As AI Giants Focus on Large Model Arms Race, Value May Lie in the 'Small'

While everyone is chasing trillion-parameter large models, I recently posed a question at our family office's investment committee meeting: If the next breakthrough in AI lies not in size, but in smallness, do we need to adjust our asset allocation strategy in advance?

Professor Herzog, an academician of the German National Academy of Science and Engineering and a foreign academician of the Chinese Academy of Engineering, stated clearly in a recent interview that the next major breakthrough in artificial intelligence will definitely not be a single large system, but rather the collaboration of numerous small, specialized agents. This view aligns highly with the investment signals I have observed over the past year.

From an asset allocation perspective, there is a clear divergence in the current valuation logic of the AI sector. Capital expenditure in the large model space has entered an "arms race" phase, with single training costs often reaching tens of millions of dollars, and diminishing marginal returns. In contrast, small agent collaboration networks present entirely different economic characteristics.

Data Comparison: Large Models vs. Small Agent Collaboration

Dimension	Large Models (e.g., GPT-4/Claude)	Small Agent Collaboration Networks
Single Training Cost	$100M-1B (estimated)	<$1M (customized fine-tuning)
Marginal Inference Cost	High (requires massive GPUs)	Low (can be deployed locally)
Vertical Domain Adaptability	Requires extensive fine-tuning	Out-of-the-box ready
Scalability	Linear growth limited by compute power	Exponential growth via network effects
Business Model Certainty	High input, low return (currently)	High ROI (verified scenarios)

Key Data: According to early-stage projects I track, a medical diagnosis-focused agent collaboration network generates 3-5 times higher accuracy improvement per dollar invested compared to general-purpose large model solutions. This is because small agents can be optimized to the extreme for specific tasks (such as image analysis, medical record interpretation), whereas large models must balance hundreds of capabilities.

Investment Logic: Why Small Agent Collaboration Fits Better with the "Risk-Reward Ratio" Principle

From the long-term allocation perspective of a family office, I summarize three logics for investing in the small agent collaboration ecosystem:

1. Capital Efficiency Advantage: No need to burn cash on foundation models; instead, focus on data, scenarios, and feedback loops.

2. Network Effect Moat: When multiple agents form collaboration protocols (such as communication standards, task scheduling), switching costs become extremely high. Early investment in companies at this protocol layer is similar to investing in TCP/IP for the internet.

3. Regulatory Arbitrage Space: Global AI regulation focuses on large models (data security, ethics); small specialized agents pass compliance reviews more easily, facing less resistance to deployment.

[!note] From a valuation perspective, the current secondary market prices AI primarily based on user growth and compute investment of large model companies, ignoring the potential "asymmetric returns" from small agent collaboration networks. Just like investing in mobile internet in 2010, you didn't invest in base stations, but in WhatsApp and Uber.

Competitive Barriers: Who is Building "Agent Collaboration Networks"?

Currently, I see three types of players positioning themselves in this track:

• Platform Companies: Providing agent communication protocols and scheduling frameworks (like AutoGPT, LangChain), but business models are still unclear.
• Vertical Domain Experts: For example, startups focused on logistics scheduling and supply chain optimization, which already have paying customers and gross margins as high as 60-70%.
• Open Source Communities: Such as Hugging Face's Agent libraries, but lacking a commercial loop.

My Judgment: When investing, prioritize companies that possess both "protocol-setting power" and "scenario implementation capabilities." For instance, an agent collaboration platform with exclusive data sources in the medical field has a moat far superior to upper-layer applications built on generic large model APIs.

Open Question: When Large Models Become Infrastructure, Who Captures Excess Returns at the Application Layer?

Over the past year, we have witnessed capital flooding into AI infrastructure (compute, models), but no "killer" products have emerged at the application layer. Professor Herzog's prophecy reminds us: True value may lie not in a single monster, but in countless collaborating "ant" legions.

I have an unresolved question: When large models become as ubiquitous as electricity, can small agent collaboration networks spawn "platform companies" worth hundreds of billions of dollars, just like the early internet? The answer to this question will determine our asset allocation direction for the next 3-5 years.

As an investor, I tend to position ahead in those "small but beautiful" agent collaboration ecosystems while everyone else is looking at large models. After all, assets with the best risk-reward ratios often appear in corners ignored by the majority.

Original link: https://www.ithome.com/0/981/813.htm