1Finity's AI-RAN Reality Check: Build Where the Customer Is Already Paying

1Finity’s AI-RAN Reality Check: Build Where the Customer Is Already Paying

Every few months the AI-RAN conversation gets a fresh coat of paint. A new field trial, a new GPU-in-the-tower announcement, a new alliance milestone. And every few months, someone quietly asks the question that the ecosystem prefers to whisper: who is actually paying for this?

This week, 1Finity—the network infrastructure business carved out of Fujitsu—gave that question its most direct answer yet. In an interview with RCR Wireless published on October 5, the company argued that operators should build AI-RAN capacity where a customer is already paying, rather than deploying GPU-accelerated radio infrastructure on speculation that AI workloads will eventually show up and justify the investment. The framing matters: 1Finity’s position is not that AI-RAN is a dead end. It is that AI-RAN performance gains, on their own, will not justify the capital outlay.

That is a much more interesting claim than the usual vendor optimism, and it deserves more attention than it will probably get. Because it reframes the entire AI-RAN debate from a technology question into a business model question—and the Open RAN community has historically been much better at the former than the latter.

The uncomfortable arithmetic of GPU-powered radios

The logic behind AI-RAN is seductive in its simplicity. Put general-purpose compute—typically NVIDIA GPUs—into the RAN footprint. Use it to run RAN workloads more efficiently. Then monetize the idle capacity by selling edge AI services to enterprises. The RAN becomes a revenue platform instead of a cost center. Operators get a second act. Vendors get a new product category. Everyone wins.

Except the arithmetic only works if the second half of that story materializes. RAN workloads are, by and large, already handled efficiently by purpose-built silicon. The performance gains from moving them onto GPUs are real but incremental, and they come with a significant increase in power draw, thermal complexity, and operational overhead. If the only justification for GPU deployment is a modest spectral efficiency improvement, the business case collapses under its own weight.

The entire bet, then, rests on the edge monetization story. And that story requires something the telecom industry has never been particularly good at: selling enterprise services at scale, at the edge, at margins that justify telecom-grade infrastructure costs.

1Finity’s argument is that operators should invert the sequence. Don’t build the AI-RAN platform and then go looking for customers. Find the customer first—the one already paying for something that AI-RAN infrastructure can serve better, cheaper, or closer—and build there. That is a profoundly unglamorous position in an industry that has spent the last three years publishing press releases about GPU deployments and spectral efficiency multipliers.

It is also, almost certainly, correct.

What the capex data actually says

The timing of 1Finity’s intervention is not accidental. Global telecom capex rose roughly 5% in the first half of 2026, according to TelecomLead, with operators shifting spending toward capacity, AI, and network modernization. That headline number sounds bullish for AI-RAN until you look at where the money is going. Operators are not writing blank checks for intelligence at the edge. They are funding specific projects with specific justifications—capacity where congestion is real, modernization where equipment is end-of-life, AI where it produces measurable operational savings.

Consider Deutsche Telekom’s AI program, reported this week by AI News. The operator is targeting €2.5 billion in savings from AI, and the use cases it highlights are revealing: a RAN Guardian Agent that detects impending network pressure from large events and cuts response times from hours to about a minute; AI agents handling 40% of customer contacts in the US market; a 55% reduction in customer service calls; a 30% drop in complaints from early AI use in new-connection setup. The company expects roughly €250 million in AI-related business revenue outside the US in 2026, scaling toward €800 million by 2030.

Notice the pattern. The savings are concrete, operational, and near-term. The revenue is smaller, slower, and further out. Deutsche Telekom is not justifying its AI investments on speculative edge monetization. It is justifying them on cost reduction and service improvement—the unsexy stuff that actually shows up in quarterly results.

That is the environment in which AI-RAN vendors are trying to sell GPU infrastructure. And it explains why 1Finity’s “build where the customer is already paying” framing is likely to resonate with the operators who actually sign the purchase orders, as opposed to the ones who appear on conference panels.

The Open RAN connection nobody wants to make

Here is the part of this story that should concern anyone who has spent the last several years arguing for Open RAN on economic grounds.

Open RAN’s original pitch was disaggregation as a cost play. Break the monolithic RAN into interoperable components, introduce competition at every layer, and let operators mix and match vendors to drive down prices. That pitch has evolved—partly because the cost savings have been harder to realize than promised, and partly because the industry needed a new narrative. AI-RAN became that narrative. The disaggregated architecture, the argument goes, is the ideal substrate for intelligence at the edge. Open interfaces let operators plug in AI applications as xApps and rApps. The RIC becomes a platform. The RAN becomes programmable.

All of that is technically true. But it quietly shifts Open RAN’s justification from a cost argument to a revenue argument—and revenue arguments require customers. If the AI-RAN business case depends on selling edge services to enterprises, then Open RAN’s fate is now tied to the telecom industry’s ability to compete in a market where it has historically struggled against cloud providers and systems integrators.

1Finity’s position implies that this sequencing is backwards. If AI-RAN only makes sense where a paying customer already exists, then the Open RAN ecosystem should be building toward specific, identifiable enterprise use cases—not toward generic platform capabilities that assume demand will materialize.

This is a hard message for an ecosystem that has spent years building toward architectural elegance. But it is the message that the capex data supports.

