Applying Aggregation Theory to AI

Applying Aggregation Theory to AI

Original Article on Medium: Applying Aggregation Theory to AI | by Sam Bobo | Sep, 2024 | Medium

Ben Thompson is a renowned self-evolved tech analyst focusing heavily on one critical element of modern business — how the advent of the internet completely disrupted current business paradigms. I’ve personally been following Ben for 10 years now, subscribing to his blog: Stratechery and listening to Podcasts from Exponent and Sharp Tech, of which, I’ve come to trust and respect as an analyst. In June 2015, Ben defined a new framework for technology organizations operating in the paradigm of the internet, dubbed Aggregation Theory. From “Defining Aggregators”

Aggregation Theory describes how platforms (i.e. aggregators) come to dominate the industries in which they compete in a systematic and predictable way. Aggregation Theory should serve as a guidebook for aspiring platform companies, a warning for industries predicated on controlling distribution, and a primer for regulators addressing the inevitable antitrust concerns that are the endgame of Aggregation Theory.

Companies achieve the classification of Aggregator once they meet the following criteria (detailed explanation can be found at the aforementioned link above, all quotes below are derived from that post unless otherwise noted):

  1. The company has a direct Relationship with users
  2. Incur zero marginal cost for serving users as goods are digital (software) with highly replicable bits of 0s and 1s that can scale in finitum world-wide; and
  3. Drive demand-driven Multi-sided Networks with Decreasing Acquisition Costs creating the flywheel of “aggregation”

Once established as one, Aggregators can be classified in one of three levels:

  • Level 1 Aggregators: Supply Acquisition (e.g Netflix) — “Level 1 Aggregators acquire their supply; their market power springs from their relationship with users, but is primarily manifested through superior buying power.”
  • Level 2 Aggregators: Supply Transaction Cost (e.g Uber)— “Level 2 Aggregators do not own their supply; however, they do incur transaction costs in bringing suppliers onto their platform.”
  • Level 3 Aggregators: Zero Supply Cost (e.g Google) — “Level 3 Aggregators do not own their supply and incur no supplier acquisition costs (either in terms of attracting suppliers or on-boarding them).”

In short, an Aggregator’s classified level is inverse to the cost of supply, meaning, as the limit of supply costs approaches zero, the aggregator’s level reaches maximum.

So why am I writing about Aggregation Theory on my blog? I’ve centered my career around Artificial Intelligence and have sought to write more analytical and educational pieces about AI here on my blog. As a loyal reader and follower of Ben, I thought it would be a worthwhile thought exercise to explore the following question: Do AI Aggregators exist?

Definition of AI Organization

In order to frame the analysis, I first need to define an AI Organization. An AI Organization is one which offers a consumer-facing solution (requirement #1 of Aggregation Theory) underpinned by Artificial Intelligence capabilities. For avoidance of ambiguity:

  • AI organizations are not ones which integrate AI into a solution as an additive feature, rather, design a solution with an AI-first mentality, i.e with AI capabilities at the center. If the former was true, nearly all organizations would be classified as an AI Organization, which is not the case.
  • AI Organizations can be existing organizations (e.g Microsoft) but offer compelling AI-native solutions that, when examined in a vacuum, would qualify as an AI Organization.

I am firmly committed to the ideal that Artificial Intelligence extends competitive advantages and does not create net new ones for existing organizations, yet, the emergence of AI-first organizations that disrupt traditional design paradigms could be disruptors in the space and gain unique advantages from lower in the value chain.

Overview of the AI Value Chain and Applying Aggregation Theory

In my piece “Vertical Hyer’scaling’ into AI Dominance” I detailed the components of a vertically integrated hyperscaler centered around AI.

For brevity and illustrative purposes, I will summarize the parts of the value chain below in order from lowest level to highest level:

  • Chips — Customized chips designed and fabricated to accelerate and optimize AI workloads
  • Infrastructure/Servers — the physical bare metal racks, servers, and underlying compute (chips) deployed to power AI model training and inference. This can be physical devices of virtualized as part of Infrastructure as a Service (IaaS)
  • Platforms — data science and machine learning oriented platforms with componentized solutions that allow for MLOps, data management and storage, model fine-tuning, and the like which enable the customization and creation of AI models
  • AI Services and Models — the actual pre-built models trained by organizations and used for inference in customer-facing solutions. This includes SaaS services such as text-to-speech as well as custom LLM models like those found on Hugging Face (could include management via Model as a Service paradigms)
  • AI Solution — the customer-facing solution that utilizes AI models. This could include agentic integrations or simply a solution that harnesses the power of AI model(s). Solutions can also invoke surrounding ecosystems to assume the role as a platform (more on that in a moment)

