2025 NVIDIA Blackwell Leads the Stargate Project: A New Era of AI Innovation

NVIDIA Blackwell Leads the Stargate Project: A New Era of AI Innovation

Introduction: The Emergence of a New AI Paradigm

NVIDIA’s latest AI chip architecture, Blackwell, has been chosen as the core engine for the ultra-scale AI infrastructure project Stargate. This groundbreaking project, driven by investments totaling $500 billion (approximately 720 trillion KRW) from OpenAI, Oracle, SoftBank, and Abu Dhabi’s AI investment firm MGX, aims to build massive data center campuses across the United States over the next four years. At the heart of this infrastructure lies the NVIDIA Blackwell-based GB200 GPU. Blackwell is celebrated as a game-changer in AI training and inference, offering up to a fivefold improvement in AI computational performance and revolutionary memory expansion compared to previous generations. Meanwhile, following the ChatGPT phenomenon in 2023, demand for ultra-large AI models surged worldwide, leading global cloud infrastructures to face shortages in AI computing resources. To address this challenge and reinforce AI supremacy, OpenAI and its partner companies have initiated massive investments—launching projects like Stargate. NVIDIA, by fulfilling this enormous demand with its Blackwell architecture, is playing a pivotal role in AI innovation. This post provides an in-depth examination of the features of the NVIDIA Blackwell architecture, the scale and significance of the Stargate project, and the global market transformations it is poised to trigger.

For further insights on NVIDIA’s latest GTC announcements, check out our detailed article on NVIDIA GTC.

NVIDIA Blackwell: The Power of Next-Generation AI Chips

Unveiled in 2024, NVIDIA’s Blackwell is the latest AI GPU architecture that boasts industry-leading performance and technological innovation. Building upon the advances of previous generations such as Volta, Ampere, and Hopper, Blackwell represents the fifth generation of AI chips.

In March 2024, during the GTC keynote, NVIDIA CEO Jensen Huang introduced the new Blackwell GPU (on the left) alongside the previous-generation Hopper GPU (on the right).

Among the Blackwell series, the top-tier B200 GPU stands out as the largest chip to date, composed of over 20.8 billion transistors. It features an innovative design that unifies two massive chiplets into a single operational unit. As a result, the Blackwell B200 achieves a staggering 20 petaflops (PFLOPs) of AI computational performance on a single chip, optimized for cutting-edge AI workloads such as large language models (LLMs) through its 6th-generation Tensor Cores and 2nd-generation Transformer Engines. Moreover, each B200 GPU is equipped with 192GB of ultra-fast HBM3E memory, quadrupling the memory capacity of its predecessors and delivering a bandwidth of 8TB/s to handle vast datasets in real time.

A particularly notable feature of the Blackwell architecture is its support for next-generation NVLink interconnect. With 5th-generation NVLink, tens of GPUs within a single server can be clustered to function as one giant GPU. By employing dedicated NVLink switches, it is possible to connect up to 576 GPUs at ultra-high speeds, forming an enormous computing domain. Structures such as NVL72, implemented in Blackwell-based systems, can combine 72 GPUs as a single node, enabling real-time inference on models with more than a trillion parameters. Thanks to this scalability, using Blackwell GPUs allows the construction of AI supercomputers on a much larger scale than before while minimizing data transfer bottlenecks between GPUs to maximize training efficiency.

In addition, NVIDIA has introduced an integrated module—the GB200 Superchip—that directly connects the Blackwell GPU with NVIDIA’s server CPU, Grace, via high-speed links. The GB200 module packages two B200 GPUs with one Grace CPU connected by NVLink, facilitating data movement between CPU and GPU much faster than PCIe and reducing bottlenecks during large-scale distributed training. The Grace CPU, based on ARM architecture, is designed for high-bandwidth memory access and is optimized to work in tandem with NVIDIA GPUs, building upon the technology proven in the GH200 (Grace-Hopper) Superchip. In summary, the Blackwell architecture elevates not only hardware performance but also the overall efficiency of system architecture, taking AI infrastructure scalability to a new level.

