Ericsson positions ASICs as AI-RAN value vehicle, keeps GPU path open

Ericsson has framed custom ASICs as the AI-RAN value capture vehicle while keeping GPU acceleration available.

The dual-track silicon stance, surfaced by RCR Wireless News, defines how the Stockholm-based vendor will approach artificial intelligence inside the radio access network. Custom chips carry the commercial case. GPUs remain an open option.

What is the strategic move?

Ericsson is signaling two parallel silicon paths for AI-RAN. ASICs anchor the value proposition. GPUs stay available as a programmable alternative for emerging model architectures.

This dual approach lets the vendor serve operators who demand power-efficient, purpose-built silicon. It also serves carriers with existing GPU infrastructure who want continuity into the radio access domain across multi-year deployments.

Why ASICs carry the value framing

Application-specific integrated circuits have defined radio access baseband processing for two decades. The economics of high-volume RAN silicon reward dedicated designs that minimize power draw per processed bit.

Ericsson placing ASICs at the center of its AI-RAN narrative continues that established pattern. The vendor is betting that radio access scale rewards chip-level specialization, even as model architectures evolve across the broader AI industry and force silicon refresh cycles on operators.

Why GPUs remain open

GPU silicon dominates AI training workloads in hyperscale data centers. Inference tasks increasingly migrate to GPU clusters at cloud providers and telecom operators building AI factories.

Ericsson keeping GPUs available acknowledges that AI model tooling, software stacks, and developer ecosystems remain GPU-centric today. Operators with existing GPU investments in adjacent networks want the same acceleration path inside the radio access domain. That continuity argument carries commercial weight in multi-year procurement cycles.

What changes for operator procurement

Carriers evaluating AI-RAN vendors now weigh silicon strategy alongside traditional RAN performance metrics. Ericsson's dual-track framing gives buyers optionality across both ASIC and GPU paths.

Procurement teams should expect AI-RAN contracts to specify which silicon architecture serves which workload class. The dual approach complicates total cost of ownership modeling. ASIC and GPU economics diverge sharply over multi-year network lifecycles, requiring vendor guidance on where each path fits across operator capex planning horizons.

What is the regulatory and standards context

The radio access network sits inside standards-defined interfaces governed by 3GPP. AI-RAN work at the O-RAN ALLIANCE and 3GPP will shape how ASIC and GPU silicon integrate with standardized RAN functions across operator networks.

Ericsson has historically favored integrated RAN stacks over fully disaggregated architectures. The vendor's ASICs-first, GPUs-open stance aligns with that integrated model, where Ericsson controls both silicon design and software layers for carrier customers across global deployments.

What to watch next

Ericsson has not disclosed AI-RAN product timelines, deployment numbers, or operator commitments in the reporting. The vendor's strategic posture suggests product launches will lead with ASIC-based AI-RAN features.

Carriers testing AI workloads in live radio networks will provide the first commercial proof points. Ericsson's dual-track framing now sets the negotiating position for AI-RAN contracts expected across the next 18 to 36 months as operators plan artificial intelligence rollouts inside the radio access layer.