AT&T Deploys AI Cloud RAN to Delay Multi-Billion Spectrum Upgrades
Ai4
AT&T is transitioning to open, virtualized baseband infrastructure and specialized network foundation models to boost spectral efficiency and defer multibillion-dollar capacity expansions under strict operational guardrails.
## Background
Published by the Ai4 channel on YouTube on August 29, 2025, this fireside chat features Adam Loddeke (AVP of RAN Technology at AT&T) and David Valdez (Vice President at CTIA). According to the session description, the discussion explores how AT&T is modernizing its Radio Access Network (RAN) infrastructure to support AI-native capabilities, why current market and technological conditions favor this transition, and how promising use cases like anomaly detection and link adaptation are operationalized.
## TL;DR
Adam Loddeke details AT&T's architectural evolution from closed, proprietary radio hardware toward virtualized Cloud RAN and centralized Service Management and Orchestration (SMO). He explains how AI applications are structured across Day 0 network planning, Day 1 site deployment, and Day 2 live operations to maximize spectral efficiency and reduce capital expenditures. Loddeke also emphasizes the critical need for deterministic guardrails in safe automation, highlights ecosystem partnerships with developers like Aira Technologies for specialized rApps, and outlines how specialized network foundation models are boosting energy efficiency across existing tabular telemetry datasets.
## Atmosphere & style
The session has the polished, pragmatic, and measured tone of an executive telecom fireside chat. David Valdez acts as an engaging facilitator, asking precise architectural and strategic questions, while Adam Loddeke provides authoritative, deeply grounded engineering explanations. The dialogue feels refreshingly realistic rather than overhyped, consistently balancing visionary 6G concepts and advanced foundation models with the strict, unyielding operational realities of carrier-grade five-nines reliability.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=dh4lOZp400k&t=4s) AT&T is transitioning its Radio Access Network from closed proprietary baseband hardware to open virtualized architectures running on commercial off-the-shelf (COTS) servers.
- [02:31](https://www.youtube.com/watch?v=dh4lOZp400k&t=151s) AI integration across the network lifecycle is organized into three operational horizons: Day 0 (planning/modeling), Day 1 (automated commissioning), and Day 2 (runtime optimization).
- [02:31](https://www.youtube.com/watch?v=dh4lOZp400k&t=151s) Maximizing spectral efficiency with AI delivers major business value by deferring multi-billion-dollar capital expenditures on spectrum acquisition and capacity upgrades.
- [06:51](https://www.youtube.com/watch?v=dh4lOZp400k&t=411s) Carrier-grade reliability requires a 'safe automation' framework with deterministic guardrails to resolve conflicting optimization loops, such as simultaneous antenna up-tilt and down-tilt commands.
- [06:51](https://www.youtube.com/watch?v=dh4lOZp400k&t=411s) AI should be treated as one tool among many, prioritizing simpler heuristic or rule-based models when they are easier to maintain and troubleshoot than complex neural networks.
- [09:44](https://www.youtube.com/watch?v=dh4lOZp400k&t=584s) Commercial 6G deployment is expected around 2028 to 2032, with Cloud RAN and centralized Service Management and Orchestration (SMO) serving as the key architectural stepping stones.
- [09:44](https://www.youtube.com/watch?v=dh4lOZp400k&t=584s) AI workloads are divided between centralized SMO data centers handling non-real-time 1-second policy loops on CPUs and edge Cloud RAN nodes executing sub-second real-time inference on GPUs.
- [13:31](https://www.youtube.com/watch?v=dh4lOZp400k&t=811s) Open RAN and standardized SMO interfaces enable a multi-vendor ecosystem, allowing operators to integrate third-party rApps—such as Aira Technologies' anomaly detection and generative code tools.
- [16:26](https://www.youtube.com/watch?v=dh4lOZp400k&t=986s) Developing specialized network foundation models tailored to tabular telemetry allows operators to unlock incremental energy efficiency gains on top of existing deep learning rApps.
- [16:26](https://www.youtube.com/watch?v=dh4lOZp400k&t=986s) Network AI deployment is an iterative continuum rather than a one-time milestone, continuously evolving as algorithms and underlying hardware accelerators mature.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=dh4lOZp400k&t=4s) Welcome and Introductions: Demystifying the RAN Architecture
David Valdez of CTIA opens the session by welcoming the audience and introducing the focus on wireless industry evolution before inviting Adam Loddeke, Assistant Vice President of RAN Technology at AT&T, to introduce his role. Loddeke explains that his responsibilities center on emerging Radio Access Network (RAN) architecture and network management, focusing primarily on emerging technologies and use cases positioned on a two-plus-year forward horizon. His team balances forward-looking research with the engineering necessary to drive practical realization of advanced architectures.
To establish a clear technical foundation, Valdez asks Loddeke to level-set on the core mechanics of a RAN. Loddeke explains that the RAN represents the critical over-the-air link connecting commercial user equipment (UE)—such as consumer smartphones—to the broader telecommunications network across operator-managed radio spectrum. A nationwide deployment involves thousands of cell sites, each featuring tower-top antennas and radio heads connected via high-speed fiber drops to base-station compute shelters at the ground level. While these base station sites historically ran on closed, proprietary hardware platforms, AT&T is actively transitioning toward open architectures, utilizing commercial off-the-shelf (COTS) servers running cloud infrastructure layers hosting virtualized RAN applications, bridging the link from cell towers back into transport and core networks.
```mermaid
flowchart LR
accTitle: Radio Access Network Overview
subgraph UE_Domain["User Domain"]
UE["User Equipment (UE)<br/>Smartphones & Devices"]
end
subgraph Core_Domain["Network Backbone"]
Transport["Transport Network"]
Core["Core Network"]
Transport --> Core
end
UE <-->|"Over-the-Air Spectrum"| Antenna
BaseCompute --> Transport
```
[02:31](https://www.youtube.com/watch?v=dh4lOZp400k&t=151s) Evolution Toward AI-Native RAN and the Three Operational Horizons
Addressing the shift toward AI-native network architecture, Loddeke describes AI-RAN as an ongoing continuum of machine learning integration rather than a sudden pivot. Early network automation relied primarily on static decision trees and deterministic rule-based systems, which gradually transitioned toward advanced deep learning algorithms. AT&T harnesses AI to optimize radio performance, streamline field operations, and drive closed-loop automation. To systematically prioritize use cases across the network lifecycle, AT&T organizes AI deployment into three distinct operational domains:
- **Day 0 (Network Planning Horizon):** Leveraging traffic forecasting models and generative AI tools to assist engineers and project managers in designing, modeling, and planning cell site additions.
- **Day 1 (Asset Deployment):** Automating site commissioning and the baseline parameter configuration of newly integrated cells and radio assets to streamline rollout velocity.
- **Day 2 (Live Network Operations):** Driving runtime optimization, which accounts for the vast majority of AI use cases—including dynamic energy efficiency controls, intelligent traffic steering, and automated outage compensation.
Loddeke notes that operational use cases are further evaluated against four business value drivers: direct operating cost reduction, capital cost deferral/avoidance, customer churn mitigation through experience enhancement, and net-new revenue generation. He places particular emphasis on capital deferral through spectral efficiency gains. Because spectrum acquisition and per-site carrier deployment represent the single largest capital expenditures in the RAN budget, maximizing spectral throughput with AI algorithms delays the necessity for multi-billion-dollar capacity upgrades and spectrum builds.
[06:51](https://www.youtube.com/watch?v=dh4lOZp400k&t=411s) Safe Automation and Reliability in Critical Telecom Infrastructure
Valdez turns to the operational reality of telecommunications as critical national infrastructure, where upgrades and automated optimizations must execute seamlessly without disrupting service. Loddeke emphasizes that maintaining carrier-grade "five nines" (99.999%) reliability requires a strict engineering framework centered on "safe automation." Network AI models must operate within deterministic boundaries and guardrails to prevent conflicting decision loops between independent optimization systems.
