NVIDIA GB300 NVL72 Server Suppliers: U.S. vs. Taiwan

Who actually builds an NVIDIA AI server?

The obvious answer is NVIDIA.

But that is only part of the story.

NVIDIA designs the GPU architecture, NVLink interconnect, networking platforms and much of the reference architecture.

Turning those technologies into thousands of operating AI racks requires another enormous ecosystem.

Server OEMs.

ODMs.

Rack manufacturers.

Power suppliers.

Cold-plate manufacturers.

CDU suppliers.

Networking companies.

And increasingly, data-center integrators capable of delivering an entire AI factory.

The NVIDIA GB300 NVL72 is perhaps the clearest example yet.

NVIDIA's GB300 NVL72 combines 72 Blackwell Ultra GPUs and 36 Grace CPUs in a fully liquid-cooled rack-scale platform.

But companies on both sides of the Pacific are building systems around that architecture.

The battle for the AI server market is often described as U.S. versus Taiwan.
The reality is increasingly U.S. plus Taiwan.

The Companies to Watch

This is a representative supplier comparison, not an exhaustive NVIDIA partner list or market-share ranking. Many of these companies manufacture and integrate products through global supply chains rather than exclusively in their headquarters country.

First, What Exactly Is GB300 NVL72?

GB300 NVL72 is not simply a server containing several GPUs.

It is a rack-scale computing architecture.

NVIDIA integrates:

• 72 NVIDIA Blackwell Ultra.

  NVIDIA integrates:

  • 72 NVIDIA Blackwell Ultra GPUs,
  • 36 NVIDIA Grace CPUs,
  • fifth-generation NVLink,
  • ConnectX-8 networking,
  • liquid cooling,
  • rack-scale power,
  • and management software

  into an architecture designed to behave much more like oning system than a collection of independent servers. citeturn844731search2

  This changes the server business.

  The basic unit of competition is moving from:

  Server → Rack → Cluster → AI Factory

  Why the Supplier Matters More Than Before

  With a conventional server, the buyer could focus heavily on CPUs, memory, storage and price.

  A GB300 deployment introduces another set of questions.

  Who integrates the rack?

  Who provides the liquid cooling?

  Who validates the manifolds and hoses?

  Who manages rack-level power?

  Who connects hundreds of racks together?

  Who commissions the system?

  Who provides global support?

  And who can deliver the infrastructure fast enough?

  NVIDIA's MGX strategy reflects this transition. MGX is a modular architecture intended to let both OEM and ODM partners build systems ranging from individual servers to complete rack-scale AI factor compute, networking, cooling, power and connectors. citeturn735938search4

  U.S. Supplier 1: Dell Technologies

  Dell has moved aggressively into GB300 rack-scale infrastructure.

  Its PowerEdge XE9712 is built around ure with 72 Blackwell Ultra GPUs and 36 Grace CPUs. citeturn844731search38

  Dell announced in July 2025 that it delivered its first integrated GB300 NVL72 racks to CoreWeave, describing the deployment aivery of Dell Integrated Racks based on GB300 NVL72. citeturn844731search1

  Dell's advantage is broader than server manufacturing.

  It can combine:

  Servers + Networking + Storage + Rack Integration + Services

  under its Dell AI Factory strategy.

  DATAAD view:

  Dell is particularly strong where large enterprises want an accountable branded supplier to integrate and support a complete AI infrastructure stack.

  U.S. Supplier 2: Hewlett Packard Enterprise

  HPE offers the product simply as NVIDIA GB300 NVL72 by HPE.

  HPE describes it as a fully integrated, liquid-cooled rack-scale system with a 72-GPU NVLink do-tuning and real-time inference. citeturn844731search23turn844731search30

  PE's direct-liquid-cooling capability and services. citeturn844731search10

  HPE also brings decades of experience from HPC and supercomputing.

  That becomes valuable as AI infrastructure begins looking increasingly like a supercomputer installed inside a commercial data center.

  DATAAD view:

  HPE is positioned particularly well for enterprises, research institutions, sovereign AI projects and customers where compute infrastructure and high-performance cooling must be engineered together.

  U.S. Supplier 3: Supermicro

  Supermicro has made liquid cooling central to its Blackwell strategy.

  Its GB300 NVL72 system combines the Blackwell with Supermicro's direct-liquid-cooling technology. citeturn844731search4

  The company has also developed what it calls Data Center Building Block Solutions, integrating racks, networking,nt and deployment blueprints around NVIDIA systems. citeturn844731search18

  This reflects an important industry change.

  Supermicro is no longer competing only on the server.

  It is competing on how quickly an entire AI data center can be assembled.

  DATAAD view:

  Supermicro's combination of rapid product cycles and its own DLC infrastructure makes it especially significant in high-density AI deployments.

  Now Look Across the Pacific

  The U.S. brands are highly visible to enterprise buyers.

  But much of the physical engineering capability behind the global server industry is concentrated in Taiwan.

