In July, Zhipu completed a HK$31.4 billion placement. Founder Tang Jie released an internal letter titled “The Big Wave Has Arrived,” announcing a strategic two-year investment in AGI, with no focus on short-term monetization. The funds will be concentrated on four core engines: long-horizon tasks, autonomous agent systems, fully self-supervised training, and extreme security governance. While cutting-edge model technologies are updated monthly, capital markets demand commercial returns—Zhipu now faces these two clocks ticking at different speeds. Historically, the company has demonstrated its ability to rapidly adapt its strategy, from the GLM general-purpose foundation model to breakthroughs in coding. Of the HK$31.4 billion, 55% will be allocated to R&D, requiring Zhipu to strike a balance between technological exploration and capital market oversight.

Article author and source: He Fan Cai Jing

The placement announcement and completion of the placement spanned four days in total, making the internal memo in between inevitably a strategic directive with implied intent.

On July 11, Tang Jie, founder of Zhipu and professor at Tsinghua University, released an internal letter titled "The Big Wave Has Arrived." In this letter, he did not discuss pricing or profitability timelines, but instead outlined the next two years' "Touch High"

("Reach for the Sky") program.

In the letter, Tang Jie first defines Zhipu as "essence, counterintuitive, focused," then explains his assessment of technological trends and outlines Zhipu’s choices for the next two years. He stated: “Over the next two years, we plan to make strategic investments—not pursuing short-term monetization of applications, but directly targeting the next frontier of AGI.”

The letter also stated that this investment will focus on four core engines: Long Horizon Task, Autonomous Agent System, Fully Self-Training, and Extreme Security Governance.

Tang Jie

On July 9, Zhipu announced the signing of a placement agreement with CICC Hong Kong. Under the agreement, Zhipu will place up to 19.78 million new H-shares at a price of HK$1,588 per share. The total proceeds from the placement are expected to amount to approximately HK$31,410.64 million, and the net proceeds (after deducting commissions and estimated expenses) are expected to total approximately HK$31,374.95 million.

On July 13, Zhipu announced the completion of the placement: all 197.8 million new H-shares were issued at HK$1,588 per share, resulting in a total proceeds of approximately HK$31.411 billion, with a net amount of approximately HK$31.375 billion after deducting commissions and expenses.

On one hand, “not pursuing short-term monetization of applications”; on the other, a billion-dollar placement just six months after going public. In an era where technical roadmaps are rewritten monthly yet capital demands rapid business returns, Zhipu is simultaneously navigating two clocks ticking at different speeds.

Looking at past performance, Zhipu has indeed correctly predicted several key questions at critical moments. For example, GLM retained its general training capabilities, MaaS successfully onboarded the first wave of API users, and Zhipu’s work, along with its associated research network at Tsinghua University, preserved reusable technical interfaces that later enabled the shift toward Coding and Agents. But this time, they need to hit all four targets at once.

Two letters in six months, facing two clocks at once

The last internal memo was on January 8, when Zhipu listed on the Hong Kong Stock Exchange, and Tang Jie titled it "Building AGI with the Spirit of Coffee."

It all began with a chance encounter on the first floor of the Hong Kong University of Science and Technology laboratory: during a short visit, Tang Jie met Yang Qiang in a café and mentioned that he was drinking too much coffee and perhaps should quit. Yang Qiang replied, “If doing research could be as addictive as drinking coffee, how could you possibly fail at it?”

This letter begins with “addiction” and then places technology and business in context. He writes, “We have always remained committed to developing AI technology that users genuinely find useful.” Later, when reviewing the period from 2023 to 2024, he directly acknowledges, “We may have made mistakes back then—both technical and business-related.”

Six months later, "The Giant Wave Has Arrived" — technology is now the sole protagonist, and the tone has shifted accordingly.

Tang Jie defined Zhipu as "essential, counterintuitive, and focused," announcing a commitment to "reach higher" at a time when the industry is generally accelerating commercial monetization. He also emphasized that they "do not pursue short-term application monetization."

Interestingly, the "placement agreement" announced by the Hong Kong Exchange on July 9 listed commercialization and scalable implementation, customer demand, business expansion, and ecosystem development as reasons for fundraising, with the proceeds also allocated for research and development, acquisitions, and working capital.

Screenshot of the placement agreement

The differences between the two internal memos and the "placement agreement" do not concern whether Zhipu should commercialize, but rather who should determine the technological direction: current orders or the upper limits of next-generation models.

In January, Tang Jie used developer scale and the company’s disclosed MaaS platform ARR to demonstrate that Coding’s bet was not mere researcher self-indulgence; in July, he requested that the organization temporarily avoid letting short-term monetization decisions limit research potential. Business feedback had not disappeared—it had simply moved from the steering wheel to the dashboard.

