AI in 6G mobile network – Benedek Kovács | Craft 2026

June 04, 2026

AI in 6G Mobile Networks – Benedek Kovács (Talk Outline)

Benedek Kovács of Ericsson (150 years old) frames 6G as AI-native networks, split into two aspects: AI for the network (AI embedded as network logic — fraud detection, false-base-station defense, intent-based management) and networks for AI (the network as a regulated, trusted platform for edge AI inference, e.g., offloading humanoid robots). Regulation, privacy, and data sovereignty / confidential computing run through the whole talk.

1. Framing & Ericsson

His first Craft Conference talk was at the very first Budapest Craft (Bálna Conference Center), where the sound failed and he had to dub the whole thing — so this deck is deliberately simpler.
Ericsson started as a telecom company in 1876 (~150 years ago) — present through pushcash telephone exchanges, wire lines, and mobile.
In the 1990s Ericsson was a major mobile-device manufacturer; today it does everything but the device — antennas, data centers, protocols, technology, business systems — everything that makes networks secure, safe, high-performing.

1.1 Eras of communication

Up to the 2000s: human-to-human — telecom meant voice / getting connected.
2000s (3G→4G): data connection. (A famous ~2002 Index article predicted people would never pull out phones in pubs — now we “live with our phones,” enabled by 4G/5G mobile broadband.)
Then human-to-machine (phone → internet/servers/cloud).
Now emerging: human-to-agent — more and more agents deployed in the network: not just fraudulent robo-callers, but helpful/guiding agents and anti-fraud agents analyzing caller/callee (the “suspected spam” label is AI). “What you see is just the tip of the iceberg.”

2. Regulation — The Backbone Constraint

Mobile networks are a regulated business: laws dictate how they’re built, implemented, operated, and what capabilities they must have.
Examples: legal interception (must be able to tap a call given a legal verdict) and GDPR (guarantee privacy, anonymity, location/number protection for everyone not under such a verdict).
Extremely important in the AI era, where a dashcam can record anyone → privacy becomes a serious issue.

3. AI for the Network

3.1 From 5G to 6G standardization (3GPP)

2018 / 5G: AI began to be part of the network — replace traditional algorithms with AI.
6G study items: develop the network based on AI. Futuristic proposals even remove deterministic protocols between phone and base station, substituting agent-to-agent communication (the network could “discuss” what connectivity you need, at what price). He doubts the extreme version happens, but some pieces make sense.
Sensible example: a YouTuber in a crowded area needs uplink bandwidth (different from downlink) — AI lets the network adapt per device.

3.2 Security use cases

False base station at a rally: an attacker mimics an operator; your phone pings all operators and tries to identify itself with the fake one → privacy injured. Networks develop algorithms to detect and cut this out on the spot.
New York incident (~6 months ago): a van full of eSIMs and Raspberry Pis drove around, devices communicating — a physical-device DDoS that “killed the network.” AI detects suspicious device behavior.

3.3 Intent-based management

Instead of calling a raw SMS-gateway API (e.g., “Voltage”), define an intent: “send this SMS only to users verified by the operator” — the network translates what “verified” means.
Another intent: for an event (like Craft), create a private network slice (like a VLAN over mobile) admitting only chosen devices for safe, reliable comms.
Other in-network agents: network positioning (indoor/outdoor) under strict regulation — with consent, e.g., verified location for bank authorization (bankomat withdrawal flagged if your phone is in another country).

4. Networks for AI — Edge Inference

Operators’ business evolution: 2G/3G sold communication → 4G/5G sold data (to the cloud) → 6G wants to sell AI services (”edge AI inference”) — building regional data centers to serve AI use cases.

4.1 Delivery-robot example (US operator)

Delivery robots on the streets of Los Angeles deliver Temu packages/food but have no arms — fine with automatic doors, but US pedestrian crossings require pushing a button.
Solution (a US operator + robot maker): the robot recognizes a nearby human and asks them to push the button. “Mixed performance” — humans aren’t always nice; robots get kicked.

4.2 Why humanoid robots

Question: do we need humanoid robots vs. purpose-built ones? Insight: the world is built for humans (door sizes, buttons, chairs, obstacles) — so humanoid behavior is often desired (sometimes just arms/upper body, sometimes legs like robot dogs).

4.3 The three-way trade-off & offload

Robotic movement metrics in three dimensions: reaction time (how fast it reacts), token rate = responsiveness/real-time, model size = sentience (how smart the decision, e.g., the angle to grab a glass). There’s a trade-off among them.
Today, on-device (no offload): a chart of max hostable model size (y) vs. token output rate (x) — a trade-off curve. Even extrapolating AI-accelerator progress to ~2030, on-device isn’t enough → companies actively pursue offloading.

