The Case for a Horizontal Federated AI operating System for Telcos

Sebastian Barros

REVIEW 3 major objections 4 minor 11 references

The Case for a Horizontal Federated AI operating System for Telcos

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims telco AI fragmentation is structural and requires a horizontal federated AI operating system with federated training, standardized agent interfaces, and neutral governance.

desk verdict A readable, honest telco AI vision paper whose federation layer is a promissory note; worth refereeing as a position piece, not as a technical result. read the letter →

arxiv 2506.17259 v1 pith:SLR6ZZWX submitted 2025-06-09 cs.NI

Sebastian Barros This is my paper

classification cs.NI

keywords federatedAIoperatingsystemtelcolearningdatasovereigntyagentsOSS/BSSintegrationO-RANTMForum

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that telco AI is fragmented not for lack of good models but for lack of a shared infrastructure layer. It proposes a horizontal federated AI operating system that abstracts the telco stack, trains models collaboratively without moving raw data, and exposes standardized agent interfaces. If this layer existed, AI models and agents could be composed, certified, and reused across operators, vendors, and operational domains instead of being locked inside vertical platforms. The argument is structural: telco AI cannot scale until a neutral coordination substrate exists.

What carries the argument

The central object is the federated AI operating system itself: a horizontal software substrate exposing tightly scoped abstractions for telemetry ingestion, agent execution, and model lifecycle management. Its kernel is a minimal runtime that orchestrates agent lifecycles, provides a shared-memory abstraction for agents to publish and retrieve insights, ingests real-time and batch telemetry, and exposes observability and audit interfaces. The design also fixes a small set of canonical agent classes (anomaly detectors, experience predictors, SLA monitors, optimization advisors) with versioned, strongly typed input/output contracts, and a federated coordination layer that schedules cross-operator training with secure aggregation and differential privacy while leaving data on-premise.

What would settle it

Run a production-scale pilot in which several regulated operators train a shared model using secure aggregation and differential privacy; if model quality drops materially under non-IID data and partial participation, or if legal review blocks cross-border gradient exchange under GDPR, or if operators opt out for competitive reasons, the architecture's load-bearing premise fails.

Watch

Extended reading notes

Core claim

The paper's central claim is a structural diagnosis: current telco AI is fragmented by design, because vertically integrated vendor stacks, data-sovereignty rules, and missing interfaces prevent cross-domain and cross-operator composition. The proposed remedy is a horizontal federated AI operating system built from five layers (infrastructure, data abstraction, model orchestration, federated coordination, and agent-facing AI abstraction) with a minimal kernel that orchestrates agent lifecycles, shared memory, and telemetry ingestion. Its load-bearing mechanism is federated learning across sovereign operators: each trains on-premise, shares only encrypted updates, and the whole is governed by a neutral body aligned with TM Forum, O-RAN, and GSMA standards. The paper concludes this layer is a necessary and viable path to scalable, composable, sovereign AI in telecom, not merely another ML platform.

Load-bearing premise

The whole architecture rests on the assumption that federated learning across sovereign operators can be made accurate and communication-efficient enough for carrier-grade use while remaining legally safe and incentive-compatible for competing operators to share model updates.

Editorial extensions

If this is right

Operators could deploy the same certified AI agent across different vendors' networks and domains, removing today's point-to-point integration for each new use case.
The industry could build shared foundation models trained on federated data, yielding broader data diversity without centralizing subscriber information or violating data-localization laws.
A neutral governance body and agent marketplace would let third-party developers contribute and monetize certified agents, creating an app-store dynamic for network intelligence.
Telcos would shift from selling connectivity to selling programmable cognitive capacity, monetizing network intelligence as APIs.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the structural diagnosis is right, it predicts that point solutions operating inside a single vendor's stack will keep failing to compose, no matter how strong the underlying model is; the binding constraint is architectural, not algorithmic.
The same horizontal-federated pattern may transfer to other regulated infrastructure sectors such as energy or healthcare, where sovereign data cannot leave the operator's premises, although the paper does not make that generalization.
A concrete test would be an open-source minimal kernel with two operators, measuring how many cross-domain workflows can be assembled purely from certified agents; the number of composable workflows, not model benchmarks, would be the success metric.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Sebastian Barros has written a genuinely readable position paper, and its best feature is that it knows it is one. The proposal — a horizontal federated AI operating system for telcos, blending federated learning, agent abstractions, TM Forum Open APIs, O-RAN RIC, and NWDAF — is not new in its components, but the synthesis is coherent and the architecture is laid out clearly. The canonical agent classes in §6, the phasing strategy in §5.1, and the governance/certification discussion in §4.5 and §5.8 are the kind of concrete framing that could actually help practitioners align on a shared vocabulary.