Loddeke highlights a classic scenario: an interference mitigation application might decide to mechanically or electrically down-tilt an antenna sector to relieve channel blocking, while a simultaneous coverage-maximization algorithm commands an up-tilt on the same sector. Without unified orchestration and governance, the network gets trapped in an endless ping-pong loop that degrades performance or causes regional outages. Reflecting on broader lessons for enterprise AI adoption, Loddeke advises that AI is merely a tool in a larger toolkit—not every operational problem requires a complex neural network. Simpler models and standard rules are often far easier for downstream maintenance teams to troubleshoot and maintain under operational stress.
[09:44](https://www.youtube.com/watch?v=dh4lOZp400k&t=584s) Architectural Roadmap to 6G: Cloud RAN and Centralized SMO Orchestration
When comparing wireless generations, Loddeke and Valdez clarify that while commercial 5G has matured across nationwide footprints over the last five to six years, 6G is still in pre-standardization and is unlikely to appear commercially until roughly 2030 (with a realistic window spanning 2028 to 2032). Like previous generational shifts, 6G will deliver wider channel bandwidths, higher peak data rates, and access to new spectrum allocations. To lay the architectural foundation for this future, Loddeke is steering two major modernization pillars across AT&T's network:
- **Cloud RAN (Open & Virtualized RAN):** Decoupling baseband software from specialized proprietary silicon and scaling containerized radio functions across open COTS computing hardware.
- **Service Management and Orchestration (SMO):** Implementing a centralized management framework incorporating the Non-Real-Time RAN Intelligent Controller (Non-RT RIC), an orchestration engine, and enterprise data management.
Loddeke breaks down how AI inference workloads map across this physical and logical network hierarchy:
| Deployment Tier | Architectural Component | Operational Latency | AI Workload Profile & Compute Targets |
|:--- |:--- |:--- |:--- |
| **Regional Core / Data Centers** | Service Management and Orchestration (SMO) / Non-RT RIC | ~1.0 second non-real-time closed loops (deployed across 8 to 12 US sites) | Global policy optimization, non-real-time rApps, offline model training, and lightweight ML models executing on standard CPUs. |
| **Far Edge / Cell Site Shelters** | Cloud RAN Distributed Unit (DU) / Centralized Unit (CU) Nodes | Sub-second real-time closed loops | Real-time link adaptation, sub-millisecond beamforming, localized continuous online learning, and heavy deep learning models utilizing specialized GPUs. |
[13:31](https://www.youtube.com/watch?v=dh4lOZp400k&t=811s) Opening the Telecom Ecosystem: Third-Party rApps and Aira Technologies
Loddeke reflects on telecom's historical reliance on closed, single-vendor proprietary stacks, highlighting AT&T's deliberate push toward a modular, multi-vendor ecosystem. Historically, AT&T's network automation was dominated by internally built solutions—such as its proprietary "Midas" optimization platform, which blends rule-based logic and deep learning models—paired with closed third-party vendor platforms.
With the rollout of the Open RAN SMO architecture, AT&T is actively trialing specialized third-party Non-Real-Time applications (rApps). As a concrete example, Loddeke points to AT&T's partnership with Silicon Valley-based Aira Technologies. AT&T is collaborating with Aira on two primary initiatives:
- **Anomaly Detection rApps:** Developing intelligent monitoring applications running on the Non-RT RIC to detect anomalous cell site degradation.
- **Generative Code Generation for rApps:** Utilizing generative AI to streamline and accelerate the automated generation of rApp software logic.
Loddeke concludes that standardizing SMO and rApp hosting interfaces provides the commercial framework necessary to test, trial, and operationalize external innovations rapidly, allowing AT&T to continuously incorporate best-in-class AI capabilities across its nationwide network.
[16:26](https://www.youtube.com/watch?v=dh4lOZp400k&t=986s) Elevating RAN Energy Efficiency with Network Foundation Models
Adam Loddeke reflects on the expanding scope of AI across telecom infrastructure, noting that while AI is not a universal fix for every network scenario, the potential opportunities far exceed what operators have harnessed so far. To illustrate this, he points to AT&T's ongoing work around network energy efficiency. The operator had previously deployed a deep learning application—operationalized as an rApp—that successfully delivered a "significant amount of savings in the network." Rather than treating that deployment as a finished milestone, AT&T evaluated how newer modeling paradigms could push performance even further.
Because the vast majority of operational radio access network data consists of tabular, numbers-based metric points, AT&T developed a specialized network foundation model tailored to ingest this data structure. By pairing this foundation model with their preexisting machine learning algorithms, the team unlocked even greater efficiency gains from the existing application. Loddeke emphasizes a forward-looking engineering mindset: "just because it's been done doesn't mean you can't do it better going forward." Looking ahead, he stresses that operators must remain flexible and eager to experiment, especially as machine learning capabilities rapidly advance through both sophisticated algorithmic architectures and underlying hardware acceleration.
```mermaid
flowchart TD
accTitle: Evolution of RAN Energy Efficiency Optimization
subgraph Data["Network Telemetry"]
A["Tabular & Numbers-Based Time-Series Data"]
end
subgraph Phase1["Initial Deployment"]
B["Deep Learning rApp"]
C["Significant Baseline Energy Savings"]
end
subgraph Phase2["Foundation Model Enhancement"]
D["Network Foundation Model"]
E["Integrated ML + Foundation Model Stack"]
F["Enhanced Energy Efficiency Improvements"]
end
A --> B
B --> C
A --> D
B --> E
D --> E
E --> F
```
- **Data Characteristics**: The telemetry powering these network models is overwhelmingly tabular, numbers-based time-series and point data.
- **Continuous Optimization**: Re-architecting existing deep learning rApps with foundational network models demonstrates that deployed AI solutions can be iteratively improved rather than left static.
- **Future Catalysts**: Ongoing evolution will be propelled both by new AI model designs and advances in underlying hardware architecture.
[17:48](https://www.youtube.com/watch?v=dh4lOZp400k&t=1068s) Session Wrap-Up and Closing Remarks
David Valdez brings the fireside chat to a close, thanking Adam Loddeke for sharing AT&T's architectural vision, operational experiences, and insights into the future of AI-driven radio access networks. He invites the audience to join him in applauding Loddeke, concluding the session on an appreciative and engaging note.
## Notable quotes
> [16:26](https://www.youtube.com/watch?v=dh4lOZp400k&t=986s) just because it's been done doesn't mean you can't do it better going forward — Adam Loddeke
## Mentioned
Adam Loddeke, David Valdez, AT&T, CTIA, Ai4, Aira Technologies, Midas, Radio Access Network (RAN), Cloud RAN, Open RAN, Non-RT RIC, SMO (Service Management and Orchestration), rApps, 5G, 6G
## Video description
Adam Loddeke, AVP - RAN Technology, AT&T
David Valdez, Vice President, Privacy & Cybersecurity Policy, CTIA
Abstract
How is AT&T evolving its RAN (Radio Access Network) infrastructure to support AI-native capabilities, and why is now the right time to invest in this transformation?
What are the most promising AI use cases in RAN today—from anomaly detection to link adaptation—and how are they being operationalized?
