The Real AI-RAN Story: US Carriers Diverge on GPUs, and That's Exactly Why Open RAN Wins

The Real AI-RAN Story: US Carriers Diverge on GPUs, and That’s Exactly Why Open RAN Wins

Something genuinely interesting happened in the AI-RAN conversation this week. It wasn’t a flashy demo. It wasn’t another hyperscaler partnership. It was a quiet, almost mundane revelation: the three largest US carriers are taking fundamentally different paths to AI-RAN, and none of them are wrong.

According to reporting from ETTelecom and Communications Today, Verizon CTO Yago Tenorio is openly skeptical about GPUs for core RAN processing, arguing that current and near-future RAN workloads—including AI inferencing for radio performance—can be handled efficiently by CPUs. T-Mobile, meanwhile, is going all-in with Nokia and NVIDIA at its AI-RAN Innovation Center, with Network CTO Ankur Kapoor expecting AI-RAN radio prototypes later this year and touting Nokia’s claims of 50% spectral efficiency gains by late 2027. AT&T, as ever, is the pragmatist in the middle: evaluating multiple GPU platforms against a power, price, and performance framework, deploying GPUs only where they fit network economics, and testing both Ericsson’s GPU-free AI-native scheduler and Intel Xeon 6 cloud RAN systems.

This is not a story about who’s right. It’s a story about why the open, disaggregated architecture that Open RAN champions is suddenly the most valuable asset in the AI-RAN era. And it’s a story that should make every Open RAN operator feel a little more confident about the road ahead.

The Great GPU Debate: A Feature, Not a Bug

Let’s be clear about what’s happening here. Verizon isn’t anti-AI. T-Mobile isn’t reckless. AT&T isn’t indecisive. They’re responding to the same technological inflection point with different business models, different network topologies, and different risk tolerances. That’s what healthy markets do.

Verizon’s position is the most technically conservative, and it deserves respect. Tenorio’s argument—that CPUs can handle AI inferencing for radio performance without the complexity of reprogramming the entire radio stack with GPUs—is grounded in a real engineering concern. GPUs are power-hungry, expensive, and require a fundamentally different software stack. If you can get the spectral efficiency gains you need from a CPU-based architecture, why introduce that complexity? Verizon has spent years building a network known for reliability and coverage. Disrupting that with a wholesale GPU swap is not obviously a winning trade.

T-Mobile’s position is the most aggressive, and it also makes sense. T-Mobile has built its brand on network leadership—it consistently wins third-party speed and coverage awards. If Nokia and NVIDIA can deliver 50% spectral efficiency gains by late 2027, as Kapoor suggests, that’s a massive competitive advantage. T-Mobile is willing to bet on the bleeding edge because it has the most to gain from being first. The AI-RAN Innovation Center isn’t a research project; it’s a strategic weapon.

AT&T’s position is the most nuanced, and frankly, the most instructive. By defining AI-RAN broadly—covering any use of AI in network operations from troubleshooting to optimization—AT&T avoids the trap of equating AI-RAN with GPUs. Its evaluation framework is explicitly economic: power, price, and performance, specific to RAN requirements. GPUs get deployed where they fit; CPUs and dedicated hardware get deployed where they don’t. This is not a lack of conviction. It’s a mature procurement strategy.

The fact that these three carriers can pursue such different paths simultaneously is a testament to the flexibility of modern RAN architectures. And it’s a direct validation of the Open RAN principle that disaggregation—of hardware, software, and now compute—creates optionality.

Why This Matters for Open RAN

For years, Open RAN skeptics have argued that disaggregation introduces complexity without commensurate benefit. They’ve pointed to integration headaches, performance gaps, and the difficulty of managing multi-vendor networks. Those criticisms weren’t entirely wrong. But the AI-RAN era is exposing a new truth: the ability to mix and match compute platforms—CPU, GPU, and specialized accelerators—is not a bug. It’s the whole point.

Consider what’s happening beneath the surface of this US carrier divergence. AT&T is testing Ericsson’s GPU-free AI-native scheduler on dedicated hardware and Intel Xeon 6 cloud RAN systems. T-Mobile is working with Nokia’s anyRAN software and NVIDIA’s Aerial RAN Computer. Verizon is evaluating whether GPUs are necessary at all. Each of these approaches relies on a disaggregated architecture where the compute layer is decoupled from the radio layer. That decoupling is what allows each carrier to optimize for its own economics.