What I value most is the candor. §4.4 openly states that secure aggregation and differential privacy lack standardized implementations in telco environments, that even anonymized cross-border updates may raise GDPR compliance concerns, and that without attribution/reward mechanisms the OS risks collapsing into silos. §5.3 lists vendor resistance, data governance, model brittleness, and ecosystem fragmentation as adoption barriers. That is honest engagement with the hard parts, and it earns the paper the benefit of the doubt.

The soft spot is the load-bearing one. The distinct value over a generic ML platform is the federated cross-operator layer, and that layer rests on capabilities the paper itself says are unresolved. The cited evidence (FedORA, the elastic O-RAN FL study) are single-domain research testbeds, not production-scale cross-sovereign deployments. The paper acknowledges the gap and then proceeds anyway. For a position paper, that is acceptable if the conclusion is framed as an open research agenda — but the abstract claims the architecture 'offers a path to break the current silos' and 'unlock ecosystem-level intelligence,' which overstates what is demonstrated. Also, §3.4 mentions 'EdgeGPT-5G' but reference [7] only covers DistilBERT; that citation mismatch should be fixed.

There is no math, no data, no implementation, so this is not a paper that produces a testable result. But it isn't trying to be. As a vision piece, it is internally consistent, standards-aware, and unusually self-aware about its own open problems. I would not cite it as evidence for federated learning viability, but I would cite it as an example of a convergent industry proposal. It deserves a serious referee if the venue handles position papers; it does not deserve to be judged as an empirical or systems contribution.

Referee Report

3 major / 4 minor

Summary. This position paper argues that AI adoption in telecommunications remains fragmented across vertical domains and vendor platforms, and that the industry needs a new horizontal "federated AI operating system" as a common execution and coordination substrate. The proposed architecture is described as a five-layer stack spanning infrastructure, data abstraction, model orchestration, federated coordination, and agent-facing abstractions, with federated learning as the core privacy-preserving mechanism for cross-operator collaboration. The paper also outlines business implications, deployment phasing, agent schemas, standards alignment with TM Forum, O-RAN, and GSMA, and a governance model centered on a neutral foundation. The contribution is architectural and strategic: it synthesizes existing trends and proposes a blueprint, but it contains no implementation, experiments, or formal specification.

Significance. If realized, a neutral, standards-aligned federated AI layer for telecom could address a genuine structural problem: AI models are currently locked in silos, and cross-operator learning is blocked by data sovereignty and competition. The paper usefully connects existing threads (O-RAN federated learning research, GSMA Open Gateway, TM Forum APIs) into a coherent vision. It also explicitly identifies key governance and incentive issues that are often overlooked. However, the significance is prospective: the paper provides no empirical evidence, no threat model, and no quantitative analysis of the federated learning feasibility on which the entire architecture depends. The central claim that this layer "offers a path to break current silos" (Abstract) is plausible but currently unsupported in a load-bearing way. As a position paper, it could serve as a useful roadmap if the unresolved technical and economic points are addressed or appropriately scoped.

major comments (3)

