Nvidia Unveils Seven Chips to Scale the World’s Largest AI Factories

Nvidia unveils seven chips for the Vera Rubin platform. These chips are designed to help build the world’s biggest AI factories. The global AI infrastructure market is growing rapidly, with experts predicting it will increase from $35.42 billion in 2023 to $45.49 billion in 2024:

Year	Market Size (USD)
2023	35.42 billion
2024	45.49 billion

Nvidia leads this growth, holding about 92% of the global GPU market. The company creates crucial technology for AI development, showcasing new ideas that support next-gen agents. These innovations are transforming AI factories around the globe.

Key Takeaways

• Nvidia's Vera Rubin platform has seven new chips. These chips help AI factories work better. They give up to ten times more power for each watt used.

• The Vera CPU and Rubin GPU work as a team. They help big AI models run faster. They also use less energy. This makes AI work quicker and costs less money.

• Nvidia's chips let people connect up to 576 GPUs. This helps train AI models much better. It also makes AI answers faster and stronger.

• The new hardware uses less energy. Each server can save up to 30% on power. This makes AI cheaper and better for the planet.

• All seven chips will be ready by late 2026. Businesses can use them to build smarter AI. This helps them stay ahead as AI keeps changing.

Nvidia Unveils Seven Chips for AI Factories

Overview of the Vera Rubin Platform

The Vera Rubin platform shows how nvidia unveils seven chips to change ai. This platform has hardware parts that work together as a team. There are racks with Rubin GPUs and Vera CPUs inside. These racks can give you up to ten times more inference power for each watt used. The NVIDIA Groq 3 LPX inference accelerator is also included. It can give up to 35 times more inference power for each megawatt. This helps with really big ai models that have trillions of parameters. The BlueField-4 STX storage racks make GPU memory bigger. This makes it easier to use large language models. Spectrum-6 SPX Ethernet racks move data fast and with little delay.

Feature/Component	Description
Vera Rubin NVL72 GPU racks	72 Rubin GPUs, 36 Vera CPUs, up to 10x higher inference throughput per watt
Vera CPU racks	256 Vera CPUs, scalable and energy-efficient for agentic ai workloads
NVIDIA Groq 3 LPX inference accelerator	Up to 35x higher inference throughput per megawatt, for trillion-parameter models
NVIDIA BlueField-4 STX storage racks	AI-native storage, enhances GPU memory for large language models
NVIDIA Spectrum-6 SPX Ethernet racks	Low-latency, high-throughput connectivity for fast data movement

Nvidia unveils seven chips to build a strong ai supercomputer. These chips work together to help big ai factories. The Vera Rubin platform hardware uses less energy. It costs less for each token and works faster than old platforms.

Feature	Vera Rubin Platform	Previous Nvidia Platform (Blackwell)
GPU Requirement	One-quarter the number of GPUs	Higher GPU requirement
Inference Throughput	Up to 10 times higher per watt	Lower throughput
Cost per Token	One-tenth the cost	Higher cost
CPU Efficiency	Twice the efficiency	Traditional CPU efficiency
Speed	50% faster at rack scale	Slower processing
Token Throughput	Five times higher with STX	Lower throughput
Energy Efficiency	Four times greater	Lower efficiency

Purpose and Vision for Next Generation of AI

Nvidia unveils seven chips to help you build new ai factories. The company wants you to use both real and virtual systems to make smart places. These factories will keep getting better at speed, energy use, and being green. There will be a big change in how power is used in ai hardware. Dion Harris from nvidia says that running big ai factories needs new ideas about power.

Huang’s vision is clear: “Every industry, every company that has factories will have two factories in the future. The factory for what they build, and the factory for the mathematics, the factory for the AI. Factory for cars, factory for AIs for the cars. Factory for smart speakers, and factories for AI for the smart speakers.”

