One day before this article was written, Nokia launched what it called the industry's first commercial AI-native radio access network platform — built entirely on Nvidia's Aerial AI-RAN software and the CUDA-powered GPU infrastructure that Qualcomm is spending billions to dislodge. The timing is not a coincidence. Qualcomm has quietly been winding down its small-cell hardware business, and what looked at first like a retreat from wireless infrastructure is better read as a reallocation: Qualcomm is abandoning a market it cannot dominate in 5G to position itself as the compute backbone of the next wireless generation. Whether that strategy succeeds depends on whether Qualcomm can own the silicon layer Nokia just handed to Nvidia.

Qualcomm Stopped Selling Small-Cell Chips to New Customers Late Last Year

Within its Dragonwing product catalog, Qualcomm offered small-cell chips under the FSM100 and FSM200 labels to OEM customers building 5G hotspots for high-traffic indoor environments like airports, stadiums, and office complexes. Late last year, the company began notifying those customers that it would no longer sell the FSM100 and FSM200 platforms to new buyers. Only one further software release is planned. Members of the Qualcomm team involved in small-cell development have already departed the company, according to a source familiar with the matter cited by Light Reading. When asked for comment, a Qualcomm spokesperson declined to address what the company characterized as "rumors or speculation."

The exit clears Qualcomm from a 5G small-cell market where Ericsson and Nokia hold dominant positions and have established multi-year customer relationships that a chip supplier cannot easily displace. But the strategic logic runs the other way: Qualcomm is not retreating because it lost — it is stepping back from a market where winning would have been incremental to position for one where winning would be structural.

Nokia's Nvidia Bet Makes Qualcomm's 6G Ambitions More Specific

The competitive dynamic sharpened on July 15, when Nokia announced its commercial AI-RAN platform alongside Nvidia. The product is built on Nokia's anyRAN software combined with Nvidia's Aerial AI-RAN stack — which runs on CUDA-powered GPUs. Nokia is targeting more than 100% spectral efficiency gains by 2028, offering operators three deployment paths: a GPU-powered plug-in to existing AirScale radios, a standalone AI-RAN node, and a CUDA-server option for cloud-native networks, according to the Futurum analyst brief on the launch. T-Mobile is the only operator with confirmed dates on Nokia's timeline, with a field trial expected before the end of 2026 and commercial service in 2027.

Nokia's announcement is not just a product launch — it is a strategic declaration. Nokia is moving from its proprietary baseband silicon to Nvidia's merchant GPU platform for AI-native radio processing. That is the same compute layer Qualcomm is trying to own in 6G, and Nokia just told the market which chip company it has chosen for it.

What Qualcomm Is Actually Building for 6G: Layer 1, Layer 2, and the Stack Problem

To understand the 6G play, it helps to know what the current one lacks. In Open RAN architecture, a base station's functions are disaggregated across three software and hardware layers. Layer 1 — the physical layer — handles the most computationally intensive work: processing the raw radio signals, running massive MIMO algorithms, performing beamforming (concentrating radio energy spatially toward individual devices), and managing the digital front-end that converts analog RF signals into digital samples. This is real-time work — it must complete in microseconds — and it requires either a specialized ASIC or a high-performance accelerator, as explained in a technical overview of Open RAN functional splits.

Qualcomm's X100 RAN Accelerator Card handles Layer 1. But it cannot be deployed alone. It must be paired with a separate CPU running the Layer 2 software — scheduling, resource allocation, and link control — and the Layer 3 network functions above that. In Viettel's deployment in Vietnam, which became the world's first commercial Open RAN 5G network to use Qualcomm's X100 platform when it launched in November 2024, Viettel designed its own solution for those upper layers. That is not a repeatable model for most operators.

"The goal is to have a CPU that is really addressing the RAN workloads more power efficiently and also use the NPU for any AI workloads," Sunil Patil, Qualcomm's vice president of product management for the RAN, said at MWC Barcelona in March. The integrated 6G telco server Qualcomm has sketched in slideware would pair the X100-successor chip for Layer 1 with an Oryon CPU — Qualcomm's own Arm-based processor — for Layers 2 and 3, plus a Hexagon NPU for AI inference workloads running in the network. For the first time, Qualcomm would own the full distributed-unit compute stack, not just the accelerator.

