NextFin News - In a decisive move to safeguard its technological sovereignty, Ericsson has officially launched its initial wave of AI-Radio Access Network (AI-RAN) products designed to operate on its own purpose-built silicon rather than relying on Nvidia’s graphics processing units (GPUs). During a pre-MWC briefing held in London on February 19, 2026, Ericsson Chief Technology Officer Erik Ekudden and Mobile Networks head Per Narvinger detailed how the company is integrating AI capabilities directly into its baseband and radio-unit silicon. This strategy is intended to improve spectral efficiency and beamforming by up to 10% without the power-hungry overhead associated with general-purpose AI chips. By prioritizing its internal Application-Specific Integrated Circuit (ASIC) development, Ericsson is positioning itself as a champion of "silicon freedom," contrasting sharply with competitors who have deepened their ties to the Nvidia ecosystem.
The timing of this announcement is critical, as the global RAN market remains under pressure, having stabilized at approximately $35 billion in 2025 after a significant post-5G-peak contraction. According to Light Reading, Ericsson’s refusal to follow Nokia’s lead—which recently accepted a $1 billion investment from Nvidia to overhaul its 5G and 6G software for GPU compatibility—stems from a desire to avoid vendor lock-in. Narvinger emphasized that while Ericsson supports multiple hardware platforms in prototype stages, including AMD and Arm, its commercial focus remains on a "one software track" approach that can eventually be deployed across various silicon architectures through abstraction layers like BBDev, rather than being natively tied to a specific proprietary platform like Nvidia’s CUDA.
This pursuit of independence is driven by the volatile nature of the semiconductor industry. Just last year, Intel, a primary supplier for Ericsson’s cloud RAN portfolio, faced significant financial turbulence and considered divesting its network and edge group (NEX). Although U.S. President Trump’s administration has emphasized domestic semiconductor stability, and Intel eventually shelved its divestment plans following a $7 billion injection from SoftBank and Nvidia, the episode underscored the risks of over-reliance on a single partner. Ericsson’s response has been to increase its R&D spending, which reached approximately 48.9 billion Swedish kronor ($5.4 billion) in 2025, with a substantial portion dedicated to in-house silicon that bridges the gap between traditional and virtualized networks.
From a technical perspective, Ericsson argues that AI-RAN is achievable without GPUs by utilizing specialized neural network accelerators integrated into the radio unit. These accelerators are designed to handle specific Layer 1 tasks—the most computationally demanding slice of RAN software—more efficiently than general-purpose hardware. While Nvidia’s "inline" acceleration approach puts the entire Layer 1 on the GPU, Ericsson’s "lookaside" strategy offloads only specific problematic functions, such as forward error correction (FEC), to discrete accelerators. This allows the bulk of the network software to remain portable across different CPU architectures, including Intel’s Xeon 6 (Granite Rapids) and potentially future Arm-based solutions.
The industry remains divided on the economic viability of these competing paths. Analysts at Dell’Oro Group have noted that while common off-the-shelf (COTS) servers are improving, they still lag behind custom silicon in terms of performance-per-watt and total cost of ownership (TCO). Ericsson’s Ekudden contends that unless a telco is co-hosting non-RAN AI workloads at the edge, the cost of deploying GPUs solely for network functions is difficult to justify. By 2030, virtualized RAN is expected to account for only 15% to 20% of the market, suggesting that the era of specialized silicon is far from over.
Looking forward, Ericsson’s strategy represents a high-stakes bet on the modularity of 6G. If the company succeeds in creating a truly hardware-agnostic software stack that performs optimally on its own efficient ASICs, it will have successfully insulated itself from the margin-squeezing power of the global chip giants. However, if the industry shifts toward a unified AI-on-everything model where GPUs become the standard for edge computing, Ericsson may find itself maintaining an expensive, isolated hardware island. For now, the Swedish vendor is banking on the fact that in the world of telecommunications, efficiency and independence are the ultimate currencies.
Explore more exclusive insights at nextfin.ai.
