA ConnectX-7 NICs and BlueField-3 DPUs in DGX GB200; Broadcom is a separate vendor
Baseboard Mgmt & Controllers	BMC microcontrollers, sensors	ASPEED (AST2500 BMC), Renesas (MCUs), NXP, Texas Instruments
Capacitors/Inductors/Resistors	Power smoothing & filtering	Murata, TDK, Taiyo Yuden (inductors); AVX, Rubycon, Samsung Electro-Mechanics (caps)
Firmware/Software	System BIOS, NVLink on-chip firmware	NVIDIA, ASPEED (BMC firmware)

Table 1: Major subcomponents of NVIDIA’s GB200 system and exemplar suppliers. Each category may involve many vendors.

This non-exhaustive table highlights that virtually every step of the GB200’s assembly chain involves companies beyond NVIDIA. Memory stacks alone involve two major DRAM makers; substrates involve multiple Taiwanese PCB firms; board-level components (passives, connectors, chips) draw from global electronics conglomerates; and large ODMs tie it all together. In the next sections we dive deeper into these areas.

NVIDIA’s multi-die package requires a sophisticated integration strategy. TSMC describes CoWoS-L as a Chip-on-Wafer-on-Substrate service using an RDL-based interposer, embedded local silicon interconnects, and a molding-based interposer; it is designed to integrate large HPC products with silicon and HBM (^[41]). Public material cited here does not establish the supplier-specific assembly sequence for an individual GB200 package.

NVLink Fabric: Inter-chip and inter-board connectivity is another packaging aspect. Each Blackwell GPU connects to the system via a high-pin-count connector. In NVL72, each GPU plugs into an Amphenol Ultrapass Paladin connector (72 differential pairs, 224 Gb/s bandwidth) on the backplane (^[23]). A corresponding 72-pair Paladin receptacle on the server backplane mates with the GPU. Internally, GPUs connect to NVSwitch chips via SkewClear EXD Gen2 twinax cables, and between NVSwitch trays NVIDIA uses OverPass and DensiLink flyover cables (^[24]). These connectors/cables are extremely high-speed and expensive; SemiAnalysis notes that “much of the cost is not from the cables but from the connectors [which] need to prevent crosstalk” (^[23]). NVIDIA’s choices here (Amphenol Paladin, DensiLink, SkewClear, etc.) drive orders for those vendors. (Samtec’s FireFly and PACMIC also compete in this space.) Future NVL36×2 setups require additional twinax 1.6 Tbps cables, further increasing cable demand (^[42]).

Networking and IO: Aside from NVLink, GB200 servers include Ethernet or InfiniBand NICs. NVIDIA completed its acquisition of Mellanox in April 2020; Broadcom remains a separate company. NVIDIA’s DGX GB200 hardware guide identifies ConnectX-7 NICs and BlueField-3 DPUs in the reference compute tray. The external-network configuration varies by deployment, so this article does not attribute a Broadcom switch or ConnectX-8 NIC to GB200 systems without a source specific to that configuration.

Advance to Panel-Level FO: Due to CoWoS limits, NVIDIA is shifting to Panel-Level Fan-Out (PFLO) packaging for Blackwell next year (^[31]). In PFLO, multiple ICs are embedded in a larger laminate panel, allowing scaling beyond wafer size. This approach uses materials like glass substrates or laminates instead of silicon interposers. WCCFtech reports that suppliers for PFLO are currently scarce, naming Powertech (PTI) and Innolux (a large TFT LCD manufacturer) as early contenders to package NVIDIA’s chips (^[34]). If adopted, this would involve new supply chains: Powertech (with experience in substrates) and Innolux (leveraging glass panel capabilities) would become NVIDIA packaging vendors.

