Can AI Make Money? The Potential Business Models Behind AI

[baoshi.rao]

Recently, various large models have been intensively released, with claims like 'catching up to GPT-4' or 'becoming China's OpenAI' scattered across articles. But let’s return to a fundamental question: Can AI actually make money? For those in the industry, this is a soul-searching question. Without the groundwork laid over the past decade, asking this might seem overly profit-driven. But with a decade of losses as context, profitability becomes a question that merges technology and business: it’s both a litmus test for the technology and proof of commercialization capability. Before answering this, we need to summarize the potential business models behind AI.

What are the potential monetization methods if AI reaches a highly mature stage? Historically, proven business models are few, and even fewer align well with AI:

1. Subscription: This is currently the most typical monetization path for AI. Essentially a form of cloud service, major cloud providers naturally integrate their self-developed AI features into their cloud product portfolios.

2. New Value-Added Services: Movies like Her exemplify this category, as do future communicable electronic pets. Compared to subscriptions, these are the 'finished dishes' rather than the 'raw ingredients,' though there’s overlap—similar to the difference between PaaS and SaaS. Traditional SaaS aligns closely with this, so we won’t list it separately (e.g., enterprise-grade assistants).

3. Hardware Sales: This operates similarly to selling computers. Multimodal large models rely on this, as success in new smart products like robots, smart speakers, and AR glasses is crucial for multimodal models. In terms of industry division, this model often overlaps with the first two, driving their growth.

4. New Advertising: Some argue large models will make search ads harder to display. I disagree—screens are large enough to accommodate recommendations like 'Click here to buy...' The key is frequency and precision.

5. Gaming and the Metaverse: While also a product, this differs significantly from the above in supporting a 'virtual central bank' model—issuing proprietary tokens (not necessarily cryptocurrencies). Only these products can sustain an independent ecosystem and monetary system.

Cutting across these models reveals two common traits:

1. AI is a 'deep well' model, updating existing systems (including humans). It creates fewer new models than the internet but impacts existing ones more deeply.

2. AI’s economic value lies in anthropomorphism, performing and surpassing human roles in economic systems (e.g., assistants).

These points are critical because they determine who profits from AI and its ultimate form.

AI’s nature as a supply chain component means companies must evolve beyond their current forms to succeed. For instance, even in gaming or the metaverse, AI firms must become experts in those domains.

This raises two questions:

1. Will AI be a standalone cloud service or part of existing ones? The answer is clear: cloud services consolidate due to economies of scale.

2. Will AI-native or domain-native companies dominate integration? It depends on domain weight—e.g., low in gaming, high in tax/healthcare. Higher domain weight favors domain-native firms.

Earlier analyses focused on commercialization from a technical perspective; this one examines pure business models. The conclusion is simple:

Pure AI firms face narrow monetization paths if stuck in the supply chain. To succeed, they must master both models and domain-specific products.

The likely outcome: Top-tier model firms may merge with cloud giants if they can’t diversify, while mid-tier firms with domain expertise will evolve and win (e.g., publishers excelling in content moderation).

All this hinges on one premise: the technology must create real value. Is it mature enough?

Not yet. Real-world applications reveal gaps, and short-term fixes are unlikely. AI’s '0-to-1' phase remains incomplete. Unlike the internet, where foundational tech (e.g., TCP/IP) stabilized early, AI’s foundations are still evolving, requiring simultaneous development and application.

Comparatively, the internet progressed step-by-step, while AI grows through self-deception—claiming to solve problems prematurely. Large models’ success, ironically, stems from non-AI-native founders.

Despite immaturity, AI reshapes what it touches, like rendering basic illustrations obsolete.

How to gauge maturity? Only through full-scenario testing. AI’s commercialization hinges on anthropomorphism, which requires handling complex, interconnected environments—demanding generalizability, not just specialized benchmarks.

AI risks becoming trapped in technical metrics, a self-deceptive cycle. True value lies in universal applicability, not isolated test-set performance.

This specialized evaluation method of technology essentially creates a fundamental hot-and-cold phenomenon: on one hand, AI seems capable of doing everything and appears incredibly magical, while on the other hand, it often proves impractical in real-world use, leading to unprofitability.

What is the full-scenario coverage approach?

Simply put, take recruitment as an example scenario. Does the technological support enable the creation of a digital employee capable of performing all the functions of a traditional recruiter, such as fulfilling a hiring request from start to finish without human intervention?

If this cannot be achieved, then beyond the first model, none of the subsequent high-value models will be viable.

This is the real challenge.

Upon closer reflection, several emerging fields in the post-internet era each face their own unique hurdles. Looking back to around 2015, three new directions emerged: artificial intelligence, blockchain, and SaaS. Many who felt that the internet had reached its limits but were unwilling to remain idle ventured into these areas. However, AI and SaaS have struggled to turn a profit for a decade, while blockchain, though profitable, has nearly vanished due to other reasons. Now, large language models seem poised to inject new life into all three, bringing them to the brink of a breakthrough. Every time I see charts like the one below, my belief in this grows stronger:

Many are curious about which field will take off first. While it's hard to pinpoint, a basic framework for judgment exists:

Assess how far the path is from technological innovation to commercial value. Midjourney represents a short path, whereas Watson exemplifies a long one. When taking real action, funding, manpower, and this distance must align.