Commercialization WG – 02.13.26

Updated Mar 2, 2026 Commercialization

Invited Pete Bernard Eric Smiley Jason Shepherd Rosina Haberl Michael Kuptz Woolley, Rob

Summary

Contact List and Document Updates
The team needs updated contact information for participating companies and confirmed that documentation updates are required, including addressing terminology shifts like avoiding the term “EAif”.

Defining Edge Taxonomy and Content
The discussion centered on defining the document’s structure, which includes ideal state definitions, a solution matrix, and a taxonomy based on LF Edge’s four-bucket management paradigms to frame decision-making for AI solutions across the continuum.

Taxonomy Approval and AI Workloads
It was decided that the proposed taxonomy needs approval from the Working Group Steering Committee via email review before a scheduled discussion, and the document must cover diverse AI types and their appropriate placement (training vs. inferencing) across different edge regions

ALIGNED

Paper Style Guide Terminology

Paper must avoid “EAIF” acronym. Use “EDGE AI ” for EDGE AI FOUNDATION, following internal style guide.

Document Reference AI Efforts

Document should reference industry efforts, such as AI RAN alliance, as specific examples of networking optimization efforts.

Infrastructure Location Terminology

Document structure uses four Mana paradigms (four buckets) to delineate where infrastructure, applications run. Paradigms provide framework for discussing tradeoffs.

Taxonomy Approval Process

Taxonomy section requires approval from Working Group Steering Committee (WGSC). Process involves emailing draft to WGSC alias followed by talk-through meeting for alignment.

Expand AI Type Scope

Document includes expansion of different AI types (Computer Vision, NLP, Federated Learning) within use cases section. Expansion helps frame specific application of infrastructure.

More details:

  • Meeting Start and Participant Status: The meeting began with Rosina Haberl and Michael Kuptz noting that Pete Bernard would be joining later due to another meeting until 10:15, and Eric would not be able to attend. The discussion temporarily shifted to Rob Woolley’s scouting background and snow conditions before Pete Bernard and Jason Shepherd joined, followed by Rob Woolley momentarily dropping from the call (00:00:00).
  • Request for Contact Information: Jason Shepherd asked for the best contacts for companies like ST and Dell, mentioning they had already contacted Raul for a security primer (00:01:24). Pete Bernard confirmed the existence of a Spab directory at “spab.edgifiation.org,” and Rosina Haberl offered to send the link, noting that any list of participating company contacts would be helpful for outreach (00:01:24).
  • Document Updates and Taxonomy Discussion: Rosina Haberl noted that the directory and some other documentation need updates, including the retirement of Moshe from SA and changes regarding “compute is not now model cat”. Jason Shepherd confirmed they are working on a paper to define the terminology for the edge world and discussed the document’s structure, which includes a solution matrix and taxonomy based on LF Edge (00:02:37) (00:06:05). Pete Bernard mentioned that the term “EAif” should be avoided in favor of simply “Aji” or “Aji Foundation” as an internal style guide preference (00:03:42).
  • Document Content and Ideal State: Jason Shepherd outlined the document’s sections, including ideal states, considerations, and a solution matrix that helps frame decisions when selecting a solution provider for AI and applications (00:03:42). They explained that the ideal state involves the portability of applications and AI models across the continuum, balancing performance, security, and privacy, while the taxonomy will help build out a table showing the types of compute and use cases at each location (00:07:03). The final section will focus on mapping specific use cases, such as on-device and generative AI, to the edge (00:08:13).
  • Service Provider Edge and AI RAN Alliance: Rob Woolley provided an update on the service provider edge section, drawing information from their colleagues working with the AI RAN Alliance (00:08:13). They questioned whether it was acceptable to reference the AI RAN Alliance directly and how detailed the content should be within the solution matrix entries. Jason Shepherd suggested mentioning industry efforts like the AI RAN Alliance as examples of optimization efforts, alongside general considerations about different types of providers, such as telcos, CSPs, and consulting services (00:09:11).
  • Nomenclature and Abstraction for Edge Regions: Rob Woolley highlighted a challenge in linking their service provider content to the document’s defined terminology, noting confusion when switching between terms like “service provider edge/user edge” and the LF Edge taxonomy terms “distributed edge/constrained device edge” (00:10:41). Jason Shepherd recommended using the four-bucket management paradigms (MANO) from LF Edge, distinguishing between service providers and end users deploying solutions within those four paradigms, such as data center edge or distributed edge (00:12:02). Rob Woolley noted a potential gap between distributed edge and data center edge in the definitions, particularly regarding security perimeters and handling complex AI models, which they plan to investigate further (00:13:22).
  • AI Workloads and Edge Regions: The discussion clarified that heavy AI training is typically confined to the data center edge due to security and compute needs, while inferencing is more likely to occur at the distributed edge or constrained device edge (00:14:36). Pete Bernard inquired about the AI RAN Alliance’s focus—specifically whether they address AI workload inferencing for network optimization or commercial GPU as a service offerings. Rob Woolley confirmed that the alliance addresses both, with working groups dedicated to using AI for radio access network efficiency, managing optimized infrastructure, and deploying AI on network equipment that is currently underutilized (00:17:42).
  • Taxonomy Approval and Communication Strategy: Pete Bernard suggested that the taxonomy section should be approved by the working group steering committee to gain community buy-in. Jason Shepherd agreed and asked for a meeting to present the high-level approach (00:19:39). They decided against presenting the draft for approval on the upcoming Monday meeting, instead opting to email the taxonomy draft to the working group steering committee alias (“WGSC”) for review and then schedule a follow-up discussion (00:20:28).
  • Inclusion of AI Types in the Document: Rob Woolley asked if the document’s scope should include expanding on different types of AI, such as machine learning versus computer vision (00:22:58). Jason Shepherd confirmed that different “flavors” or implementations of AI, such as voice analytics, computer vision, and generative AI, should be included in the use cases section. They explained this content would describe where these AI types are seen across the continuum and in different markets, like natural language processing on consumer devices for privacy or federated learning for hospitals (00:24:03).

Suggested next steps

  • uncheckedJason Shepherd will package and email the draft focused on the taxonomy to the working group leads (using the WS WGSC alias) on Monday morning for review.
  • uncheckedJason Shepherd will reach out directly to a couple of people and include an ask in the Monday morning email.
  • uncheckedRosina Haberl will send Jason Shepherd the link to the spab directory.
  • uncheckedJason Shepherd will ask Raul to do a primer for the security section.
  • uncheckedRob Woolley will follow up and determine if there is a gap in the definition of the distributed edge.
  • uncheckedJason Shepherd and Pete Bernard will set up a meeting with the working group steering committee for a talk-through session to get approval on the taxonomy.