Will GPUs Eventually Make Their Way to the Top of the Cell Phone Tower

By Brian Newman, Contributing Editor

Insight at the Intersection

Ericsson lost roughly 6 percent of its market value on June 9 over a chip NVIDIA has not confirmed building. The trigger was a Light Reading report, from correspondent Iain Morris, describing a dedicated, sub-100W chip NVIDIA is reportedly developing for the 6G radio unit, the equipment mounted at the top of the cell tower rather than in the baseband cabinet at the bottom.

A correction is worth stating plainly before going further. NVIDIA later told Light Reading it had not confirmed developing a GPU specifically for the 6G radio unit. The outlet revised its own language from GPUs "would" become essential in advanced radios to GPUs "could" become essential, at NVIDIA's request. The stock move was real. The underlying claim is still sourced to unnamed people close to the matter, not to an NVIDIA product announcement.

The obvious read is an engineering story which is general-purpose compute finally muscling into a domain long ring-fenced by custom silicon, because the radio unit runs deterministic, real-time processing in a harsh, passively cooled environment at the top of the mast. That framing undersells what is actually happening. This is an economics story wearing an engineering costume.

A Market Too Small to Fund What It Needs

Four data points make the case, in order.

First, the market. Operators worldwide spent about $35 billion on RAN products last year, down from $45 billion in 2022, according to Omdia figures reported by Light Reading. A shrinking market is a hard place to fund bespoke silicon.

Second, the supplier math. Marvell earned just $542 million in revenue from all telecom customers combined in the twelve months to last November. Designing custom silicon for a single RAN vendor no longer clears the bar against a market that size, regardless of how good the chip is.

Third, the workload. NVIDIA's own justification is quantified rather than aspirational. Basic radios use four transmitters and receivers. 5G-Advanced and 6G could push that to 128, which NVIDIA says demands 32 times the low-PHY compute. Ultra-MIMO in 6G's higher spectrum bands puts radios with up to 1,024 transmitters and receivers on the table. That trajectory outruns what a single-purpose ASIC, designed years in advance for a fixed antenna count, can absorb.

Fourth, the vendor split. Ericsson is defending its internally developed custom silicon and rejecting the argument that the investment is unsustainable, and its investors punished the stock anyway. Nokia already took a $1 billion NVIDIA investment, introduced the ARC-Pro accelerated computing platform as a software-upgradable path from 5G-Advanced to 6G, and is now weighing GPUs for future radios rather than ruling them out. Marvell is repositioning itself from a single-customer silicon supplier into a baseband chip vendor for the industry at large, outside Huawei and ZTE.

The Discipline Every RAN Vendor Now Faces

Each vendor is making a distinct bet, not reacting to the same pressure the same way. Ericsson is betting that differentiated silicon still earns a premium large enough to justify owning it outright. Nokia is hedging, keeping custom silicon in play while opening the door to GPUs and to O-RAN Alliance open fronthaul interfaces that let the distributed unit and radio unit run on different silicon rather than the same platform by default. Marvell is betting that a shared, industry-wide chip beats a shrinking exclusive contract. None of these are technology bets first. They are capital-allocation bets, made by leadership teams reading the same $35 billion market and drawing three different conclusions about how to survive its size.

Is NVIDIA disrupting a healthy market, or filling a vacuum the market's own economics already created? And when the vendor everyone once dismissed as too small to matter starts pitching itself as the chip that serves the whole industry, has the RAN already stopped being a hardware business?

Strategic Signal

This connects directly to the AI Gateway story covered in the prior edition. AT&T's bet is that the model layer is commoditizing and that the durable advantage sits in the routing architecture built around it. NVIDIA's radio unit push extends the identical logic one layer down the stack. The silicon layer, historically protected by proprietary hardware and long design cycles, is now facing the same pressure that has already reached the model layer. Two separate fronts in AI and telecom, the data center and the radio, are converging on the same argument: owning purpose-built infrastructure is getting harder to justify everywhere at once, not because the infrastructure stopped mattering, but because the economics that used to protect it stopped cooperating.

Watch third and fourth quarter earnings from Ericsson, Nokia, and Marvell for how each frames capital spending on this bet, and watch for whether NVIDIA moves from an unconfirmed report to an actual product announcement. That confirmation has not happened yet, and until it does, this remains a signal worth tracking rather than a fact worth building a strategy around.

Brian C. Newman is a telecom and AI strategy consultant, course creator, and former Verizon technology leader with more than 30 years of experience across wireless networks, 5G, network operations, infrastructure modernization, and emerging technologies. He helps organizations understand how AI, connectivity, edge computing, and digital infrastructure are reshaping business operations, real estate, public safety, and customer experience.