In summary, the advanced packaging and interconnect stack involves:

• Die-to-die integration: TSMC's CoWoS-L advanced-packaging service uses RDL-based interposer technology and embedded local silicon interconnects; public sources cited here do not confirm individual GB200 assembly, test, or PFLO supplier relationships (^[41]).
• GPU-to-backplane connectors: NVIDIA documents an NVLink passive copper cable cartridge backplane, but the cited public material does not identify its component suppliers.
• NVLink switch fabric: NVIDIA documents nine switch trays in an NVL72 rack, with two NVLink switches per tray.
• InfiniBand/Ethernet ports: NVIDIA networking components and high-speed I/O hardware; the exact configuration depends on the system.
• Riser connectors and midplane: Public material cited here does not identify component-level suppliers for these interfaces.

This layer alone accounts for dozens of supplier relationships. For example, SemiAnalysis estimates that the NVLink backplane assembly by itself costs several thousand dollars per board because of premium connectors (^[23]). The cited material does not identify NVIDIA’s vendor-qualification decisions for individual connector or cable interfaces.

Power delivery for GB200 NVL72 is rack-scale: NVIDIA’s published power-budget guidance lists 1.2 kW maximum per GB200 GPU and 120 kW designed rack power. The NVL72 architecture uses 48 V DC busbars at the rack level. Busbars (often aluminum or copper conductors) run vertically, bringing 48 V from the room UPS/PDU to each 1U server. In each server, a power distribution board (PDB) steps down 48 V to 12 V for the GPUs/SOC, via large VRM arrays. According to supply-chain reports, each GB200 card receives power through four 12V DC and four ground high-current connectors (RapidLock style) around the GPUs (^[26]). These TE Connectivity “RapidLock” connectors (or similar) carry ~675 A per cable.

NVIDIA’s public hardware guide documents power shelves, a bus bar, and liquid cooling for the DGX GB rack system, but it does not identify the suppliers of VRM controllers, MOSFETs, busbars, passives, connectors, or cable harnesses. Those vendor relationships should not be inferred from a general component market or from prior NVIDIA systems.

NVIDIA documents that power shelves convert AC power to nominal 50–51 V DC and distribute it through the rack bus bar. Its public hardware guide does not provide a GB200 board-level VRM phase count, connector-current rating, or component supplier list.

Power Infrastructure at Rack Scale: In addition to the cards, the power distribution units and power shelves factor in. NVIDIA lists approximate rack power consumption of 120 kW and documents redundant power shelves and a rack bus bar; it does not identify PDU, chiller, or busbar suppliers.

Overall, the power delivery sub-system adds tens of suppliers for everything from semiconductors to metalwork. While none of these companies are as famous as NVIDIA or TSMC, they are essential: without them, the GB200 modules cannot be powered or controlled.

Once the individual components (chips, memory, boards) exist, final assembly and integration involve large OEMs (Original Equipment Manufacturers) and CSPs (Cloud Service Providers). These organizations source GB200 modules and related parts to build complete servers or clusters.

• Foxconn (Hon Hai) and Pegatron: Foxconn’s server business (often under Foxconn Industrial Internet) has claimed to be the first to ship GB200 servers (^[28]). Analysts expect Foxconn to build a substantial share of Blackwell racks; one brokerage projects Foxconn could ship 17,000–30,000 racks (40–70% market share) by 2026 (^[20]). Foxconn and Taiwanese ODMs like Quanta and Wistron receive GB200 boards or modules from NVIDIA/OEMs and assemble them into branded servers (e.g. Foxconn’s “Star Gate” AI servers). They also manage customization (e.g. adding custom cooling loops or integrated cabling).

• Supermicro, Dell, HPE, Lenovo: Western OEMs likewise incorporate GB200. Dell announced at NVIDIA GTC 2024 new PowerEdge servers (“B200” nodes) with NVIDIA Blackwell GPUs. HPE offers ProLiant and Apollo systems with Blackwell acceleration (as per HPE community posts). These companies collaborate with NVIDIA on system design; e.g., they decide on chassis layouts, which card slot schemes to use, etc. While they largely function as system integrators, they also have purchasing influence: NVIDIA builds baseline GB200 boards, then these OEMs may buy or license them for their server designs.