[3.3 and 5.6] The paper's defining feature is federated learning across sovereign operators, yet its feasibility evidence is limited to single-domain research testbeds. Section 3.3 cites FedORA and an elastic O-RAN federated learning study as validating "production-grade telecom environments," but those are early-stage academic xApp experiments, not cross-operator production deployments. Section 5.6 then merely mandates that secure aggregation, differential privacy, and encrypted computation "must be used," with no protocol design, threat model, or analysis of the utility/communication/latency trade-offs. Section 4.4 itself concedes that these techniques "lack standardized implementation across Telco environments" and that sharing even anonymized updates across borders "may raise compliance concerns." This is an internal contradiction that undermines the central claim. The authors should either provide evidence or a concrete design for cross-operator federated training, or substantially soften the claim that the proposed OS is a "viable path" (Abstract).
[4.4] The incentive and governance problem is acknowledged but not solved. Section 4.4 states that without attribution and reward mechanisms "the federated OS risks collapsing into isolated silos," but the discussion of Flower, Clara, and blockchain-based ledgers is only suggestive. No concrete incentive model, cost-sharing rule, or mechanism design is proposed. Since the entire shared-learning benefit depends on competing operators contributing high-quality updates, this is a load-bearing gap. The authors should specify at least a strawman incentive protocol or explain why this problem can be deferred to a future governance body.
[5.4] The "minimal viable kernel" is described only as a bullet list of components (orchestrator, shared memory, telemetry ingestion, observability) with no definition of the kernel's API surface, data model, or fault isolation boundaries. Section 5.4 says a "lightweight reference implementation will be critical," but none is provided. As an architecture paper, this is acceptable for a vision, but the term "operating system" implies a defined set of system calls or interfaces; without those, the claimed "standardized abstractions" (Section 1) are not falsifiable. The authors should provide a minimal interface specification or state clearly that this is a research agenda rather than an architecture proposal.

minor comments (4)

[Abstract and throughout] The text contains numerous typographical artifacts, such as "T elecommunications" (Section 2.1), "V endor" (Section 2.2), "F rom" (Section 4.1), and missing spaces in headings. These should be corrected.
[References] Reference [1] is duplicated (GSMA Intelligence is listed twice, once in the [1] entry and again after [3]). Also, several URLs are split across lines with spaces, which may break hyperlinks; please fix.
[3.3] The phrase "production-grade telecom environments" in reference to FedORA overstates the maturity of the cited study; consider using "research-grade O-RAN testbeds" to align with the actual evidence.
[5.2] Open data model identifiers such as "TMF620" and "TMF638" are mentioned without explanation; a one-line description of each would help readers unfamiliar with TM Forum catalogs.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an architecture proposal whose premises (data sovereignty, fragmentation, O-RAN federated-learning experiments) are external evidence; the proposal is not derived from its own definitions.

full rationale

This paper contains no equations, no fitted parameters, and no predictions, so the main circularity patterns (self-definitional derivations, fitted inputs renamed as predictions, ansatz smuggled via citation) do not apply. The central claim—that telco AI fragmentation requires a horizontal federated AI operating system—is argued from externally grounded observations: GSMA adoption statistics [1], SoftBank's proprietary LTM [2], GDPR and data-localization constraints [3], and experimental O-RAN federated-learning work (FedORA and elastic FL) [4][5]. The 'federated' design property is defined as preserving data locality and then justified by citing sovereignty regulations; that is an inference from an external constraint, not a tautology in which the term is defined as its own conclusion. The paper also does not rename a known result: it proposes a new system layer assembled from standards (TM Forum, O-RAN, NWDAF, Kubeflow) and prior research, rather than presenting those elements as a derived prediction. There are no self-citations by the author, so self-citation load-bearing and uniqueness-imported-from-authors patterns are absent. The paper does acknowledge limitations that weaken its feasibility case, e.g. §4.4: 'techniques like differential privacy and secure aggregation are still under active research and lack standardized implementation across Telco environments,' and 'even sharing anonymized updates across borders may raise compliance concerns.' That is an honest statement of missing support, not a circular step; per the rules, absence of external consensus or production-scale evidence is a correctness risk, not circularity. Because the paper's derivation chain is self-contained as an argument and does not reduce any conclusion to its inputs, the circularity score is 0.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

This position paper introduces no fitted constants or derived quantities. It depends on several domain assumptions about federated learning, governance, and standards mapping. The central claim is a proposal, so most technical feasibility is pulled from cited experimental work rather than demonstrated here.

