Nvidia’s billion-dollar bet on Nokia, Samsung’s parallel gambit with SK Telecom, and the race to make 6G manage itself.
On July 15 2026, Nokia turned two years of AI-RAN promises into an order form. Nine months after Nvidia committed one billion dollars for a stake near three percent of the Finnish vendor, Nokia announced the industry’s first commercial AI-RAN platform: a GPU-accelerated radio built on its anyRAN software and Nvidia’s Aerial system, sold as a subscription rather than a hardware refresh. The platform has already shown spectral efficiency gains above twenty percent (Nokia claims), with fifty percent targeted for 2027 and more than one hundred percent by 2028, enough to double the usable capacity of spectrum operators already own. That number deserves attention: spectrum is the single most expensive, least liquid asset on a carrier’s balance sheet, and a software upgrade that doubles its yield is a different order of claim than the five to ten percent gains each chip generation typically delivers.
The pitch addresses three objections that stalled AI-RAN for years: GPUs draw too much power for a cell site’s thermal budget, cost too much against purpose-built silicon, and force operators to gut infrastructure that already works. Nokia’s fix is a GPU-powered plug-in card that fits its existing AirScale power envelope, alongside a standalone node and a cloud-native option on commodity servers. Yet the headline numbers are targets, not results, built on software still being written. Ericsson, by contrast, has sold a non-GPU AI-in-RAN subscription since June 2026, live across fifteen deployments and delivering roughly ten to twenty percent gains today. Two architectural bets are now running side by side: one wagers GPU compute unlocks a step change worth waiting for, the other that modest, immediate gains on existing silicon beat a speculative multiple.
Samsung and SK Telecom are pursuing a third path. In late 2025, the two companies deepened a long-standing AI-RAN Alliance partnership into a formal agreement covering channel estimation, distributed MIMO, and AI-based scheduling for 6G, with Samsung developing the models and SK Telecom supplying nationwide operational data and live test environments. Where Nokia outsources its computational core to Nvidia, Samsung is building the full stack itself: the memory silicon inside Nvidia’s own accelerators, and now the network intelligence meant to run alongside them. For a country whose industrial policy treats semiconductor and telecom capability as strategic assets, that is not a footnote. It is the point.
It is tempting to fold Samsung’s extraordinary 2026 into this same AI-RAN narrative, but the numbers resist it. Samsung’s market value crossed one trillion dollars in May 2026 and its shares have posted triple-digit gains the same year, driven almost entirely by the memory chip supercycle: HBM4 and HBM4E qualification, DRAM and NAND prices climbing on hyperscaler demand, and quarterly profit that grew nineteenfold in the most recent quarter. What it is actually moving the needle is the reinforcement of the the underlying thesis that Samsung is assembling a horizontally integrated AI-infrastructure business spanning memory, foundry, and network intelligence, rather than riding a single commodity cycle. That distinction matters for anyone reading the stock as a proxy for AI-RAN progress.
The question these three moves raise is not whether networks will become self-managing, but which economics get there first. If Nokia’s GPU bet pays off, the RAN stops being a fixed cost and becomes a programmable compute layer telcos can rent out for edge AI inference, turning idle radio infrastructure into a genuine second revenue line. If it does not, Ericsson’s narrower, silicon-agnostic approach, and Korea’s vertically integrated one, will define the next decade of 6G economics instead. Omdia already prices the cumulative AI-RAN opportunity above two hundred billion dollars by 2030. Whoever gets there will have answered a harder question than whether a network can optimize itself: whether it can do so without another decade of hardware cycles funding the answer.