AT&T部署AI Cloud RAN以推迟数十亿美元频谱升级
AT&T正在转向开放式虚拟化基带基础设施和专用网络基础模型，以在严格的运营护栏下提升频谱效率并推迟数十亿美元的扩容支出。
Adam Loddeke详细介绍了AT&T从封闭专有的无线电硬件向虚拟化Cloud RAN及集中式服务管理与编排（SMO）的架构演进。他解释了如何将AI应用贯穿布局于Day 0网络规划、Day 1基站部署和Day 2实时运维中，以最大化频谱效率并降低资本支出。Loddeke还强调了安全自动化中确定性护栏的关键必要性，重点提及了与Aira Technologies等开发者在专用rApps方面的生态合作，并概述了专用网络基础模型如何依托现有的表格遥测数据集提升能效。
AT&T 正在将其无线接入网（RAN）从专有硬件过渡至基于通用商用服务器（COTS）的开放虚拟化架构，并将 AI 部署划分为 Day 0（规划建模）、Day 1（自动调试）和 Day 2（运行期优化）三个周期。 (2:31)
通过 AI 提高频谱效率有助于延缓价值数十亿美元的频谱购置与扩容资本支出；同时，保障电信级可靠性必须依赖具备确定性护栏的“安全自动化”框架，以解决不同优化算法间的冲突。 (6:51)
面向预计在 2028 至 2032 年商用的 6G 网络，AT&T 以 Cloud RAN 和集中式服务管理与编排（SMO）为基础，将 AI 负载分流至集中式数据中心（1 秒级非实时策略控制）与边缘节点（毫秒级 GPU 实时推理）。 (9:44)
依托 Open RAN 和标准化 SMO 接口，AT&T 积极引入第三方生态，正与 Aira Technologies 合作开展异常检测 rApp 以及生成式代码构建工具的试验。 (13:31)
针对以数值表格为主的 RAN 遥测数据，AT&T 开发了专用网络基础模型，将其与现有能效 rApp 相结合，进一步提升了基站节能成效。 (16:26)
Reading summary
AT&T is transitioning its Radio Access Network (RAN) toward open, cloud-based, and AI-native architectures to lower operational overhead and defer costly physical infrastructure investments. According to Adam Loddeke, Assistant Vice President of RAN Technology at AT&T, realizing these efficiencies requires deterministic guardrails to prevent conflicting automations, a tiered compute distribution, and an opening of the software ecosystem to modular third-party applications.
AI Lifecycle Domains and Capital Deferral (2:31)
Loddeke frames the transition to AI-RAN as an ongoing continuum rather than a sudden shift, organizing deployments across three operational horizons: Day 0 for traffic forecasting and site planning, Day 1 for automated commissioning and baseline parameter configuration, and Day 2 for live network runtime operations, which accounts for the vast majority of use cases.
Operational use cases are prioritized across four business drivers: operating cost reduction, capital cost deferral, customer churn mitigation, and new revenue. Loddeke emphasizes that spectral efficiency gains provide the highest capital impact by delaying multi-billion-dollar spectrum acquisitions and carrier deployments, which represent the single largest expenditure category in the RAN budget.
Deterministic Guardrails for Safe Carrier Automation (6:51)
To preserve carrier-grade "five nines" (99.999%) reliability across critical infrastructure, AT&T restricts network AI to deterministic boundaries under a "safe automation" framework. Without centralized governance, uncoordinated optimization applications can trigger contradictory commands—such as an interference-mitigation tool down-tilting an antenna sector while a coverage application simultaneously commands an up-tilt—causing unstable ping-pong loops that degrade regional performance.
To resolve these conflicts, operational telemetry passes through policy engines that arbitrate commands before cell-site execution occurs. Loddeke also cautions against overusing complex neural networks, noting that deterministic rules and simpler models are frequently preferable because field maintenance teams can troubleshoot them more effectively under operational stress.
Hierarchical Compute from Cloud RAN to 6G (9:44)
AT&T's architectural modernization relies on Cloud RAN—which decouples baseband software from proprietary silicon onto commercial off-the-shelf servers—and centralized Service Management and Orchestration (SMO). While Loddeke forecasts that commercial 6G will not arrive until roughly 2030 (with an anticipated window of 2028 to 2032), modernizing this compute hierarchy is necessary to accommodate split-latency AI workloads.
Workloads are divided into two physical tiers. At regional core data centers spanning 8 to 12 US sites, the Non-Real-Time RAN Intelligent Controller (Non-RT RIC) executes global policy optimization, offline model training, and non-real-time rApps on standard CPUs with one-second control loops. At the far-edge cell site shelters, Distributed Unit (DU) and Centralized Unit (CU) nodes leverage specialized GPUs to handle sub-second, real-time closed loops such as beamforming and link adaptation.
Modular rApps and Network Foundation Models (13:31)
AT&T is moving away from historically closed, single-vendor systems and purely internal software—such as its proprietary Midas platform—toward an open ecosystem of modular Non-Real-Time applications (rApps). Through standardized SMO interfaces, the operator is collaborating with Aira Technologies to trial rApps for cell degradation anomaly detection as well as generative AI tools designed to automate rApp code creation.
AT&T is also pairing established machine learning applications with domain-specific network foundation models tailored to tabular, numerical time-series metrics. By layering a foundation model onto an existing deep learning rApp that had already achieved substantial energy savings, the operator unlocked further reductions, demonstrating that deployed network AI models can be continuously improved through new algorithmic architectures.
AT&T's operational roadmap indicates that achieving carrier-scale AI requires balancing cutting-edge modeling with deterministic rule engines and multi-tiered compute. However, because commercial 6G remains years from standardization, the near-term success of this strategy depends entirely on scaling Cloud RAN and successfully integrating third-party rApps without compromising five-nines network reliability.