Key Technical Features of the Blackwell B200/GB200:

  • Ultra-Large Chiplet Structure: A GPU composed of 20.8 billion transistors across two chiplets, the largest chip available.
  • Massive Memory Capacity: Equipped with 192GB HBM3E memory delivering an 8TB/s bandwidth for handling vast data in real time.
  • Exponential Computational Performance: Achieves 20 PFLOPs in AI computations, optimized for LLMs with support for ultra-low precision FP4/FP8 operations.
  • Scalable Architecture: Utilizes 5th-generation NVLink to connect up to 576 GPUs, with configurations such as NVL72 supporting massive cluster setups.
  • CPU-GPU Integration: The GB200 Superchip configuration with Grace CPU minimizes CPU-GPU bottlenecks.
  • Reliability and Security: Enhanced by dedicated RAS engines and Confidential Computing technology.

Thanks to these overwhelming specifications, the Blackwell GPU has captured the market’s attention immediately upon release, driving record growth in NVIDIA’s Data Center revenue. In fact, NVIDIA achieved a 142% year-over-year increase in Data Center revenue for the fiscal year 2025, reaching $115.2 billion, with new Blackwell chip sales alone accounting for $11 billion in Q4 2025 (October–December)—the fastest product sales growth in the company’s history. NVIDIA CFO Colette Kress stated, “Blackwell is being adopted at an unprecedented pace and scale in our history, and with its integration, building clusters of 100,000 GPUs will soon become commonplace.” Thus, the advent of Blackwell is positioned as a game-changer that will rapidly transform the AI computing industry.

The Stargate Project: The Dawn of Ultra-Scale AI Infrastructure

Announced in January 2025, the Stargate project is an ultra-scale AI data center initiative designed to meet the explosive AI compute demand from OpenAI. Launched as a joint venture by SoftBank Chairman Son Jeong-ui, OpenAI, Oracle, and Abu Dhabi’s investment firm MGX—with SoftBank providing financial backing and OpenAI leading operations—this project is unprecedented in scale, with an overall investment of $500 billion. Immediately following the announcement, an initial $100 billion was injected, triggering the construction of the first data center campus. The U.S. government has also taken notice; during a White House event, President Donald Trump declared, “We will build the world’s most powerful AI system in America,” expressing full support for Stargate. This strategic investment aims to outpace competing nations like China in the AI arms race.

The first data center campus under the Stargate project is located in Abilene, Texas, with site preparation and construction already underway since the second half of 2024. Reports indicate that this campus will eventually feature between 10 and 20 individual data center buildings constructed in phases, making it the largest AI supercomputing center in the world when completed. Larry Ellison, Oracle Chairman, has mentioned on site that “ten data centers are already under construction,” while OpenAI is reportedly evaluating additional sites across 16 U.S. states. Each campus is planned as a massive facility consuming gigawatts (GW) of power; with just 5–10 campuses, a colossal compute infrastructure spanning several GW will be established.

The uniqueness of the Stargate project lies not only in the physical scale of its data centers but also in the qualitative level of AI computing power they will house. OpenAI and its partners plan to invest the world’s best AI compute capabilities into this enormous infrastructure, which will be used for training and serving next-generation GPT models. To that end, tens of thousands of NVIDIA Blackwell GB200 GPUs will be deployed across the Stargate data centers. Market research suggests that the first phase facility in Abilene, Texas, will have approximately 64,000 NVIDIA GB200 chips installed and operational by the end of 2026. Considering that a single Blackwell GPU costs between $30,000 and $40,000 (roughly 40–50 million KRW), the hardware cost for GPUs in one facility is estimated to be in the trillions of KRW. Factoring in future campus expansions and additional site developments, the Stargate project is set to become the largest global AI computing infrastructure, with hundreds of thousands of GPUs deployed. The vast compute resources will primarily support OpenAI’s model training and inference and could also be made available to external companies or research institutions via Oracle’s cloud services. Given that NVIDIA and Oracle, both with extensive experience in operating massive GPU clusters, will manage this infrastructure, high reliability in operating and utilizing such an ultra-scale system is expected.