Assessing Zero Marginal Supply Costs (Requirement #2)

Within each part of the value chain, one must understand the digital supply costs per marginal user:

  • Chips — a physical product, a wafer, that gets fabricated en masse via processing plants (“Fabs”) which require massive up-front investment and highly subject to faults. This section of the value chain would not qualify as an aggregator.
  • Infrastructure — Again, a physical product of bare metal servers and chips, only scalable in physical locations which also have physical constraints of square footage. Not an aggregation play.
  • Platforms — Once the services are designed, developed, and deployed, additional marginal users cost nothing and can scale based on geographic availability.
  • AI Services and Models — AI models, depending on the complexity of the transformation and number of parameters required, can incur a massive up front cost. For example, think about training an LLM or speech recognition engine, however, once deployed, the cost of inference is significantly lower and approaches zero at scale. For the case of LLMs, time will eventually cause the cost curve to approach 0 for each marginal user and should remain part of this analysis.
  • AI Solutions — Pursuant with all software services, AI based solutions are software products delivered over the internet and incur 0 marginal cost per user (notwithstanding inference costs mentioned previously).

Aptitude for Multi-Sided Markets (Requirement #3)

With the above analysis complete on costs, we can eliminate Chips and Infrastructure. Finally, we must now delve into the aptitude to create multi-sided markets.

  • Platforms — Platforms, under a microscope, behave similarly to marketplaces in the sense that the catalog of available services (SaaS based) are a mixture of first party developed services and support for third party offerings as well. Take the IBM Cloud as an example. The IBM Cloud includes managed open-source solutions and partnered solutions for data science workstreams alongside in-house developed services under its purview. For platforms that are a one-stop shop for all ML workloads, platforms have the aptitude to create a multi-sided market through soliciting partnerships from third party vendor integrations.
  • AI Services and Models — Generalized AI models (Conversational AI Services, LLMs, etc) are commoditized. Organizations that have obtained a data-driven advantage and create unique models could leverage Model as a Service (MaaS) offerings. Siloed, Models can be monetized via APIs as building blocks but emerging Model as a Service platforms are eyeing a place to be the single destination for developers and data scientists to the ability to have models available to piecemeal model outputs to form complex AI-powered solutions may seek to form hosting partnerships and monetization agreements to make those newly created models available on a marketplace, creating a multi-sided market.
  • AI Solutions — These AI solutions are offered to the market with a specific set of use cases and intended value provided to end customers. While AI solutions themselves would represent a single sided market between the company and end user, the creation of a broader integration ecosystem, transforming the solution into a platform, could yield multi-sided markets. The primary lead example would be OpenAI’s ChatGPT with custom GPT plugins and functions.

With the above analysis, this begs the question: do any AI aggregators exist today?

Examination of possible AI Aggregators

OpenAI

OpenAI, founder of the GPT family of models, revolutionized the market and set pace for the “era of AI” by popularizing Generative AI capabilities, even though GPT models were around back in 2017. OpenAI’s models, themselves, are commoditized. OpenAI pursued another monetization avenue through employing Reinforcement Learning through Human Feedback (RLHF) and building ChatGPT as a consumer facing chat solution. Furthermore, OpenAI sought to monetize ChatGPT by building in the ability for custom GPTs and I noted in “Platform vs Pipeline — the Difficult Path to Monetization for Foundation Models” that this created a multi-sided market:

Companies who seek to outsource functionality employ a platform play by creating a two-sided market of consumers and developers. This is seen all too common with the Apple iPhone and app store as an easy example. OpenAI sought the platform paradigm through announcing the GPT Store. The GPT store, in effect, allowed people and organizations to create custom “GPTs” or plugins that worked with ChatGPT. These GPTs could range from a simple API passthrough such as Wolfram Alpha for answering mathematical questions or the NewYorkTimes for querying news stories to custom GPTs build by hobbyists and developers looking to make money and building novel GPTs using clever prompting, ranging from resume builders to virtual girl(boy)friends and the like. With most stores, 80% is garbage but the 20% is highly useful and create a significant amount of extensible value. This then, attracts new customers and the originating company taking a rake as a platform fee. The two-sided market approach requires incentivization of one party, typically the developers, to then create a flywheel where more value = more users = more incentive to develop more, etc.