## Background
本内容由频道 Ai4 于 2025 年 8 月 29 日在 YouTube 平台发布。视频记录了 AT&T 无线接入网技术助理副总裁 Adam Loddeke 与 CTIA 隐私及网络安全政策副总裁 David Valdez 的炉边对话，重点探讨了 AT&T 如何推进其 RAN（无线接入网）基础设施向 AI 原生演进、当前的投资时机，以及从异常检测到链路自适应等关键 AI 用例的落地实践。
## TL;DR
对话中，Adam Loddeke 拆解了从传统基站到基于 COTS 通用服务器与云化架构的 RAN 演进历程，并阐述了 AI 在网络规划（Day 0）、部署（Day 1）及运维（Day 2）三大生命周期中的商业价值与频谱效率提升。他强调电信关键基础设施必须坚持“安全自动化”以防止多智能体冲突死循环，并指出 AI 应作为务实工具而非万能药。此外，Adam 深入介绍了 AT&T 提前布局 6G 的两大核心工程——Cloud RAN 与包含 Non-RT RIC 的集中式 SMO 架构，并分享了与 Aira Technologies 的生态合作以及利用网络基础模型（Network Foundation Model）持续升级基站能效优化的实战经验。
## Atmosphere & style
整场炉边对话展现出专业、务实且极具前瞻性的技术研讨氛围。主持人提问精准聚焦产业核心痛点，而 Adam 的回答则深入浅出，既能用生动的物理机房图景与天线上倾/下倾冲突案例化解复杂通信概念，又能站在运营商顶层架构的高度给出冷静克制的判断。两位专家的互动节奏明快、默契融洽，观众不仅能切身感受到顶级电信巨头在推进架构变革时的严谨与魄力，也能体会到面对未来 6G 与网络大模型演进时的探索热情。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=dh4lOZp400k&t=4s) RAN（无线接入网）计算架构正经历从封闭专有硬件向基于通用 COTS 服务器和云基础设施的开放化转型。
- [02:31](https://www.youtube.com/watch?v=dh4lOZp400k&t=151s) AI-RAN 是网络机器学习长期演进的延续，核心目标聚焦于网络性能优化、端到端自动化以及日常运维的极致简化。
- [02:31](https://www.youtube.com/watch?v=dh4lOZp400k&t=151s) AI 在无线网络中划分为 Day 0（规划）、Day 1（部署）和 Day 2（运维）三大生命周期阶段，当前绝大多数落地用例集中在 Day 2 阶段。
- [02:31](https://www.youtube.com/watch?v=dh4lOZp400k&t=151s) AI 赋能 RAN 的四大商业价值衡量标准包括直接削减运营成本、延缓/规避资本支出、降低用户流失率以及开拓全新增收机会。
- [02:31](https://www.youtube.com/watch?v=dh4lOZp400k&t=151s) 提升频谱效率（Spectral Efficiency）是运营商的核心抓手，能显著延缓在数万基站加装硬件的高昂资本支出，优化投资节奏。
- [06:51](https://www.youtube.com/watch?v=dh4lOZp400k&t=411s) 关键电信网络对“五个九”可用性的要求极高，AI 引入必须建立“安全自动化”防护边界（Guardrails），防止不同应用产生对抗冲突与震荡死循环。
- [08:31](https://www.youtube.com/watch?v=dh4lOZp400k&t=511s) AI 只是工程工具箱中的一种工具，应审慎评估落地场景，避免过度复杂的算法设计导致一线运维团队难以快速排障和定位根因。
- [09:48](https://www.youtube.com/watch?v=dh4lOZp400k&t=588s) 6G 预计将于 2030 年前后落地，AT&T 正通过 Cloud RAN 与 SMO（服务管理与编排）两大核心工程提前 3 至 5 年完成基础架构储备。
- [09:48](https://www.youtube.com/watch?v=dh4lOZp400k&t=588s) SMO 架构采用分层闭环设计：集中层负责反应延迟约 1 秒的非实时策略，极低时延实时闭环则直接下沉至 Far Edge 远端边缘层协作。
- [13:31](https://www.youtube.com/watch?v=dh4lOZp400k&t=811s) 通过 SMO 和 Non-RT RIC 开放架构，运营商可打破传统主设备商封闭生态，引入如 Aira Technologies 等初创公司快速开发与验证 rApps 微服务应用。
- [15:44](https://www.youtube.com/watch?v=dh4lOZp400k&t=944s) 针对电信网络特有的海量结构化表格数据（Tabular Data），AT&T 研发了专属网络基础模型（Network Foundation Model），推动 AI 成为底层通用能力。
- [15:44](https://www.youtube.com/watch?v=dh4lOZp400k&t=944s) 基站能效优化的实践表明，将网络基础模型叠加至成熟深度学习算法之上能够实现二次能效飞跃，凸显持续迭代算法架构的长期红利。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=dh4lOZp400k&t=4s) 开场介绍与 RAN（无线接入网）的基础架构解构
活动由代表无线通信产业组织 CTIA 的 David Valdez 主持开场，随后邀请 AT&T 新兴无线接入网技术与网络管理团队负责人 Adam Loddeke 进行自我介绍。Adam 阐述了自己的职责范围，他主要聚焦于未来两年及以上的下一代新技术研发与应用场景落地，团队一部分侧重前沿技术研究，另一部分则专注于推动这些未来场景的商业化与网络实现。
为了让非专业背景的听众统一认知基准，Adam 对 RAN（Radio Access Network，无线接入网）的基础概念进行了直观普及。手机（UE 终端）通过运营商管理的无线频谱与蜂窝基站进行空中接口通信。整个网络由分布在各地的成千上万个基站铁塔组成。基站的物理形态通常是：塔顶安装有天线和射频单元（Radio），光纤线路从塔顶垂直到塔底的机房/机柜中，机房内配备有计算单元。过去这些计算单元多为专有定制硬件，而现在正向开放化演进，逐渐被运行在云基础设施层和 RAN 应用程序上的通用 COTS（商用现货）服务器所替代。从机房一直延伸到用户终端这一整段，即构成了团队核心聚焦的 RAN 范畴，其后则是传输网与核心网部分。
```mermaid
flowchart LR
accTitle: 无线接入网（RAN）基础物理与架构分层
subgraph UE_Layer["终端层"]
UE["用户设备 (UE / 手机)"]
end
subgraph Tower_Layer["基站塔顶"]
Antenna["天线 (Antennas)"]
Radio["射频单元 (Radios)"]
end
subgraph Base_Layer["塔底机房 (Shelter)"]
Fiber["光纤直连 (Fiber)"]
Compute["COTS 通用服务器 / 云基础设施"]
end
subgraph Core_Layer["后端网络"]
Transport["传输网 (Transport)"]
CoreNet["核心网 (Core Network)"]
end
UE <-->|"空中接口 (无线频谱)"| Antenna
Antenna --- Radio
Radio --- Fiber
Fiber --- Compute
Compute --> Transport --> CoreNet
```
---
### [02:31](https://www.youtube.com/watch?v=dh4lOZp400k&t=151s) AI 与 RAN 的演进关系及三大生命周期应用阶段
谈及 AI 原生（AI-Native）无线接入网基础设施的设计时，Adam 指出当前的 "AI-RAN" 本质上是网络中机器学习（ML）技术长期演进的延续。早期网络中部署的 AI 更多是基于决策树和规则的初级系统，随后逐步演变为更高级的深度学习能力。在网络中引入 AI 的核心目的在于：网络性能优化、赋能全流程自动化，以及大幅精简网络日常运维。
```mermaid
flowchart TD
accTitle: RAN 生命周期三大应用阶段与商业价值分类
subgraph Lifecycle["RAN 生命周期阶段"]
D0["Day 0：规划阶段<br/>• 流量预测<br/>• 基站规划与网络设计<br/>• 生成式 AI 辅助工程与项目管理"]
D1["Day 1：部署阶段<br/>• 新基站/小区自动配置<br/>• 开局与部署流程精简"]
D2["Day 2：日常运营与维护<br/>• 节能调度优化<br/>• 流量智能调度<br/>• 故障自愈与中断补偿"]
end
subgraph Business["四大商业驱动力"]
B1["成本削减 (Cost Savings)"]
B2["成本延缓与规避 (Cost Deferral)"]
B3["改善体验降低流失 (Churn Reduction)"]
B4["新业务增收机会 (Revenue Generation)"]
end
D0 --> Business
D1 --> Business
D2 --> Business
```
Adam 将 AI 在无线网络中的应用全面细分为三大生命周期阶段（Day 0、Day 1、Day 2）与四大商业价值维度：
- **Day 0（规划阶段）**：聚焦于中长期网络容量规划。利用机器学习模型进行流量预测、基站选址设计，甚至引入生成式 AI（Generative AI）来协助工程师和项目经理自动化梳理规划流程。
- **Day 1（部署阶段）**：着重于网络资产的入网部署，通过自动化手段实现新站点与新小区的零接触或一键式配置上线。
- **Day 2（运维阶段）**：承载了当前绝大多数的 AI 实际用例，包括基站能效动态调节（节能关断）、智能流量导流（Traffic Steering）、基站突发故障时的网络自愈与覆盖补偿等。
- **价值归类维度**：所有应用均被纳入四类商业目标衡量——直接削减运营成本（Cost Savings）、延缓或规避资本支出（Cost Deferral/Avoidance）、通过提升网络体验减少用户流失（Churn Reduction），以及开拓全新的业务收入来源（Revenue Generation）。
- **频谱效率的核心地位**：在所有方向中，Adam 最为关注的是**提升频谱效率（Spectral Efficiency）**。无线频谱的竞拍购置是运营商最高昂的支出之一，且每引入一段频谱都需要在全网数万个站点进行大规模硬件加装与升级。通过 AI 提升现有频谱的使用效率，虽然无法彻底消除未来的资本投入，但能够显著延缓大规模扩容投资，将其平滑推迟到预算更可控的时间节点。
---
### [06:51](https://www.youtube.com/watch?v=dh4lOZp400k&t=411s) 关键基础设施的升级挑战：“安全自动化”与防死循环
电信网络作为承载社会运行的关键基础设施，必须严格维持“五个九”（99.999%）的高可用性。因此在网络中引入任何 AI 或自动化能力时，核心红线是确保“安全自动化（Safe Automation）”，设置严密的防护边界（Guardrails），精确评估每一项自动化操作的影响。
