In the legend of 6G, it's not Qualcomm but NVIDIA that takes the lead.
Cloud is where AI is born;
Device is where AI acts;
Edge will be where AI lives.
If you were to look for the most "sexy" rather than the most "tangible" story about NVIDIA at the GTC DC conference, different people might have different answers. However, there's a technology brand that might just be added to the list of candidates. It's the very reason NVIDIA recently invested $1 billion in the veteran telecom provider Nokia. That brand is the core technology brand of NVIDIA's AI - RAN:
Aerial.
Aerial is a term with an industrial flavor today. Its etymology comes from the Latin word "aerius," which means "airy, light, and lofty." In ancient times, it was also used to describe illusory, light, and elegant beauty. It reminds people of another product from a "leader" in Silicon Valley, OpenAI's video product:
Sora.
The original meaning of Sora in Japanese is "sky," which can refer to the actual sky or the state of mind when imagining a distant place.
It's no coincidence that both NVIDIA and OpenAI have pointed their flagship products towards the sky, and it's not just a convergence of aesthetic preferences in Silicon Valley. Perhaps they are telling a similar story, but let's put that aside for now.
So - called AI - RAN, in simple terms, means adding computing power to communication base stations. This allows future large - scale scenario computing to be completed directly at the base station without going to the cloud and then be returned to the terminal devices in the scenario.
This approach has two obvious advantages:
One is to save a significant amount of transmission costs and reduce the pressure on cloud computing centers. Microsoft CEO Satya Nadella mentioned this AGI dilemma during a recent conversation with OpenAI CEO Sam Altman. He said that he still has GPUs in his warehouse, but there aren't enough computing centers to accommodate them, and there's also a shortage of sufficient power. In fact, distributing computing power across base stations worldwide through AI - RAN can be one of the solutions to the above problems.
The second is to significantly reduce the latency of computing power. Readers who have played online games know that a latency of less than 50 milliseconds is considered "smooth." However, if this latency is applied to autonomous driving, the highways would be filled with wrecked cars. This is also an important reason why the high - precision map solution has been marginalized today. But if the computing center is moved closer to the user, the latency can be reduced by more than 90%, from milliseconds to microseconds, enabling many scenarios with high latency requirements to function properly.
According to available public information, Aerial first appeared in NVIDIA's materials at the 2019 Mobile World Congress in Los Angeles. At that time, it was just a set of SDK tools aimed at supporting GPU acceleration and software - defined 5G radio access networks.
Yes, 2019 was the dream year when "He Tongxue" shot a 5G video that received 30 million views.
From this perspective, Aerial can probably be considered a "5G - native" tool. It was the rapid growth of communication bandwidth that allowed NVIDIA to glimpse the future world, and thus Aerial was born.
In the past five years, 5G has disappointed many people, but Aerial has been quietly developing.
During this period, Aerial's positioning shifted from the initial SDK toolbox data to a platform serving "industry, academia, and research."
In March 2021, NVIDIA published a paper titled "NVIDIA Aerial GPU Hosted AI - on - 5G" at the IEEE 5G sub - forum, introducing NVIDIA's hyper - converged platform for 5G connectivity and mobile edge computing (MEC). At that time, their case target was Industry 4.0.
In the abstract of this paper, NVIDIA clearly stated:
Aerial is an open platform that aims to be industry transformational by providing researchers with a platform for next - generation wireless and AI research.
One month later, NVIDIA launched the Aerial A100 at the 2021 GTC conference, which is the NVIDIA AI - on - 5G computing platform .
The Aerial A100 is actually a product that combines the NVIDIA Aerial software development kit with the NVIDIA BlueField - 2 A100 chip. The latter is largely derived from Mellanox's DPU product line, which NVIDIA acquired in 2020. Although it is a pieced - together product, a fusion card that includes the "5T FOR 5G" solution and integrates GPUs and DPUs, it provides a valuable computing platform for AI - RAN and has received support from companies such as Google Cloud and Fujitsu.
Yes, they didn't even include "AI" in the name. The so - called 5T refers to "Time - Triggered Transmission Technology for Telco." As the name suggests, it aims to solve the problems of precise timestamps and high clock accuracy, which are actually the foundation of large - scale edge computing.
Interestingly, the media's headline for the 2021 GTC was: "NVIDIA Launches Its First CPU and Vigorously Promotes the ARM Ecosystem." Although the server business is still important, few people think about CPUs today, let alone view NVIDIA from an ARM perspective.
Subsequently, Aerial reached a small peak, which can be seen from NVIDIA's "technology blog." The number of technology blogs about Aerial increased rapidly after 2022.
