The Week the Network Became the AI
I’ve been watching this space long enough to know that the biggest shifts don’t always come with a flashy product launch. Sometimes they come quietly, in a press release from Berlin, with a sentence that sounds almost boring until you realize what it actually means.
This week, that sentence was: “T-Mobile has completed the world’s first AI-RAN call on a live network.”
On the surface, it’s a technical milestone. But what it represents is bigger. For years, AI has lived in data centers, accessed through apps. This week, it moved into the radio network itself—into the cell towers, the spectrum, the actual infrastructure that connects every phone call and every text.
And that changes the economics of telecom in a way that’s hard to overstate.
The First AI Call on a Live Network
On October 21 at GTC Berlin, NVIDIA and Nokia announced that T-Mobile had completed the world’s first AI-RAN call on a commercial network in Bellevue, Washington. At the same time, BT, Deutsche Telekom, and Elisa announced their own AI-RAN pilots on live networks.
Here’s why this matters. AI-RAN means applying AI directly within the radio signal-processing chain—channel estimation, beamforming, multi-user MIMO pairing, scheduling. Instead of running AI as a separate layer on top of the network, it becomes part of how the network processes radio signals in real time.
The early numbers are modest but real: Deutsche Telekom’s lab measurements showed 10–15% uplink throughput improvements, with greater gains under more complex radio conditions. Elisa ran voice, mobile data, and video analytics simultaneously on the same AI-RAN platform over 5G—proving that AI workloads and radio workloads can coexist on shared infrastructure.
“We’re extending accelerated computing into the radio network itself,” Jensen Huang said, “creating the foundation for AI-native 5G”.
The implication is that telecom operators are no longer just selling bandwidth. They’re selling compute at the edge. And the network becomes a distributed intelligence platform, not just a pipe.
MODEL-4: Intelligence as a Network Service
On the same week, Personal AI launched MODEL-4, an operating system designed to let telecom operators deliver personal intelligence to every subscriber—billed alongside talk, text, and data.
The technical specs are striking: 16ms time-to-first-token, 405ms end-to-end voice latency, and $0.02 per million tokens, with operator gross margins above 90% at scale. Tier-1 small language models handle about 95% of traffic inside the operator’s network, routing only complex requests to larger models.
But the strategic idea is what’s interesting. Carriers already have what AI companies spend billions to acquire: direct relationships with nearly every individual and business, a trusted identity for every line, an existing billing relationship, and a distributed network footprint.
MODEL-4 uses the SIM card and phone number as the agent’s identity—no app, no login. And it packages AI as a token plan, the same way operators sell data today.
Deutsche Telekom’s board member Rodrigo Diehl put it plainly: “AI won’t just run on our network. It will be discovered, orchestrated, and executed on the network—in the right place, with the right performance, optimal latency, security, and the right level of sovereignty”.
That’s a different vision of AI than the one most of us are used to. Not a chatbot you open. A service that’s always there, built into the connection itself.
The $1.5 Billion Bet on Open-Source Chips
While the networks grabbed headlines, the silicon behind them told a different story. Tenstorrent, the AI chip startup led by legendary chip architect Jim Keller, raised $1.5 billion in a Series E round, more than doubling its valuation to $6.2 billion.
The company develops AI chips and software using RISC-V, the open-source chip architecture that eliminates license fees and reduces dependence on proprietary ecosystems like Arm. The funding will scale production of its Galaxy servers, which combine dozens of its chips to power chatbots, coding assistants, and video generators.
The timing matters. As AI inference moves to the edge—into networks, factories, and devices—the demand for open, customizable silicon grows. Tenstorrent’s bet is that the next wave of AI won’t be dominated by a single vendor’s architecture.
MiniMax M3.1 Flash: The Efficiency Breakthrough
On the model side, MiniMax launched M3.1 Flash, a model that’s nearly three times faster and 70% cheaper than its predecessor.
The benchmarks are competitive: 73.7% on DeepSWE v1.1 for software engineering, 61.7 on OSWorld V2 for computer use, and 82% on CyberGym for vulnerability discovery. But the real story is the cost. Input pricing is 1 yuan per million tokens, output is 4 yuan, and cached reads are 0.1 yuan.
For businesses building agentic workflows—where a single task might involve dozens of model calls—that cost difference compounds fast. The model is already available on MiniMax’s API and will open-source its weights later.
Efficiency is becoming the new frontier. Not just raw capability.
Claude Duet: An Agent for the Rest of Us
And on October 22, Anthropic launched Claude Duet, a standalone AI agent application designed to proactively support the work of professionals and small business owners.
The framing is deliberate. “Proactively support” means the agent doesn’t wait for a prompt—it anticipates, suggests, and acts. And “small business owners” means Anthropic is targeting the millions of businesses that don’t have IT departments or AI teams.
This is the other side of the infrastructure story. While the networks and chips get built, the tools are getting simpler. The goal is to make agentic AI something a solo founder can use, not just a Fortune 500.
The Pattern Worth Watching
Three stories from one week. The first AI call on a live network. Intelligence packaged as a network service. A $1.5 billion bet on open-source silicon. A model that cuts costs by 70%. An agent built for small businesses.
What ties them together?
AI is moving from the app layer to the infrastructure layer.
Not just running on networks. Becoming part of them. Not just consuming compute. Generating it at the edge. Not just for enterprises with dedicated teams. For anyone with a connection.
The companies winning aren’t the ones with the flashiest demos. They’re the ones embedding AI where the work already happens—in the signal chain, in the SIM card, in the chip architecture, in the tools people already use.
This was the week AI stopped being an app and started being the network itself.
#AI, #Telecom, #AIRAN, #TMobile, #Nokia, #NVIDIA, #PersonalAI, #Tenstorrent, #MiniMax, #ClaudeDuet, #AgenticAI, #AIIinfrastructure, #NetworkAI, #DigitalTransformation, #Innovation, #TechNews