  NVIDIA itself has repeatedly highlighted companies GIGABYTE and Wiwynn within its Blackwell ecosystem. citeturn735938search0

  This is where the comparison becomes interesting.

  The Taiwanese ecosystem is no longer simply manufacturing someone else's server design.

  It increasingly designs:

  Rack + Compute Tray + Power + Cooling + Networking + Manufacturing

  Taiwan Supplier 1: Foxconn

  Foxconn — Hon Hai Technology Group is one of the most important manufacturers in the AI-server supply chain.

  Foxconn publicly demonstrated GB300 NVL72 infrastructure at NVIDIA GTC and has desc200 and GB300 rack-scale systems. citeturn976793search4turn976793search12

  Its role is also expanding beyond assembly.

  Foxconn's subsidiary Ingrasys has demonstrated GB300 NVL72 integrated with in-row CDU cooling, while ths work in high-efficiency power and thermal design. citeturn976793search20

  Foxconn is also deploying GB300 infraereign AI and supercomputing initiatives in Taiwan. citeturn976793search16

  DATAAD view:

  Foxconn shows how the traditional electronics-manufacturing model is evolving into full AI-infrastructure integration.

  Taiwan Supplier 2: QCT

  Quanta Cloud Technology — QCT is part of the Quanta ecosystem and has long focused heavily on hyperscale cloud infrastructure.

  QCT has moved its GB300 NVL72 architecture into production preparation a00 rack built by QCT at COMPUTEX. citeturn976793search1turn976793search17

  Its AI POD architecture also combines rack-snetworking and management for large AI deployments. citeturn976793search21

  DATAAD view:

  QCT represents the hyperscale ODM model particularly well.

  Instead of selling primarily one branded box at a time, the focus is on designing infrastructure that cloud and AI operators can deploy at very large scale.

  Taiwan Supplier 3: Wiwynn

  Wiwynn is another company deeply rooted in cloud data-center infrastructure.

  Wiwynn and Wistron stated that they were among the early companies offering GB300 NVL7platform with advanced liquid-cooling technologies. citeturn976793search10

  Its GB300 system is a fully liquid-cooled rack-level architecture incorra GPUs and ConnectX-8 800Gb/s SuperNIC technology. citeturn976793search18

  Wiwynn's demonstrations have also highlighted broader coothan treating the GPU server as an isolated product. citeturn976793search2

  DATAAD view:

  Wiwynn is particularly relevant to hyperscale buyers because its traditional strength lies at the intersection of cloud infrastructure, rack engineering and mass deployment.

  Taiwan Supplier 4: GIGABYTE / Giga Computing

  GIGABYTE's Giga Computing takes a somewhat different route.

  It combines the manufacturing strengths of the Taiwan ecosystem with a visible global server brand.

  Its GB300 NVL72 GIGAPOD rack integrates 72 Blace CPUs, ConnectX-8 networking and liquid cooling. citeturn976793search15

  GIGABYTE also offers a pod-scale architecture built around 16 GB300 NVL72 racks, representing 1,152 Blackwell Ultra GPUs and 576 Grace CPUs. The company lists rack power up to roughly 140 kW for that desin cluster-level cooling and management architecture. citeturn976793search7

  DATAAD view:

  GIGABYTE illustrates how Taiwanese suppliers are moving beyond the traditional ODM model toward complete branded AI-factory platforms.

  U.S. vs. Taiwan: The Business Models Are Different

  The easiest way to understand the market is not by asking which country makes a better server.

  The historical strengths are different.

  Area	U.S. OEM Model	Taiwan Ecosystem Model
  Enterprise Brand	Very strong	Growing
  Global Enterprise Sales	Strong	Often partner/hyperscale focused
  Services & Support	Major strength	Varies by supplier
  Hyperscale ODM Capability	More limited	Major strength
  High-Volume Manufacturing	Global supply chain	Core historical strength
  Rack-Level Engineering	Strong	Very strong
  Liquid Cooling Integration	Rapidly expanding	Rapidly expanding
  Direct Hyperscaler Relationships	Strong	Historically very important

  This comparison describes broad industry tendencies rather than strict boundaries. All of the companies discussed operate through international supply chains and increasingly compete across traditional OEM/ODM categories.

  The Old OEM / ODM Boundary Is Disappearing

  Historically the server industry was easier to understand.

  OEM: brand, sell and support the server.

  ODM: design and manufacture the hardware behind the scenes.

  AI is blurring that line.

  Foxconn is building AI factories.

  QCT sells rack-scale cloud infrastructure directly.

  Wiwynn designs sophisticated liquid-cooled AI systems.

  GIGABYTE sells complete AI pods.

  At the same time, Dell, HPE and Supermicro are moving deeper into rack, power, networking and cooling integration.

  The market is converging toward the same destination:

  End-to-end AI infrastructure.

  GB300 Is Also a Liquid-Cooling Competition

  Another important point is easily missed.a fully liquid-cooled architecture. citeturn844731search2

  That means a supplier cannot win simply by knowing how to mount GPUs into a chassis.