In a traditional, vertically integrated RAN, you buy a baseband unit from one vendor and you get whatever compute architecture that vendor has chosen. Your only lever is the vendor’s roadmap. In an open, disaggregated RAN, you can choose your compute platform independently. You can run AI workloads on CPUs today and migrate to GPUs tomorrow if the economics change. You can deploy GPUs in some parts of your network and CPUs in others. You can hedge your bets.

This is exactly what AT&T is doing. And it’s exactly what Open RAN was designed to enable.

The Nokia-NVIDIA Juggernaut and the Microsoft Wildcard

While the US carriers debate GPU strategies, Nokia is quietly assembling an AI-RAN ecosystem that spans the globe. On September 16, Nokia announced that eight named operators—A1 Group, Chunghwa Telecom, du, e&, Mobily, stc, TPG Telecom, and Zain Saudi—are advancing proofs of concept or live field trials using its AI-RAN platform and NVIDIA’s Aerial RAN Computer. Earlier work with T-Mobile, SoftBank, Indosat Ooredoo Hutchison, and NTT DOCOMO continues.

Nokia says its AI-powered radio technology has demonstrated spectral efficiency gains of more than 20% in testing, with further software improvements planned for 2027 and 2028. The company is targeting pilot deployments before the end of 2026 and a commercial rollout next year. That’s an aggressive timeline, but Nokia has been methodical: it launched its commercial AI-RAN platform in July 2026, and the NVIDIA partnership—announced in October 2025, cemented by a $1 billion NVIDIA investment—has moved from research to productization in under a year.

Then there’s the Microsoft angle. Nokia expanded its partnership with Microsoft to build a unified, agent-based data foundation for network operators. The goal is to give carriers a common platform for automating mobile networks, with use cases like autonomous VoNR assurance, RAN geo experience, and predictive maintenance. Microsoft supplies the software and cloud layer; Nokia contributes its network expertise. The pitch is that data access times drop from weeks to minutes, enabling AI agents to act on network data in real time.

This is a smart move by Nokia. It positions the company not just as a RAN vendor but as an AI-native network automation platform. And it gives operators a path to AI-RAN that doesn’t require them to rip and replace their entire RAN. The Microsoft partnership is about layering intelligence on top of existing infrastructure—a software-first approach that complements the hardware-heavy GPU narrative.

But let’s be skeptical for a moment. Nokia’s stock jumped on the AI-RAN news, and the company is preparing to rejoin the Euro STOXX 50. There’s a lot of hype swirling around AI-RAN right now. The spectral efficiency claims—20%, 30%, 50%—are impressive, but they’re coming from vendors with a vested interest in the narrative. The real test will be commercial deployments in 2027 and 2028, when operators have to prove that these gains translate into real-world economics.

The CPU vs. GPU Question: A False Dichotomy?

Here’s where I think the industry conversation is getting a little lazy. The framing of “CPU vs. GPU” for AI-RAN is misleading. The real question isn’t which chip is better. It’s which chip is better for which workload, at which layer of the network, at which point in time.

Verizon’s Tenorio is right that CPUs can handle a lot of AI inferencing for radio performance. Modern x86 processors, especially Intel’s Xeon 6 line, are incredibly capable. They’re also more power-efficient and easier to program than GPUs for many RAN workloads. If you’re running a small cell with modest AI requirements, a CPU-based architecture might be perfectly adequate.

But T-Mobile’s Kapoor is also right that GPUs unlock capabilities that CPUs can’t match. NVIDIA’s Aerial RAN Computer is designed specifically for AI-RAN workloads, and the company’s CUDA ecosystem gives it a software moat that’s hard to replicate. If you’re trying to push spectral efficiency to its theoretical limits, GPUs may be the only way to get there.

The beauty of Open RAN is that operators don’t have to choose. They can deploy CPUs in some scenarios and GPUs in others. They can start with CPUs and migrate to GPUs as the software matures. They can run AI workloads on general-purpose hardware and offload specific functions to accelerators. The architecture supports heterogeneity.

This is a fundamentally different value proposition than the one offered by traditional RAN vendors, who tend to push a single, vertically integrated solution. Open RAN says: here’s the interface, here’s the standard, now go build the best network you can with whatever components make sense. That’s a powerful idea, and it’s becoming more powerful as the compute landscape diversifies.