In fact, leading model companies are simultaneously facing two clocks ticking at different speeds.

On one side, the "technology clock" is ticking faster: DeepSeek-R1 was released in January 2025, followed by Qwen3 in April of the same year, and GLM-4.5 just three months later in July; since then, significant updates to leading domestic models have occurred on a monthly, rather than annual, basis. On the other side, the "business clock" demands increasingly specific answers: When will contracts be signed? Has inference cost decreased? Can API pricing be raised? When will gross margins cover more R&D expenses?

The former requires bearing the research risk when evidence is insufficient, while the latter requires translating this risk into observable business outcomes.

The internal memo on January 8 already revealed the dilemma of being tested by dual timelines.

On that day, Zhipu went public at an offering price of HK$116.2 per share. Tang Jie did not treat the bell-ringing ceremony as an endpoint, but instead announced the next-generation GLM-5, a new model architecture, more general-purpose reinforcement learning, continual learning, and autonomous evolution, while establishing the X-Lab for frontier exploration. Additionally, the same letter highlighted feedback received on the previous Coding investment through metrics such as developer engagement, platform revenue, and overseas business growth.

At this important capital milestone, Tang Jie’s narrative did not feel like a celebration, but rather the beginning of resetting technical challenges.

The actions in the July internal memo were similar. This time, however, four pathways are being pursued in parallel, with a larger capital scale, and the external requirement has shifted from valuation by pre-IPO investors to ongoing pricing by a public company.

However, leading AI companies have no blueprint to follow.

Internet companies can build their products first and scale infrastructure later; traditional software companies can prioritize releases based on customer needs; but frontier model companies must invest in training clusters, data, and talent before the capabilities or applications even exist.

The two clocks are not negating each other but must run simultaneously. What they truly compete over is not merely the score of a particular generation, but a decision-making mechanism that can continuously correct itself even in the absence of answers.

Don't shut the door before you get the answer.

Zhipu's earliest critical decision was not betting on coding or agents.

On March 18, 2021, Tang Jie and other authors submitted a paper titled "GLM: General Language Model Pretraining with Autoregressive Blank Infilling."

At the time, pre-trained models still exhibited clear distinctions: BERT resembled a student skilled in reading comprehension, GPT acted more like an author continuing text from context, T5 handled conditional generation through an encoder-decoder architecture, while GLM was originally designed to enable a single model to both fill in blanks and continue writing, as well as complete articles according to specific requirements.

GLM uses "autoregressive blank filling" to attempt enabling a single framework to handle natural language understanding, conditional generation, and unconditional generation. This is like equipping the technical foundation with a set of universal outlets rather than guessing in advance which appliance will be the bestseller. Of course, this paper belongs to the Tsinghua research lineage and cannot be directly attributed to Zhipu AI, but it still reflects the early technical orientation of Tang Jie and his team.

From the end of 2021 to 2022, this orientation became evident in company decisions, as reflected in academic research.

The GLM-130B project records show that the team began training a 130-billion-parameter model on May 6, 2022, using 96 servers, each equipped with eight A100 40GB GPUs. Zhipu provided the computing power and participated in the evaluation and demonstration system development.

Source: "GLM-130B: An Open Bilingual Pre-Trained Model"

To avoid disrupting the overall development timeline, Zhipu simultaneously established two small teams: one focused on training GLM-130B and the other on building the MaaS platform.

Under this setup, the same adventure is equipped with two stopwatches. Whether or not this was intentionally designed by management, the two teams have effectively created two feedback loops: one validating the technical limits and another validating market demand.

Tang Jie left a detail in his letter: the two small teams may not have known about each other’s existence at the time. In mid-2022, GLM-130B achieved阶段性 training results, after which the project was publicly released and open-sourced, while bigmodel.cn went live and acquired its first batch of API users.

However, this arrangement did not achieve the desired balance or correct the errors made by Zhipu.

From 2023 to 2024, the "hundred models battle," price wars, and the AI assistant boom occurred simultaneously, and Zhipu sought revenue across MaaS, on-premises deployment, enterprise models, and Agents. By 2024, the company's revenue reached RMB 312 million, with on-premises deployment accounting for RMB 264 million, or 84.5%. At the time, the revenue structure remained heavily weighted toward on-premises deployment, with a more resource-intensive delivery model.

In his listing letter, Tang Jie did not portray this history as a series of correct decisions, but instead acknowledged mistakes in both technology and business.

He wrote, “We may have made mistakes back then—both technical and commercial.” He directly pointed out “short-term gains and short-term hype.” Immediately after, he wrote another sentence: “Then the emergence of DeepSeek woke us up.” At that time, the performance of Zhipu’s models had not met expectations, and Zhipu needed to find a sufficiently narrow breakthrough.