4.4 Why offload must often be local, not cloud

Latency + jitter: you normally ping a server and get ~30–50 ms round trip, but AI inference (even audio-visual/camera feed) needs all packets present before the GPU’s next cycle — jitter (packets at 100–200 ms) causes delay even on the best networks if bandwidth is short — worse on the internet. So cloud offload is often not feasible; download models locally.
Data sovereignty (the bigger reason): humanoid robots run in factories with sensitive production data that enterprises won’t let leave the premises; a city surveillance system should be processed in-country (e.g., Hungary), not US/China. EU countries regulate that such data must not leave the country; India — not leave its regions. So models run on the regional edge, not the cloud.

5. Evolving the Exposure Layer

Differentiated connectivity is already implemented — network APIs let you ad-hoc request uplink bandwidth (perhaps for a price) for influencer streaming.
Tomorrow: more AI-enabled interfaces — service exposure via Model Context Protocol and agent-to-agent protocols, treating the mobile network as an agent whose services are represented in an “MCP.” (Exact design still in progress.)

5.1 The next-generation network platform

Left side: a regulated, trusted platform for AI workloads providing coverage/uplink, location (verify user location), and sensing.
Sensing — a headline 6G feature: use the network as a radar (detect a drone, a speeding vehicle). Analogy: university papers claim Wi-Fi can sense a heartbeat / heart attack in the next room; he’s seen Wi-Fi detect hand positions through a wall. Mobile networks would work similarly — big innovation area, many startups.
But: if the network can sense, you don’t want to be sensed from outside — everything exposed via network agents that keep GDPR while keeping interfaces open for innovation.

6. Conclusion — AI-Native Networks

Moving from AI added to the network → AI-native networks (6G). Two concepts: AI for the network (logic built in) and networks for AI (network designed for AI).
Traffic types differ — voice, video download, and agent-to-agent each have different characteristics; the network can be optimized per traffic type.
6G as AI-network solutions for secure, reliable, trusted, sovereign AI applications — a very important new area.
Confidential computing case: with a Swedish truck manufacturer, Ericsson verified the entire software stack touched by a packet (device → server → device) so the operator can prove no eavesdropping or malicious content. “No cloud provider today provides this” — they can offer tenant separation and claim sovereignty but can’t prove it with software tools.

7. Q&A

Reticulum protocol / mesh networks? Hasn’t heard of it — interested (offline chat).
What problem does 6G solve that consumers aren’t asking for? Mainly enterprises/providers, not consumers — consumers want a “safe city,” authorities want secure/private computing.
Will users notice 6G? Short answer: invisible. Indirectly, once your phone’s AI agent negotiates with the network you’ll see the benefit.
Do we need 6G or are we underutilizing 5G? Depends on the region. 4G/LTE was designed for the long term — 6G radio is compatible with 4G radio, cores compatible with cores; the distinct new feature is sensing (used by apps, not a direct consumer feature). Many regions face frequency/bandwidth shortages (China, US, Japan, India); Hungary doesn’t.
VPN-like ability to virtually change operator? Yes — VPN-like features exist (operators run hundreds of VPNs); with eSIM consumers can already switch quickly; 6G brings widely available enterprise device VPNs (e.g., for service robots).
Building AI into the network or networks for AI? Both.
Edge AI vs. cloud AI importance? Edge is mainly B2B with specific models and privacy/sovereignty needs; centralized cloud AI runs things like fraud-detection agents.
Most exciting 6G use case? Turning base stations / edge entities into AI contexts — sense a drone, point a camera, gather device info to build an environmental context for smarter app decisions. Sensing will be the “aha, this is 6G not 5G” moment.

People, Companies & References Cited

Benedek Kovács — speaker; Ericsson.
Ericsson — founded 1876; builds everything but the mobile device.
3GPP — standardization (5G AI in 2018; 6G AI-native study items).
Model Context Protocol & agent-to-agent protocols — planned network service exposure.
Examples/cases: US delivery robots (Temu/LA), humanoid-robot manufacturers, a Swedish truck manufacturer (confidential computing), the New York eSIM-van DDoS, false base stations, “Voltage” SMS gateway, the ~2002 Index pub-phone prediction.
Concepts: AI for the network vs. networks for AI, edge AI inference, intent-based management, network slicing, network sensing/radar, data sovereignty, GDPR / legal interception, confidential computing, reaction time / token rate / model size trade-off, jitter vs. inference.

Video: https://www.youtube.com/watch?v=vARsbVTiDo0 — Transcript via yt-transcript.sh; exhaustive outline generated from the full transcript.