assumptions (4)

domain assumption Federated learning with secure aggregation and differential privacy can achieve cross-operator model improvements without leaking sensitive data and within latency and bandwidth constraints.
Sections 3.3 and 5.6 assert this by citing O-RAN research [4][5] and standard privacy techniques; no production-scale evidence is provided.
domain assumption A neutral foundation-style governance body can maintain versioning, schema evolution, and certification in a way that prevents ecosystem fragmentation.
Section 4.5 analogizes to Linux Foundation and GSMA, but provides no mechanism or case study showing this works for AI agents across competing telcos.
domain assumption Existing telecom standards (TM Forum Open APIs, O-RAN RIC, 3GPP NWDAF) can be mapped into a common horizontal layer without major re-architecture of legacy OSS/BSS.
Section 3.6 and 5.2 assume API mapping is low-friction; the paper itself lists integration difficulty as a risk in Section 5.3.
domain assumption Kubernetes-native ML stacks such as Kubeflow and KServe provide sufficient reproducibility, CI/CD, and GPU-aware scheduling for carrier-scale AI.
Sections 3.6 and 5.7 rely on this to argue deployability, without validating on telco infrastructure.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Case for a Horizontal Federated AI operating System for Telcos." pith.science (2026). https://pith.science/paper/SLR6ZZWX

@misc{pith2026250617259,
  author       = {Pith},
  title        = {Pith review of: The Case for a Horizontal Federated AI operating System for Telcos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SLR6ZZWX}},
  note         = {Machine review of arXiv:2506.17259}
}

read the original abstract

As artificial intelligence capabilities rapidly advance, Telco operators face a growing need to unify fragmented AI efforts across customer experience, network operations, and service orchestration. This paper proposes the design and deployment of a horizontal federated AI operating system tailored for the telecommunications domain. Unlike vertical vendor-driven platforms, this system acts as a common execution and coordination layer, enabling Telcos to deploy AI agents at scale while preserving data locality, regulatory compliance, and architectural heterogeneity. We argue that such an operating system must expose tightly scoped abstractions for telemetry ingestion, agent execution, and model lifecycle management. It should support federated training across sovereign operators, offer integration hooks into existing OSS and BSS systems, and comply with TM Forum and O-RAN standards. Importantly, the platform must be governed through a neutral foundation model to ensure portability, compatibility, and multi-vendor extensibility. This architecture offers a path to break the current silos, unlock ecosystem-level intelligence, and provide a foundation for agent-based automation across the Telco stack. The case for this horizontal layer is not only technical but structural, redefining how intelligence is deployed and composed in a distributed network environment.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

11 extracted references · 9 canonical work pages

[7]

Alvi et al., DistilBERT for Telecom Support Systems: Compressing Domain-Specific LLMs for Edge Deployment , in Proceedings of the IEEE ICC 2023, Rome, Italy, May 2023

S. Alvi et al., DistilBERT for Telecom Support Systems: Compressing Domain-Specific LLMs for Edge Deployment , in Proceedings of the IEEE ICC 2023, Rome, Italy, May 2023

work page 2023
[1]

GSMA Intelligence, Telco AI: State of the Market Q2 2024 , GSMA Intelligence, September 2024

work page 2024
[2]

SoftBank Corp., SoftBank Develops a Foundational Large Telecom Model (LTM), Press Release, March 19, 2025

work page 2025
[3]

GSMA Foundry, Open-Telco LLM Benchmarks Launch to Advance AI in Telecoms, Press Release, March 2025. Available at: https://www.gs ma.com/newsroom/press-release/gsma-open-telco-llm-benchma rks-launches-to-advance-ai-in-telecoms/ GSMA Intelligence, Telco AI: State of the Market, Q2 2024 , GSMA Intelligence Report, September 2024

work page 2025
[4]

et al., Synergies Between Federated Learning and O-RAN: Towards an Elastic Architecture for Distributed ML Services , arXiv preprint arXiv:2305.02109, May 2023

