NVIDIA, Google, Meta, AMD and Cisco Unveil Open AI Data Center Blueprints

NVIDIA, Google, Meta, AMD and Cisco have announced new open infrastructure initiatives aimed at making it easier to build and scale next-generation artificial intelligence data centers. The announcements, made around the Open Compute Project (OCP) Global Summit on October 12, 2026, focus on standardized rack designs, high-speed networking, advanced cooling and power delivery for increasingly demanding AI workloads. The effort reflects a broader industry shift toward shared infrastructure specifications that allow hardware manufacturers and cloud operators to develop compatible systems more efficiently.

NVIDIA is contributing key Vera Rubin rack and compute tray designs to the OCP through its modular MGX architecture, while AMD has introduced its own rack-scale infrastructure platform for agentic AI. Google, Meta and Cisco are contributing to related networking and interoperability initiatives. The objective is to reduce engineering complexity and make it easier for data center operators to deploy AI infrastructure at scale. Although the designs are being opened to broader industry participation, this does not mean that every component or software layer will become open source or interchangeable across all vendors.

Key Takeaways

• Open infrastructure: NVIDIA is contributing Vera Rubin rack and compute tray designs to the Open Compute Project through its MGX architecture.
• AMD’s alternative: AMD has introduced an open rack-scale platform called Helion for demanding AI workloads.
• Meta’s networking upgrade: Meta is adopting NVIDIA Spectrum-6 Ethernet switches for its next-generation AI infrastructure.
• Google partnership: Google and NVIDIA are advancing Falcon Transport support for future networking products.
• Cisco and NVIDIA optics: The companies are developing an open specification for a new 400-gigabit-per-lane optical module design.
• Higher-density racks: NVIDIA’s updated MGX 1.2 architecture supports rack deployments of up to 250 kilowatts.
• Industry objective: Shared designs could shorten deployment timelines, improve interoperability and simplify cooling, power delivery and maintenance.

What Are the Open MGX Blueprints?

NVIDIA MGX is a modular reference architecture for designing AI servers and data center systems. It provides a common framework that manufacturers can use to develop computing platforms with different combinations of processors, accelerators, memory, networking, power delivery and cooling.

Instead of requiring every hardware manufacturer to engineer an entire rack from scratch, MGX provides standardized building blocks and reference designs. System manufacturers can adapt those designs to specific workloads while retaining compatibility with a broader infrastructure ecosystem.

At the OCP Global Summit, NVIDIA contributed key Vera Rubin rack and compute tray designs through its open MGX architecture. The contribution forms part of the OCP’s Open DC Initiative, which aims to encourage shared approaches to data center infrastructure.

The updated MGX 1.2 specification introduces changes designed for higher-density AI systems. These include rack-scale liquid cooling, interoperable rack connections, standardized power and cooling interfaces, configurable input/output bays and support for higher-power infrastructure.

The architecture also incorporates a liquid-cooled 5,000-amp busbar designed for next-generation systems. A busbar distributes electrical power within equipment, and its capacity becomes increasingly important as accelerator racks consume more electricity.

According to NVIDIA, MGX 1.2 supports rack deployments of up to 250 kilowatts. The architecture is intended to help manufacturers build, test, service and deploy systems more efficiently while preserving flexibility across different configurations.

Why AI Data Centers Need Standardized Designs

Traditional data centers were built around comparatively conventional server racks. Modern AI infrastructure increasingly combines large numbers of accelerators, high-bandwidth memory, specialized networking and advanced cooling systems.

Training and operating large AI models can require thousands of processors to work together. AI agents add another layer of demand because they may execute repeated inference requests, call tools, process information and run multiple tasks concurrently.

As a result, data center performance depends on more than individual chips. Power delivery, cooling capacity, networking bandwidth and the physical arrangement of hardware all influence how effectively a facility can operate.

Standardized blueprints could help address several challenges:

• Faster deployment: Manufacturers can reuse validated design components instead of developing every rack configuration independently.
• Simplified maintenance: Standard interfaces and modular components can make repairs and replacements easier.
• Cooling efficiency: Shared specifications can help operators integrate liquid-cooling systems with high-density computing equipment.
• Supply-chain flexibility: Common interfaces can make it easier to source compatible components from multiple suppliers.
• Infrastructure scalability: Repeatable designs can help data center operators expand capacity as AI demand increases.

However, open reference designs do not automatically guarantee compatibility. Manufacturers must still validate electrical, thermal, mechanical and software requirements before combining components from different suppliers.

NVIDIA Opens Vera Rubin Rack and Compute Tray Designs

NVIDIA’s contribution centers on the infrastructure surrounding its Vera Rubin AI computing platform. Vera Rubin is designed for large-scale AI training and inference, and its rack architecture integrates computing, networking, power delivery and cooling into a coordinated system.