Between 2022 and 2023, NVIDIA successively launched a series of tools to accelerate AI - RAN, such as DOCA GPUNetIO, the Sionna library, and the Aerial Research Cloud.
DOCA GPUNetIO allows GPUs to directly connect to the network bypassing the CPU, reducing latency and costs while increasing throughput.
The Sionna library is an open - source library accelerated by GPUs for communication system research. The official says it can implement an "automatic differentiation framework" and "backpropagate gradients through the entire communication system," making it very suitable for "neural network integration."
The Aerial Research Cloud is the first fully programmable research sandbox for 5G and 6G networks.
But you can still trace all of NVIDIA's efforts on Aerial back to the title of that 2021 paper:
GPU Hosted AI - on - 5G
2024 was a big year for NVIDIA's Aerial ecosystem.
In February, the well - known AI - RAN Alliance was established, led by NVIDIA and SoftBank. Other founding members include Ericsson, Nokia, Samsung, T - Mobile, Microsoft, AWS, Arm, DeepSig, and Northeastern University in the United States.
The AI - RAN Alliance is an important step for NVIDIA to redefine 6G because it includes almost all the most important communication - related companies (except Huawei and ZTE). Its organizational goal is to combine AI with RAN to make 6G a truly AI - native network.
In March, NVIDIA also launched a 6G research platform that includes the Omniverse ecosystem and Aerial CUDA.
It wasn't until September that NVIDIA officially announced the launch of NVIDIA AI Aerial at GTC Paris 2024.
In its official introduction, it has become a one - stop platform for optimizing wireless networks and providing new generative AI experiences.
This is how NVIDIA introduced AI Aerial in its official press release .
The NVIDIA AI Aerial platform offers a full range of capabilities, including high - performance software - defined RAN, as well as training, simulation, and inference, enabling telecom operators to participate in all stages of the development and deployment of next - generation wireless networks.
The capabilities provided by the NVIDIA AI Aerial platform include:
NVIDIA Aerial CUDA Accelerated RAN: It includes software libraries that enable partners to develop and deploy high - performance virtualized RAN workloads on NVIDIA accelerated computing platforms.
NVIDIA Aerial AI Radio Framework: It includes software libraries based on PyTorch and TensorFlow for developing and training models that can improve spectrum efficiency and add new features to 5G and 6G radio signal processing. This framework also includes NVIDIA Sionna, a link - level simulator that can be used to develop and train neural network - based 5G and 6G radio algorithms.
The NVIDIA Aerial Omniverse Digital Twin (AODT) is a system - level network digital twin development platform. AODT can simulate wireless systems with physical accuracy, whether it's a single base station or an integrated network composed of a large number of base stations covering an entire city. It includes a software - defined RAN (Aerial - CUDA Accelerated RAN), a user equipment simulator, as well as the real terrain and object attributes of the physical world.
This includes many definitions and imaginations of the next 6G by NVIDIA:
High - performance computing power, the distributed network capabilities of neural networks, and the development capabilities of the virtual world through digital twins.
From this point on, NVIDIA began to accelerate the expansion of the entire AI - RAN ecosystem. It not only established experimental fields with edge computing service providers like Vapor IO in Las Vegas but also continuously open - sourced new tools and promoted new leading partners.
This series of ecological actions reached a phased peak at the GTC DC conference in October 2025.
1
Aerial's Sprint to the CUDA Moment
During the GTC DC conference in October, NVIDIA finally threw two big stones into the surging lake of AI - RAN:
First, open - source the Aerial software. After being open - sourced, the software can run on various NVIDIA platforms, including the NVIDIA DGX Spark.
The DGX Spark is the small "lunch box" that Jensen Huang personally delivered to Elon Musk and Lee Jae - yong. It's currently the world's smallest "AI supercomputer" that can perform AI model inference with up to 200 billion parameters and model fine - tuning with 70 billion parameters locally. It perfectly matches the diverse development ecosystem requirements of scenarios like AI - RAN and is very suitable for university researchers and individual developers to participate in.
Second, NVIDIA announced a strategic cooperation by investing $1 billion in Nokia. After the investment, NVIDIA will hold a 2.9% stake. NVIDIA's AI - RAN ecosystem products will be integrated into Nokia's RAN product portfolio, and the two sides will jointly promote the implementation of 6G AI - RAN.
