Skip to content
5G/6G Academy
AI & RAN

AI-RAN & the AI-Native Network

AI inside the radio itself — the 3GPP hooks that went normative, RIC-hosted apps running on live networks, the shared-compute argument, and honest energy numbers with their denominators.

Watch the free lessonJoin for free
31 lessons≈3.6h of video6 modulesEarns TELCOMA AI-RAN Specialist

What you'll learn

  • Open with the radio's real tension, re-teach the air-interface and RAN-compute frames a Foundation-only learner needs, and lock the three-strand discipline.
  • Exactly the ML mechanics the rest of the course spends — fitted functions, time-series prediction, one-sided vs two-sided models, life-cycle management and fallback discipline — all through radio data.
  • The 3GPP strand with exact provenance — the frozen one-sided framework and its three use cases, the life-cycle signalling, and the two-sided story told at its true status.
  • Deepen the two-clock picture into working knowledge — loop budgets, the named wires, where models are trained and shipped, and a scorecard moved honestly forward.
  • Cash the shared-compute promise honestly — the three prepositions, what running the physical layer on accelerated compute takes, what has genuinely run versus been announced, the open silicon debate, and the economics.
  • The network-side 3GPP track that went normative a release earlier, sleep mechanics and honest per-denominator savings, the three operating models that run production RAN AI, and a truthful cash-out of "AI-native".

Module 1 — Why AI in the RAN

4 lessons · 1 lab · ≈24 min

Open with the radio's real tension, re-teach the air-interface and RAN-compute frames a Foundation-only learner needs, and lock the three-strand discipline.

  1. 1
    Where AI meets the radio — the network's hardest caseWatch free
    5:29
  2. 2
    The air interface in four frames: beams, channel, scheduling, positionFree with an account
    6:21
  3. 3
    RU, DU, CU: where the RAN computesFree with an account
    7:15
  4. 4
    AI-RAN, defined — three strands, one disciplineFree with an account
    7:18
  5. Name the StrandCheckpoint

    Sort six real claims into hooks in the standard, applications on a controller, and shared computers — then demand the evidence each strand can actually produce.

    ~5 min

Module 2 — ML mechanics for RAN engineers

5 lessons · 1 lab · ≈33 min

Exactly the ML mechanics the rest of the course spends — fitted functions, time-series prediction, one-sided vs two-sided models, life-cycle management and fallback discipline — all through radio data.

  1. 1
    Training vs inference: a model is a fitted functionRequires subscription
    5:59
  2. 2
    The channel is a time-seriesRequires subscription
    6:07
  3. 3
    One-sided or two-sided: where the model standsRequires subscription
    6:43
  4. 4
    Life-cycle management: a model in serviceRequires subscription
    6:44
  5. 5
    Fallback: the floor the network never losesRequires subscription
    6:30
  6. Will the Decision Still Be True?Interactive lab

    Doppler decides how long the channel stays itself. Set speed and band, then choose where the model runs — and find out whether its answer survives the trip.

    ~5 min

Module 3 — AI on the air interface — the 3GPP strand

6 lessons · 1 lab · ≈43 min

The 3GPP strand with exact provenance — the frozen one-sided framework and its three use cases, the life-cycle signalling, and the two-sided story told at its true status.

  1. 1
    Rel-19: one-sided models go normativeRequires subscription
    7:23
  2. 2
    Beam prediction: measure a few, know them allRequires subscription
    6:52
  3. 3
    CSI prediction & AI positioning: the other two hooksRequires subscription
    7:15
  4. 4
    Life-cycle signalling: activate, monitor, fall backRequires subscription
    6:56
  5. 5
    Two-sided CSI compression: why it had to waitRequires subscription
    7:34
  6. 6
    The two-sided exchange: encoder, air gap, decoderRequires subscription
    5:58
  7. Predict the BeamInteractive lab

    Cut a 32-beam sweep down and watch a predictor guess the rest. Find where the saving stops being free — the answer is a number, not an opinion.

    ~6 min

Module 4 — RIC-hosted intelligence — the O-RAN strand

5 lessons · 1 lab · ≈33 min

Deepen the two-clock picture into working knowledge — loop budgets, the named wires, where models are trained and shipped, and a scorecard moved honestly forward.

  1. 1
    Near-RT and non-RT: two clocks, now with numbersRequires subscription
    5:06
  2. 2
    E2 and A1: the wires get namesRequires subscription
    5:38
  3. 3
    The O-RAN AI/ML loop: where the models liveRequires subscription
    6:16
  4. 4
    rApps in production: the scorecard movesRequires subscription
    6:01
  5. 5
    An energy rApp, end to end: one full circuitRequires subscription
    6:05
  6. Route It on the Right WireCheckpoint

    Match six exchanges to the wires that carry them, and find the whole circuit that rides the oldest, dullest path in the building.

    ~5 min

Module 5 — AI-RAN infrastructure — the shared-compute strand

6 lessons · 1 lab · ≈43 min

Cash the shared-compute promise honestly — the three prepositions, what running the physical layer on accelerated compute takes, what has genuinely run versus been announced, the open silicon debate, and the economics.

  1. 1
    AI-for-RAN, AI-on-RAN, AI-and-RAN: three prepositions, kept honestRequires subscription
    7:08
  2. 2
    L1 on GPUs: the physical layer on a general-purpose machineRequires subscription
    7:48
  3. 3
    GPU AI-RAN today: what has actually runRequires subscription
    8:00
  4. 4
    AI without GPUs: the purpose-built counter-argumentRequires subscription
    8:30
  5. 5
    AI-and-RAN economics: does sharing the machine pay?Requires subscription
    8:32
  6. 6
    Research vs shipping: the AI-RAN scorecardRequires subscription
    8:07
  7. Share the MachineInteractive lab

    One accelerator runs a site’s physical layer. Sell the rest as AI inference over a real 24-hour load curve — without ever taking a slot from the radio.

    ~6 min

Module 6 — Energy savings & operating models in production

5 lessons · 1 lab · ≈33 min

The network-side 3GPP track that went normative a release earlier, sleep mechanics and honest per-denominator savings, the three operating models that run production RAN AI, and a truthful cash-out of "AI-native".

  1. 1
    AI/ML for NG-RAN: the oldest AI in the RANRequires subscription
    7:39
  2. 2
    Cell sleep, by levels: DTX, DRX and deeperRequires subscription
    7:33
  3. 3
    Energy savings: the numbers, with their denominatorsRequires subscription
    7:08
  4. 4
    Embedded, rApp, or GPU: who runs the loopRequires subscription
    8:29
  5. 5
    AI-native, honestlyRequires subscription
    7:49
  6. Read the Energy ClaimCheckpoint

    Five real percentages, five different denominators. Work out what each number is actually a percentage of — and catch the two the industry tells backwards.

    ~5 min

New to the terminology? Look up any acronym in the telecom glossary.

Take the certification exam

Earns a certificate

42 questions · 60 min · 65% to pass. Score 65%+ to earn your TELCOMA Certified AI-RAN Specialist — a QR-verifiable certificate.

Not ready yet? Take a free 20-question mock exam first.

7-day money-back guarantee

Unlock every lesson in AI-RAN & the AI-Native Network

Stream all 31 lessons, follow the 6-module path, and earn the TELCOMA AI-RAN Specialist certificate.

  • Every module unlocked
  • Labs & full-length practice exams
  • Verifiable certificate
  • TELCOMA since 2009
See Pro plans Or watch a free lesson — no signup