The company says the open MGX approach can help partners adapt the underlying infrastructure to different deployment requirements. This is particularly relevant as AI data centers move toward racks that consume substantially more electricity than conventional enterprise server configurations.

The updated architecture includes liquid-cooled compute trays, standardized rack-scale connections and higher-density configurations. It also supports native 19-inch equipment integration, allowing the design to accommodate compatible equipment used in existing data center environments.

NVIDIA is also extending the MGX architecture beyond GPU-heavy systems. Its Vera CPU rack design is intended to support workloads associated with AI agents, including tool execution, orchestration, analytics, cloud services and storage.

These tasks can require substantial CPU resources alongside accelerator capacity. A specialized CPU rack could help operators allocate different types of computing resources to the workloads that need them.

Opening reference designs through OCP could encourage system manufacturers to build compatible infrastructure without independently recreating every element of NVIDIA’s rack architecture. However, access to the physical design specifications does not mean NVIDIA’s complete hardware and software stack is open source.

AMD Introduces Helion for Rack-Scale AI Infrastructure

AMD is also expanding its role in AI infrastructure through its Helion rack-scale platform. Announced at the OCP Global Summit, Helion is designed to support agentic AI workloads using AMD EPYC processors and Instinct accelerators.

The platform follows a rack-level approach, combining computing hardware and infrastructure components into a coordinated system. AMD is contributing the design to the Open Compute Project, allowing partners to build on a common framework.

AMD says the platform is intended to accommodate different deployment configurations, including CPU-only and hybrid systems. This flexibility could help data center operators adapt infrastructure to a mixture of AI inference, model execution and general-purpose computing workloads.

The platform is also designed to support air- and liquid-cooled configurations, reflecting the different thermal requirements of AI deployments.

Infrastructure initiative	Main focus	Potential benefit
NVIDIA MGX 1.2	Modular rack architecture, cooling and power	Standardized systems for high-density AI
AMD Helion	Rack-scale systems built around EPYC CPUs and Instinct accelerators	Flexible AI infrastructure configurations
Meta Spectrum-6 adoption	High-capacity Ethernet networking	Greater bandwidth for large AI clusters
Google-NVIDIA Falcon Transport	Networking support for future SuperNICs and DPUs	Improved transport options for AI infrastructure
Cisco-NVIDIA USFP	High-density optical connectivity	Greater bandwidth density in networking systems

AMD’s approach gives system manufacturers another rack-scale design to consider as they develop AI infrastructure. Whether it gains broad adoption will depend on hardware availability, software compatibility, system performance and the willingness of data center operators to deploy the platform.

Meta Adopts NVIDIA Spectrum-6 for AI Networking

Meta is adopting NVIDIA Spectrum-6 Ethernet switches for its next-generation AI infrastructure. The networking platform is designed to support large AI clusters in which thousands of processors must exchange data efficiently.

NVIDIA says Spectrum-6 delivers 102.4 terabits per second of switching capacity and supports 200-gigabit-per-lane Ethernet connectivity. The system is intended to improve the bandwidth and reliability required by large-scale AI training and inference.

Meta’s implementation includes a switch called N-IceCube, designed to operate within its IcePack liquid-cooled networking rack. The switch runs Meta’s Facebook Open Switching System, or FBOSS, an open-source network operating system contributed to OCP.

Meta and NVIDIA are also advancing Spectrum-X Multiplane networking technology. The approach distributes traffic across multiple independent network planes, creating additional paths between computing resources.

Multiple network paths can improve utilization and resilience. If a component fails, traffic may be redirected through other available paths, helping reduce the risk that a single failure disrupts an entire AI workload.

Meta’s infrastructure operates at a scale where networking problems can affect the productivity of large numbers of accelerators. If processors spend too much time waiting for data, the overall system may deliver less useful computing output even when the processors themselves are powerful.

The companies say the combination of Spectrum-6, ConnectX networking and Multiplane technology is intended to help scale AI factories beyond 100,000 GPUs. That figure describes the intended infrastructure scale, not a guarantee that every deployment will achieve the same performance.

Google and NVIDIA Advance Falcon Transport

Google and NVIDIA are also collaborating on Falcon Transport support for future NVIDIA ConnectX-10 SuperNICs and BlueField data processing units.

SuperNICs are specialized network interface devices designed to improve communication between servers and accelerators. Data processing units, or DPUs, can offload networking, security and infrastructure tasks from general-purpose CPUs.

These components are increasingly important in AI systems because distributed computing workloads can generate substantial network traffic. The network must move data between processors, storage systems and other infrastructure components without creating excessive latency or congestion.

The Google-NVIDIA collaboration is part of the wider push to develop transport technologies that can support increasingly large AI clusters.

However, the announcement does not mean that Google is abandoning its own AI accelerators or networking systems. Google has developed its own Tensor Processing Units and operates a large cloud infrastructure ecosystem.

The partnership instead illustrates how companies can collaborate on selected infrastructure layers while continuing to compete in other parts of the AI market.