Abdisarabshali, P. et al., Synergies Between Federated Learning and O-RAN: Towards an Elastic Architecture for Distributed ML Services , arXiv preprint arXiv:2305.02109, May 2023. Available at: https://ar xiv.org/abs/2305.02109

arXiv 2023
[5]

et al., FedORA: Resource Allocation for Federated Learn- ing in O-RAN using Radio Intelligent Controllers , arXiv preprint arXiv:2505.19211, May 2025

Salama, A. et al., FedORA: Resource Allocation for Federated Learn- ing in O-RAN using Radio Intelligent Controllers , arXiv preprint arXiv:2505.19211, May 2025. Available at: https://arxiv.org/ab s/2505.19211

arXiv 2025
[6]

R. Tso, H. Zhang, and L. Sun, TeleNet: Graph Neural Network for Root Cause Analysis in Communication Networks , IEEE Transactions on Network and Service Management, vol. 21, no. 1, March 2024. 21

work page 2024
[8]

TM Forum, Open APIs for Digital Service Providers , Accessed April

work page

Show all 11 references

[9]

Available at: https://www.3g pp.org/ftp/Specs/html-info/23288.htm

3GPP Technical Report 23.288, Network Data Analytics Function (NWDAF) for 5G , Release 17, 2023. Available at: https://www.3g pp.org/ftp/Specs/html-info/23288.htm

2023
[10]

Available at: https://www.gsma.com/new sroom/resources/open-gateway-and-telco-llms 22

GSMA Intelligence, Open Gateway and the Future of Telecom LLMs , Whitepaper, February 2024. Available at: https://www.gsma.com/new sroom/resources/open-gateway-and-telco-llms 22

2024
[2025]

Available at: https://www.tmforum.org/open-apis/

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [7]

Alvi et al., DistilBERT for Telecom Support Systems: Compressing Domain-Specific LLMs for Edge Deployment , in Proceedings of the IEEE ICC 2023, Rome, Italy, May 2023

S. Alvi et al., DistilBERT for Telecom Support Systems: Compressing Domain-Specific LLMs for Edge Deployment , in Proceedings of the IEEE ICC 2023, Rome, Italy, May 2023

work page 2023

[2] [1]

GSMA Intelligence, Telco AI: State of the Market Q2 2024 , GSMA Intelligence, September 2024

work page 2024

[3] [2]

SoftBank Corp., SoftBank Develops a Foundational Large Telecom Model (LTM), Press Release, March 19, 2025

work page 2025

[4] [3]

GSMA Foundry, Open-Telco LLM Benchmarks Launch to Advance AI in Telecoms, Press Release, March 2025. Available at: https://www.gs ma.com/newsroom/press-release/gsma-open-telco-llm-benchma rks-launches-to-advance-ai-in-telecoms/ GSMA Intelligence, Telco AI: State of the Market, Q2 2024 , GSMA Intelligence Report, September 2024

work page 2025

[5] [4]

et al., Synergies Between Federated Learning and O-RAN: Towards an Elastic Architecture for Distributed ML Services , arXiv preprint arXiv:2305.02109, May 2023

Abdisarabshali, P. et al., Synergies Between Federated Learning and O-RAN: Towards an Elastic Architecture for Distributed ML Services , arXiv preprint arXiv:2305.02109, May 2023. Available at: https://ar xiv.org/abs/2305.02109

arXiv 2023

[6] [5]

et al., FedORA: Resource Allocation for Federated Learn- ing in O-RAN using Radio Intelligent Controllers , arXiv preprint arXiv:2505.19211, May 2025

Salama, A. et al., FedORA: Resource Allocation for Federated Learn- ing in O-RAN using Radio Intelligent Controllers , arXiv preprint arXiv:2505.19211, May 2025. Available at: https://arxiv.org/ab s/2505.19211

arXiv 2025

[7] [6]

R. Tso, H. Zhang, and L. Sun, TeleNet: Graph Neural Network for Root Cause Analysis in Communication Networks , IEEE Transactions on Network and Service Management, vol. 21, no. 1, March 2024. 21

work page 2024

[8] [8]

TM Forum, Open APIs for Digital Service Providers , Accessed April

work page