With nvidia unveils seven chips, you get hardware that helps this vision. These new chips bring fresh ideas to ai hardware. You can now grow your ai projects faster and use less energy. Nvidia gives you tools to lead in the next wave of ai.

The Seven Chips Powering AI Supercomputers

Vera CPU: High-Performance AI Processing

The Vera CPU is a strong processor for AI projects. It gives you great single-thread performance and lots of memory bandwidth. This chip uses half the energy of older CPUs. One rack can support over 22,500 environments at the same time. It finishes tasks 50% faster than regular CPUs. Nvidia made this chip for new AI jobs like reinforcement learning.

Feature	Description
Performance	Great single-thread performance and lots of memory bandwidth
Energy Efficiency	Uses half the energy of older CPUs
Scalability	Supports over 22,500 environments in one rack
Speed	Finishes tasks 50% faster than regular CPUs
Design	Made for new AI jobs like reinforcement learning

Rubin GPU: Accelerating AI Workloads

The Rubin GPU helps make AI work faster. It has 336 billion transistors, which is more than before. You get 288GB of HBM4 memory and 22 TB/s of memory bandwidth. The NVLink bandwidth is now 3.6 TB/s, which is 50% better. The transformer engines are fourth-generation and can scale up or down. Speculative decoding makes conversational AI three to four times faster. You do not need to copy tensors because memory is shared.

Feature	Rubin GPU	Previous Nvidia GPUs
Transistor Count	336 billion	208 billion (Blackwell)
Memory Capacity	288GB HBM4	Lower capacity
Memory Bandwidth	22 TB/s	Lower bandwidth
NVLink Bandwidth	3.6 TB/s (50% better)	NVLink 5
Transformer Engines	Fourth-generation, can scale	Previous generation
Speculative Decoding	3-4x faster for conversational AI	Not available
Memory Coherency	No need to copy tensors	Needed explicit transfers

NVLink 6 Switch: High-Speed AI Connectivity

The NVLink 6 Switch connects your AI chips fast. It has built-in compute to speed up group tasks. You get better service features. The switch also makes your AI factory more reliable.

• Built-in compute speeds up group tasks

• Better service features

• More reliable for your AI factory

ConnectX-9 SuperNIC: Advanced Networking for AI

The ConnectX-9 SuperNIC gives your AI fast networking. Each GPU gets 1.6 terabits per second of bandwidth. Data moves quickly because of low latency. The SuperNIC supports programmable RDMA. You can use GPU-direct networking for big operations.

Feature	Description
Bandwidth	1.6 terabits per second for each GPU
Latency	Low latency for fast data
RDMA Support	Programmable remote direct-memory access
Networking Capability	GPU-direct networking for big operations

BlueField-4 DPU: Data Processing for AI Factories

The BlueField-4 DPU boosts data processing in your AI factory. It helps your supercomputer by making storage, networking, and security faster. The chip also supports elastic scaling. It makes your AI setup stronger and more powerful.

Spectrum-6 Ethernet Switch: Scalable AI Networking

The Spectrum-6 Ethernet Switch helps your AI network grow. It improves traffic between racks in your AI factory. You can set it up with Spectrum-X Ethernet or NVIDIA Quantum-X800 InfiniBand switches. The switch gives you fast and smooth connections. Spectrum-X Ethernet Photonics uses less power and is ten times more reliable than old transceivers.

• Improves traffic between racks in AI factories

• Can use Spectrum-X Ethernet or NVIDIA Quantum-X800 InfiniBand switches

• Gives fast and smooth connections

• Spectrum-X Ethernet Photonics uses less power and is more reliable

Groq 3 LPU: Specialized AI Acceleration

The Groq 3 LPU gives special speed for AI. It has 128 GB of on-chip SRAM for big models. The chip is made for low-latency inference, so answers come fast. You can handle trillion-parameter models with almost no delay. This makes real-time AI possible for hard jobs.

