How Can the AI-RAN Alliance Deliver on Early Industry Hype?

There is considerable excitement around Artificial Intelligence in Radio Access Networks (AI-RAN), with NVIDIA and SoftBank already conducting pilot programs in Shenzhen, as well as developing cloud platforms (e.g., AITRAS from SoftBank, Japan), with the AI-RAN Alliance predicting commercial deployments as early as 1Q 2026. However, the current reality does not fit into this publicized hype from the AI-RAN Alliance, as the rest of the market is adopting a more cautious “wait-and-see approach.” This slower adoption is demonstrated in ABI Research’s AI-RAN revenue forecast, which projects revenue of US$1.22 billion by 2027, with considerable growth forecast to start only from 2029 onward, culminating in US$6.18 billion by 2032. For more information on the AI-RAN market, see ABI Research’s AI-RAN Market Developments report published in 2Q 2025 (PT-3651).

One of the biggest drivers of this disconnect is the unproven value proposition, especially around Artificial Intelligence (AI) workloads at the edge. SoftBank’s independent claims around large operational savings (US$5 of operational savings for every dollar invested) remain unsubstantiated and not validated by credible third parties as of May 2025. This specific financial uncertainty resonates with operators, which need clear business cases to approve large Capital Expenditure (CAPEX). The fact that the AI-RAN Alliance has grown from 11 founding members to over 80 members within a year is a positive development for collaborative development, but the lack of comparable growth in terms of operator members (still only 8) indicates that they remain hesitant about investing until the financial benefit has been proven.

AI-RAN holds a lot of potential to transform the telecoms industry by improving performance, reducing latency, and unlocking new revenue through optimized and automated network operations, thus lowering CAPEX and Operational Expenditure (OPEX) for operators. Furthermore, AI-RAN systems can predict traffic, change network capacity dynamically, and reduce over-provisioning, which equates to operational costs savings and improved resilience in the network. AI-RAN will also help enable next-gen features such as live network traffic prediction, dynamic resource allocation, predictive and proactive maintenance, and network slices; this means that AI will convert traditional Radio Access Networks (RANs) from a CAPEX investment into an additional revenue stream.

However, deploying AI-RAN on a global scale is not without significant challenges. The lack of independently verified Return on Investment (ROI) claims makes large AI-RAN investments risky. Operators cannot justify CAPEX-heavy transitions without hard evidence of performance gains and cost savings. Another major issue is the ongoing technical argument over the use of Graphics Processing Units (GPUs) and Central Processing Units (CPUs) in AI-RAN architectures (read the recent ABI Insight, “Are GPUs Worth Considering for RAN Today?”). NVIDIA and SoftBank have invested significantly to develop GPUs, with NVIDIA releasing Aerial RAN Computer-1 (ARC-1) in September 2024. In May 2025, NVIDIA also introduced ARC-Compact, which is based on Grace 1 CPU, and aimed at addressing concerns surrounding cost and flexibility with its ARC-1. However, the ARC-Compact is still GPU-accelerated. Despite these advancements, there exists little market demand for GPU architectures, with many operators seeing their CPUs being more than capable of handling current AI workloads.

While GPUs may be essential for more complicated AI-RAN workloads like massive Multiple Input, Multiple Output (mMIMO) and beamforming, they have a much larger Total Cost of Ownership (TCO) compared to CPUs, particularly surrounding higher energy costs and creating operator reliance on proprietary platforms such as NVIDIA’s CUDA, adding extra hurdles for operators. This disconnect means that the heavy emphasis on GPU-based architectures by the AI-RAN Alliance may not align with the current needs of the industry. Operators ultimately want—at a minimum—performance metrics that demonstrate AI-RAN and GPU solutions are more cost-efficient than similarly effective CPU-based solutions, or even custom silicon solutions. Additionally, concerns over vendor lock-in complicate adoption, especially given the strategic priority placed on long-term flexibility and open standards.

To prove the value of AI-RAN and turn early hype into commercial deployments, the industry will need to committ to a unified strategy that is oriented around standardization, flexibility, and innovation. Earlier vendor contributions have been positive, but the Alliance now needs to think bigger, on the ecosystem level, and provide compelling value propositions to drive broader adoption.

First, the Alliance should start leading a series of roundtables and field validation workshops with Tier One operators to demonstrate pilot deployments of AI-RAN that utilize GPU-based infrastructure. It is critical that these evaluations show that the AI-RAN can deliver expected system performance outcomes under real network conditions, not just small-scale field trials. Operators should have direct input into performance benchmarks and participate in side-by-side performance comparisons with incumbent vendor basebands under identical configurations. Findings, especially where GPU-based systems meet or exceed custom silicon, should be shared transparently to inform public case studies and vendor roadmaps. Creating this feedback loop will help surface deployment obstacles, validate performance claims, and build operator trust in AI-RAN as a carrier-grade alternative.

To support procurement decisions, the Alliance should develop a comprehensive whitepaper that compares the financial and overall operational impact of GPU-based AI-RAN deployments against/alongside traditional CPU or custom silicon-based solutions, with projections through 2030 (the expected timeline for operators to start deploying these systems at scale, rather than solely looking at 2025). Having a long-term perspective is essential because telco infrastructure investments generally require many years and must adapt to the constantly changing AI and network demands. The whitepaper should also include a cost and performance modeling perspective in urban, rural, and remote specific contexts to highlight the relative initial cost and also the long-term value derived from flexibility, faster iterations, and new revenue channels. By 2030, GPU energy efficiency improvements will be a necessary factor for operators to achieve sustainability targets and lower OPEX. The programmability of GPUs allows for improved performance and quick innovation cycles, creating opportunities for new revenue streams that fixed hardware does not. By considering depreciation and refresh cycles aligned with the carrier's timelines, future investments can be protected. This long-term analysis will allow operators to feel confident transitioning GPU-based AI-RAN systems from trial to more standard deployments by 2030.