• ZT Systems, Penguin Solutions, Cirrascale, etc.: Smaller companies that specialize in HPC appliances have also launched Blackwell-based platforms. ZT Systems, which AMD acquired in March 2025 (^[43]), introduced the ACX200 rack, a turnkey liquid-cooled platform using GB200 modules and Grace CPUs (^[7]). The AWS blog similarly brands its ultraservers but acknowledges that “the P6e-GB200 UltraServers represent [AWS’s] most powerful GPU offering to date” with 72 GPUs per server (^[8]). These examples illustrate that integrators configure the final product around NVIDIA’s raw modules.

• Cloud Providers (Hyperscalers): AWS (Amazon), Microsoft Azure, Google Cloud, Meta, and others are major end-users. They partner with ODMs or deploy Nvidia’s DGX-branded systems. For example, AWS’s UltraClusters integrate hundreds or thousands of Blackwell GPUs, which are deployed in GB200 Superchips. The AWS blog emphasizes that their new P6e-GB200 machines have 72 GPUs linked by NVLink and can be liquid-cooled or air-cooled (^[8]) (^[44]). These cloud customers influence the supply chain too: they place massive orders that ripple down through NVIDIA, TSMC, and OSAT, affecting supplier demand planning.

Implications: The diversity of OEMs means multiple supply channels exist. NVIDIA can sell GB200 modules to Foxconn for one customer, while directly supplying Dell for another. Some OEMs like Dell/HPE might design their own PCBs and simply purchase Blackwell GPUs from NVIDIA. Others (like ZT Systems) might buy GB200 cards and mount them on their own motherboard (as indicated in the ZT press release). In any case, practically all major server vendors worldwide are involved. This broad ecosystem further multiplies the number of vendors: for each component in our table above, the customer market is global.

In mid-2024, The Register reported on Foxconn’s plans for GB200: Foxconn’s AI server unit claimed it would begin shipping “small volumes” of GB200 systems in Q4 2024, ramping in Q1 2025 (^[27]). According to Foxconn, their system (“DGX NVL72”) packs 36 GB200 superchips (72 GPUs + 36 CPUs) into 18 1U servers, delivering 1.44 exaFLOPS (FP4) and 13.5 TB HBM3e total (^[6]). The article noted that Foxconn would be “the first supplier” to ship GB200 accelerators. This implies that NVIDIA (who co-brands some systems) was relying on Foxconn’s manufacturing capabilities to fulfill initial orders.

Foxconn’s involvement exemplifies how OEM supplier networks operate: The cited reporting does not identify Foxconn’s specific bill of materials, component suppliers, or the allocation of assembly work; those relationships are therefore not attributed here.

The Register piece also highlights supply issues: it says NVIDIA and TSMC encountered “challenges with the advanced packaging tech” for Blackwell, and that CoWoS capacity remained limited (^[13]), causing NVIDIA to prioritize GB200 over smaller B100/B200 parts. Foxconn, however, was optimistic about ramping. This underscores that multiple companies (NVIDIA HQ, TSMC fabs, Foxconn/ODM factories) each play a role in launching a product. If any one stock has issues (e.g. packaging yields at TSMC’s OSATs), it reverberates across the chain.

Understanding how much each part contributes to cost can reveal supply dependencies, but current NVIDIA GB200 specifications do not validate a component-level bill of materials or pricing. NVIDIA publishes 372 GB of HBM3E per Grace Blackwell Superchip and 13.4 TB per NVL72; any cost model must identify its product configuration, date, and price assumptions.

No NVIDIA primary source cited here publishes a current GB200 component-level BOM, HBM price, or rack DRAM-cost estimate. Those figures are therefore omitted rather than extrapolated from a 192 GB-per-GPU assumption.

