Zhipu secures 3 billion yuan in financing, and its commercialization revenue is expected to double in 2024 | Exclusive from 36Kr.

Written by Zhou Xinyu

Edited by Su Jianxun

"Intelligent Emergence" exclusively learned that the AI unicorn "Zhipu" has recently completed a new round of financing, with an amount of 3 billion yuan. The new investors include multiple strategic investors and state-owned capital, and old shareholders such as Legend Capital continue to follow the investment.

It is understood that this round of financing will be used for the further research and development of Zhipu's base model, upgrading from answering questions to solving complex, multimodal tasks, to better support the development of the industry ecosystem.

According to "Intelligent Emergence", Since 2024, Zhipu's commercialization income has achieved a growth of more than 100%. "If an enterprise service company wants to achieve an ARR (Annual Recurring Revenue) of several hundred million yuan, it usually takes nearly 10 years." Zhipu COO Zhang Fan once told "Intelligent Emergence", "Last year, we completed this in just a few months."

In the B-end, with a significant reduction in API prices, the annual income of the Zhipu MaaS open platform bigmodel.cn API has increased by more than 30 times year-on-year; in the C-end, since the launch of the paid function in the third quarter of 2024, "Zhipu Qingyan" is expected to have an annual income of more than ten million yuan. Currently, Zhipu Qingyan has more than 25 million users.

Currently, selling model APIs and providing services through MaaS are the main To B business models of model-level enterprises.

The so-called MaaS (Model as a Service) can be understood as a platform that provides model APIs and various development tools to developers.

For example, on the Zhipu MaaS open platform bigmodel.cn, there are model product APIs such as the language model GLM-4 series, the multimodal model GLM-4V series, and the autonomous agent AutoGLM, as well as a model fine-tuning platform that can be deployed in just three steps.

It is understood that The average daily Tokens consumption of bigmodel.cn has increased by 150 times, and the paid Tokens have increased by more than 40 times. Currently, the platform has 700,000 enterprise and developer users, with a year-on-year growth of more than 30 times, and the number of paying customers has increased by more than 20 times.

Nowadays, large model enterprises have also seen the certainty of commercialization in the B-end with clear downstream demands.

However, with the outbreak of the price war, the commercialization of MaaS is beginning to face challenges as the model capabilities of various companies have not yet拉开差距 (This phrase remains untranslated as its meaning is unclear without more context).

From the beginning of 2023 to the second half of 2024, the price of the Zhipu open platform has decreased from 500 yuan per million Tokens to 0.1 yuan per million Tokens.

A senior executive of Zhipu told "Intelligent Emergence": "This is a price reduction brought about by technological progress. For example, a large number of technical optimizations have been carried out in the model architecture and engineering. Even if the price has dropped by 5,000 times, we still do not lose money, and this is not a price war."

The significant growth in users and Tokens is on the one hand due to the price reduction, and on the other hand, the scenarios are becoming more and more complex. The aforementioned senior executive mentioned: "Last year, when everyone used the model, it was mainly for simple conversation applications. This year, it has begun to shift to solving more complex problems and multimodal tasks."

One of the sufficiently complex scenarios is Agent.

Worldwide, a consensus in the AI industry is that 2025 will be the year of the explosion of Agent. Whether it is Zuckerberg, Bill Gates, or Ilya, the former chief scientist of OpenAI, they have all expressed an optimistic attitude towards the application prospects of Agent in public speeches.

In 2024, Zhipu has also made an active layout in Agent. In October, Zhipu conducted an internal test of the Agent assistant AutoGLM and received more than 1 million user visits within one month.

On November 29, at the Agent OpenDay held by Zhipu, the officially released AutoGLM already supports autonomous execution of long-step operations of more than 50 steps, as well as cross-App task execution.

In addition, Zhipu also released the Web-end agent AutoGLM-web and the PC-end agent GLM-PC. The former can autonomously play the last episode of "Alley Family" on Mango TV and send bullet comments to check in; the latter can remotely operate the computer in an "unmanned" manner, similar to the "computer use" function released by Anthropic.

At the same time, these functions will be the first to be implemented in end-side devices with clear consumer demand. At the Agent OpenDay, manufacturers such as Honor, ASUS, Xiaopeng, Qualcomm, and Intel were present to introduce their relevant cooperation with Zhipu.

As the application of AI has been put on the agenda, the monetization ability of the C-end is also the focus of AI companies. Currently, AI C-end applications are still generally in the customer acquisition stage. Except for a few products with first-mover advantages or technical advantages such as ChatGPT and Character.ai, most companies still find it difficult to convert user volume into payment.

But in the C-end, Zhipu still has a remarkable performance: Less than half a year after the launch of the paid service, the expected annual income of Zhipu Qingyan exceeds ten million yuan.

With technology and funds reaching a bottleneck, there are fewer and fewer companies in China that can keep up with the pace of OpenAI - GPT-4o and Sora are a clear watershed.

Zhipu has become one of the few manufacturers whose model products can still be fully benchmarked against OpenAI. Since 2024, Zhipu has successively released the video generation model CogVideoX benchmarked against Sora, and the end-to-end speech models GLM-4-Voice and GLM-4-VideoCall benchmarked against GPT-4o.

These models provide video generation, video calling, and emotional speech functions, all of which are available on Zhipu Qingyan. Qingyan has also become the largest model To C product with the most complete modalities in China at present.

"We believe that value verification is the only way to AGI, that is, users are willing to use it and pay for it." A senior executive of Zhipu summarized to "Intelligent Emergence".

Welcome to follow!

Welcome to communicate!