智能交易市场: How Companies Buy and Sell AI

Just a couple of decades ago, discussions around artificial intelligence were largely confined to academic papers and algorithm design. Fast forward to today, and AI has transformed into a full-fledged business. We're seeing cloud providers selling model APIs, consulting firms offering implementation solutions, and even traditional enterprises licensing their internally trained models to competitors. This burgeoning market for intelligence has naturally caught the attention of economists.

Recently, the prestigious economic journal, Journal of Economic Perspectives, published a seminal article titled "The Emerging Market for Intelligence: How Firms Buy and Sell AI." This research attempts to integrate the AI industry into a robust economic analysis framework, addressing a seemingly simple yet profoundly complex question: How exactly do companies buy and sell AI capabilities in practice?

From Tech Race to Tradable Commodity

For years, the AI landscape was dominated by a relentless focus on benchmark performance and the latest model breakthroughs. However, the paper argues that the real commercial revolution is happening at the transactional layer. Increasingly, companies aren't training foundational models from scratch; instead, they're purchasing access and inference capabilities from the market. Concurrently, organizations with significant compute power and proprietary data are packaging their intelligence into standardized products for external sale. This dynamic has given rise to new roles: model wholesalers, niche industry intermediaries, and specialized service providers focusing on data labeling and fine-tuning.

One particularly intriguing phenomenon highlighted in the paper is the vast price disparity for similar AI services. Some are billed per token, others per API call, and some even by subscription seats. While this isn't entirely new to the traditional software industry, AI models introduce unique marginal cost structures that complicate pricing significantly. Training costs are largely sunk, but inference costs are not zero. This economic characteristic inherently grants sellers a scale advantage, while buyers must remain vigilant about potential vendor lock-in.

Practical Implications for Key Stakeholders

If you're a technology decision-maker, this research offers a valuable lens to move beyond short-term questions like "which model should we choose?" and instead focus on long-term procurement strategies and cost structures. For those in the investment community, it provides analytical tools to evaluate AI companies' business models, helping to discern whether an AI startup truly possesses a defensible moat or is merely an intermediary.

• For enterprises (buyers): Clearly define your needs. Avoid overpaying for unnecessary "general intelligence." Understand the significant cost differences between consumption-based APIs and private deployments.
• For vendors (sellers): Be clear about what you're selling—a model, a service, or a specific outcome. Your pricing logic will vary dramatically based on this distinction.
• For industry observers: The maturity of the AI trading market serves as a crucial indicator of the overall health and stability of the entire AI sector.

Key Signals to Watch

The paper points out that the AI trading market exhibits a blend of competition and potential monopolization. Major players control the underlying compute infrastructure, while the application layer remains highly fragmented. This upstream concentration and downstream dispersion pattern bears a striking resemblance to the cloud computing market a decade ago.

Another critical aspect is the evolving status of data. While models can be replicated, data is inherently difficult to duplicate. Many enterprises are, in essence, selling not just models but rather meticulously cleaned and annotated industry-specific datasets. The valuation of such data assets currently lacks universally accepted standards.

When intelligence becomes a tradable commodity, the market demands more than just programmers who can call APIs; it requires mature commercial rules and frameworks.

As someone who has closely followed the AI industry, I believe the emergence of such economic research is a significant signal in itself. AI has definitively moved from the lab into the marketplace, and that market is now beginning to define its own rules. Whether you're an enterprise procuring AI services or a vendor providing AI capabilities, taking the time to digest this paper will provide a much-needed, grounded perspective on the "buying and selling of intelligence."