Despite the advanced planning, supply constraints have repeatedly emerged. Two key bottlenecks have been HBM memory and CoWoS packaging capacity:

• HBM Memory Shortage: In March 2024, NVIDIA’s major memory partner Micron announced that its entire HBM3E output for 2024 was pre-sold (mostly to NVIDIA) and that it expected to “capture a sizable chunk” of the HBM market (^[38]). Micron’s CEO stated HBM3E was “sold out for calendar 2024, and the overwhelming majority of 2025 supply has already been allocated” (^[38]). This implies extremely tight supply for other players (like AMD or Chinese AI chip makers). Meanwhile, SK Hynix and Samsung will ramp their own HBM3e production, but bridging that gap could take years. The result is that HBM manufacturers hold bargaining power and NVIDIA likely had to prioritize who gets early shipments (Cloud Service Providers over enterprises).

• CoWoS Packaging Capacity: TSMC’s CoWoS lines have been oversubscribed. After the H100 launch in 2022, even CEO C.C. Wei warned of an AI chip “shortage through 2025” due to CoWoS limits. By 2024, NVIDIA began exploring alternatives. In mid-2024 WCCFtech reported that NVIDIA had “decided to resolve CoWoS issues” by developing a new “PFLO” (panel-level fan-out) packaging for later Blackwell units (^[2]). Indeed, by late 2024/25 NVIDIA is reportedly shifting new Blackwell (GB200) units to this PFLO method (^[31]), in partnership with substrate firms.

Taiwan-based commentary echoes these issues. One semiconductor newsletter notes that individual component “supply chains get reworked” for each server OEM; specialized evaluation of over 50 subcomponents shows many had to ramp suddenly for GB200 (^[1]). Digitimes reported that GB200 shipments were delayed due to substrate/material shortfalls (^[40]) (the snippet indicates “key materials for chip substrates are in short supply”). All these signals point to the fact that even Tier-1 vendors cannot produce GB200 at will; they are constrained by the limited production capabilities of their upstream partners.

To overcome these, NVIDIA and the industry are taking steps:

• Onshoring: Building the Arizona fab and planning US packaging lines (Amkor’s OSAT by 2027) reduce reliance on Taiwanese facilities (^[3]) (^[29]).
• New packaging tech: Panel-level packaging is an industry development, but this article does not attribute a GB200 transition or supplier relationship to Powertech, Innolux, or another company without direct confirmation.
• Prioritization: NVIDIA explicitly prioritized GB200 (high-end racks) over lower-tier B100/B200 chips to maximize the use of scarce resources (^[13]) (^[45]).

Hyperscale cloud providers illustrate the system-scale integration of GB200 and its supply chain. Amazon Web Services (AWS) announced new EC2 instances leveraging NVIDIA’s GB200. The P6e-GB200 UltraServers offer up to 72 NVIDIA Blackwell GPUs in one system, with 360 PFLOPS of dense FP8 compute (^[8]). AWS highlights that they aggregated five generations of Nitro networking and liquid cooling to achieve these specs. Notably, AWS’s deployment underscores two supply-chain facets:

1. NVLink Integration: AWS’s description of “72 NVIDIA Blackwell GPUs interconnected using 5th-generation NVLink” (^[8]) confirms how AWS is adopting NVIDIA’s supplied NVL72 architecture at scale. To implement this, AWS needed tens of thousands of NVLink switch chips and associated high-speed cables/connectors. (International Datacenters and network equipment from Accton or Arista supply AWS with 800G Ethernet switching, but the core GPU domain is NVIDIA/Amphenol technology.)

2. Cooling Solutions: AWS explicitly distinguishes their two offerings: P6-B200 instances (“air-cooled”) versus P6e-GB200 Ultra (“liquid-cooled”) (^[44]). AWS describes the P6e-GB200 UltraServer as liquid-cooled but does not identify the component-level cooling suppliers in the cited announcement. This article therefore does not attribute its chillers, cold plates, pumps, or coolant to specific companies.

AWS's public announcement demonstrates the scale of the P6e-GB200 UltraServer configuration, but it does not identify component-level procurement relationships for GPUs, boards, networking, chassis, or cooling. Those supplier relationships should not be inferred from the configuration alone (^[46]).