Feature	Description
On-chip SRAM	128 GB of SRAM for big models
Latency	Made for low-latency inference, so answers come fast
Model Handling	Handles trillion-parameter models with almost no delay

When you use all seven Nvidia chips together, you get a strong AI system. The Vera Rubin NVL72 system has 72 Rubin GPUs and 36 Vera CPUs with fast NVLink 6 links. You can train big models with fewer GPUs and get ten times more throughput for less money. The Vera CPU Rack makes reasoning tasks faster, and the BlueField-4 STX storage rack helps by moving cache data. This setup builds a strong AI factory that does hard jobs well.

Impact on AI Infrastructure and Industry

Performance and Efficiency Gains

The new Nvidia hardware makes AI work much better. You get more speed and use less energy. Large language models run four times faster with the same GPUs. If you use 10,000 GPUs, you get seventy-three times more speed. The new TensorRT-LLM software makes Blackwell Ultra GPUs up to 2.7 times faster.

Benchmark Type	Performance Gain
LLMs	4 times faster with same GPUs
LLMs	73 times faster with 10,000 GPUs

Software Update	Performance Gain
TensorRT-LLM	Up to 2.7 times faster on Blackwell Ultra GPUs

AI factories now save more energy. Using DPUs can lower energy use by thirty percent for each server. Liquid cooling works twenty percent better than air cooling. Nvidia GPUs use twenty times less energy for some AI and HPC jobs than old CPUs. Blackwell chips give three to five times more AI work in places with power limits. Even small energy savings can save a lot of money.

Scalability for AI Supercomputers

You can make your AI setup bigger to build huge AI supercomputers. NVLink lets you join up to 576 GPUs for better AI model speed. In a group of 72 GPUs, you get 130TB/s GPU bandwidth. SHARP FP8 support gives four times more bandwidth efficiency. Multi-server support lets you build clusters that go past one server with 1.8TB/s interconnect. The NVL72 system gives nine times more GPU speed than an eight-GPU system.

Feature	Description
NVLink Scalability	Grows up to 576 GPUs for better performance
Bandwidth	130TB/s GPU bandwidth in a 72-GPU group
Efficiency	4 times better bandwidth with SHARP FP8 support
Multi-server Support	Makes clusters bigger than one server
Throughput	NVL72 gives 9 times more GPU speed

You can spread AI jobs across many data centers. The NVIDIA Vera Rubin NVL72 system has 72 Rubin GPUs and 36 Vera CPUs for big AI jobs. HGX Rubin NVL8 links eight Rubin GPUs to help with training and inference. Spectrum-6 Ethernet makes networking and backup better in cloud data centers. Spectrum-XGS Ethernet lets many data centers work together as one AI system.

Industry Adoption and Ecosystem Support

Many companies now use Nvidia’s new ideas. Hardware, storage, system, and cloud companies use the Vera Rubin platform.

Partner Type	Company Name
Hardware Vendor	Supermicro
Storage Vendor	Vast Data
Storage Vendor	DDN
System Vendor	Dell Technologies
System Vendor	Lenovo
Cloud Provider	CoreWeave
Cloud Provider	Nebius
Cloud Provider	Microsoft
Cloud Provider	Red Hat

Red Hat supports the Vera Rubin AI platform right away. Cloud companies like CoreWeave, Nebius, Microsoft, and Red Hat grow their systems to handle more AI jobs. You get help from a growing group that lets you build and grow AI factories.

"We made the most advanced AI chips ever, in the best factory, here in America for the first time. This is just the start."

Nvidia’s CEO, Jensen Huang, says these new AI chips will change every industry. You can use them to build smarter factories and make AI agents better at thinking, working, and learning. The new processor and hardware help you lead in AI infrastructure.