In theory, OpenAI would be classified as a Level 3 aggregator assuming ChatGPT was disjoint from OpenAI GPT itself as there is no purchase of supply (GPTs), simply early incentives to attract suppliers, which is typical in the creation of a multi-sided market. I’d argue, however, that the flywheel has yet to be started with GPTs given the lackluster adoption of the GPT store.

Hyperscalers Broadly

Peering in to the world of Hyperscalers at the likes of Google, Amazon, Microsoft, and IBM, many have sought to create Models as a Service (“MaaS”) marketplaces. These marketplaces have initialized with open source model partnerships such as Llama (Meta) but these hyper scalers are simply adding models on a “marketplace” to amoritize the fixed costs associated with scaling up infrastructure and chips:

Furthermore, hyperscalers are tackling the Open Source market for LLMs, including broad announcements to host LLaMA 2 on Azure or Google cloud? Why would these companies host competitor models, whether proprietary or not, when the hosting company offers one of their own? Simple — two primary reasons (1) cost amortization — (more later in the piece) and (2) switching cost — offering a popular model adds additional value to companies already on the platform and entices others. Why move when there is already a contract in place and/or many workloads already operate on that cloud, its hard to switch.

Until a marketplace materializes for custom models, these platforms will be highly limited.

The Ultimate AI Aggregator Vision / Recipe

There exists the ability for one of the Hyperscaler’s to emerge as an AI aggregator, specifically a Level 3 Aggregator with the following “recipe”

The above verticalized stack creates both a multi-sided market with suppliers, incentivized to supply models and data, and fueled by consumers who arrive at the platform for its robust data science capabilities differentiated by design.

Personally, I believe this company does not exist but could be a recipe for a truly exceptional aggregator for AI workloads. This interpretation of Aggregation Theory for AI Organizations should withstand the litmus test of the readers and Ben Thompson himself but puts into motion an application of AI into Aggregation Theory!

要查看或添加评论,请登录

Sam Bobo的更多文章

  • A LLM Trying to “Catch’em All” Playing Pokémon

    A LLM Trying to “Catch’em All” Playing Pokémon

    Original Post on Medium: A LLM Trying to “Catch’em All” Playing Pokémon | by Sam Bobo | Mar, 2025 | Medium Growing up…

  • Preserving AI for Social Good — How AI can Preserve History

    Preserving AI for Social Good — How AI can Preserve History

    Original publication on Medium: Preserving AI for Social Good — How AI can Preserve History | by Sam Bobo | Feb, 2025 |…

  • Adaptation Artifacts for Tuning Models

    Adaptation Artifacts for Tuning Models

    Originally posted on Medium: Adaptation Artifacts for Tuning Models | by Sam Bobo | Feb, 2025 | Medium Perfect theories…

  • My Journey into AI

    My Journey into AI

    Original Post on Medium: My Journey into AI. A short post detailing my entry into AI… | by Sam Bobo | Feb, 2025 |…

    1 条评论
  • The Opportunities Lost with Human Replacement

    The Opportunities Lost with Human Replacement

    Original post on Medium: The Opportunities Lost with Human Replacement | by Sam Bobo | Jan, 2025 | Medium “The Giver”…

  • Grounding the World Around Us — Inference on the Edge

    Grounding the World Around Us — Inference on the Edge

    Original Post on Medium: Grounding the World Around Us — Inference on the Edge | by Sam Bobo | Jan, 2025 | Medium Take…

  • AI Model Management and Lock-In Potentials

    AI Model Management and Lock-In Potentials

    Original Post in Medium: AI Model Management and Lock-In Potentials | by Sam Bobo | Jan, 2025 | Medium Entering the…

  • Blogging Year in Review 2024

    Blogging Year in Review 2024

    Originally posted on Medium: 2024 Blogging Year in Review. Humbled by my readership, I recap 2024… | by Sam Bobo | Dec,…

  • Standardizing for Commoditization — Anthropic’s Model Context Protocol (MCP)

    Standardizing for Commoditization — Anthropic’s Model Context Protocol (MCP)

    Originally posted on Medium: Standardizing for Commoditization — Anthropic’s Model Context Protocol (MCP) | by Sam Bobo…

    1 条评论
  • A Thesis on Artificial Intelligence

    A Thesis on Artificial Intelligence

    Original post on Medium: A Thesis on Artificial Intelligence | by Sam Bobo | Dec, 2024 | Medium Artificial Intelligence…

    1 条评论

社区洞察

其他会员也浏览了