Earlier last week, on October 28, Nvidia held its GPU Technology Conference (GTC) in Washington, DC, where it made a number of announcements. One of these was a $1 Billion investment in Nokia, which sparked excitement across the telecom sector. This post takes a deeper look at that investment. For context, I would note the following five important points:
- Nvidia invested $1 billion in Nokia, taking a position at $6.01/share, or about a 3% stake. For perspective on the size of the investment, earlier this month Nvidia became the first company to reach a $5 trillion market cap.
- Nokia is the weakest of the major RAN vendors, having failed to secure major deals. For instance, it lost its deployment base with AT&T.
- The announcement centers on 6G. Standardization is expected to conclude in 2029, so limited deployment may begin in 2030 at the earliest.
- Nvidia emphasized that this move would bring RAN development and production to the United States, aligning with the objectives of the current US administration. Given Nokia’s position, this is not an unreasonable strategy, though it does not guarantee success. It is also worth noting that Cloud RAN and Virtualized RAN solutions were built on Intel’s technology, which, like Nvidia’s, is American.
- Nokia will be using Nvidia’s Aerial RAN Computer Pro (ARC-Pro), an edge or cell site platform. This marks a further evolution in Nvidia’s approach to supporting telecom infrastructure, building on earlier ARC solutions that focused on RAN deployments in data centers.
The rest of the post will explore the broader aspects of AI RAN, as I see them.
We recently published our paper on the AI infrastructure market. It addresses investments in AI data center and highlights weak links in the investment value chain. You can download it here.
| |
ARC-Pro Signals Nvidia’s Shift Toward Edge RAN
The ARC-Pro is Nvidia’s third RAN product, following the ARC-1, released in 2024, and the ARC-Compact, announced in mid-2025. The ARC-Pro is designed for base station baseband deployments located at the base of the tower. It shares similar physical specifications with the ARC Compact but includes upgrades to interfaces and processing power.
In addition to supporting RAN functions, the ARC-Pro is well suited for AI edge inference applications. This contrasts with Nvidia’s first RAN product, the ARC-1 platform, which is intended for data center deployments and optimized for AI training workloads. The ARC-Pro consumes approximately 300 W of power, compared to 3500 W for the ARC-1.
| ARC-1 | ARC-Compact | AC-Pro | |
| Launch date | Oct-24 | May-25 | Oct-25 |
| Platform | GB200 NVL2: 2x Grace CPU (144 ARMv9 cores) + 2x Blackwell dual-die GPU (40 PFLOPS FP4) | Grace C1 CPU (72 Arm Neoverse V2 cores) + L4 GPU (485 TFLOPS FP8) | Grace CPU (72 Arm Neoverse V2 cores) + Blackwell RTX PRO GPPU (1 PFLOPS/1000 TOPS FP4) |
| RAN carries & system throughput | ~36x 100 MHz 64T64R; 170 Gbps total | Up to 30 carriers; 25 Gbps | Not publicly specified |
| Power consumption | ~3,500 W (full node) | GPU: 72W; system: < 300 W | ~300 W (system) |
| Use case | High-density AI data centers (C-RAN aggregation) | Telecom cell site (Distributed RAN edge) | Telecom cell site (D-RAN/6G upgrades) |
| AI applications | AI training or inference | Edge AI inference, RAN optimization | Edge AI inference, RAN optimization (with 6G sensing) |
The ARC-Pro is also optimized for low precision operations using 4 and 8-bit floating point formats (FP4 and FP8). This benefits AI inference applications, where speed is often prioritized over accuracy. In RAN baseband applications, physical layer functions perform best across a range of floating and fixed point precisions, which ultimately affects system performance and operational efficiency.
The ARC-Pro and Nokia partnership announcement confirms the realignment of Nvidia’s RAN products toward edge deployments. This marks a departure from the original AI RAN concept of running RAN workloads on large AI training GPUs. The shift reflects months of collaboration with telecom vendors and operators.
AI RAN Revisits the Cloud RAN Playbook
To frame AI RAN within the broader evolution of telecom infrastructure, it offers a new take on a familiar theme: using programmable devices for RAN functions. Over the years, various types of programmable hardware have been used, including DSPs, FPGAs, CPUs, and NPUs. The most recent attempt coincided with the rise of cloud computing and virtualization, with Intel positioning its x86 CPU technology for RAN baseband processing. This became known as “Cloud” or “Virtual” RAN. Intel invested in both hardware and software development and worked to build an ecosystem to gain market traction. It eventually became clear that CPU-based platforms were too expensive and lacked the performance needed to compete with dedicated solutions.