That completeness matters. Ericsson has no interest in buying Qualcomm's L2 and L3 software — Patil has said as much directly. But Ericsson's CEO, Börje Ekholm, named Qualcomm on a recent earnings call as a potential future hardware supplier for the layers where Qualcomm would offer only silicon. The established vendors could buy the chip without ceding the software. Whether that kind of partial engagement is what Qualcomm actually wants is a different question.

Viettel's current network is targeting 5,000 deployed sites by the end of 2026, with some already in commercial operation, according to Caleb Banke, Qualcomm's staff manager for product marketing, speaking at MWC Barcelona. On live networks under heavy traffic loads, the deployment has demonstrated up to 24% better power efficiency compared with traditional network hardware, a figure Qualcomm and Viettel have cited jointly — though it has not been independently audited by a third party.

Why Qualcomm Is Buying a Compiler, Not Just a Chip

Qualcomm's most consequential move in the 6G silicon race may not be a chip at all. In June, the company announced the acquisition of Modular — the AI software startup behind the Mojo programming language and the MAX inference engine — in an all-stock deal valued at approximately $3.92 billion. The deal has not yet closed; regulatory review is expected to conclude in the second half of 2026, according to the official Qualcomm announcement.

Modular was founded by Chris Lattner, the engineer who built LLVM (the open-source compiler infrastructure underlying most modern programming languages), created Apple's Swift language, and previously led Tesla's Autopilot software. At Modular, Lattner spent four years building Mojo — a programming language designed to match C-level performance with Python's usability — and MAX, an inference engine that runs the same AI model code on chips from Nvidia, AMD, Intel, Apple, and Qualcomm without per-chip rewrites, as NAND Research analyst Steve McDowell explains.

The reason this matters for 6G radio is CUDA. Nvidia's Compute Unified Device Architecture, introduced in 2006, has accrued roughly four million developers over two decades and created a software moat that makes changing GPU vendors far more expensive than the hardware difference alone would suggest — because code tuned for CUDA cannot run on AMD, Qualcomm, or Intel silicon without a rewrite. Nokia's decision to build its AI-RAN platform on Nvidia's Aerial stack means Nokia's operator customers are now writing or adopting CUDA-dependent workloads for their networks. Operators who go deep on CUDA will need a compelling software reason to consider Qualcomm silicon later.

Modular's MAX is designed to be that reason: a hardware-agnostic inference layer that lets developers write once and run anywhere, including Qualcomm's Hexagon NPU. If it works as promised, it reduces the switching cost that makes CUDA so sticky. But that promise depends entirely on what Qualcomm does with the platform after the acquisition closes. The risk, noted by analysts following the deal, is that commercial incentives will push Qualcomm to optimize MAX specifically for its own silicon rather than maintaining true hardware-agnosticism. Dango Daily's analysis of the acquisition specifically flags that if Qualcomm closes off MAX post-acquisition, it could trigger industry pushback. A developer who migrates from CUDA to MAX only to find MAX performs best on Qualcomm hardware has not escaped lock-in — they have simply changed the company holding the key. Modular has committed to keeping MAX open to third-party hardware, but that commitment belongs to a pre-acquisition company.

The Pressure Behind the Pivot: Smartphones, Apple, and China

Qualcomm's core handset business faces structural pressure on multiple fronts. Device shipments grew just 1% year over year in the most recent quarter, as AI-driven demand for memory and other components has raised smartphone prices enough to delay consumer upgrades, according to Omdia analyst Runar Bjorhovde. Apple, Qualcomm's largest US customer, is in the process of shifting iPhone modems to in-house chips — a transition that analysts estimate could remove $7 to $8 billion in annual Qualcomm revenue beginning in 2027, when the current supply agreement is expected to expire.

Geographic concentration compounds the vulnerability. Chinese customers accounted for 46% of Qualcomm's revenues in its most recent fiscal year, up from 37% two years earlier, at a time of sustained US-China trade tension. Qualcomm generated $44.3 billion in revenues and $12.4 billion in operating profits for that fiscal year, making its China exposure substantial in absolute terms. Meanwhile, anyone following Qualcomm's executive team at its June investor day had to wait two hours into the presentation to hear about 6G — because AI infrastructure and data centers dominated the agenda. CEO Cristiano Amon has made clear that those markets, not smartphones, are where he expects Qualcomm's next decade of growth to come from.

The RAN infrastructure play fits that frame: it is a higher-margin, longer-cycle business than handset chips, with deeper technical lock-in once adopted and less exposure to the consumer upgrade cycle that makes handset revenues volatile.

