Ericsson’s CEO Bets on Silicon: The RAN Chip War Behind AI-Native Networks

On the morning of October 20, 2026, Per Narvinger took a question that every RAN vendor chief executive now has to answer sooner or later. The newly installed CEO of Ericsson was asked, in effect, whether the GPU juggernaut bearing down on the radio access network—Nvidia’s, and by extension Nokia’s—had made him reconsider the Swedish vendor’s long commitment to designing its own custom silicon. His answer, as reported by Light Reading’s Iain Morris, was unambiguous: Ericsson’s ASICs are an “extraordinary asset,” and he is not steering the company away from them.

It was a small moment in a single interview, but it landed at a revealing juncture. In the same news cycle, Nokia and SoftBank were lining up a commercial AI-RAN field trial in Japan for 2027. AT&T and Verizon were telling an RCR Wireless forum about the measurable throughput and spectral-efficiency gains they are seeing from AI features embedded in the RAN. And a 3GPP debate over how much AI belongs in the 6G core was quietly compressing. Behind all of it sits a question that rarely makes the keynote slides: what kind of chip actually runs the intelligence, and who gets to own that layer of the network.

The claim that started the argument

The immediate backdrop to Narvinger’s remark is a numbers fight. Nokia has claimed it can double the spectral efficiency of a network by 2028 through its GPU strategy. Ericsson’s published figures, which assume continued use of custom silicon, are considerably more modest. That gap—roughly a doubling versus something noticeably smaller—has become the most consequential marketing battleground in RAN hardware since the first Open RAN interoperability demos.

The problem, as several observers have pointed out, is that the comparison is not apples to apples. The baseline for the Nokia claim has not been fully specified, and as Chetan Sharma, founder of Chetan Sharma Consulting, noted in a recently published white paper, significant gains are naturally easier to achieve on a badly performing network. A site already running near its theoretical efficiency ceiling under classical processing has far less headroom for AI-driven improvement than a site operating well below that ceiling because of interference, channel-estimation errors, or suboptimal scheduling. In other words, a doubling is plausible in some conditions and fanciful in others, and neither vendor’s headline number tells you which condition you are in.

Marcus Weldon, the former Nokia chief technology officer, has weighed in too, reportedly reckoning that the huge improvements Nokia has talked about will be limited to a small number of use cases. That is a striking thing for a former Nokia CTO to say about his old employer’s flagship claim, and it captures the mood: the industry is no longer asking whether AI can improve the RAN, but where, and at what cost.

Two philosophies, one fork in the road

What makes the Ericsson-Nokia divergence so interesting is that it is not really a dispute about AI. Both companies believe AI will be woven into the RAN. The disagreement is about the substrate.

Ericsson’s position is that purpose-built ASICs remain the most efficient way to move bits. Custom silicon can be tuned for the specific math of radio processing—channel estimation, link adaptation, scheduling—and it can do so within tight power and thermal envelopes that matter enormously at a cell site. Narvinger’s framing of ASICs as an “extraordinary asset” is a claim about engineering discipline and supply-chain control as much as raw performance. If you own the chip, you own the optimization curve.

Nokia’s bet, by contrast, leans on GPUs and the merchant-silicon ecosystem around them. The logic is that general-purpose accelerators improve on a cadence set by the broader computing industry, that software can be iterated faster on them, and that AI workloads—especially the training and inference of large models—map naturally onto GPU architectures. Nokia’s anyRAN software architecture playbook, which will underpin the SoftBank trial, is explicitly designed to let radio sites deploy hybrid RAN methodologies combining proprietary, cloud, and open technology. That is a software-first philosophy in which the hardware becomes, at least in principle, more interchangeable.

For an Open RAN audience, this fork is not academic. Open RAN’s promise has always rested on disaggregation and multivendor interoperability, and the compute layer is now the least disaggregated part of the stack. If intelligence migrates onto GPUs, the number of viable silicon suppliers could widen—but it could also concentrate around whoever controls the accelerator ecosystem. If it stays on custom ASICs, efficiency may win out, but the vendor lock-in that Open RAN was meant to dissolve may simply relocate from the radio to the chip.

Meanwhile, the operators are voting with their field trials

The silicon argument would be mere inside baseball if operators were not already generating evidence. At RCR Wireless News’s Telco AI Forum 2026, Adam Loddeke, AVP of RAN technology at AT&T, and Cindy Dong, associate director of RAN planning at Verizon, described deployments producing measurable results.

AT&T, according to Loddeke, is already using AI at scale, primarily through applications and rApps spanning the network lifecycle—planning and forecasting, deployment and configuration, optimization, and lifecycle management. The operator has also been trialing Ericsson’s AI-native link adaptation feature and, Loddeke said, has seen “significant improvements” in throughput and spectral efficiency. Verizon, per Dong, is pursuing a dual RAN AI strategy: centralized rApps in a non-real-time environment combined with local AI integrated into the gNodeB for real-time optimization. Verizon’s work with Ericsson has demonstrated downlink spectral-efficiency improvements from AI-based link adaptation, and another field trial produced coverage improvements.

Note the vendor in both cases: Ericsson. That is not a coincidence, and it is a useful counterweight to the GPU narrative. The AI gains operators are reporting today are being delivered, at least in these instances, on custom silicon. It does not prove ASICs are superior for every future workload, but it does show that the incumbent architecture has real headroom—and that operators are measuring it in the field rather than in a lab.

Both executives also stressed a discipline that should hearten anyone worried about AI hype: worthiness sets the gate. AI must deliver meaningful performance benefits that justify additional compute, cost, and operational complexity. Scaling, they said, changes the challenge—moving beyond trials will require open architectures, automation, and disciplined lifecycle management across distributed networks. That is the language of procurement, not prophecy.

