Who Actually Made Money in Large Model Entrepreneurship?
-
Years ago, when asked if autonomous driving was profitable, a CEO replied:
It's like asking if elementary school students make money - it might force them to work at McDonald's.
Now, many want to know if large model entrepreneurship is profitable, and the answer might be:
It's like fresh college graduates working jobs that barely support themselves.
Large model entrepreneurship appears to be thriving, but turning a profit is difficult. A CEO from a large model startup attended an offline salon where over ten companies were present—only two had any revenue. Factoring in R&D costs, none were profitable.
Investors have high expectations. Zhu Xiaohu, Managing Partner at GSR Ventures, stated that AI startups in China must identify viable application scenarios and be profitable from day one.
By this standard, most entrepreneurs fall short.
Some listed internet giants have engaged in "reverse marketing." For instance, 360 claimed its large model products generated 20 million yuan in related revenue; SenseTime reported a 670% increase in generative AI-related income; and Meitu announced that large models drove a 3.2-fold profit growth.
These companies are in a hurry to tell the outside world that they are making money, using various modified statements. But upon closer analysis, you'll find that the business is still the same as before—just with a new name and a fancy label.
An investor bluntly stated that they are still trying to figure out which companies will turn the promise of artificial intelligence into long-term profits. Using a hockey analogy, they said, "The puck in the middle isn't under control, and no one knows where it's going."
How do large models make money? This is a core question. Only by understanding this can we see where entrepreneurs are heading and how capital is flowing.
We break this question down into four smaller ones—Whose money is being made? How is the money made? Who is making the money? And how long can it last?
Here’s the main content.
Who pays the bill?
From the perspective of the ultimate payer, the business models of large AI models can be divided into two categories—to C (consumer) and to B (business). Strictly speaking, there is also to G (government), which is included here under to B.
In the tech and internet industry, to C is a lucrative business with significant marginal effects. Well-known products like WeChat, Didi Chuxing, Meituan Waimai, and Douyin (TikTok) are all to C products. Developing such hit applications is the dream of many entrepreneurs.
Are there any to C hit applications in the large model industry?
Yes, ChatGPT.
In late November last year, US AI startup OpenAI launched ChatGPT, which gained 100 million monthly active users within two months, astonishing the entire tech industry. Four months later, monthly active users surpassed 1 billion, making it the fastest-growing website in history.
Charging these users subscription fees has proven to be a lucrative business—OpenAI introduced the paid subscription version ChatGPT Plus in early February (when monthly active users had just exceeded 100 million), priced at $20 per month.
Subsequently, a large number of similar products emerged in the US, primarily targeting consumer use with subscription-based payment models.
According to data from app store monitoring platform Sensor Tower, AI app downloads in the first half of 2023 increased by 114% year-over-year, exceeding 300 million, surpassing the total for all of 2022. Additionally, in-app purchase revenue for AI apps surged by 175% year-over-year, approaching $400 million—while individual user payments may be small, with enough users, the business becomes substantial.
Most general AI model products launched by domestic manufacturers are currently free. The first major company to break this trend was Baidu, which introduced a paid professional version of Wenxin Yiyan on November 1st, priced at 59.9 yuan per month.
Unlike subscription models, another approach involves providing free products to end-users (C-side) while charging business clients (B-side, such as advertisers)—a model often described as "the sheep's wool is paid for by the pig while the dog foots the bill." This is a common practice in the internet industry, though few companies currently have the capability to implement it effectively.
The B2B market is vast but fragmented. The most straightforward way for large AI models to monetize is by offering API access.
Back in the summer of 2020, OpenAI launched GPT-3. In January the following year, a company called Jasper was founded. By integrating the GPT-3 model and fine-tuning it for marketing scenarios, Jasper automated the generation of various marketing copy styles and earned $30 million in its first year. It only needed to pay OpenAI for API usage rights.
Therefore, OpenAI's earliest revenue actually came from the B2B sector. Companies like Jasper, which need to call APIs from foundational large models—essentially a large number of 'developers'—are clients of OpenAI and other large model companies.