Adam 举了一个生动的反面场景：如果网络中缺乏全局协调机制，一个应用因为检测到下行阻塞为了性能优化而执行天线下倾（Down-tilt），而另一个以扩大覆盖为目标的优化应用同时感知到信号不足并执行天线上倾（Up-tilt），两者就会陷入无休止的“一下一上”对抗震荡与死循环（Endless Loop）。这不仅无法优化网络，还会引发区域性能劣化甚至网络瘫痪。因此，所有多智能体或自动化控制必须具备统一策略仲裁，杜绝造成用户体验下降或非预期中断。
- 关键基础设施的核心考量：在保障业务不中断的前提下实现 AI 算法与硬件架构的平滑迭代。
- 安全防护机制（Guardrails）：避免不同控制算法在并发运行时出现目标冲突与震荡死循环。
---
### [08:31](https://www.youtube.com/watch?v=dh4lOZp400k&t=511s) AI 落地通信网络的经验与务实哲学
在总结给其他行业部署 AI 的经验教训时，Adam 将 AI 定义为“工具箱中的一种工具”。尽管 AI 热度极高且潜力巨大，但绝不是包治百病的灵丹妙药，企业应当克制且审慎地评估使用场景。
Adam 强调，并非网络中的所有问题都需要套用复杂的算法模型来解决。有时人为把系统设计得过于复杂，反而会导致下游的一线网络运维和排障团队在出现故障时难以定位根因、无法快速恢复服务。因此，最聪明的做法是明确业务边界，仅在最能产生实际效能的场景中合理部署 AI，兼顾系统健壮性与可维护性。
### [09:48](https://www.youtube.com/watch?v=dh4lOZp400k&t=588s) 5G 与 6G 的演进对比及 AT&T 的两大核心网络工程
面对主持人关于 6G、元宇宙应用及政府关键任务用例的提问，Adam Loddeke 首先向现场观众通俗地梳理了 5G 与 6G 的代际差异与时间线。5G 是如今大众手机上广泛使用的网络，已经规模部署了大约五到六年；而 6G 预计要到 2030 年左右（可能在 2028 至 2032 年浮动）才会初现端倪，目前甚至尚未进入正式标准制定阶段。每一次移动通信代际更迭，都伴随着更宽的频谱带宽、更快的数据传输速率以及全新的应用场景需求。
为了提前 3 至 5 年为 6G 奠定坚实基础，Adam 透露他目前跳脱出平时“着眼未来两年”的常规研发角色，全身心投入推进两项重大基础设施战略工程：
1. **Cloud RAN（云化无线接入网）**：将开放（Open RAN）与虚拟化（vRAN）架构大规模引入并扩展至 AT&T 的商用网络中。
2. **SMO 架构（服务管理与编排）**：构建全网的管理与编排中枢，其核心包含非实时无线智能控制器（Non-Real-Time RIC），可运行各类优化应用（rApps），同时具备强大的工作流编排引擎和全网数据使能管理能力。
```mermaid
flowchart TD
accTitle: AT&T RAN 架构与 AI 闭环部署分层示意图
subgraph SMO_Layer ["SMO 集中层 (全美 8-12 个核心节点)"]
SMO["SMO 平台 / Non-RT RIC"]
rApps["rApps (异常检测 / 能效优化)"]
NFM["网络基础大模型 (Network Foundation Model)"]
SMO --- rApps
SMO --- NFM
end
subgraph Edge_Layer ["远端边缘层 (Far Edge / 站点基站)"]
CloudRAN["Cloud RAN 节点 (vCU / vDU)"]
NearRT["Near-RT / 实时闭环引擎"]
ComputeMix["弹性算力: CPU (轻量) / GPU (重载训练)"]
CloudRAN --- NearRT
CloudRAN --- ComputeMix
end
SMO_Layer -- "非实时闭环控制 (延迟 ~1s)" --> Edge_Layer
Edge_Layer -- "表格化网络指标与遥测数据" --> SMO_Layer
```
- **部署周期预估**：5G 历经约 5-6 年规模应用，6G 预计在 2030 年前后落地（范围约 2028-2032 年）。
- **SMO 节点布局**：全美预计建设 8 到 12 个集中式 SMO 部署点，负责处理非实时闭环策略，端到端反应延迟可压低至 1 秒以内，足以敏捷应对全网性能调整。
- **边缘层设计**：极低时延的实时闭环用例则直接下沉至远端边缘（Far Edge），与 Cloud RAN 节点原生协作运行。
- **算力弹性配置**：轻量级推断以标准通用 CPU 为主；涉及大模型处理或复杂在线训练时，按需在网络特定节点灵活挂载 GPU。
---
### [13:31](https://www.youtube.com/watch?v=dh4lOZp400k&t=811s) 打破电信封闭生态：与 Aira Technologies 的 rApp 合作创新
针对电信行业长期以来相对封闭的技术生态，Adam 指出 AT&T 正致力于通过引入开放架构逐步破局。构建 SMO 的核心战略目标之一，就是让第三方创新开发者和初创公司能够直接在电信网络之上开发、部署和测试微服务化应用程序（rApps），打破以往单纯依赖传统 RAN 设备主设备商或仅靠运营商内部自研的局限。
在具体生态合作上，Adam 举例介绍了 AT&T 与来自硅谷的创新企业 Aira Technologies 的深度合作。双方目前正基于 Non-RT RIC 展开两个关键应用方向的联合验证：
- **异常检测（Anomaly Detection）rApp**：利用 AI 自动识别并诊断无线接入网中的隐蔽故障与流量突变。
- **代码自动生成（Code Generation）用例**：通过生成式 AI 技术辅助自动化生成 rApp 的逻辑代码与部署配置，进一步缩减开发迭代周期。
AT&T 现网中虽然已稳定运行着名为 **Midas** 的自研智能化管控平台（集成了基于规则的系统、传统机器学习以及深度学习模型），同时也运行着部分第三方商业套件，但未来的方向是建立标准化应用试验机制：让外部厂商进场快速试点验证，评估商业与网络收益，加速推动整个无线产业向开放软件生态演进。
---
### [15:44](https://www.youtube.com/watch?v=dh4lOZp400k&t=944s) 无处不在的 AI：网络基础模型与能效优化实战演进
在谈及未来几年在现有庞大存量（Legacy）与新型网络中落地 AI 的挑战与前景时，Adam 强调 AT&T 的核心理念是将 AI 打造为一种“无处不在”（ubiquitous）的网络底层能力。通信网络中的绝大部分遥测数据与性能指标都是以数值点、统计日志为主的结构化表格数据（Tabular Data）。针对这一特征，AT&T 自主构建了专门面向通信域的**网络基础模型（Network Foundation Model）**。
Adam 分享了一个极具代表性的能效优化（Energy Efficiency）演进实例：
- **第一阶段（传统深度学习）**：此前网络中部署的深度学习节能 rApp 已经通过精准关断低负载载波等手段，为 AT&T 带来了非常显著的电力消耗与运营成本节省。
- **第二阶段（网络基础模型赋能）**：将新型网络基础模型与原有的机器学习算法结合后，系统的预测精度与决策效率进一步提升，实现了比上一代方案更为出色的二次能效提升。
正如 Adam 总结所言：“曾经做过的优化，并不意味着未来不能做得更好。”面对不断演进的机器学习算法与日新月异的硬件计算架构，电信运营商必须保持高度的架构灵活性，勇于拥抱新技术并积极探索未知的网络智能化红利。会谈在主持人的致谢与全场掌声中圆满收尾。
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=dh4lOZp400k&t=4s) 团队一部分侧重前沿技术研究，另一部分则专注于推动这些未来场景的商业化与网络实现。 — Adam Loddeke
> [02:31](https://www.youtube.com/watch?v=dh4lOZp400k&t=151s) 当前的“AI-RAN”本质上是网络中机器学习（ML）技术长期演进的延续。 — Adam Loddeke
> [02:31](https://www.youtube.com/watch?v=dh4lOZp400k&t=151s) 通过 AI 提升现有频谱的使用效率，虽然无法彻底消除未来的资本投入，但能够显著延缓大规模扩容投资，将其平滑推迟到预算更可控的时间节点。 — Adam Loddeke
> [06:51](https://www.youtube.com/watch?v=dh4lOZp400k&t=411s) 在网络中引入任何 AI 或自动化能力时，核心红线是确保“安全自动化（Safe Automation）”，设置严密的防护边界（Guardrails），精确评估每一项自动化操作的影响。 — Adam Loddeke
> [06:51](https://www.youtube.com/watch?v=dh4lOZp400k&t=411s) 如果网络中缺乏全局协调机制，一个应用因为检测到下行阻塞为了性能优化而执行天线下倾（Down-tilt），而另一个以扩大覆盖为目标的优化应用同时感知到信号不足并执行天线上倾（Up-tilt），两者就会陷入无休止的“一下一上”对抗震荡与死循环（Endless Loop）。 — Adam Loddeke
> [08:31](https://www.youtube.com/watch?v=dh4lOZp400k&t=511s) 将 AI 定义为“工具箱中的一种工具”。尽管 AI 热度极高且潜力巨大，但绝不是包治百病的灵丹妙药，企业应当克制且审慎地评估使用场景。 — Adam Loddeke
> [08:31](https://www.youtube.com/watch?v=dh4lOZp400k&t=511s) 并非网络中的所有问题都需要套用复杂的算法模型来解决。有时人为把系统设计得过于复杂，反而会导致下游的一线网络运维和排障团队在出现故障时难以定位根因、无法快速恢复服务。 — Adam Loddeke
> [13:31](https://www.youtube.com/watch?v=dh4lOZp400k&t=811s) 构建 SMO 的核心战略目标之一，就是让第三方创新开发者和初创公司能够直接在电信网络之上开发、部署和测试微服务化应用程序（rApps），打破以往单纯依赖传统 RAN 设备主设备商或仅靠运营商内部自研的局限。 — Adam Loddeke
> [15:44](https://www.youtube.com/watch?v=dh4lOZp400k&t=944s) AT&T 的核心理念是将 AI 打造为一种“无处不在”（ubiquitous）的网络底层能力。 — Adam Loddeke
> [15:44](https://www.youtube.com/watch?v=dh4lOZp400k&t=944s) 曾经做过的优化，并不意味着未来不能做得更好。 — Adam Loddeke
## Mentioned
Adam Loddeke, David Valdez, AT&T, CTIA, Aira Technologies, Midas, Cloud RAN, Open RAN, vRAN, SMO (Service Management and Orchestration), Non-Real-Time RIC, rApps, 5G, 6G, Network Foundation Model (网络基础大模型)
## Video description
Adam Loddeke, AVP - RAN Technology, AT&T
David Valdez, Vice President, Privacy & Cybersecurity Policy, CTIA
Abstract
How is AT&T evolving its RAN (Radio Access Network) infrastructure to support AI-native capabilities, and why is now the right time to invest in this transformation?
