Saturday, December 7, 2024

If Generative AI is "Winner Take All," That Dictates Investment Bets

Contestants in the generative artificial intelligence model business are following a “winner takes all” approach to the market, investing at a pace that gets criticized by financial analysts (and rightly so, in some respects) for exceeding what seems immediate and tangible financial payback.


But the “winner takes all” strategy has worked often in the internet era, across many different segments of the market. So we already can predict some outcomes. 


“Wild” levels of spending are going to pay off for at least one and perhaps two providers of GenAI models, in terms of leadership of the market (ecosystems built on them). Some of the contestants will eventually “break even” for their investors, neither gaining or losing equity value. 


But most will fail, losing their investors most to all of the invested capital. That is simply what has happened in winner-take-all markets. 


As we saw early on, even pre-internet, for operating systems, and later saws for search engines, mobile operating systems, e-commerce, social media, ridesharing, peer-to-peer lodging and other app segments, market leadership is highly concentrated with one or sometimes two providers dominating. 


So the companies developing GenAI frontier models, such as OpenAI, Google DeepMind, and Anthropic, are assuming the market  will eventually shake out with a “winner takes all” structure, with market leadership highly concentrated and creating a new ecosystem of value around the leading models, with the hundreds of other would-be contestants relegated to history.


Market/Product Category

Dominant Players

Search Engines

Google

Social Media

Facebook, Instagram, TikTok

E-commerce

Amazon

Mobile Operating Systems

Apple iOS, Google Android

Desktop Operating Systems

Microsoft Windows, macOS

Web Browsers

Google Chrome, Safari

Streaming Video

Netflix, YouTube

Cloud Computing

Amazon AWS, Microsoft Azure, Google Cloud

Online Advertising

Google, Meta (Facebook)

Smartphones

Apple (iPhone)

Productivity Software

Microsoft Office, Google Workspace

Mobile Payment Systems

Apple Pay, Google Pay, PayPal

Digital Maps

Google Maps

Food Delivery Apps

DoorDash, Uber Eats

Ride-Hailing Apps

Uber, Lyft (US), Didi (China)

Video Conferencing

Zoom, Microsoft Teams


By way of contrast, end users of GenAI--especially enterprises--will have to answer for their choices, as they would for any other capital investments made to support their businesses. And that goes for suppliers fo hardware and software incorporating AI functionality. 


GenAI model leadership will undoubtedly shake out into a “winner takes all” pattern. But that also means hardware or software suppliers might “bet on the wrong horse,” as might their customers. That generally means high danger for model suppliers and those firms that are direct parts of those ecosystems. 


Hardware and software participants might face moderate to high impact if they are in the “wrong” ecosystem or platform. End users--whether consumers or businesses--might face low repercussions if their chosen GenAI supplier does not emerge as the eventual market leader (either provider one or two). 


Basic or firm-specific functionality might not be the big problem, as niche solutions might be perfectly viable, especially if a model emerges as providing superior industry-specific functionality. 


Value Chain Segment

Example Participants

Degree of Danger if Supplier/Partner is Not a Winner

Reason for Risk

Foundation Model Providers

OpenAI, Google DeepMind, Anthropic, Meta

High

If the chosen provider loses, end-users face high costs in switching models due to integration dependencies.

Model Fine-Tuners & Adaptation

Hugging Face, Scale AI, Stability AI

High

Fine-tuning investments and adaptations may not be transferrable to new, winning models.

Hardware Providers

Nvidia, AMD, Intel, custom chip startups

Moderate to High

Dependence on specialized chips may limit flexibility; non-dominant players may face R&D cutbacks.

AI Infrastructure Platforms

AWS, Microsoft Azure, Google Cloud

High

Non-winner platforms may lack interoperability, leading to high switching costs and stranded data assets.

Middleware Providers

Databricks, Snowflake, API integration services

Moderate to High

Middleware built for a specific provider may need reconfiguration, reducing compatibility with other models.

Application & Service Developers

Jasper, ChatGPT plugins, custom AI tools

Moderate to High

Apps tightly integrated with a non-winning model may struggle with compatibility and require costly rewrites.

Consulting & Implementation

Deloitte, Accenture, AI-specific consultancies

Moderate

Models selected may lose support, limiting longevity of solutions delivered to clients.

AI Tooling & DevOps

MLOps tools, AI pipeline solutions (e.g., Weights & Biases)

Moderate

DevOps tools tied to a specific model may lose relevance if clients pivot to dominant providers.

Enterprise & Custom Solutions

SAP AI, Oracle AI, Salesforce Einstein

Moderate

Custom solutions based on non-winning models may have compatibility and feature limitations.

Data Providers

Web-scraping firms, domain-specific datasets

Low to Moderate

Dependence on any one model is less critical, though dataset compatibilities may shift.

Consumer-facing Applications

Grammarly, Canva, AI-powered image and video tools

Low to Moderate

User-facing tools may need updates for compatibility but can adapt faster than core infrastructure layers.

End Users & Businesses

Large enterprises, SMEs, individual users

Low to Moderate

Users may face inconvenience, but model-agnostic interfaces could minimize direct switching costs.


Still, longer term, the specific GenAI ecosystem should matter, as network effects should eventually emerge.


No comments:

AI Increases Data Center Energy, Water E-Waste Impact, But Perhaps Only by 10% to 12%

An argument can be made that artificial intelligence operations will consume vast quantities of electricity and water, as well as create lot...