Nokia’s dApp Gamble: E3, Nvidia GPUs, and the Fight for Open RAN’s Next App Economy

Executive briefing — October 1, 2026

The Bottom Line

Nokia is making a deliberate, public bet that the next phase of Open RAN value creation runs through a new class of network application — the dApp — built on the O-RAN-compliant E3 interface and, over time, on Nvidia GPUs as a general-purpose substitute for custom silicon. According to reporting from ETTelecom, CEO Justin Hotard claims 19 operators have signed up for “initial deployments,” and startups including Cohere Technologies (Pulsone), DeepSig, and ZaiNar are already writing RAN software for Nokia-built networks.

The strategic contrast is the story. Ericsson and Samsung are reported to be resistant to the dApp model. That means operators are no longer choosing only between vendors — they are choosing between two competing visions of how intelligence gets injected into the radio access network. This briefing lays out what is actually being proposed, what is proven, what is not, and what procurement and technology leaders should do about it.

What Nokia Is Actually Proposing

The mechanics matter more than the marketing. Nokia’s CTO Pallavi Mahajan framed the shift in unusually direct terms: “We’ve taken the whole concept of APIs and decoupling layers in a very serious fashion, and what it now enables us to do is bring in an O-RAN-compliant E3 interface.” That interface, per the reporting, opens the door to sensing, geo-precise location monitoring, and improved spectral efficiency.

The architectural logic runs like this. The first generation of Open RAN applications — xApps — required a near-real-time RAN intelligent controller (RIC) and the E2 interface. Interest in xApps was real but ultimately underwhelming; the use cases rarely justified the integration overhead. dApps are designed for tasks that demand faster execution and a real-time RIC, riding on the newer E3 interface. Nokia now supports that interface.

The second half of the bet is silicon. Nokia’s long-term vision involves a transition toward Nvidia GPUs as a general-purpose alternative to custom ASICs. Nokia has co-developed its latest custom silicon with Marvell Technology, but the company anticipates a future in which Nvidia chips become the primary alternative. Nvidia’s Aerial framework, which runs on the CUDA software platform, provides the software foundation. Mahajan noted it “provides a base for our teams and Nvidia to go out and test some of the models.”

That combination — an open, real-time application interface plus merchant GPU compute — is the substance behind the AI-RAN messaging. It is also, notably, a bet against the vertically integrated model that Ericsson and Samsung have defended.

Why the dApp Framing Is a Genuine Break From xApps

It is worth being precise about why this is not simply a rebrand. Cohere CEO Ray Dolan, speaking on a Telecoms.com podcast cited in the Light Reading coverage, argued that a dApp would be closer to a direct integration than the earlier server-based approach and would likely come with service benefits. His claim is that dApps would achieve what xApps did not.

That is a strong claim, and it should be treated as one. The honest read is that dApps are architecturally more ambitious — real-time execution, tighter integration, a newer interface — but the commercial track record does not yet exist. Operators evaluating this should separate three questions: Does the E3 interface deliver measurably better latency and control than E2-based xApps? Do the initial dApp use cases (sensing, location, spectral efficiency) map to revenue or cost problems an operator actually has? And does the dApp ecosystem have enough independent software vendors to avoid re-creating single-vendor lock-in under a new name?

On the third question, the early signal is encouraging. DeepSig, ZaiNar, and Cohere Technologies are all named as active developers. That is a small but real ecosystem, and it is more than xApps attracted at a comparable stage.

The GPU Question Is the Real Fork in the Road

The silicon debate is where this stops being an interface standards discussion and becomes a capital allocation decision.

Nvidia’s positioning is straightforward: GPUs as a general-purpose substitute for custom silicon in the RAN, with the Aerial framework as the software stack and CUDA as the developer platform. Nokia’s receptiveness to this is what many RAN observers regard as a massive gamble — Light Reading’s characterization, not ours. The upside is flexibility, a broader developer base, and the ability to run multiple workloads on shared accelerated infrastructure. The downside is power, cost per unit of RAN processing, and dependence on a vendor whose primary market is not telecom.

Ericsson’s demurral is telling. The company has built its AI-RAN narrative around purpose-built silicon — the AI-powered RAN Processor 6672, commercially deployed with Vodafone Egypt, which enables AI-native operations at the radio site with up to four times higher processing capacity and more than 50 percent lower energy use. That is a concrete, shipped product with published efficiency numbers. Nokia’s GPU-forward vision is a direction of travel with 19 signed-up operators and lab and field trials, not a fleet-wide deployment.

For operators, the practical question is not “GPUs or ASICs” in the abstract. It is whether the workload mix in a given market — AI inference at the edge, sensing, spectrum sharing, energy optimization — justifies general-purpose compute at the cell site, and whether the power and cooling budget at that site can absorb it.

What the Independent Evidence Says

The strongest third-party data point in this news cycle comes not from Nokia or Nvidia but from GlobalLogic, the Hitachi Group digital engineering company. Its proof-of-concept trials in a 5G Open RAN environment produced processing time improvements of as much as 15 times under certain conditions, with power consumption falling by around 40 percent. A second experiment using deep-learning models to distinguish genuine network-access attempts from noise or interference reported a 99.98 percent reduction in false detections.

Two caveats apply, and GlobalLogic itself is careful to flag them. These are controlled-environment results, not commercial network outcomes, and they do not prove AI-RAN is ready to transform commercial mobile networks overnight. But they do establish something important: the underlying technologies can produce measurable gains when AI and GPU acceleration are brought into an Open RAN environment. That is a materially different claim from vendor marketing, and it is the kind of evidence a skeptical CTO can take to a board.