Why is the Stargate Project Deploying NVIDIA Blackwell GB200 at Scale?

Why did the Stargate project choose to deploy the NVIDIA Blackwell-based GB200 at such scale? The answer lies in Blackwell’s ability to provide the most powerful and scalable AI hardware platform currently available. OpenAI’s next-generation models consist of hundreds of billions to trillions of parameters, requiring GPUs capable of processing enormous computational loads. As previously mentioned, the Blackwell B200/GB200 achieves up to a fivefold performance increase over the previous Hopper (H100) generation, with quadrupled GPU memory capacity, making it easier to load entire large models into GPU memory. This translates to the ability to efficiently train larger batches and deeper models—improving training efficiency for ultra-large models.

Furthermore, the introduction of FP8 and FP4 low-precision operations along with enhanced transformer operation optimizations in the Blackwell architecture can dramatically boost the inference speed of large-scale language models such as ChatGPT. For example, servers based on Blackwell can process tens of thousands of tokens per second, significantly reducing latency in real-time conversational AI services and lowering operational costs. Even if OpenAI develops a next-generation model far larger than GPT-4, an infrastructure comprising tens of thousands of Blackwell GPUs is expected to sufficiently support it. According to NVIDIA’s announcement, an NVL72 node consisting of 72 Blackwell GPUs demonstrated over 30 times the real-time inference performance compared to the existing 8-GPU DGX servers, implying a substantial reduction in the number of physical servers required for large-scale model deployment. In other words, deploying Blackwell in Stargate will reduce the number of servers needed for equivalent computational tasks, thereby enhancing power and space efficiency and ultimately lowering operating costs.

Support from cutting-edge networking technologies such as NVIDIA Spectrum-X is another critical factor driving the large-scale adoption of Blackwell in Stargate. In these ultra-large GPU clusters, inter-GPU communication performance greatly influences overall system performance. NVIDIA, alongside launching Blackwell, introduced its high-performance Ethernet technology, Spectrum-X, which offers an optimized network solution for massive AI clusters. NVIDIA has announced that it will supply this Spectrum-X networking platform to the first data centers in the Stargate project. This will minimize communication latency between GPUs and, when combined with NVLink switches, will enable the formation of an ultra-fast AI fabric connecting tens of thousands of GPUs. The turnkey nature of this total solution provided by NVIDIA is a significant advantage for Stargate. In short, it is the comprehensive blend of top performance, remarkable scalability, and efficient large-scale operation—embodied in the Blackwell GB200 platform—that has made it the clear choice as the core of the Stargate project.

Global Impact on the AI Industry and Future Outlook

The adoption of NVIDIA Blackwell in the Stargate project carries significant implications for the global AI industry. Firstly, this case symbolizes the rapid expansion in the scale of investments in AI infrastructure. Whereas in the past, individual companies or institutions invested billions of dollars in AI infrastructure, the landscape is now evolving toward mega-projects involving consortia with investments totaling hundreds of billions of dollars, as demonstrated by Stargate. NVIDIA CEO Jensen Huang has described these enormous AI data centers as “AI factories” and emphasized that they will be at the heart of the next industrial revolution. This signals that the competition in AI model development has become a battle among ultra-large players, and those lacking immense capital and technology risk being left behind. In fact, some industry experts have even speculated that Samsung might join Stargate, potentially forming another major AI alliance—illustrating the disruptive impact of the announcement. Consequently, other global tech giants are expected to accelerate their own efforts in building AI supercomputers. Microsoft, Google, Meta, and even Chinese state-owned enterprises are already expanding or planning to expand data centers with tens of thousands of GPUs. Startups such as CoreWeave in the United States have also begun securing over 250,000 GPUs to challenge the AI cloud service market. To meet this explosive demand for Blackwell GPUs, NVIDIA is ramping up production, and if this trend continues, its dominant position in the AI semiconductor market is expected to remain firmly in place for some time.