In the end, Coding became that narrow gate.

In September 2022, the joint team from Tsinghua University, Zhipu AI, and Huawei opened the source code, weights, API, and editor plugins of CodeGeeX.

The CodeGeeX paper submitted in 2023 showed that this 13-billion-parameter model was trained on 850 billion tokens up to June 2022, covering 23 programming languages. Additionally, the team expanded the originally Python-focused HumanEval benchmark to include C++, Java, JavaScript, and Go.

The paper reveals that, at the time, the relevant plugin generated approximately 4.7 billion tokens per week for tens of thousands of active users, and 83.4% of respondents believed it improved programming efficiency.

In other words, Coding was not an impulsive decision. Before DeepSeek shifted industry focus, the Zhipu team had already integrated code models, cross-language evaluation, editor plugins, and user feedback onto a single pipeline.

Coding is not a fruit that naturally grows on the GLM tree; it is a side road that had already been partially constructed and later consolidated for heavy traffic. After testing GLM-4.1 in April 2025, Zhipu released GLM-4.5 on July 28, integrating reasoning, coding, and agent capabilities into a single unified model architecture.

Connect these nodes, and Zhipu's past decision-making methods become increasingly clear.

When trends are unclear, maintain general capabilities and task interfaces; narrow your focus once external signals emerge, then use evaluations, user feedback, and revenue to determine whether to increase your investment. But these steps aren’t about always predicting the future correctly—they’re more about the ability to quickly switch answers.

Illustration: He Fan Cai Jing

Ability to swap answers

If the papers from Zhipu, THUDM at Tsinghua University, and their partner institutions were listed in a long list, “technical options” could easily become a numbers game. But what truly matters for company decisions is not how many papers there are, but whether they can be reassembled after trends shift.

Based on publicly available materials, the technological capabilities accumulated by Zhipu and its associated research network resemble those of a factory with foundational capabilities.

GLM, GLM-130B, and subsequent versions like GLM-4.5/5 serve as the engine and production line, addressing general modeling, large-scale training, MoE, and post-training; CodeGeeX, WebGLM, CogAgent, LongWriter, and others function as different workstations, enabling the model to handle code, search, GUI, and long-form outputs; HumanEval-X, AgentBench, and LongBench act as instruments and quality control lines, measuring whether capabilities have truly improved.

Illustration: He Fan Cai Jing

The agent route best explains this "reassembling."

The WebGLM paper submitted in 2023 connected search, retrieval, answer generation, and preference scoring, effectively granting the model a pass to a resource room. In the same year, the CogAgent paper employed high- and low-resolution visual encoders to identify small text, buttons, and page elements on screens, enabling the model to see its workspace clearly. The AutoGLM paper, submitted in October 2024, further linked planning, visual localization, and device operations into an actionable chain.

In other words, WebGLM helps the model find the door, CogAgent enables it to see the doorknob, and only AutoGLM begins attempting to push the door open.

But being able to open a door doesn't mean you can work independently.

In papers related to AutoGLM, a task success rate of 36.2% was reported on the more challenging AndroidLab benchmark. This figure demonstrates that device operation capabilities have emerged, while also bringing to light issues of failure, anomaly recovery, permissions, and responsibility in open environments.

This is precisely the true starting point of the four pathways described by Tang Jie in "The Great Wave Has Arrived"—they do not begin at the same elevation.

Illustration: He Fan Cai Jing

For example, long-range tasks are most easily misled by the illusion created by "million-context."

A long context is like giving the model a bigger desk, allowing it to spread out contracts, code repositories, and lab notes all at once. But a long-range task requires it to step away from the desk, work on the project continuously for days, remember mistakes, invoke tools, observe outcomes, and replan.

According to relevant papers, LongBench measures the ability to read long sequences, LongWriter explains how to write long sequences, and AutoGLM, AgentRL, and GLM-5 are only beginning to address enduring performance. These three types of "long" are interconnected and cannot replace one another.

For example, fully self-supervised training has not yet been closed.

ReST-MCTS* enables the model to generate multiple reasoning trajectories, then selects paths through tree search and process rewards, while GLM-4.5/5 reveals synthetic agent trajectories, self-distillation, and agent reinforcement learning.

The model behaves like a student who can generate some practice problems and try multiple solutions, but the exam subjects, testing environment, and grading criteria are still primarily controlled by humans. Stating that “AI can now fully train itself” would exceed the current evidence. However, ReST-MCTS* still relies on externally provided final correctness criteria.

The answer is based on the premise that self-training has not yet covered goal setting and final validation.

Among the four pathways, the one with the weakest public evidence is precisely the security governance that Tang Jie emphasized the most.

LongSafety demonstrates that the relevant team has researched behavioral safety evaluation in long contexts; Self-Contrast involves model alignment without requiring additional preference feedback; VisionReward involves multidimensional preference modeling in image and video generation.