The IMC 2026 signal

If you want to see the gap between AI-RAN rhetoric and AI-RAN reality, look at what was actually announced at India Mobile Congress this week.

HCLTech used the opening day to outline an AI-native telecom showcase covering autonomous networks, OSS modernization, network monetization, AI-powered service assurance, Open RAN engineering, and the transition toward 6G. Notably, Communications Today characterized it as “an announced exhibition programme rather than a confirmed product launch.” That distinction matters. The industry is full of AI-native positioning that has not yet converted into deployable products with paying customers.

Ericsson’s IMC presence followed a similar pattern: demonstrations of AI-powered RAN architecture, intelligent operations with agentic AI, and a set of 5G applications spanning drone logistics, oral health screening, agriculture, and assistive technology. These are interesting proofs of concept. They are not, by and large, commercial-scale deployments with established revenue models.

To be fair, IMC is a showcase event, and showcases are supposed to be forward-looking. But the pattern across the industry is consistent: AI-RAN capabilities are being demonstrated faster than they are being monetized. And that is precisely the gap that 1Finity is pointing at.

The counterargument, and why it is weaker than it sounds

The obvious rebuttal is that infrastructure investment always precedes demand. You build the network, and the applications follow. This is the “if you build it, they will come” theory of telecom capex, and it has an uneven track record. It worked for 4G, where the smartphone ecosystem was already forming and the demand curve was visible. It worked less well for 5G, where operators spent heavily on infrastructure and then waited years for enterprise use cases to materialize at scale.

The AI-RAN bet is structurally more like 5G than like 4G. The demand curve for edge AI services is not yet visible. The enterprise buyers are not yet lining up. And the competitive landscape—cloud providers with mature edge offerings, systems integrators with enterprise relationships, hyperscalers with AI expertise—is considerably more crowded than the smartphone market was in 2010.

Building AI-RAN infrastructure on speculation means accepting that risk. Building it where a customer is already paying means transferring that risk to the customer relationship, where it belongs.

There is also a technical argument for 1Finity’s position. GPU-accelerated RAN infrastructure is expensive to deploy, power-hungry to operate, and complex to manage. Deploying it broadly, in the hope that edge workloads will eventually justify the cost, creates a stranded asset risk that operators are increasingly unwilling to accept. Deploying it narrowly, in locations where a specific customer’s workload already justifies the investment, is a much more defensible capital allocation.

What this means for the Open RAN ecosystem

If 1Finity is right—and the capex data suggests it is—then the Open RAN ecosystem needs to rethink its AI-RAN go-to-market strategy. Not the technology. The strategy.

First, the industry should stop selling AI-RAN as a platform play and start selling it as a solutions play. Operators do not need a programmable RAN substrate. They need specific problems solved: congestion at specific sites, latency-sensitive applications for specific customers, operational savings in specific processes. The platform is the means, not the end.

Second, the ecosystem should be honest about the timeline. Deutsche Telekom’s numbers are instructive: operational savings are real and near-term, but AI-related revenue is small and back-loaded. Any AI-RAN business case that depends on near-term edge revenue is built on sand.

Third, the Open RAN community should recognize that its economic argument is now intertwined with its AI argument. If AI-RAN monetization disappoints, the entire disaggregation thesis takes a hit. That is a risk worth managing explicitly, rather than assuming that architectural openness will eventually pay for itself.

The uncomfortable conclusion

The AI-RAN narrative has been driven, for the last several years, by what is technically possible. GPUs in the RAN. AI applications on the RIC. Spectral efficiency gains in field trials. Multi-vendor interoperability at scale. All of it real, all of it impressive, and all of it insufficient on its own.

What 1Finity has done is reintroduce the question that the ecosystem has been avoiding: who pays? Not eventually, not in the 6G era, not once the applications mature. Who pays now, for the infrastructure being deployed today?

The answer, for most operators, is the customer who is already paying for something else. The enterprise with a latency-sensitive workload. The venue with a congestion problem. The industrial site with a connectivity requirement that justifies dedicated compute. Build there. Prove the model. Then expand.

This is not the story that AI-RAN vendors want to tell. It is not the story that fills keynote stages or generates glowing press releases. But it is the story that the capex numbers support, that operator CFOs will recognize, and that the Open RAN ecosystem ignores at its peril.

The AI-RAN hype cycle has been running for three years. The monetization cycle has barely started. 1Finity’s message is that the gap between them is not a temporary condition to be waited out—it is a structural feature of the market that operators should design around.

Build where the customer is already paying. It is a less exciting slogan than “AI-native networks,” but it is a considerably better business plan. And in a market where operators are being asked to fund the next generation of RAN infrastructure on the promise of future revenue, the boring answer may be the only one that survives contact with a boardroom.

Sources

Related Posts

Fujitsu's 1Finity Bets Big on India: 200 Engineers, AI-RAN and the Open RAN Waiting Game

Fujitsu's 1Finity expands its India team to 200+ engineers, targeting AI-RAN, 6G and optical networking as it waits for the Open RAN market to finally take off.

WiSig Opens Commercial Orders for Made-in-India O-RAN Massive MIMO Radio: What It Means for Open RAN Economics

WiSig Networks opens commercial orders for its Made-in-India O-RAN 32TR Massive MIMO ULPI radio, with volume shipments planned for 2027. A decision memo for Open RAN operators and vendors.