Zaowu Cloud is a startup specializing in AI design solutions. During the development of their proprietary system, they rely on external foundational large model APIs, including GPT-4, Baichuan-13B, and ChatGLM2-6B, paying based on usage volume.
When SenseTime launched its 'Rixin' large model earlier this year, the product was not made available to C-end consumers, nor did it offer beta testing opportunities like tech giants such as Baidu and Alibaba. Instead, it directly opened API interfaces targeting government and enterprise clients.
Another approach to B2B monetization is the SaaS model.
AI vendors leverage large models' capabilities to provide solutions, revamp systems, and streamline processes for enterprises, ultimately achieving cost reduction and efficiency gains—naturally, for a fee. This parallels the fervor around industrial internet and enterprise digitization in recent years.
Qiu Yiwu, founder of Zaowu Cloud, shared an example with Dingjiao: They assisted an e-cigarette brand with product design. Traditionally, spending 1 million yuan would yield 100+ design proposals from a conventional firm. Now, using AI large models, 800 design proposals were generated at a computational cost of just 10 yuan.
Similar logic applies to enterprises seeking AI replacements for roles like sales, customer service, and financial advisors—they’re willing to pay for it.
How to monetize?
After identifying who pays, the next question is: How to actually secure the revenue?
In the C-end market, money is made through applications.
In this wave of large models, the first company to make money in the C-end market was Jasper, as mentioned earlier.
Jasper's business is built on OpenAI's platform. It seized a timing advantage—being among the first to participate in the small-scale internal testing of GPT-3, gaining access to the API, and launching its product before ChatGPT.
Copywriting is a market with clear demand. As long as AI-generated copy performs better than humans, people will pay for it. At one point, over three-quarters of Jasper's users paid $80 or more per month to access various writing template kits. In 2021, its revenue even surpassed that of OpenAI, which provided the underlying technology.
This has inspired the industry. In the U.S., there are numerous startups that utilize large model APIs to develop new applications, with the most popular being AI chatbots and AI image generation products like Midjourney, accounting for 49% and 31% of app store downloads, respectively.
In the U.S., a situation has emerged where foundational large models struggle to generate revenue, while upper-layer applications easily make money. A survey in June this year of the top 50 AIGC websites by global monthly traffic revealed that 90% of the listed applications generate income, with almost all companies adopting a subscription model.
However, this path has not yet fully materialized in China.
The most fiercely contested battlefield in China is the foundational large models, with the 'Hundred Models War' focusing on universal large models rather than applications.
The 9.9 yuan-priced Miaoya Camera briefly gained popularity in July. Wenxin Yiyan, launched in March, only started charging for subscriptions in November. According to Sensor Tower data, in the first half of 2023, the US market contributed 55% of the total revenue from AI applications, the European market accounted for 20%, and other markets, including China, made up only 25% combined.
There are many reasons for this, such as China's late start in foundational large models and the immature conditions for application-layer development; Chinese consumers are less willing to pay for AI applications; and the uncontrollable nature of AI-generated content inevitably faces regulatory challenges—China only opened the first batch of large model filings in early September, before which they could only be tested internally.
Zhang Peng, CEO of Zhipu AI, said that given the differences in market environments between China and the US, the opportunities for large model companies lie in vertical applications for enterprises.
For the B2B market, the most lucrative opportunity is in developing industry-specific large models.
Using large models to enable intelligent transformation in retail, finance, manufacturing and other industries has become a widely accepted approach among Chinese enterprises. Latecomers in large model development like Tencent, Huawei, and JD are all vigorously promoting industry-specific large models.
This is based on a consensus: Industry-specific models fine-tuned from general large models using domain-specific data perform better in particular fields.
Chinese internet giants are building foundational large models while deploying industry-specific versions to capture vertical markets. For example, after releasing its Pangu model, Huawei quickly rolled out specialized versions for finance, manufacturing, mining, meteorology and other sectors to achieve broad coverage.