To put numbers on the supply chain, analysts have made bold forecasts. WCCFtech reported (citing Taiwan Economic Daily) that NVIDIA expects to ship ~420,000 Blackwell GPUs in the second half of 2024, rising to 1.5–2 million units in 2025 (^[47]). These figures dwarf the ~100k H100 GPUs shipped in all of 2022. If true, by the end of 2025 NVIDIA will have deployed roughly 2 million Blackwell GPUs into the field. A GB200 NVL72 reference rack contains 18 1RU compute trays with four GPUs each, so a shipment forecast expressed in GPUs cannot be converted into a server or rack count without identifying the mix of products shipped. Sources similarly project tens of thousands of racks by 2026 – one analysis even suggests up to 60,000 racks by 2026 (^[48]) (consistent with 4.3 million GPUs under such a scenario).

Driving this scale requires enormous ramping of suppliers. For example:

• Memory: NVIDIA specifies 372 GB HBM3E per Grace Blackwell Superchip and 13.4 TB for an NVL72 rack. A shipment-wide memory or DRAM-cost estimate requires a documented product mix and pricing assumptions.
• Copper and Substrate: Tens of thousands of 700+mm² substrates needed, each made of specialized materials (copper foil, epoxy, glass cloth from suppliers like Kolon or Ibiden).
• Electronics Components: Millions of high-end capacitors/inductors.
• Cabling: If each rack uses ~1000 feet of specialized cable, the total cable length is enormous.
• Power Supplies: Gigawatts of DC-to-DC conversion capacity installed.

No single supplier can handle such volumes alone. Even TSMC had to scale: it reportedly booked over half of its CoWoS capacity for 2026-27 for NVIDIA’s chips (^[40]). Likewise, Micron has been expanding HBM fabs. OSAT players (Amkor, Powertech, PTI) are building new fab lines with backing from TSMC, driven by NVIDIA’s orders. Meanwhile, PCB fabricators (Unimicron, Ibiden) are adding layers and copper volume capacity for AI substrates.

In effect, Blackwell has become an anchor tenant for the high-end semiconductor ecosystem. As one industry commentator put it, “NVIDIA’s Blackwell: a milestone for US chips but Taiwan’s packaging power” (^[49]). The entire chain – from raw silicon through final server assembly – is now targeted on this product.

The supply chain realities of the GB200 have broad implications:

• Concentration of Suppliers: The high-end AI chip supply chain is heavily concentrated in a few regions and firms. TSMC (Taiwan) and now Samsung/Intel (South Korea/USA) control the most advanced wafer production. Only a handful of OSATs can do CoWoS. Major DRAM makers (SKH, Micron) dominate HBM. This concentration means geopolitical and capacity risks. For example, a policy decision affecting Taiwan or export controls on HBM could ripple through NVIDIA’s deliveries. This has led to initiatives like U.S. onshoring of fabs and packaging (through CHIPS Act funding) to mitigate risk (^[15]) (^[29]).

• Cost Structure and Profit Margins: As noted, the BOM for GB200 is heavily weighted toward memory and packaging costs (^[4]). NVIDIA’s selling price per chip ($30k–40k reported (^[50])) yields very large gross margins on silicon itself, but system prices ($2–3M per rack) ensure NVIDIA’s partners also see healthy revenue. However, rapid scaling may compress margins for some vendors if competition increases (e.g. if Samsung enters HBM3e at scale). The capital intensity is huge, so suppliers need steady orders to justify new fabs and lines.

• Innovations in Packaging: The GB200 has spurred packaging R&D. Panel-level FO and other techniques (embedded die laminates, fan-out wafer level) are being fast-tracked. If PFLO proves superior (as NVIDIA believes), it could displace CoWoS for future products. Already, specialized companies are emerging. For example, Cohu and KLA have accredited equipment for panel-level packaging aimed at NVIDIA’s roadmap (^[34]). Over the next 2–3 years, new vendors in the packaging segment (beyond the traditional OSATs) may emerge, drawn by NVIDIA’s demand.