Third, NVIDIA launched the Aerial RAN Computer Pro (ARC - PRO) platform. If the previous Aerial platform was a small platform in the early stages of research and development, the ARC - PRO is an industrial - grade runway prepared for top partners. It can be directly integrated with base stations, enabling AI - RAN functions for 5G and facilitating the smooth transition from 5G to 6G. The title in the introduction has also officially changed from AI - RAN to "6G AI." NVIDIA's tone regarding this 6G AI platform has become quite bold and might even seem aggressive to Chinese readers:
"Drive the United States back to the leadership position in telecommunications."
Readers who are familiar with NVIDIA's history might have a sense of deja vu at this point. The development process of Aerial is very similar to that of CUDA, almost a replication of the business aesthetics of technology.
Jensen Huang strongly promoted CUDA because he saw that GPUs were not limited to gaming graphics but had the potential for general computing. He promotes Aerial because he believes that RAN is not just about communication but could also be the future AI infrastructure.
CUDA started from the university ecosystem and took ten years to cultivate, ultimately forming its own unique ecological moat. Aerial also started as a small research and development platform and gradually evolved into an AI - RAN ecosystem, taking a total of five years to become the moat for NVIDIA's base station computing power business.
From a single - point tool → a general platform → an ecological moat, this is a classic story of NVIDIA and time.
However, upon closer examination, it can be found that the two are not exactly the same in terms of business details.
The difference is not in the pace of development. Aerial was born with a silver spoon in its mouth, so its pace is naturally faster. The biggest difference lies in NVIDIA's attitude towards ecological partners.
Taking the algorithm ecosystem as an example, CUDA is mainly closed - source, while Aerial is mainly open - source.
In terms of cooperation, CUDA's openness is mainly to cultivate developers' habits and build a good reputation. Although Aerial also welcomes a large number of small and medium - sized developers, its core strategic goal is aimed at the top players in the ecosystem.
If NVIDIA wants to replicate the CUDA moment, the first "North Star metric" is actually to have its standards and ecosystem accepted by the "big players." Without the support of the complex interest ecosystem in the communication industry, AI - RAN would be out of the question. In terms of general computing, GPUs have a dimensionality reduction attack on CPUs from scratch. However, Aerial's cooperation with RAN is an upgrade from scratch.
From this perspective, although NVIDIA is the new darling of Silicon Valley with a valuation of $5 trillion, it still plays the role of an "underdog" and a "persuader" in the communication market.
The "investment of $1 billion" is one of the most powerful ways to persuade.
Nokia might be the most perfect partner Aerial can find today.
In the European and American markets excluding Huawei, Nokia's market share in 5G base stations has fallen behind Ericsson, and the gap with the latter is widening. Especially in the US base station market, after losing the contract with Verizon, there were even rumors that Nokia might lose its remaining contract with T - Mobile. Cooperation with NVIDIA might help Nokia regain its footing in the US market.
In 2023, Nokia introduced a "2030 Plan." This plan highly aligns with NVIDIA's preferences. It places AI at the top of all strategies and "cloud continuum" at the second place. The so - called cloud continuum actually means seamlessly integrating the "cloud - edge - terminal," and the biggest market opportunity behind it is actually the "edge cloud" market.
According to NVIDIA's technological vision, today's communication companies could all become new cloud computing companies in the future. We'll discuss this in detail later.
Finally, Nokia is a listed company with a dispersed equity structure. The major shareholder is a fund under the Finnish government, which does not directly interfere in the company's operations. Therefore, NVIDIA's 2.9% stake will neither overly arouse the resistance and concerns of other base station companies nor prevent NVIDIA from directly influencing Nokia.
For NVIDIA, investing in Nokia has two major strategic advantages.
First, the deployment speed of 6G AI will lead the industry. If NVIDIA is the first to launch base station equipment with Aerial deployed in the market and then form an application ecosystem on the basis of Aerial, it will be the first to create a growth flywheel for the ecosystem.
This is a bit like Android's advantage over systems like Symbian and Windows Mobile. The first - mover advantage will form an important advantage for the operating system.
Second, it's about the very important voting mechanism in the traditional RAN market landscape: 3GPP.
3GPP is the core organization for formulating global mobile communication technology standards, with hundreds of key enterprises from different industrial chains within the organization. 3GPP is somewhat of a celebrity in China because a few years ago, it was involved in a very well - known news in the technology industry, the so - called "voting gate" between Huawei's proposed Polar code and Qualcomm's proposed LDPC code. The voting process was sensationalized, and the result caused great public relations damage to the voting enterprises.
In fact, a coding proposal requires a 71% vote to pass, so there is no such thing as a "critical minority" in the strict sense. It's a decision - making form based mainly on technological consensus and supplemented by voting. However, the image of 3