The Japan trial and the software-first counterargument

Nokia’s counterpoint arrives in the form of the SoftBank agreement, reported by Telecompaper and others on October 20. The two companies have agreed to trial AI-RAN technology under commercial traffic conditions in Japan, with a field trial planned for 2027 using Nokia’s anyRAN software architecture playbook. The phrase “commercial traffic conditions” matters: this is not a lab demo, and the 2027 timeline gives Nokia room to mature the software while GPU supply chains and tooling catch up.

It is also a reminder that the AI-RAN race is increasingly global and operator-led. SoftBank has been among the more aggressive Japanese players in AI infrastructure, and pairing it with Nokia’s hybrid architecture gives the Finnish vendor a high-profile venue to demonstrate that GPU-based intelligence can work in a live commercial setting. If the trial produces credible spectral-efficiency numbers under real traffic, the burden shifts back to the ASIC camp to show its own scaling story.

But 2027 is also a long way off in a market that moves quarterly. Between now and then, Ericsson will be shipping AI features on silicon it controls, AT&T and Verizon will be measuring them, and the industry will be watching whether the gains compound or plateau.

The 6G clock is ticking

Complicating all of this is 6G. As RCR Wireless reported on October 20, the AI architecture for 6G remains unsettled even as the core timeline compresses. Speakers noted that companies differ on how much AI should be integrated into the network and how extensively the architecture should change, with concerns about AI reliability and sustainability complicating discussions. A compromise is expected through the 3GPP consensus process in the coming months, but not immediately.

That uncertainty cascades into silicon. If 6G is designed around AI-native air interfaces, intelligent resource allocation, and predictive network management—as market researchers at SNS Insider describe the emerging consensus—then the compute substrate becomes a first-order standards question, not a vendor preference. A network built for AI-native operation will need to run inference close to the radio, at low latency, under tight power budgets. Whether that inference runs on an ASIC, a GPU, or some heterogeneous mix is precisely the kind of decision that gets locked in years before deployment.

The VIAVI testing story is instructive here. The company’s AI RSG platform supports RAN simulation, training, testing, and validation across 4G, 5G, emerging 6G, Open RAN, and NTN environments, including Massive MIMO optimization, SON, rApps, xApps, and RIC functions—scaling from thousands of cells and UEs to city-scale scenarios. VIAVI also tests AI-native components to enable neuromorphic architectures and AI-driven PHY-layer innovation, and its AI Inference Test measures whether responses stay accurate, consistent, and timely under live RAN load. The message is that the industry now needs to validate AI behavior itself, not just radio behavior—and that validation has to happen across whatever silicon the vendor chose. Testing, in other words, is becoming silicon-agnostic even as vendors are not.

What Open RAN operators should take away

For operators building or evaluating Open RAN, the Ericsson-Nokia silicon split offers a few practical lessons.

First, treat spectral-efficiency claims as conditional. A doubling is not a property of a chip; it is a property of a chip, a workload, a baseline, and a radio environment. Ask for the baseline, the traffic profile, and the site conditions. Chetan Sharma’s point about headroom is the right lens: the worse your starting efficiency, the more dramatic the AI gain will look.

Second, watch where the intelligence lives. Verizon’s dual strategy—centralized rApps plus local gNodeB AI—is a pragmatic acknowledgment that not all AI belongs in the same place. Real-time optimization needs to be close to the radio; non-real-time policy and orchestration can live centrally. That split has architectural implications for the RIC, for fronthaul, and for compute placement, and it is exactly the kind of decision Open RAN’s disaggregated design was meant to make explicit.

Third, do not mistake a silicon debate for a standards debate. Open RAN’s value proposition is interoperability and multivendor choice. If the compute layer consolidates around one accelerator ecosystem, that value erodes regardless of how open the interfaces are. Operators should be asking vendors hard questions about portability, about software that runs across silicon, and about exit costs.

Fourth, expect the argument to be settled by economics, not benchmarks. Both AT&T and Verizon emphasized that AI must justify its compute, cost, and operational complexity. That is the gate that will ultimately decide whether custom ASICs or GPUs dominate the AI-native RAN—not a white paper, and not a CEO’s confidence.

The long game

Per Narvinger’s defense of Ericsson’s ASICs is, in the end, a bet that engineering depth beats ecosystem momentum. Nokia’s GPU strategy is a bet that the opposite is true, and that software velocity will outrun custom silicon. SoftBank will help test the second proposition in 2027; AT&T and Verizon are already testing the first in 2026.

What is clear is that the AI-RAN conversation has matured past the question of whether intelligence belongs in the network. It is now a fight over where that intelligence runs, who controls the substrate, and whether the gains are real enough to pay for. For Open RAN operators, that fight is not a spectator sport. The silicon choices being made today will determine how open the network actually is tomorrow—and whether the intelligence at the edge is a competitive market or a new kind of lock-in.

Narvinger, for his part, seems untroubled. His silicon, he suggests, is an extraordinary asset. The next eighteen months will tell us whether the market agrees.

Sources

• Ericsson CEO champions his silicon over Nvidia chips in RAN — Light Reading
• AI in the RAN moves toward the scale test — RCR Wireless News
• SoftBank and Nokia line up Japanese AI-RAN field trial for 2027 — Telecompaper
• Nokia, SoftBank Plan Joint Commercial AI-RAN Field Trial in Japan in 2027 — Moomoo
• AI architecture remains unsettled as 6G core timeline compresses — RCR Wireless News
• AI-Driven Telecom Networks: VIAVI on Testing and Network Automation — Electronics Media
• 6G Network Infrastructure Market Size, Share & Growth Report, 2035 — SNS Insider