What are the most promising AI use cases in RAN today—from anomaly detection to link adaptation—and how are they being operationalized?
[Source](https://www.youtube.com/watch?v=dh4lOZp400k)
Source quotes
[0:04] Adam Loddeke: 团队一部分侧重前沿技术研究，另一部分则专注于推动这些未来场景的商业化与网络实现。
[2:31] Adam Loddeke: 当前的“AI-RAN”本质上是网络中机器学习（ML）技术长期演进的延续。
[2:31] Adam Loddeke: 通过 AI 提升现有频谱的使用效率，虽然无法彻底消除未来的资本投入，但能够显著延缓大规模扩容投资，将其平滑推迟到预算更可控的时间节点。
[6:51] Adam Loddeke: 在网络中引入任何 AI 或自动化能力时，核心红线是确保“安全自动化（Safe Automation）”，设置严密的防护边界（Guardrails），精确评估每一项自动化操作的影响。
[6:51] Adam Loddeke: 如果网络中缺乏全局协调机制，一个应用因为检测到下行阻塞为了性能优化而执行天线下倾（Down-tilt），而另一个以扩大覆盖为目标的优化应用同时感知到信号不足并执行天线上倾（Up-tilt），两者就会陷入无休止的“一下一上”对抗震荡与死循环（Endless Loop）。
[8:31] Adam Loddeke: 将 AI 定义为“工具箱中的一种工具”。尽管 AI 热度极高且潜力巨大，但绝不是包治百病的灵丹妙药，企业应当克制且审慎地评估使用场景。
[8:31] Adam Loddeke: 并非网络中的所有问题都需要套用复杂的算法模型来解决。有时人为把系统设计得过于复杂，反而会导致下游的一线网络运维和排障团队在出现故障时难以定位根因、无法快速恢复服务。
[13:31] Adam Loddeke: 构建 SMO 的核心战略目标之一，就是让第三方创新开发者和初创公司能够直接在电信网络之上开发、部署和测试微服务化应用程序（rApps），打破以往单纯依赖传统 RAN 设备主设备商或仅靠运营商内部自研的局限。
[15:44] Adam Loddeke: AT&T 的核心理念是将 AI 打造为一种“无处不在”（ubiquitous）的网络底层能力。
[15:44] Adam Loddeke: 曾经做过的优化，并不意味着未来不能做得更好。
Source quotes
[16:26] Adam Loddeke: just because it's been done doesn't mean you can't do it better going forward
Original transcript
[0:04] So, uh, good afternoon everyone. Thank
[0:05] you so much for joining us. Uh, as you
[0:07] mentioned, I'm David Valdez. I'm with
[0:09] CTIA and we represent the wireless, uh,
[0:12] industry. And Adam, do you want Adam, do
[0:14] you want to give a short introduction?
[0:15] >> Yeah, I'm Adam Ladaki. I work for AT&T.
[0:18] Uh, and I have kind of responsibilities
[0:20] for the emerging RAN architecture and
[0:23] network management uh, team within AT&T.
[0:26] I tend to focus on new technology and
[0:29] use cases that are tend to be in that
[0:31] two plus years into the future uh time
[0:34] frame. So I'm tend to be on uh part of
[0:38] my team's a little bit more research
[0:39] focused. The other part of it is trying
[0:41] to figure out how to drive realization
[0:42] of those use cases.
[0:45] Okay. Excellent. So I was thinking that
[0:47] in terms of level setting the
[0:48] conversation I wanted Adam to provide us
[0:50] just with an overview of RAN and what we
[0:52] mean by RAND and then we could just
[0:54] start talking about some of the uh
[0:56] dynamics of what RAN means. So Adam do
[0:58] you want to just give us a short review?
[1:01] >> I'll do a quick introduction what RAN is
[1:03] because I'm not sure if everybody
[1:04] understands what RAND it stands for
[1:06] radio access network. So those of you
[1:08] that may have some familiar with your
[1:11] cell phones and how they work, you
[1:13] obviously have a have a have a device a
[1:15] handset that has communicates over the
[1:18] air via the spectrum that's managed by
[1:21] the various operators.
[1:23] uh and typically there's multiple
[1:27] thousands of cell sites that exist uh
[1:30] within the network that transmit that
[1:33] spectrum uh and communicate with the
[1:35] multiple UE devices. And so think about
[1:38] you have that overtheair communication.
[1:41] You have within the cell site you have
[1:43] at the top of the tower there's
[1:44] typically antennas
[1:46] and radios and then from there you have
[1:49] fiber connections that run down into the
[1:51] base of the tower. and there's some sort
[1:53] of compute. It could be uh I'd say
[1:57] proprietary based, which is most of
[1:58] what's out there today, but I think
[2:00] you're starting to see more openness uh
[2:02] transpire and start you're seeing more
[2:04] what I would say cot servers that are
[2:05] going to be at the at at the base of the
[2:07] tower that'll be running on um some sort
[2:11] of a cloud infrastructure layer along
[2:13] with the RAN application. And so when I
[2:16] think about from that shelter out to the
[2:19] UEE, that's kind of what we're looking
[2:21] at from a RAN perspective. And then
[2:22] obviously you got the transport and the
[2:24] core part of the network that exists,
[2:26] but but my team focuses primarily on
[2:28] that RAN piece of it.
[2:31] >> Thank you. So um Adam and I were talking
[2:33] earlier about RAN and AI, right? and
[2:37] talking about RAN infrastructure and how
[2:40] RAN infrastructure is being designed to
[2:43] have AI native capabilities and I wanted
[2:45] Adam to kind of walk us through that a
[2:47] little bit.
[2:48] >> Yeah. So I mean I think today you know
[2:50] there's a lot of talk about AI ran and I
[2:52] think you know we're at this point where
[2:55] really it's a continuation of just the
[2:57] evolvement of machine learning uh within
[3:01] within the network and the machine
[3:02] learning techniques that are applied.
[3:04] And so, you know, there's, you know, I'd
[3:07] say in the early days, there was a lot
[3:08] more um decision tree, more rulebased
[3:11] type of AI that would have been, if you
[3:14] want to call it AI, that would have been
[3:15] deployed in our networks. And then it
[3:17] evolved more towards um
[3:21] more advanced deep learning type
[3:23] capabilities. And so, again, at this
[3:25] point, you know, you're starting to see
[3:26] an evolvement of of the types of AI that
[3:28] exists in the network in different uh
[3:30] capabilities. and and really we utilize
[3:32] AI in our networks to to optimize the
[3:35] network. It enables automation and it
[3:37] also is is what we do to kind of really
[3:39] streamline uh what we do from an
[3:41] operational perspective in the network.