The Chip Wars Shaping 6G Radio

At MWC Barcelona in March, the industry's strategic divide over 6G compute was visible in real time. Nokia and Samsung moved toward merchant silicon — GPUs and general-purpose CPUs that operators can share between radio workloads and AI inference jobs, potentially generating a second revenue stream by renting idle capacity to AI customers, according to 650 Group's analysis of MWC 2026. Ericsson, Huawei, and ZTE moved in the opposite direction, committing to purpose-built ASICs that they argue deliver superior power efficiency and total cost of ownership when optimized for the specific real-time constraints of Layer 1 processing.

Qualcomm is attempting a third path: an NPU-based merchant silicon architecture that claims power efficiency competitive with custom ASICs while maintaining the flexibility and dual-use economics of a general-purpose platform. The Hexagon NPU — which Qualcomm has also been integrating into its AI200 and AI250 inference ASICs — is the technological bet behind both the 6G RAN play and the broader data center campaign. Whether it can deliver power-efficient Layer 1 processing and AI inference in a single basestation package, at the scale operators will require, is the technical question the Viettel deployment is supposed to begin answering.

The historical parallel that Qualcomm's leadership keeps invoking is Nvidia's transformation from a graphics card company into the default compute infrastructure for AI. Nvidia's five-year stock surge of approximately 911% reflects the leverage that comes from owning the compute fabric for a dominant new workload. Qualcomm is betting that 6G radio networks will generate a similarly large and durable compute demand — and that it can occupy the silicon position in that network before Nvidia or Ericsson locks it down.

Qualcomm's first 6G silicon is targeted for 2028, with early commercial launches in 2029 and deployment at scale by 2030, CEO Cristiano Amon said on the company's Q2 FY2026 earnings call. The 3GPP standards body locked the Release 21 specification timeline in June, setting a March 2029 freeze date — the point at which chip vendors can begin designing certified 6G silicon. That timeline is tight, and it is the same window Nvidia's Nokia alliance is building toward.

The small-cell exit, seen in this light, is not an admission of defeat. It is a conservation of engineering resources for a bet that carries far higher stakes — and far less margin for error.

Frequently Asked Questions

Why did Qualcomm exit the small-cell market?

Qualcomm stopped selling its FSM100 and FSM200 small-cell chips to new customers starting in late 2025, keeping support in place for existing clients but ending new-buyer engagement. The exit frees engineering resources for 6G RAN silicon, where Qualcomm believes the compute market is substantially larger and structurally more defensible than in 5G small-cell hardware, where Ericsson and Nokia hold commanding positions.

What is Nokia's AI-RAN platform and why does it matter for Qualcomm?

Nokia launched its commercial AI-native RAN platform on July 15, 2026, built on Nvidia's Aerial AI-RAN software and CUDA-powered GPU hardware. The platform targets more than 100% spectral efficiency gains by 2028 and will enter T-Mobile field trials before year-end, with commercial service in 2027. Its significance for Qualcomm is competitive: Nokia chose Nvidia's silicon for the Layer 1 compute role that Qualcomm is positioning to fill in 6G. The more operators build AI-RAN workflows on CUDA, the stronger Nvidia's software moat becomes — which is exactly what the Modular acquisition is designed to counter.

How does Qualcomm's Modular acquisition address Nvidia's software advantage?

Modular's Mojo programming language and MAX inference engine let developers write AI code once and deploy it across chips from Nvidia, AMD, Intel, and Qualcomm without hardware-specific rewrites. This directly targets CUDA lock-in, which has kept roughly four million developers tied to Nvidia hardware for nearly two decades. The deal, valued at approximately $3.92 billion and announced June 24, is expected to close in the second half of 2026. The critical open question is whether Qualcomm will keep MAX genuinely hardware-agnostic after the acquisition closes or optimize it primarily for Qualcomm silicon — which would replace one form of lock-in with another.

Will Qualcomm compete directly with Ericsson and Nokia in 6G infrastructure?

Not exactly. Qualcomm's stated 6G strategy involves providing the silicon and compute platform — the Layer 1 accelerator and eventually a full distributed-unit server including CPU and NPU — while leaving Layer 2 and Layer 3 software to established RAN vendors. Ericsson's CEO has named Qualcomm as a possible future hardware supplier, which suggests a co-existence model is at least plausible. What Qualcomm is betting against is a future where traditional vendors lock operators into proprietary compute stacks the way they have in 5G — and in that bet, its most direct competitor is not Ericsson but Nvidia.