  The rack requires competence in:

  • cold plates,
  • hoses,
  • quick disconnects,
  • manifolds,
  • CDUs,
  • water quality,
  • leak detection,
  • flow balancing,
  • and facility integration.

  This is why nearly every major GB300 supplier is now talking about cooling as part of the product rather than an accessory.

  The Rack Is Becoming a Product

  Traditional IT purchasing treated the rack largely as a cabinet containing independent servers.

  GB300 changes that.

  The rack itself is now an engineered computing product.

  Compute trays must connect correctly.

  NVLink must operate as one domain.

  Networking must support enormous east-west traffic.

  Power must be distributed at unprecedented density.

  Coolant must reach each cold plate at the correct temperature and flow.

  Management software must see the whole system.

  So the competitive unit becomes:

  Rack Engineering

  not merely:

  Server Engineering.

  And the Next Step Is the Pod

  One rack is no longer enough.

  GIGABYTides a useful example of where the market is moving. citeturn976793search7

  The next layer is:

  Multiple Racks → Network Fabric → Shared Cooling → Shared Power → Cluster Management

  This means tomorrow's server supplier may increasingly be judged on how well it designs a 5 MW or 20 MW AI cluster, rather than merely whether its individual server benchmark is slightly faster.

  Why Taiwan Matters So Much

  The importance of Taiwan in AI infrastructure is not accidental.

  The region has spent decades developing expertise across:

  server manufacturing,

  motherboards,

  power supplies,

  connectors,

  thermal management,

  rack integration,

  electronics manufacturing,

  and advanced semiconductor supply chains.

  NVIDIA's own Blackwell ecosystem includes a broad group of Taiwanese and global partners s systems, connectors and data-center infrastructure. citeturn735938search0

  AI now pulls those capabilities together into one product.

  Why the U.S. OEMs Still Matter

  Manufacturing capability alone does not determine the enterprise market.

  Large corporations and governments often want:

  • one accountable supplier,
  • global support,
  • installation and commissioning,
  • software integration,
  • financing,
  • security certification,
  • and long-term service contracts.

  This is exactly where companies such as Dell and HPE remain powerful.

  They translate an extremely complex NVIDIA ecosystem into something an enterprise procurement organization can actually purchase and operate.

  So Who Really Builds the GB300?

  The technically correct answer is:

  an ecosystem.

  NVIDIA provides the fundamental accelerated-computing architecture.

  OEM and ODM partners engineer and manufacture the systems.

  Thermal companies help build the liquid infrastructure.

  Networking companies connect the racks.

  Data-center companies provide power and heat rejection.

  Cloud operators finally turn the hardware into usable AI compute.

  DATAAD Supplier View

  If You Prioritize...	Companies Worth Watching
  Enterprise integration and global support	Dell / HPE
  Rapid high-density DLC deployment	Supermicro
  Mass-scale manufacturing	Foxconn
  Hyperscale ODM architecture	QCT / Wiwynn
  Branded rack and pod-scale platform	GIGABYTE

  This is an editorial characterization of supplier positioning, not a product recommendation. Actual selection depends on geography, availability, system configuration, support requirements and procurement model.

  The Bigger Story: The AI Server Is Becoming Infrastructure

  GB300 reveals something bigger than a new NVIDIA product cycle.

  The server itself is disappearing into the infrastructure.

  When one rack requires approximately 100 kW-class power, liquid cooling, specialized networking and coordinated software, it begins to look less like conventional IT equipment and more like industrial machinery.

  That brings new industries into the server ecosystem:

  Power + Cooling + Fluid Interfaces + HVAC + Controls.

  This is also why companies outside traditional IT — from CDU manufacturers to electrical-equipment companies — are moving aggressively into AI data centers.

  DATAAD Insight

  The title of this article asks:

  U.S. or Taiwan?

  But GB300 may show why that question is becoming outdated.

  The U.S. remains dominant in AI compute architecture, software and enterprise technology platforms.

  Taiwan remains extraordinarily important in turning those architectures into physical servers and racks at scale.

  And both sides are expanding into the other's traditional territory.

  OEMs are becoming infrastructure integrators.

  ODMs are becoming AI-factory solution providers.

  Server companies are becoming cooling companies.

  Manufacturers are becoming system architects.

  The future AI server may be designed in America, engineered across both sides of the Pacific, manufactured through a global supply chain and delivered as an integrated liquid-cooled AI factory.

  So the real competition may not ultimately be:

  U.S. vs. Taiwan.

  It may be:

  Which ecosystem can integrate compute, power, cooling and manufacturing fastest?

  ---

  DATAAD Analysis
  Source basis includes official NVIDIA, Dell Technologies, HPE, Supermicro, Foxconn, QCT, Wiwynn and GIGABYTE product and company materials. Country references indicate corporate base and ecosystem positioning rather than the physical manufacturing location of every component or system.