However, the three approaches address different problems and none of them demonstrate that Zhipu has established continuous interpretability capabilities for the advanced GLM internal mechanisms.

These papers and technical options explain why Zhipu was able to pivot quickly after the trend shifted, but cannot determine where it will turn next.

After 3.14 billion, Touch High needs intermediate scales even more.

The second hands of two clocks will not change their "tick-tock" rhythm for any excuse.

On the evening of July 13, Zhipu announced the completion of this placement. All 197.8 million new H-shares were issued at a price of HK$1,588 per share, raising a total of approximately HK$31.411 billion, with a net amount of approximately HK$31.375 billion after deducting commissions and related expenses.

More important than the funds arriving is that Zhipu has disclosed a more specific allocation plan.

According to the announcement of the placement completion, approximately 55% of the placement proceeds will be allocated to research and development, including advancing the GLM foundational model, investing in cutting-edge areas such as long-horizon tasks, autonomous agents, self-evolution, security mechanisms, and interpretability, as well as expanding the R&D team and deploying computing resources; approximately 15% will be used for business expansion, strategic investments, and acquisitions; the remaining approximately 30% will be utilized for ongoing business operations, debt repayment, capital structure optimization, and other general corporate purposes. The company expects to fully utilize these funds by the end of 2027.

Screenshot of the placement announcement

This means that Touch High is no longer just an internal mobilization slogan. Once a listed company includes its technological direction in the use of proceeds, that direction simultaneously becomes a financial commitment: how many resources have been invested, what stage has been reached, which paths will continue to receive increased investment, and which paths need to be scaled back—all become trackable issues for the capital market.

This is also a necessary change in Zhipu’s next decision-making mechanism.

Startups can initially use technology to make bets and explain the process only after the results emerge; listed companies cannot package their next two years into a black box waiting for a final answer. The HK$3.14 billion purchased extended time for technological exploration, but it did not buy a two-year silence.

The 2025 annual report showed that, as of April 10, 2026, Zhipu had utilized HK$1.828 billion of the IPO proceeds; the announcement of the completion of the placement on July 13 further disclosed that, as of June 30, the cumulative utilization of IPO proceeds amounted to HK$4.588 billion, with HK$308 million remaining, representing a utilization rate of approximately 93.71%. Within 81 days, an additional HK$2.76 billion was utilized.

"Utilization" is not equivalent to "burning," but the speed of capital deployment has become the public backdrop for Zhipu's completion of a new funding round six months after its listing.

In 2025, Zhipu reported an audited revenue of RMB 724 million, a gross profit of RMB 297 million, R&D expenses of RMB 3.18 billion, and an adjusted net loss of RMB 3.182 billion. R&D spending was approximately 4.39 times the revenue, while gross profit accounted for only 9.3% of R&D expenses. Revenue from the open platform and APIs reached RMB 190 million, a year-over-year increase of 292.6%; enterprise-grade Agent revenue reached RMB 166 million, up 248.8% year-over-year. The light platform business grew rapidly, but on-premises deployment still contributed RMB 534 million, accounting for 73.7% of total revenue.

MiniMax provides a peer scale.

Revenue in 2025 was $79.04 million, with 67.2% coming from AI-native products and 73% from outside mainland China; however, gross profit during the same period was $20.08 million, representing only about 7.9% of the $253 million spent on R&D. Even with deeper productization and higher overseas revenue, leading independent model companies still lack a proven self-sustaining business model.

The Dark Side of the Moon offers another exploration path. Investors disclosed that it completed a $500 million funding round by the end of 2025. In May 2026, Bloomberg, citing financial advisors, reported that it raised approximately $2 billion more, with annual recurring revenue (ARR) exceeding $200 million in April.

Coding itself left Zhipu with over 242,000 paying developers and a price increase of at least 30%. Meanwhile, Zhipu’s open platform and API revenue reached RMB 190 million, and enterprise-level Agent revenue reached RMB 166 million. The latter two figures demonstrate growth in the platform and Agent businesses, but cannot be attributed solely to Coding revenue. More critical operational metrics—Coding user retention, ARPU, renewal rate, and unit task cost—have yet to be disclosed.

Touch High also requires intermediate metrics. For example, long-term tasks can disclose task duration, success rate, human takeover rate, and recovery rate; autonomous agents can monitor real-world calls, enterprise deployments, per-task cost, and anomaly recovery; self-training can compare the efficiency of synthetic data, training cost, and capability gains.

Zhipu’s most compelling historical shift was narrowing its focus from multiple exploratory paths to the narrow gate of coding. Now, Touch High is simultaneously opening four doors, with each path differing in maturity, validation cycle, and resource requirements.