[3:43] >> And can you talk a little bit about what
[3:45] are some of the more promising um AI use
[3:48] cases in the RAN uh space?
[3:51] >> I I think there's really the ability to
[3:53] ubiquously apply AI throughout the RAN.
[3:56] Um I don't think it needs to be isolated
[3:58] any particular area. I think I think of
[4:00] really I think a lot of the things we
[4:02] look at we we kind of bucketize them in
[4:05] in three domains. You I think of day
[4:07] zero which is kind of the planning
[4:09] horizon. And so we're looking at um you
[4:13] know potentially traffic forecasting you
[4:16] know how do we how do we build and and
[4:19] design sites into the network. So you
[4:20] kind of have the planning stage of
[4:22] things and how we utilize machine
[4:23] learning and AI in that space and even
[4:26] in some cases use generative AI uh to
[4:29] help streamline some of the activities
[4:30] that the engineers and project managers
[4:32] are doing. Then we kind of have what I
[4:34] call day two or day one which is the
[4:37] deployment of of these network assets.
[4:39] And so think about automating how we
[4:42] configure new sites, new cells into the
[4:44] network and how do we streamline that
[4:46] piece of it. And then I think the
[4:48] thirdly there's kind of what I call day
[4:49] two and that's kind of the operations of
[4:51] the network. And probably the the vast
[4:53] majority of use cases tend to fall that
[4:55] we think of fall in that day two
[4:57] category where you're operating the
[4:58] network itself. Um and again you think
[5:02] about energy efficiency. You think about
[5:04] um traffic steering. What do you do for
[5:07] outage compensation? There there's a lot
[5:09] of use cases that exist. I think in that
[5:11] third category, you know, there's
[5:14] there's and then I think when I think of
[5:15] those three categories, then there's
[5:16] buckets. You know, within those buckets,
[5:18] you have, okay, is this a cost savings,
[5:22] you know, use case? Is this a
[5:25] cost deferral, you know, or cost
[5:28] avoidance type of use case? Does this
[5:30] drive some sort of customer
[5:33] churn reduction where it it's you're
[5:35] improving the performance of the
[5:37] customer experience, but you're not
[5:38] necessarily driving any cost out of the
[5:40] business, but you're you're effectively
[5:42] doing it by by reducing the amount of
[5:44] churn that would exist. And then I think
[5:46] the fourth category that we tend to see
[5:48] is um new new revenue generation
[5:51] opportunities. You know, how how do you
[5:53] incorporate opportunities to drive
[5:55] increased revenue into the network? And
[5:56] so you look at all those different
[5:58] areas. Because I think the area that we
[5:59] that I like to focus on a little bit is
[6:01] is really how do I drive spectral
[6:03] efficiency gains into the network. And
[6:05] the reason I say that is because one of
[6:07] the the most the one thing that I spend
[6:10] probably the most amount of money on in
[6:11] the network on a day in dayout basis at
[6:13] least in the in the RAN space is really
[6:14] it's it's the spectrum that we acquire.
[6:17] There's a high cost around spectrum and
[6:19] then beyond the spectrum cost it's
[6:21] deploying those that spectrum into the
[6:24] network on a on a siteby-sight basis
[6:26] throughout throughout the network. And
[6:27] so that's a significant amount of cost
[6:29] um to manage that and continually to to
[6:32] upgrade. And so if we can defer that
[6:33] cost through spectral efficiency gains,
[6:36] get more efficient with the spectrum
[6:37] that we have deployed in the network, it
[6:39] really drives a lot of savings to AT&T,
[6:42] it may not avoid the need to spend, but
[6:45] it delays it to a point where it becomes
[6:47] more manageable um within the budget
[6:50] itself.
[6:51] What's unique about uh AT&T and other
[6:54] carriers is that these are critical
[6:55] infrastructures. So the idea that as
[6:59] you're doing upgrades to the to to to
[7:02] the infrastructure, you got to continue
[7:04] to continuously have continuous service.
[7:06] Can you talk about some of the
[7:07] challenges of uh upgrading the network
[7:10] uh through by scaling AI AI and how you
[7:14] do that through hardware and software
[7:15] and some of the challenges of upgrading
[7:17] the system? I think for us it's you know
[7:19] again we're we're critical
[7:22] infrastructure you know and so we we
[7:24] drive towards trying to achieve 59's
[7:26] reliability and so when we think about
[7:28] introducing AI or any capability to the
[7:30] network we always have to think about
[7:32] how do we do it safe and so a lot of
[7:34] times we'll talk about safe automation
[7:36] you know how do you how do you introduce
[7:38] automation into the network safely so
[7:40] that you have the right guardrails in
[7:41] place and you understand what changes
[7:44] are being done and and you you can
[7:47] manage what that impact is because at
[7:49] the end of the day I don't want things
[7:51] running into a continuous loop where
[7:53] this application is for example because
[7:56] it's it's a great idea it down tilts a a
[7:59] sector in the site because it's for
[8:01] whatever reason might be blocking and
[8:03] then I have a use case over here that's
[8:04] trying to maximize coverage in the
[8:06] network and and so then it's up tilting
[8:08] and what you don't want to get is is in
[8:09] the cycle of one's down tilting one's up
[8:12] tilting and you get into this endless
[8:13] loop and so you know we really have to
[8:16] think about collectively when we when we
[8:19] introduce all of this automation that
[8:20] we're doing it safely and and making
[8:22] sure that it doesn't create a
[8:24] degradation in customer experience or an
[8:27] outage which obviously is something that
[8:28] we're highly um focused on.
[8:31] >> So what are some of the lessons learned
[8:34] um that you think may apply to other
[8:36] industries when you're deploying AI in
[8:39] the network? I
[8:41] >> mean I I look at AI as a kind of the
[8:43] tool in a toolkit. It's an exciting new
[8:45] and obviously we're everybody's here to
[8:47] talk about AI and so I think it's it
[8:50] it's
[8:51] there's a lot of potential and promise
[8:53] here um around AI. I think it it does
[8:57] require the right guard wells to be put
[8:59] in place. I mean we we need to be smart
[9:01] about where we deploy it um because at
[9:04] the end of the day we are responsible
[9:06] for the the experience for the customer
[9:09] and and so we want to make sure we're
[9:10] doing the right things at the right
[9:12] time. Um and AI is popular, but it does
[9:16] doesn't need to be the solution for
[9:17] everything. You know, you don't always
[9:19] need to have a complex
[9:22] um algorithm to solve a problem in in
[9:24] the network. And sometimes the more
[9:26] complex you make things, the more
[9:28] difficult you make it for all the
[9:29] surrounding the downstream teams to
[9:31] troubleshoot the issues and resolve
[9:33] them. And so I think it's being smart
[9:35] about where we put AI uh and making sure
[9:38] that we take advantage of it where it
[9:39] makes sense. and um how is uh AT&T
[9:44] preparing the RAN infrastructure for the
[9:48] demands of 6G metaverse application and
[9:50] other other missionritical government
[9:52] use cases and can you just mention the
[9:54] difference between 5G and 6G for people
[9:56] in the audience?