The progress of the Stargate project is also poised to accelerate innovation across the data center industry as a whole. Data center campuses rated at 1GW will present new technical challenges in power supply, cooling, and networking. For instance, innovative large-scale liquid cooling systems or advanced air conditioning systems may be introduced to efficiently manage the heat generated by hundreds of thousands of high-performance GPUs. Stable power supplies will be ensured by integrating renewable energy generation, energy storage systems, and high-voltage direct current (HVDC) power distribution, using state-of-the-art electrical infrastructure. Ultimately, these advances will drive forward the evolution of data center technologies, benefiting not only AI but also general cloud services. Moreover, the construction of vast AI infrastructure in the United States is expected to have a positive impact on local economies by stimulating regional economic activity and generating high-tech jobs. Already, hundreds of data center operation personnel are being hired in the Abilene region, and each new campus is projected to create thousands of new jobs.

From a technological perspective, the advent of ultra-large clusters like Stargate could fundamentally change the paradigm of AI research and development. Whereas past efforts focused on overcoming model limitations through algorithm optimization and distributed learning techniques, the alleviation of hardware constraints now enables the direct implementation and experimentation of models with an unprecedented number of parameters. This could ultimately bring us one step closer to achieving AGI (Artificial General Intelligence) and provide the fertile ground for a quantum leap in AI model performance. Researchers at OpenAI and elsewhere may soon be able to attempt massive projects that were previously unthinkable, paving the way for scientific breakthroughs and the solution of complex problems through AI.

The coming years will require close attention to the progress of NVIDIA and the Stargate project. NVIDIA has already announced plans for its next-generation chip—Blackwell Ultra—and subsequent architectures (codenamed Vera Rubin), outlining a roadmap for exponential growth in AI computational power. Similarly, following the launch of its first campus, Stargate is expected to gradually expand additional campuses, completing its infrastructure within four years. In this process, the initial Blackwell-based equipment will be sequentially replaced or augmented with next-generation upgrades, ensuring that the evolving AI infrastructure remains at the forefront of technology. For example, after 2026, Blackwell Ultra chips or enhanced Grace CPUs might be introduced, further boosting cluster performance. Through the co-evolution of hardware and infrastructure, the AI industry is set to sustain rapid growth for the foreseeable future, and the close collaboration between NVIDIA and OpenAI is poised to wield significant influence across the industry.

Conclusion: The New Era of AI Innovation Unleashed by NVIDIA Blackwell and Stargate

The advent of the NVIDIA Blackwell-based GB200 has established a new milestone in AI hardware, and its integration with the Stargate project is set to unlock its full potential. This project—where the world’s leading AI chips meet unprecedented levels of capital and infrastructure—is expected not only to build massive data centers but also to trigger a paradigm shift in AI innovation. With NVIDIA’s years of expertise, the deep collaboration with OpenAI, and the strong reputation both companies hold in the industry, there is immense confidence in the success of this ultra-scale AI system.

Of course, enormous projects such as this come with their own set of challenges. Early reports of thermal issues in Blackwell-based systems, for instance, underscore the possibility of unforeseen hurdles accompanying the adoption of new technology. Additionally, sustainability issues related to the power consumption and heat generated by hundreds of thousands of AI chips must be addressed. Nonetheless, NVIDIA, OpenAI, and other industry players are actively pursuing further technological advancements and infrastructure innovations to overcome these challenges.

Once the Stargate infrastructure—comprising hundreds of thousands of NVIDIA Blackwell GPUs—becomes fully operational, we will witness a new era where the speed and capability of AI model training and inference will be unparalleled. This will lead to a substantial increase in AI utilization across industries, spurring innovative services and products in fields such as autonomous driving, healthcare, finance, and manufacturing. The convergence of NVIDIA Blackwell and the Stargate project is a landmark event in AI history, accelerating the pace of AI development to unprecedented levels. As we continue to follow the progress and results of this colossal project, it becomes imperative to closely monitor how this AI infrastructure will transform the future of humanity. The Stargate project and NVIDIA’s endeavors in pushing AI frontiers prepare us for the next chapter in the AI revolution.