Tech companies capable of developing general large models prefer integrating them with their cloud services to penetrate industries and create diversified revenue streams.
Cloud providers like Baidu, Tencent, Alibaba, and Huawei are hosting multiple large models (both self-developed and third-party) on their cloud platforms, then packaging models, computing power, and tools as AI development platforms for external services.
They operate like shopping mall owners, preparing the infrastructure—such as space, utilities, and equipment—for businesses (developers and enterprises) to set up shop, offering services and charging fees. At the same time, they also run their own stores.
For example, Baidu's Wenxin Qianfan large model platform allows enterprises to select foundational models, utilize various tools, perform cloud-based inference, fine-tuning, and hosting, and create their own large models for customized product development. This approach binds customers more effectively than simply offering API calls.
To build influence and attract customers, some vendors open-source their large models while commercializing closed-source versions. Examples include Baichuan Intelligence, Zhipu AI, and Alibaba.
Baichuan Intelligence initially released several open-source large models, free for commercial use. After gaining attention, it launched two closed-source models with larger parameters and stronger performance, offering API interfaces to B2B clients for monetization.
This strategy resembles the free-sample model in cosmetics: free trial versions lead to paid commercial products. "Additionally, if the formula is disclosed, manufacturers wanting to create new products based on it would need to pay licensing fees," said Chen Ran, founder and CEO of OpenCSG, an AI company.
Who's making money?
Large model companies are eager to demonstrate their profitability to the outside world, but in reality, few have successfully turned a profit.
According to OpenAI's latest disclosures, ChatGPT now boasts 100 million weekly active users, with 2 million developers utilizing OpenAI's API. Additionally, 92% of Fortune 500 companies are using OpenAI's products to build services. By targeting both B2B and B2C markets, OpenAI is projected to generate over $1.3 billion in revenue this year—a massive leap from last year's tens of millions. However, despite this growth, OpenAI remains unprofitable due to high R&D and computing costs.
Fortunately, OpenAI continues to reduce costs through technological advancements. The GPT-3.5 Turbo model, released on March 1, offers API pricing 10 times cheaper than GPT-3.5. In August, OpenAI further lowered costs by improving API call speeds. The newly launched GPT-4 Turbo is priced at more than 2.75 times lower than GPT-4.
Many companies are learning from OpenAI. OpenAI's rival Anthropic has launched the paid version Claude Pro, charging $20 per month (same as ChatGPT Plus); Baidu's release of the paid version of Ernie Bot also aims to monetize in the consumer market.
Consumer monetization requires scale. The high cost of underlying computing power necessitates massive product adoption. In China, there hasn't yet emerged a truly viral consumer application. This means the mobile internet profit model - earning attention from consumers and advertising revenue from businesses - isn't yet viable.
Compared to AI chatbots, office software currently represents the clearest monetization path and the most heavily invested scenario by major tech companies.
Microsoft, which has invested over $10 billion in OpenAI, has integrated ChatGPT features into workplace collaboration software Teams, Bing search engine, Edge browser, and Office suite Copilot, creating an all-in-one AI-powered office ecosystem.
This presents new revenue opportunities for Microsoft. ChatGPT-powered Teams charges $7 per month, while Office 365 Copilot costs $30 monthly, with 1 million users already paying for AI-enhanced Copilot features. The business version launched in November. Market expectations suggest Microsoft's revenue will see significant growth.
DingTalk has rapidly followed suit in China. Within Alibaba Group, DingTalk became the earliest implementation scenario for the Tongyi Qianwen model, integrating AI into group chats, documents, video conferences, and app development. DingTalk then began charging enterprises, with various plans adding tens of thousands of yuan to existing annual fees.
Kingsoft Office also announced WPS now incorporates large models from Baidu, Zhipu AI, and Minimax via API calls, offering features like text summarization, expansion, rewriting, and automatic document generation. Currently free, these features may become paid services next year.
