Ericsson Goes with Custom Silicon (Rather than Nvidia GPUs) for AI RAN

Why It Matters

Ericsson’s silicon strategy reinforces a proprietary hardware ecosystem, shaping cost, performance and vendor‑lock‑in dynamics for 5G and AI‑enabled networks worldwide.

Key Takeaways

• Ericsson selects custom ASICs, not Nvidia GPUs, for AI‑RAN
• Intel remains sole commercial silicon partner; others prototype only
• AI‑RAN delivers 10% spectral efficiency boost in field tests
• Hardware abstraction layers aim to decouple software from silicon
• L1 processing still requires platform‑specific optimization despite higher‑layer portability

Pulse Analysis

The race to virtualize radio access networks has pushed most operators toward cloud‑RAN solutions built on general‑purpose CPUs, yet Ericsson is doubling down on custom ASICs to extract maximum performance from AI‑driven workloads. By integrating neural‑network accelerators directly into its baseband ASICs, the Swedish vendor sidesteps the power and cost penalties of high‑TDP GPUs while preserving the low‑latency characteristics essential for massive‑MIMO and real‑time beamforming. This approach also lets Ericsson leverage Intel’s SoC‑level FEC acceleration, keeping the critical PHY layer tightly coupled to silicon that can meet 5G’s stringent throughput demands.

At the same time, Ericsson is laying the groundwork for broader software portability through hardware abstraction layers such as BBDev and potential CUDA interfaces. These abstractions aim to isolate the AI‑RAN stack from underlying silicon, allowing the same L2/L3 codebase to run on Intel, AMD or future Arm‑based platforms. However, the L1 processing block—where forward error correction and channel estimation reside—still demands hand‑tuned implementations to hit parity across architectures. The company’s recent field results, showing a 10% uplift in spectral efficiency without external accelerators, illustrate how AI‑native algorithms can compensate for hardware diversity when tightly integrated into purpose‑built silicon.

For the telecom ecosystem, Ericsson’s stance signals a bifurcated future: open‑source, cloud‑native software layers coexisting with vendor‑specific silicon for the most compute‑intensive tasks. Competitors that bet on GPU‑centric AI‑RAN may face higher power budgets, while operators seeking cost‑effective upgrades could gravitate toward Ericsson’s ASIC‑first model. As AI becomes a core control loop for network optimization, the balance between hardware flexibility and performance will dictate market share, influencing everything from equipment procurement to the rollout speed of 5G and beyond.