The Policy Wildcard: AI for Secure Networks Act

While the industry focuses on GPUs and spectral efficiency, a quiet piece of legislation is moving through Congress that could have a significant impact on AI-RAN adoption. The House Committee on Energy and Commerce recently passed the AI for Secure Networks Act (H.R. 7294), which directs the NTIA to study how AI affects the security and resilience of telecom infrastructure.

The bill specifically calls out Open RAN security, along with AI-enhanced threat detection, malware blocking, zero-trust architectures, and energy efficiency. It also directs the study to examine the cyber risks posed by advanced AI and autonomous agents to critical communication systems.

This is a big deal for two reasons. First, it signals that policymakers are taking AI-RAN security seriously—which is good, because the attack surface of an AI-native network is fundamentally different from a traditional one. Second, it creates a potential funding and research pipeline for Open RAN security, which could accelerate the development of standards and best practices.

For Open RAN operators, the message is clear: security can’t be an afterthought in the AI-RAN era. If you’re deploying AI agents to manage your network, you need to think about how those agents are authenticated, how their decisions are audited, and how they’re protected from adversarial manipulation. The AI for Secure Networks Act is a step in the right direction, but it’s just a study. The real work will happen in the standards bodies and the vendor labs.

The Huawei Ban and the Open RAN Opportunity

Finally, it’s worth noting the broader geopolitical context. Seventeen European telecom CEOs recently pushed back against the European Commission’s revised Cybersecurity Act, which would mandate the removal of Chinese equipment from Huawei and ZTE. They argue that full replacement would cost up to €40 billion, draining resources earmarked for fiber, 5G, and 6G investment.

This is a complex issue with no easy answers. But it’s relevant to the AI-RAN conversation because it highlights the stakes of vendor choice. If European operators are forced to replace Huawei equipment, they’ll be making massive procurement decisions that will shape their networks for a decade. The temptation will be to replace like-for-like—traditional equipment with traditional equipment—rather than embracing Open RAN.

But that would be a mistake. The Connect Europe report released in early September concluded that “the Open RAN market in the EU is still nascent, with most operators relying on single-vendor end-to-end solutions.” That’s a problem. If European operators replace Huawei with Ericsson or Nokia in a single-vendor configuration, they’ll be locking themselves into the same closed architecture that created the Huawei dependency in the first place.

The AI-RAN era offers a chance to break that cycle. By adopting open, disaggregated architectures, European operators can diversify their vendor base, avoid lock-in, and position themselves for the AI-native future. The US carriers’ divergent GPU strategies show that this is possible. It’s not easy, and it requires a level of technical sophistication that not every operator has. But it’s the right path.

The Closing Take: Divergence Is Strength

Here’s my take: the fact that AT&T, T-Mobile, and Verizon are taking different paths to AI-RAN is the best news the Open RAN community has had in years. It proves that disaggregation works. It proves that operators can make independent technology choices without sacrificing performance. It proves that the AI-RAN market is maturing beyond a single vendor’s roadmap.

For too long, the RAN market has been defined by a handful of vendors selling vertically integrated solutions. Open RAN promised to change that, but progress has been slow. The AI-RAN era is accelerating the change. When compute becomes the differentiator, the ability to mix and match compute platforms becomes a competitive advantage. And that’s exactly what Open RAN delivers.

So let’s stop asking which carrier is right about GPUs. The better question is: how quickly can the industry build the open interfaces, standards, and ecosystems that make this divergence sustainable? The answer will determine whether AI-RAN becomes a driver of innovation or just another vendor lock-in story. I’m betting on innovation. But it’s going to take work.

Sources

Related Posts

AI-RAN's Commercial Call: 5 Things the Nokia-NVIDIA GTC Berlin Wave Means for Open RAN Operators

Nokia and NVIDIA now count nearly 20 operators on AI-RAN. Five takeaways from T-Mobile's live call, Elisa's field trial and the 2027 pipeline.

AI-RAN Leaves the Lab: A Skeptical Exec's Q&A on T-Mobile's Live Call, Elisa's Field Trial, and Nokia-NVIDIA's 2027 Playbook

A skeptical exec's Q&A on T-Mobile's live AI-RAN call, Elisa's European field trial, and what Nokia and NVIDIA's 2027 commercial timeline really means.