Interpreting Zhipu AI's Internal Letter: The Tide Has Arrived, No Monetization After Listing, Betting on the Most 'Cash-Intensive' AGI Path

Author: KK.aWSB

On July 11, Zhipu AI's founder, Tang Jie, sent an internal letter to all employees titled "The Tide Has Arrived."

The letter itself is not long, but the timing is extremely subtle—just a few days before it was sent, Zhipu had experienced a significant stock price fluctuation.

Six months ago, on January 8 of this year, Zhipu debuted on the Hong Kong Stock Exchange at an issue price of HKD 116.2, becoming the "first stock of large models." Since then, with the continuous iteration of its flagship model, the stock price skyrocketed, peaking at HKD 2980—an increase of over 24 times compared to the issue price, with a market capitalization that once exceeded HKD 1.3 trillion, surpassing Xiaomi and approaching three times that of Baidu.

However, just before and after this letter was sent, with the first batch of restricted shares being unlocked, the stock price fell by more than 19%, leading to intense market discussions about whether Zhipu is a bubble and whether its valuation can hold.

Tang Jie did not directly respond to the stock price, the pressure of the unlock, or the valuation controversy. Instead, he chose this moment to convey something that almost all investors did not expect: the listing is not the end; for the next two years, the company will not allocate resources towards short-term monetization but will continue to bet on foundational research directions that have no immediate returns and are extremely cash-intensive.

In this article, I will thoroughly outline the core content of this letter, along with the market background that is crucial for understanding this matter, and finally provide my own independent interpretation.

Part One: What Does This Letter Actually Say?

The letter is divided into five parts, which I will summarize one by one.

1. Who We Are: A "Counterintuitive" Methodology

Tang Jie begins not with products but with Zhipu's underlying methodology—summarized in his own words as three terms: "essence, counterintuitive, focus": only by thinking deeply enough can one make sufficiently unconventional choices; once a choice is made, one must be able to persist long enough.

He provided three examples that connect Zhipu's key decisions over the past twenty years:

In 2006, Zhipu's predecessor team was operating a single-machine academic search system, which seemed unremarkable, but they judged that it was related to the "mining of disciplinary evolution mechanisms," a question worth spending ten years to answer.

From 2021 to 2022, when "making machines think like humans" still seemed like a fantasy to most, the team redirected resources to bet on a model with hundreds of billions of parameters, resulting in GLM-130B—this was a year and a half before ChatGPT exploded globally.

On January 8, 2026, the day Zhipu rang the bell on the Hong Kong Stock Exchange, Tang Jie said the team did not celebrate this as an endpoint but regarded it as a new starting point—while others were celebrating, they chose to "reset" and return to foundational model research.

These three examples point to one attitude: short-term commercial interests and industry trends are, in Tang Jie's view, merely "scenery along the way"; the true endpoint is AGI (Artificial General Intelligence).

2. Understanding This Era: The "Ceiling" of Intelligence is Being Redefined

The second part contains the letter's most critical judgment, which Tang Jie states very plainly: the biggest lesson from the past twenty years is that real business opportunities never lie in minor tweaks to products and business models but in the moment when the "ceiling of intelligent capability" itself is raised.

He describes the evolution of AI capabilities as a clear path—from "perceptual intelligence" (being able to see and hear) to "cognitive intelligence" (being able to understand and reason), ultimately pointing towards AGI. He provides a rather stringent definition of AGI: it is not the intelligence of a single genius but the sum of human wisdom, capable of creating original knowledge at the level of relativity to be considered truly at the top.

On the road to this goal, he believes there are three mountains that must be crossed:

The first mountain is long-term task capability. AI should no longer just be able to "answer questions in seconds" but should be able to plan and execute complex projects that span weeks, months, or even years—like a top security expert tirelessly digging for vulnerabilities in software.

The second mountain is fully autonomous intelligent systems. This upgrades from the previously discussed "one-person company" (where one person with AI can perform the work of an entire company) to a "no-person company"—a group of intelligent agents, each with specialized capabilities, that can collaborate, operate autonomously 24/7, and self-manage. He specifically mentions that challenges previously thought to require a paradigm shift—memory, continuous learning, and self-evaluation—are gradually being overcome.

The third mountain, the most difficult and imaginative one, is self-evolution. AI begins to train AI—writing code, cleaning and synthesizing data, and training itself. Tang Jie believes that while this will consume more computing power, it saves the most precious resources: human energy and time.

He also referenced an important external benchmark—a report from Google's DeepMind, which suggests that even if the capabilities of a single model remain at human levels, if the number of AGI instances can be increased tenfold each year, theoretically, within five years, there could be one hundred million instances that share underlying intelligence, collaborate at a hundred times efficiency, and replicate experiences at nearly zero cost—on a collective level, this amounts to superintelligence (ASI).