Availability and Future of AI Factories

Production Status and Market Rollout

All seven Nvidia chips for big ai factories will be ready by late 2026. Each chip, like the Vera Rubin NVL72 GPU racks and ConnectX-9 SuperNIC, is being made as planned. Here is a simple chart that shows their progress:

Chip Type	Status	Availability
Vera Rubin NVL72 GPU racks	Full Production	Second half of 2026
Vera CPU racks	Full Production	Second half of 2026
NVIDIA Groq 3 LPX inference	Full Production	Second half of 2026
NVIDIA BlueField-4 STX storage	Full Production	Second half of 2026
NVIDIA Spectrum-6 SPX Ethernet	Full Production	Second half of 2026
NVIDIA NVLink 6 Switch	Full Production	Second half of 2026
NVIDIA ConnectX-9 SuperNIC	Full Production	Second half of 2026

• The Vera Rubin platform and its chips are taking longer to come out. There might be more delays in the next three months.

• These delays could make Nvidia lose some market share.

• Nvidia is moving from H200 chips to the Vera Rubin design, which means fewer people want the old chips.

Note: If you want these new ai systems, plan for late 2026 to use them in your next ai data centers.

Partnerships and Ecosystem Expansion

Nvidia is teaming up with many partners to grow the ai world. Microsoft will use Vera Rubin NVL72 systems in its new ai data centers. CoreWeave will add Rubin systems to its ai cloud in 2026. Nebius will use Rubin for its AI Cloud and Token Factory. Server companies like Cisco, Dell, HPE, Lenovo, and Supermicro are making new servers with Rubin chips. Big ai labs like OpenAI and Meta are using Rubin for advanced ai models.

• Microsoft is making Azure’s ai stronger with Rubin.

• CoreWeave will start using Rubin in 2026.

• Nebius is putting Rubin in its ai setup.

Nvidia is also working with research labs and cloud companies. For example, Together AI and 5C are running ai factories with Nvidia GPUs in Maryland and Memphis. Argonne National Laboratory uses Nvidia systems for research. The Omniverse DSX project helps build smart and flexible buildings.

Outlook for the Next Generation of AI

The Vera Rubin platform will change how you build and run ai factories. Experts think this platform will make ai better and cost less. You will see new things in robotics, self-driving factories, and national security. The hardware will need new software to work its best, especially for jobs that need to happen at different times. Trillion-parameter models will soon handle real-time, multi-modal data with almost no wait.

Companies like HPE and Equinix are getting ready for what’s next. HPE is adding new Nvidia GPUs to help with big ai jobs. Equinix is starting the NVIDIA Instant AI Factory, which gives you ready-to-use ai tools for fast setup and better results.

The future of ai will depend on how well you can grow, change, and create with Nvidia’s newest hardware. Countries and companies are building Rubin-based clusters to stay ahead in a world powered by ai.

Nvidia’s seven new chips and the Vera Rubin platform make ai much better. You get faster answers, spend less money, and can use bigger models. The table below shows some important features:

Feature	Specification
Memory Bandwidth	Up to 13 TB/s (HBM4)
Transistor Count	500 billion per chip
Tensor Cores	20,000 per chip
Cooling Technology	Two-phase liquid cooling

Now, you can train smart models with fewer GPUs. Running ai factories is easier and uses less energy. Nvidia works with Corning to help US factories and make the ai supply chain stronger. The Vera Rubin platform is a new top choice for fast and green ai supercomputers.

These tools let you help shape the future of ai. Nvidia leads the way, so you can build smarter models and better ai factories.

 

 

 

 

 

Written by Jack Elliott from AIChipLink.

 

AIChipLink, one of the fastest-growing global independent electronic   components distributors in the world, offers millions of products from thousands of manufacturers, and many of our in-stock parts is available to ship same day.

 

We mainly source and distribute integrated circuit (IC) products of brands such as Broadcom, Microchip, Texas Instruments, Infineon, NXP, Analog Devices, Qualcomm, Intel, etc., which are widely used in communication & network, telecom, industrial control, new energy and automotive electronics. 

 

Empowered by AI, Linked to the Future. Get started on AIChipLink and submit your RFQ online today!