Cisco and NVIDIA Develop a New Optical Module Specification

Cisco and NVIDIA have announced plans to develop an open industry specification for a new Ultra-Dense Small Form Factor Pluggable optical module, known as USFP.

The proposed design targets 400-gigabit-per-lane connectivity. Optical modules convert electrical signals into optical signals and are essential for transmitting large volumes of data across networking equipment.

As AI clusters grow, network bandwidth requirements are increasing rapidly. Conventional approaches can require more physical modules, ports and cabling as systems scale, creating challenges involving space, power consumption and cooling.

Cisco and NVIDIA say the proposed USFP design could deliver four times the bandwidth density and approximately 1.3 times the power efficiency of 200-gigabit-per-lane optical modules.

These figures are company-reported expectations for the planned design. The technology will need to progress through specification development, implementation and validation before its practical advantages can be assessed independently.

The companies say the objective is to preserve the serviceability and deployment flexibility of pluggable optics while increasing bandwidth. An open specification could encourage additional manufacturers to develop compatible components, although broad adoption is not guaranteed.

Cisco’s broader OCP announcements also include advances in switching silicon, networking systems and infrastructure management. Together, the initiatives illustrate how networking has become a central part of AI data center design rather than a supporting component added after the computing hardware is selected.

Power and Cooling Become Critical AI Infrastructure Challenges

Power delivery and thermal management are among the most important constraints facing next-generation AI facilities.

High-performance accelerators consume substantial electricity and generate considerable heat. When many accelerators operate together in a densely packed rack, conventional cooling methods may not be sufficient to maintain operating temperatures.

Liquid cooling can transfer heat more effectively in high-density systems, but it introduces additional engineering requirements involving plumbing, leak prevention, maintenance and compatibility between components.

NVIDIA’s MGX 1.2 design addresses these issues through standardized liquid-cooling interfaces and rack-scale integration. Its support for deployments of up to 250 kilowatts also reflects the direction of AI infrastructure, where individual racks can consume far more power than traditional server racks.

The company is also promoting an 800-volt direct-current power architecture for future AI factories. Higher-voltage distribution can reduce current for a given power level, potentially lowering electrical losses and simplifying parts of the power-delivery system.

Such infrastructure changes require coordination across data center operators, electrical equipment suppliers and computing hardware manufacturers. Open specifications could help those groups plan compatible systems, but actual efficiency gains will depend on facility design and operating conditions.

What Open AI Infrastructure Means for the Technology Industry

The announcements demonstrate that competition in AI infrastructure is increasingly taking place at the system level. Companies are not only competing to build faster processors; they are also developing the networking, cooling, power delivery and software infrastructure needed to deploy those processors efficiently.

For manufacturers, open blueprints can reduce duplicated engineering work and provide a shared starting point for new products. For cloud operators, common designs could make it easier to compare configurations, expand capacity and work with a broader range of suppliers.

For NVIDIA, opening selected infrastructure specifications may help extend adoption of its architecture across the ecosystem. The company can encourage system manufacturers to build around its computing and networking technologies while contributing to common standards.

AMD’s Helion initiative gives manufacturers another route to building rack-scale AI systems. Meta’s Spectrum-6 adoption demonstrates the importance of networking, while Google’s and Cisco’s partnerships show that interoperability can extend beyond server racks to network transport and optical connectivity.

The industry will still need to balance openness with proprietary differentiation. A reference design may be publicly available while important accelerators, software stacks and system optimizations remain controlled by individual companies.

The Bigger Picture

The move toward open AI infrastructure reflects the changing economics of data center construction. As AI models and agents require more computing capacity, companies must coordinate processors, memory, networking, cooling and power as parts of one system. Standardized designs could help reduce deployment complexity and support a wider supplier ecosystem, while allowing operators to adapt infrastructure to different workloads.

NVIDIA, AMD, Google, Meta and Cisco are approaching the challenge from different positions, but their announcements point toward greater collaboration on selected infrastructure standards. The long-term impact will depend on whether the specifications gain adoption, whether compatible components become widely available and whether the resulting systems deliver measurable improvements in cost, reliability and energy efficiency.

Looking Ahead

The next stage will involve manufacturers and infrastructure operators translating reference designs into validated products and production deployments. Important milestones include adoption of the updated MGX specification, the development of AMD’s Helion ecosystem, implementation of the proposed USFP optical module and the rollout of next-generation networking systems. The announced specifications provide a foundation, but deployment timelines and performance will depend on manufacturing readiness, validation and customer demand.

For the AI industry, the shift toward shared blueprints could make it easier to expand computing infrastructure as demand grows. Open designs may lower engineering barriers and improve interoperability, but they will not eliminate the need for specialized chips, software optimization or reliable power and cooling. The companies that successfully combine open standards with dependable, high-performance systems will be better positioned to support the next generation of AI services.