The GlobalLogic work also reinforces a structural point that operators should internalize. AI-RAN is unlikely to be a single piece of equipment that an operator installs. It is better understood as an architectural shift bringing together radio systems, software, AI models, accelerated computing, and distributed processing. That framing has direct procurement implications: you cannot buy AI-RAN as a SKU, and any business case built around a single line item is likely wrong.

The Broader Momentum — and the Hype Risk

Context matters here. Nokia has reported growing global momentum for AI-RAN, with operators across North America, Europe, Asia-Pacific, and the Middle East advancing from early evaluation to lab and live field trials alongside partners such as Nvidia. The platform has delivered spectral efficiency improvements of more than 20 percent, according to the company. Nokia’s target of doubling spectral efficiency by 2028 on much of the existing infrastructure looks more achievable with its Cohere partnership, and Vodafone, Telstra, and Bell Canada have publicly endorsed the USM technology after field trials — all three have invested in the startup.

That is a credible set of endorsements from named, serious operators. It is also not the same as commercial scale. The 19 “initial deployments” figure deserves scrutiny: Nokia’s own CEO identified three companies in his LinkedIn post, and per Light Reading, none of them is a telco. That gap between the headline number and the named participants is exactly the kind of detail an executive briefing should flag rather than smooth over.

Meanwhile, the industry is converging on the same architectural thesis from multiple directions. The Linux Foundation’s ONE Summit 2026 program in Tokyo dedicates tracks to open source RAN and next-generation wireless including OCUDU, AI-RAN, A2A-T, 6G and non-terrestrial networks, GPU/NPU acceleration, and mission-critical private networks. Analyst coverage continues to point to strong growth in adjacent categories — Analysys Mason estimates the market for Harmonic’s cOS SensAI intelligence fabric at $2.2 billion, a useful proxy for how much operators are willing to spend on intelligence layered over the RAN.

What This Means for Operators: Five Decisions

1. Treat the E3/dApp question as a procurement gate, not a roadmap slide. If you are writing a multi-year RAN contract in the next four quarters, the presence or absence of O-RAN-compliant E3 support is now a material differentiator between vendors. Ask each bidder directly whether they support E3, what dApps are commercially available on their platform today, and what the integration path looks like for third-party software. Ericsson’s and Samsung’s reported resistance is a legitimate strategic position — but you should know you are buying it.

2. Separate the interface decision from the silicon decision. These are being bundled in vendor messaging, and they should not be. E3 support is an interoperability and ecosystem question. GPU-based compute at the RAN is a power, cost, and site-engineering question. It is entirely coherent to want the former while being skeptical of the latter in dense urban sites with constrained power budgets.

3. Demand controlled-environment numbers be replicated in your network. The GlobalLogic results — up to 15x processing improvement, roughly 40 percent lower power, 99.98 percent false-detection reduction — are the right benchmark to hold vendors to. Ask for a field trial design that measures the same variables on your traffic mix before you commit capex.

4. Price the ecosystem risk. A dApp economy is only valuable if enough independent developers show up. Three named startups is a start, not a market. Build contract terms that protect you if the dApp ecosystem stalls — for example, commitments on interface stability and the right to fall back to conventional RAN software without penalty.

5. Watch the energy math above all else. Ericsson’s 6672 claims more than 50 percent lower energy use at the radio site; GlobalLogic measured roughly 40 percent lower power consumption in its Open RAN proof of concept. Different tests, different baselines, but the same direction. In a market where opex increasingly determines RAN economics, energy efficiency is the metric most likely to survive contact with a CFO.

The Skeptical View

The case against Nokia’s bet is not that it is wrong — it is that it is early and expensive. GPUs at the RAN edge carry power and cost penalties that purpose-built silicon avoids. The xApp experience showed that Open RAN application ecosystems can attract interest without achieving commercial scale. And the AI-RAN narrative has been running ahead of deployable revenue for several years now.

The case for it is that the alternative — closed, vendor-specific intelligence — has a well-documented history of slow innovation and limited operator leverage. If AI workloads in the RAN are going to grow, and the evidence from multiple independent tests says they will, then an open application interface plus merchant compute is the more adaptable architecture. Nokia is not guaranteeing it wins. It is guaranteeing that it is in the race.

What to Watch Next

• Whether Ericsson or Samsung softens its position on dApps and E3, or doubles down on integrated silicon.
• Whether the 19 “initial deployments” convert into named, commercial telco deployments with published results.
• Whether Cohere’s USM technology delivers on the doubling-of-spectral-efficiency target by 2028, and whether Vodafone, Telstra, and Bell Canada expand beyond trials.
• Whether the ONE Summit Tokyo discussions around OCUDU and open source RAN produce reusable dApp frameworks that lower integration cost for everyone.
• Whether independent test houses replicate the GlobalLogic results on live commercial networks.

The decision window is open now. Operators signing multi-year RAN contracts in the next twelve months will effectively choose which of these two architectures they are betting on for the rest of the decade.

Sources

• Nokia bets on dApps, Nvidia GPUs for 5G open RAN competitiveness: Report — ETTelecom
• Nokia puts new open RAN tech in 5G revival bid as Ericsson demurs — Light Reading
• LATAM mobile networks could become smarter with AI-RAN and GPU acceleration — Intelligent CIO LATAM
• Enterprise IT & Communications Technology Industry News — Pipeline Magazine
• AI Takes Center Stage at Open Networking & Edge (ONE) Summit 2026 Program — PR Newswire