In response to the shortcomings of early platforms, Intel integrated hardware acceleration to improve cost and performance. This approach gained better traction in the market when vendors such as Ericsson and Samsung announced Cloud RAN or virtualized RAN products built on Intel’s platform. A few operators, including AT&T and Telus, adopted these products in parts of their networks.
AI RAN follows a similar path. The key difference is that GPUs can perform the mathematical operations required for RAN functions more efficiently than CPUs, allowing them to replace some of the hardware accelerators built for CPU platforms. However, GPUs designed for data centers are too expensive and consume excessive power in RAN applications.
Reminder of Nokia’s FPGA Decision
For its first 5G base station design, Nokia adopted an FPGA platform from Altera/Intel, a decision made prior to 2018. This platform created significant challenges. It was expensive, consumed high power, and introduced other technical issues. Some of these were outside Nokia’s control, including Intel’s failure to meet its 10 nanometer process milestone. This setback impacted Nokia’s profitability in 2019 and 2020 and, in my view, contributed to its current struggles by missing a critical window during the rollout of 5G networks.
Nokia later switched to a system-on-chip developed in partnership with Marvell, but the damage had already been done.
Why SoCs Dominated Baseband Processing in Telecom
Telecom infrastructure products rely on a range of computing devices; however, for baseband processing, system-on-chips (SoCs) have remained dominant. The idea of using programmable devices such as DSPs, NPUs, and FPGAs has always been present, yet it has gained only limited traction for baseband workloads. The extent of operator commitment to Ericsson’s Cloud RAN and Samsung’s Virtualized RAN remains unclear. This stands in contrast to other areas of telecom where programmable devices have successfully replaced ASICs.
The dominance of SoCs in baseband platforms stems from a mix of technical and market factors. On the technical side, wireless baseband processing is highly complex, and its complexity is often underestimated by non-experts. Designing a high-performance baseband solution requires a series of architectural decisions that balance trade-offs across precision (16 vs. 64 bit), throughput (Gbps), latency (micro-second scale), timing and frequency synchronization (order of 0.1 ppm), and energy efficiency.
The wireless protocol stack includes a wide range of functions, each with distinct requirements for precision, timing, and synchronization. For example, reducing precision can lower latency and power consumption, but it comes at the cost of signal fidelity. This is where companies often stumble, as Nokia did with its initial 5G design despite having a highly experienced engineering team. In fact, only a handful of teams globally are capable of producing a well-optimized baseband engine.
SoCs offer a blend of engines. Some functions are implemented in hardware for power efficiency and processing speed (e.g. FFT/IFFT, encryption, etc.). Embedded DSPs and CPUs provide programmability and flexibility for functions that can tolerate broader performance ranges (e.g. equalization on DSP and MAC on CPU). Memory access is optimized for the target application, and interfaces are often integrated to reduce overall system cost.
This brings the economic and financial dimensions into focus. As the vendor ecosystem consolidated, only the largest players with sufficient volume could support SoC development. If the design is successful, their products benefit from lower unit costs, as most of the SoC development expense is capitalized and amortized over large volumes. In contrast, companies relying on programmable devices face relatively high development costs and must share a larger portion of profits with the device vendor. Overall, the cost structure favors vertical integration to compress margins. Disaggregation, where a vendor builds its product on a third-party programmable device, remains a niche approach in macro base stations. This analysis applies primarily to macro deployments, as the economics for small cells differ and warrant a separate discussion.
Concluding thoughts on AI RAN integration
AI RAN represents another iteration in the use of programmable devices, with a focus on combining AI and RAN workloads. AI will undoubtedly play a role in RAN functions, often referred to as “AI for RAN.” Research already shows benefits in specific applications such as power management, massive MIMO and beamforming implementation, and channel estimation, along with other areas related to operations and maintenance. These AI applications are use-case driven and will be selectively implemented and provisioned through both hardware and software.
Looking ahead, GPU cores for AI inference may play a larger role in future RAN infrastructure. In that case, using Nvidia GPUs to power the entire RAN is one possibility among others. Another likely path is to integrate GPU cores into SoC designs. There are several advantages and drawbacks to this approach, which are beyond the scope of this post.
A more complex issue is what the industry refers to as “AI on RAN,” which involves offering AI services over the RAN infrastructure. In this model, the baseband engine runs both RAN and AI workloads, likely focused on inference, although Nvidia has also targeted training. This raises broader considerations that I will address in a future post.