• Environmental and Infrastructure: A side effect of GB200’s power density is increased interest in liquid cooling. Analysts predict liquid cooling penetration in AI datacenters will jump to 30% by 2025 on NVIDIA’s order (^[51]). This means vendors of cooling infrastructure (cooling distribution units, rack-level chillers, etc.) will see booming demand. Companies like Asetek, CoolIT, Altec, and Metz are likely to expand. Conversely, ordinary air-cooled designs (with fewer GPUs per rack) may coexist, as indicated by NVIDIA shipping a trimmed-down B200A variant for customers unsuited to liquid cooling (^[52]).

• Supply Chain Coordination: Meeting Blackwell demand requires unprecedented coordination. NVIDIA and government are pushing supply chain mapping, strategic materials stockpiling, and flexible manufacturing. For instance, TrendForce noted companies are securing wafer and substrate orders years in advance. NVIDIA’s detailed briefings to investors mention dozens of upstream issues being tracked (from quick-disconnect leakage to PCB trace design) (^[53]). This level of granularity suggests that NVIDIA’s own supply team must manage hundreds of vendor relationships.

Future Outlook: Through 2026, the GB200 is already giving way to the GB300 NVL72 (Blackwell Ultra) rack — which NVIDIA describes as integrating 72 Blackwell Ultra GPUs and 36 Grace CPUs, higher FP4 throughput, and is ramping in volume this year — with NVIDIA's next-generation Rubin platform on the roadmap for 2026–2027. The industry is already talking about “beyond trillion parameter models” requiring even denser interconnect. If NVIDIA pushes to, say, 100 GPUs per rack, supply chain demands will intensify further. Meanwhile, competitors (AMD, Intel, specialized chips) will drive complementary chains (e.g. memory for AMD’s MI300, Intel’s 4th-Gen Gaudi). But for now, NVIDIA’s Blackwell is a “demand shock” that has financiers and supply chain analysts projecting strong growth for substrate makers, packaging companies, and integrators in 2026–2027 (^[19]) (^[48]).

NVIDIA’s GB200 NVL72 is a rack-scale system that combines 36 Grace CPUs, 72 Blackwell GPUs, fifth-generation NVLink, liquid cooling, and supporting infrastructure. Public documentation establishes these product-level facts, but it does not establish a complete component-level supplier roster. The article therefore treats only directly evidenced relationships as GB200 supplier relationships and does not infer them from companies’ roles in adjacent markets. (^[22])

Despite this complexity, the supply chain is strained by demand. Analysts report potential multi-year shortages in CoWoS capacity and HBM supply (^[13]) (^[38]). In response, NVIDIA and its partners have accelerated new technologies: panel-level fan-out packaging, US-based assembly facilities, and alternative cooling formats. Throughout, market data confirm how central NVIDIA’s AI chips have become: TrendForce projects Blackwell to be 80% of NVIDIA’s high-end sales in 2025 (^[17]), and demand drivers like Foxconn’s Star Gate and AWS’s UltraClusters push supplier revenues higher.

In sum, the GB200 exemplifies modern hardware’s globalized production. No single country or company can make a GB200 board alone – it is a system-of-systems spanning continents. Our report has cataloged the principal elements of this chain, with evidence and numbers to back each category. Any detailed claim above is grounded in cited sources; for example, we cite SemiAnalysis on the “50 different subcomponents” (^[1]), NVIDIA/AWS on system scale (^[8]) (^[10]), and supply-chain analyses on packaging and memory limits (^[13]) (^[47]).

Going forward, the GB200’s supply network will likely expand further as NVIDIA and others scale AI. Future architectures may require even more advanced packaging and cooling, drawing in new vendors (e.g. Intel’s foray into glass substrates (^[54])). However, even under best-case scaling, the complexity will remain enormous. As NVIDIA’s own engineering team remarked, deploying Blackwell has “huge ramifications on dozens of downstream and upstream suppliers” (^[55]). Our deep dive here has sought to illuminate exactly who those suppliers are and how they fit together. In doing so, we underscore that the story of NVIDIA’s GB200 is as much about industry-wide collaboration as it is about a single product.