His conclusion is that this wave will penetrate the entire technology stack from top to bottom—operating systems may be rewritten as "LLM OS," and even the von Neumann architecture may face challenges; industries such as finance, law, and e-commerce will not escape.

3. Strategic Direction: The "Touch High" Plan, Four Core Engines

After clarifying the judgment of the era, Tang Jie provided Zhipu's specific strategy for the next two years, dubbed the "Touch High" plan—focusing not on short-term application monetization but directly aiming for the next high ground of AGI, with significant strategic investments in four core directions:

The first engine: long-term tasks. Developing a new generation of memory architecture that allows models to learn, act, and accumulate knowledge throughout the entire lifecycle of a project. He provided a vivid example—breaking down a grand goal like "designing a new cancer drug molecule" into thousands of independently executable sub-tasks.

The second engine: autonomous intelligent systems. Upgrading from "intelligent assistants" to "digital employees," the goal is to build a society of hundreds or thousands of intelligent agents, each with independent professional personalities and skills, capable of debating, collaborating, reviewing each other's code, and allocating resources autonomously.

The third engine: complete self-training. In the context of high-quality human data nearing depletion, transforming computing power into fuel for evolution—building synthetic data factories that generate new knowledge from zero through adversarial interactions between AI and AI, and enabling systems to reconstruct their own code in a safe sandbox.

The fourth engine, which Tang Jie particularly emphasized and elaborated on in the letter, is extreme safety governance. His principle is that the stronger the AI capability, the tighter the safety constraints must be. Zhipu rejects "patchwork" safety solutions and hopes to directly encode human ethics, social norms, and national laws into the model's value function as axioms. He revealed plans to invest hundreds of billions in research on "mechanical interpretability"—in simple terms, understanding the neural logic behind every decision made by the model, transforming black-box systems into transparent, interpretable systems. He also stated that he would actively participate in international AI governance to prevent the misuse of technology.

In this section, Tang Jie specifically mentioned a judgment: when leading overseas models are delayed in public release due to safety concerns, and when leaders of those companies publicly warn that the profound impact of AI will reshape the global power landscape, the development of superintelligence and the research on "super alignment" must proceed in tandem—safety is no longer an optional component of technology but a prerequisite for whether technology can be allowed to exist.

4. Open Ecosystem: Touching High While Paving the Way

In the fourth part, Tang Jie discusses the open-source strategy and clearly aligns it with the "Touch High" plan as two sides of the same strategic coin.

He provided the reasoning: true safety is not built through technical closure and barriers but through extensive participation, sharing, co-construction, and public oversight.

In terms of products, this translates to Zhipu's recent release of GLM-5.2—the company's most powerful open-source model, supporting truly usable million-level contexts, continuing to lead in long-term task capabilities, and being fully open-sourced under the most permissive MIT license: anyone can download, deploy, and commercialize it, regardless of the type or nature of the user organization.

His logical statement is: cutting-edge intelligence should not belong to a few people, nor should it be subject to the arbitrary revocation of permissions by a few rule-makers. It should be open, usable, and buildable, serving every developer. One hand reaches up to touch high, challenging the limits of intelligence; the other hand paves the way down, making the most advanced capabilities as open and inclusive as possible—the heights reached belong to all humanity, and the roads built belong to everyone.

5. Conclusion: Turning Mountains into Roads

At the end of the letter, Tang Jie addresses a question that almost everyone would ask: why, after going public, should core resources continue to be invested in the most uncertain directions?

His answer is a belief: "Those who truly reach the top will turn mountains into roads." He recalls Zhipu's early participation in the Wudao large model project, where that cognition coalesced into a shared belief among hundreds of scientists, which later transformed through Zhipu's industrial investments and ecosystem into a foundation that the next generation of entrepreneurs can leverage to take off. He hopes to build this road higher and wider today—high enough to maintain safety boundaries, and wide enough to allow humanity to explore more unknowns; wide enough for every developer and every team to find a path upward.

At the end of the letter, he uses a weighty phrase as a summary: "Not reaching the top is failure." He emphasizes that this pursuit of height belongs to all humanity.

Part Two: Background Not Mentioned in the Letter but You Must Know

If one only looks at the letter itself, it is easy to regard it as a purely inspiring strategic declaration. However, placing it back in the real market environment reveals more layers.

First, the timing of this letter closely follows a real stock price crisis. Zhipu's stock price rose from the issue price of HKD 116.2 to a peak of HKD 2980, with a 24-fold increase reflecting the market's extreme optimism about the story of "the next OpenAI." However, after the unlocking of restricted shares, the price fell by more than 19%, indicating that this optimism was being tested by reality at that moment.