[9:57] >> So 5G is probably what you're what you
[9:59] have on your phone today. Um it's mostly
[10:02] what's been deployed I think since 20
[10:06] make sure I'm over the last couple
[10:08] years. years. I'm trying to think at
[10:09] least at least the last uh
[10:12] >> five six years
[10:12] >> five six years I think we've had 5G
[10:14] deployed. Uh I think 6G is probably not
[10:18] going to start showing its face until
[10:19] probably 2030 give or take. Now I say
[10:22] that and and then it may move to 28 or
[10:24] it could move to 2032 but I think that
[10:26] it's still very much in this not even in
[10:28] the standards yet. So I think we're
[10:29] going to see that probably in the
[10:30] 2030ish time frame. Um so it's the next
[10:34] evolution. um probably new spectrum, new
[10:37] capabilities. Typically with every new
[10:38] G, there's wider bandwidths, better
[10:42] better speeds, different use cases that
[10:44] you're trying to solve. Um, but I think,
[10:47] you know, I think at the end of the day,
[10:49] it's, you know, when we think about the
[10:52] what we're doing today. So, one of the
[10:53] the two main projects that I have on my
[10:55] plate that's that kind of deviates from
[10:57] my normal stay in two years into the
[11:00] future type role is is getting cloud
[11:03] RAN, which is kind of open and and
[11:05] virtualized RAN into the network and
[11:07] scaled into AT&T's network. And then the
[11:09] second project that's on my plate at the
[11:11] moment is around SMO architecture. So,
[11:14] service management orchestration. So
[11:15] think of that as a network management a
[11:21] nonreal time ran intelligent controller
[11:24] that you can run applications on and
[11:25] then it also has kind of an
[11:26] orchestration engine on that along with
[11:28] data enables some data management
[11:30] capabilities as well and so as we roll
[11:33] both of those new initiatives into the
[11:35] network and scale them out over the next
[11:37] three to five years as we get in
[11:39] preparation for 6G
[11:42] it gives us a lot of flexibility it you
[11:44] know this infrastructure that we're
[11:45] putting out there gives us ultimately
[11:47] flexibility into the network. Um, and so
[11:49] when we think about different AI use
[11:51] cases that will be out there, you know,
[11:53] things that that are more non-real time
[11:56] that are closed looped, you know, we
[11:58] could probably address through the SMO
[12:01] architecture. And so, think of the SMO
[12:03] probably being deployed at anywhere from
[12:05] 8 to 12 locations across the US. Doesn't
[12:08] doesn't need to be real time, but but
[12:10] again, you're going to be able to have
[12:11] latencies as low as potentially 1
[12:13] second. So you're still going to be able
[12:14] to react relatively quick to to problems
[12:17] in the network or address issues in the
[12:19] network. Things that require real time
[12:23] closed loop um use cases, those types of
[12:25] use cases, you're probably going to have
[12:26] to run them on the far edge um
[12:28] potentially with cloud RAM, maybe run
[12:31] either within in conjunction with the
[12:33] application itself or you know as a
[12:36] partner to kind of that cloud RAM node
[12:38] that would be deployed on the far edge.
[12:40] And so you have these different use
[12:42] cases. Um and I think for us as a
[12:45] network operator, the infrastructure
[12:47] flexibility is key. Um you know, if it's
[12:51] a large,
[12:53] you know, machine learning model, yeah,
[12:55] maybe maybe we need GPUs somewhere in
[12:57] the network to support that if but I
[12:59] think a lot of the use cases that we'll
[13:00] see will probably be more lightweight.
[13:02] Um and maybe CPUs are are adequate for
[13:05] that. And then you're going to have
[13:06] models that are going to require
[13:08] continuous learning. uh and so that has
[13:10] a different cadence than than than a use
[13:13] case that requires offline training. And
[13:15] so, you know, for us, it's really about
[13:17] having flexibility with the
[13:19] infrastructure so that we can put GPUs,
[13:22] CPUs as needed, where needed within the
[13:24] network to support whatever the use case
[13:26] that ultimately evolves and triggers uh
[13:29] within the network.
[13:31] >> Thank you, Adam. One of the things I
[13:32] like about the AI4 conference is that
[13:34] it's opportunity for developers and
[13:37] manufacturers to kind of come together
[13:39] and look at opportunities. Can you give
[13:42] an example of uh companies you may have
[13:44] partnered with on AI and uh uh ran
[13:49] infrastructure improvements? I'm going
[13:51] to say historically the the telecom
[13:53] industry has been fairly um
[13:57] closed but I think one of the things
[13:59] that we are obviously pushing towards is
[14:02] is trying to open up the open up the
[14:03] network as much as possible and again
[14:05] it's not going to happen overnight but
[14:06] we are working towards this and I think
[14:08] one of the one of the companies for
[14:10] example that we're working on uh with is
[14:12] is called IRA IRA technologies and and
[14:15] they're based in Silicon Valley and and
[14:18] so we're working on an R app use case
[14:20] with them around anomaly detection. Um,
[14:23] and they're also looking at a use case
[14:25] around code generation u to support R
[14:28] apps. And so these R apps are these
[14:31] non-realtime
[14:33] applications that ultimately will reside
[14:36] on that SMO uh within AT&T's network. Uh
[14:40] and and so again the point of for us
[14:42] rolling out the SMO architecture is
[14:44] really to be able to enable third
[14:45] parties to innovate and enable these R
[14:48] apps to be something beyond just
[14:52] being developed by AT&T or or just
[14:54] developed by the RAN vendor. Um, if you
[14:57] look at most of what I have in my
[14:59] network today up to this point, it it's
[15:01] been developed by AT&T, we have we have
[15:04] a platform called Midas that that's AT&T
[15:07] homegrown um applications. Some of them
[15:10] are rule-based
[15:12] um some of them are machine learning,
[15:13] deep learning, uh different capabilities
[15:15] within within that. And then we have
[15:18] kind of a platform that's by a third
[15:20] party that that runs within the network
[15:21] that runs applications on it. The goal
[15:23] is really to get to the point where we
[15:25] can have third parties come in, us trial
[15:28] um what the capabilities of that
[15:29] application is and then figure out does
[15:32] it make sense for AT&T or not. And
[15:35] it's going to take time and a little bit
[15:36] of momentum, but I do see this as a
[15:38] space that will open up over the coming
[15:40] years because I I do think you're start
[15:41] see momentum in the space.
[15:44] >> So, as you're planning, I think when you
[15:46] were you and I were talking earlier, you
[15:47] had mentioned that you plan a couple of
[15:49] years out. So, as you're planning a
[15:52] couple of years out, how do you see AT&T
[15:56] deploying AI in this legacy system and
[16:00] what are some of the challenges? You've
[16:01] hit on some of those, but I think as our
[16:03] last question, just, you know, what do
[16:05] you see on the horizon and how are you
[16:07] preparing for some of the challenges
[16:08] that may be there as you're trying to
[16:10] implement AI across? Yeah, I mean I
[16:12] think again I think AI is is is really I
[16:16] look at it trying to put it ubiquitously
[16:18] throughout the network and and so I
[16:20] think we're still in this I'm going to
[16:23] call constant evolution of of the
[16:26] inclusion of AI. Um it doesn't fit all
[16:28] use cases but I think there's a lot more
[16:30] opportunity than than what we've had. I
[16:32] think there's scenarios I'll point to an
[16:34] example where we've had AI in the
[16:36] network around energy efficiency for
[16:38] example um and and we've seen
[16:42] significant amount of savings in the
[16:43] network from from that particular
[16:46] application that was that that utilized
[16:48] deep learning and
[16:51] from from an R app and and from a from
[16:53] an AI perspective but I think what we've
[16:55] seen is we built more of a network
[16:57] foundation model where we're you know
[16:58] again a lot of the a lot of data that we
[17:00] deal with is tabular um numbersbased uh
[17:04] points data. And so we've been we've
[17:07] built a network foundation model and
[17:10] with the use of the network foundation
[17:11] model along with the some of the machine
[17:14] learning that was previously created, we
[17:15] actually have been able to see even
[17:16] better improvements in the energy
[17:18] efficiency app that we have in the
[17:20] network. And so I think there's a lot of
[17:22] opportunity. Um just because it's been
[17:24] done doesn't mean you can't do it better
[17:26] going forward. And I think we have to be
[17:28] aware that there's going to continue to
[17:29] be evolution in the capabilities around
[17:32] machine learning and whether it be new
[17:35] models that that will move things
[17:36] forward could be advances in hardware
[17:39] architecture that will enable
[17:40] capabilities as well. And I think we
[17:42] just have to be flexible and be willing
[17:43] to try new things and and explore what
[17:46] the potential opportunities are out
[17:47] there.
[17:48] >> Excellent. Well, I think we're uh at an
[17:51] end. Please uh help me and applaud Adam
[17:54] for joining us today. Thank you so much,
[17:56] Adam. I really enjoyed that.