These products aren't entirely new innovations but rather upgrades to existing offerings. DingTalk resembles Microsoft Teams, and WPS mirrors Office - both are embedding large model capabilities into their existing product lines to enhance monetization. As Qiu Yiwu describes it, this wave of AI models essentially represents an engine replacement, with built-in AI empowering various business functions.
In China, whether for individuals or enterprises, the willingness to pay for a product heavily depends on its stickiness.
Min Kerui, CEO of MetaTech, believes that many companies are reluctant to pay tens of thousands of yuan annually for software that merely offers management functions. Instead, they prioritize whether the software can deliver measurable new user growth.
Therefore, whether domestic enterprises will accept pricing models like DingTalk's remains to be seen over time.
An entrepreneur in the large model field told DingJiao that current B2B large model products are not yet standardized and can easily become advanced human outsourcing. If they are too standardized, they lose flexibility. At this stage, only well-funded medium and large enterprises eager to embrace new technologies are willing to pay for such products.
How Long Can the Profits Last?
Discussing profitability in the early stages of an industry boom might be a luxury, as the rules of the game can change overnight.
Jasper was once a sensation in the market, with profitability that made the industry envious. In October last year, it secured $125 million in funding, reaching a valuation of $1.5 billion.
A month later, OpenAI launched ChatGPT, which was free to use and delivered stunning results. This left Jasper in an awkward position, rapidly diluting its value. Zhu Xiaohu once remarked that Jasper might soon be worth nothing, unable to hold its ground.
When OpenAI steps into application development, companies using its APIs to build products may face rapid replacement if their offerings are similar. Recently, OpenAI held its first developer conference, announcing GPTs and Assistants API, effectively replacing much of the work previously done by developers.
Li Di, CEO of Xiaoice, believes that large model API companies pose a significant threat to startups, as they naturally extend their reach into downstream application layers, creating competition with their 'customers.'
Qiu Yiwu also recognized early on that entrepreneurship in the AI industry would eventually face competition and threats from tech giants. Whether developing AI tools for consumers or services for businesses, it's challenging to build barriers. "Tools are easily replicated, and in the end, you can only become a part of the ecosystem of these giants," he said.
Zaowuyun has already secured clients like Starbucks, Haier, and Supor, generating stable revenue. However, he believes that to sustain this business long-term, it must evolve into a platform. "Connecting upstream developers with downstream clients is the only way to build a moat."
Many domestic large model manufacturers see industry-specific large models as a breakthrough for commercialization, while most companies in the industry remain in a wait-and-see mode, especially small and medium-sized enterprises, which are cautious about paying for such services.
An employee from a system integration company told "Dingjiao" that they had early access to Baidu's Qianfan large model platform, which allows them to freely utilize the platform's large model capabilities or develop and deploy their own industry-specific models. "If it required payment from the start, we might not have used it, as there are many open-source alternatives available now."
For large model companies to make money, the fundamental goal is to deliver incremental value to businesses in the industry. Whether it's reorganizing key business processes using AGI's reasoning capabilities or reshaping product forms and human-computer interactions, the ultimate objective is to enhance productivity. Only when businesses start making money can large model companies benefit as well.
All of this hinges on the premise that the technology is mature and stable enough—which it currently isn't. Large model companies have yet to truly integrate into industries, and there remains a gap between the technology and the actual needs of enterprise applications. "It's like a fresh college graduate with strong foundational skills but lacking professional expertise—still in the internship phase, not yet formally hired," commented Qiu Yiwu.
He gave an example where some vendors present project proposals to enterprises with stunning case studies in their PPTs, but the actual implementation often involves significant randomness. For instance, when AI generates a product display image, the 10 images showcased in the PPT might be selected from 100 attempts. "This is like an engine that hasn't been finalized—its output is unstable."
Even so, companies are actively competing for clients. On one hand, they aim to secure their position in the emerging market; on the other, they need to learn industry know-how from corporate clients to refine their model capabilities.
Overall, large model technology is evolving rapidly, with business models and industry competition still in flux. However, the commercialization process has already begun—some companies are leading the way, while others are just getting started.