Frequently Asked Questions (FAQ)

What is the Stargate Project?

The Stargate Project is an ultra-scale AI data center initiative launched in 2025 to support the explosive AI computing demands of OpenAI. Jointly invested in by SoftBank, OpenAI, Oracle, and the investment firm MGX from Abu Dhabi—with SoftBank providing financial backing and OpenAI taking the lead in operations—the project will build massive data center campuses across various U.S. locations. The first campus is under construction in Abilene, Texas, and the entire project, valued at $500 billion, represents an unprecedented level of investment in a single initiative. This strategic effort aims to reinforce U.S. AI leadership and support the development of AGI.

What is the NVIDIA Blackwell GB200, and how powerful is it?

Blackwell is NVIDIA’s latest GPU architecture unveiled in 2024. The GB200 refers to a superchip that integrates two B200 GPUs with one Grace CPU. The B200 GPU contains 20.8 billion transistors, delivers up to 20 PFLOPs of AI computational performance, and is equipped with 192GB of ultra-fast HBM3E memory. Connected by a custom interconnect boasting 10TB/s bandwidth between the chiplets, and featuring 6th-generation Tensor Cores and 2nd-generation Transformer Engines, Blackwell GB200 offers nearly five times the inference speed and four times the memory capacity of its predecessor, the H100 (Hopper). With 5th-generation NVLink, it can cluster hundreds of GPUs together, providing unparalleled scalability.

Why is OpenAI building its own data centers in addition to using Azure Cloud?

Despite receiving massive cloud resources from Microsoft Azure, the explosive success of ChatGPT has led to an exponential increase in compute demands that may eventually exceed what Azure can provide both in scale and flexibility, while also potentially driving up costs. Through the Stargate Project, OpenAI aims to secure its own dedicated AI infrastructure to support large-scale computations and reduce dependence on Azure. Additionally, owning distributed data centers across the U.S. offers advantages in data sovereignty and security. This strategy not only enhances control over data and infrastructure but is also expected to lower long-term cloud expenditure. OpenAI will continue to collaborate with Azure as part of a dual-track approach.

Who are the competitors to NVIDIA Blackwell?

In the current AI accelerator market, major competitors challenging NVIDIA include AMD and Google. AMD is pursuing competition with data center GPUs like the MI300, and Google employs its proprietary TPUs for large-scale AI computing. However, many industry experts—including figures such as Elon Musk—have stated that “there is nothing better than NVIDIA hardware at this point,” reflecting the dominant performance of Blackwell. NVIDIA’s comprehensive strengths in hardware, its CUDA ecosystem, software optimizations, and interconnect technologies like NVLink ensure that the Blackwell architecture remains the leader in AI training and inference infrastructure for the foreseeable future. Moreover, U.S. government policies regulating the export of cutting-edge AI chips, particularly with respect to China, further cement NVIDIA’s competitive position.

Sources and References:

  • OpenAI Official Blog Announcement (January 21, 2025)
  • DatacenterDynamics News Article (February 27, 2025)
  • Bloomberg Report (Rachel Metz et al., March 6, 2025)
  • NVIDIA GTC 2024 Keynote and Technical Documents (summarized from TechPowerUp, etc.)
  • Entrepreneur/Bloomberg Interview (October 4, 2024)
  1. NVIDIA Official Website: https://www.nvidia.com/en-us/
  2. NVIDIA GTC Page: https://www.nvidia.com/en-us/gtc/
  3. OpenAI Blog: https://openai.com/blog/
  4. DatacenterDynamics: https://www.datacenterdynamics.com/
  5. Bloomberg Technology: https://www.bloomberg.com/technology
  6. TechPowerUp: https://www.techpowerup.com/
Scroll to Top