Second, the strategic direction of this letter did not emerge from thin air. Tang Jie's bet on coding capabilities started after the release of DeepSeek R1 in early 2025, when he judged that "the exploration of the dialogue paradigm has basically peaked," and thus redirected resources towards programming and reasoning capabilities—this decision led to a 60-fold increase in the annual recurring revenue (ARR) of Zhipu's MaaS platform over the past year. GLM-5.2 has already entered the top three in international authoritative evaluation rankings. These specific commercial achievements are precisely the confidence behind this letter's declaration of "not pursuing short-term commercialization."

Third, in a horizontal comparison, one can see that this is not a choice unique to Zhipu. Around the same time, OpenAI continued to launch various agent products, Anthropic focused on programming capabilities like Claude Code, Google advanced the Gemini Agent, and Meta laid out personal AI assistants comprehensively. Leading AI companies globally are almost unanimously treating "autonomous intelligent agents" as the next battleground—the concept of "no-person company" (NPC) proposed by Tang Jie essentially competes for the same thing: the only interaction entry point between humans and the digital world in the future.

Part Three: My Interpretation

Looking at all this information together, I believe the real question this letter seeks to address is not "what Zhipu is doing," but a sharper question: after going public and obtaining real capital, why continue to invest massive resources in directions that show no immediate commercial returns?

This is essentially a redefinition of the "valuation anchor point."

Tang Jie provides the answer, and the core logic is actually a very simple principle in the financial market: how much a high-growth tech company is worth today is never determined by how much profit it made this quarter, but by whether the capital market believes it can become the next platform-level company of the era.

If the market believes this story, then all the massive R&D investments today can be reclassified as "reasonable costs for the future"; if the market does not believe it, then even if current revenues are impressive, they may not support the valuation of a trillion-dollar company.

This letter is essentially an attempt to persuade the market again during the window of unlock pressure and stock price volatility: please continue to believe in this grand narrative about AGI, rather than viewing us through the short-term lens of quarterly financial reports.

One point worth pondering is the tension: are "Touch High" and "Openness" truly not contradictory?

The letter places "Touch High" (conquering cutting-edge technology) and "Openness" (making the strongest current models open-source for free) as two sides of the same coin, reasoning that "true safety cannot be built through closure."

This logic holds, but it is worth thinking one layer deeper: while Zhipu claims that cutting-edge intelligence "should not be monopolized by a few people, nor should it be subject to the arbitrary revocation of permissions by a few rule-makers," it is simultaneously racing towards goals that, if achieved—such as "complete self-training" and "autonomous intelligent agent society"—would essentially make Zhipu itself the few who possess the strongest technology and the greatest ability to define rules.

I believe a more realistic interpretation is that "Touch High" and "Openness" are indeed two sides of the same coin, but these two sides correspond to two different monetization methods, rather than pure idealism. By open-sourcing the current generation of models, Zhipu gains a developer ecosystem, technical reputation, and market share—this is a "using the present to exchange for the future" strategy: the more you use it, the more you cannot do without this ecosystem. When Zhipu truly establishes a competitive advantage in frontier areas such as long-term tasks and autonomous intelligent agents, this batch of ecosystem users will become its most natural commercialization foundation.

The real target of this letter is not the stock price, but the "credibility of the narrative."

I believe the key to understanding this letter is that it is not written for ordinary users; it is essentially a "confirmation of belief" letter directed at the capital market and the core team.

To the capital market, it says: short-term stock price fluctuations are unimportant; what matters is our judgment on the path to AGI and our historical record of every "counterintuitive" choice being validated—ten years of dormancy in 2006, a year and a half ahead of ChatGPT in 2021, and the 60-fold growth brought by the bet on coding after DeepSeek—this track record is the real collateral behind this promise of "continuing to burn money."

To the internal team, it conveys: going public is not the end; the company's self-positioning is not "a company that makes good products," but "a builder of the next generation of intelligent infrastructure"—this is a grander narrative framework that can also retain top talent.

An unresolved question remains.

This letter speaks grandly and with a sense of belief, but it does not, and cannot, answer several real questions that determine success or failure: can the technological lead be sustained? Can massive R&D investments ultimately translate into a stable, sustainable commercial loop? Can the open ecosystem and self-developed chip layout truly form a competitive barrier that others cannot replicate?

These questions are ones Tang Jie chose not to address in the letter, not because they are unimportant, but because at this stage, "clarifying a sufficiently credible direction" is more effective for stabilizing market confidence than "precisely answering every realistic detail." As for whether this direction can ultimately be realized, it will require at least two years of time to answer with real technological breakthroughs and commercialization data, rather than relying on this letter.

In Conclusion

This over two-thousand-word internal letter superficially discusses AGI, long-term tasks, intelligent agents, and safety governance, but its real battlefield is actually "how to persuade everyone to believe in a sufficiently grand future amid a round of stock price fluctuations."

This is itself the most typical scene of this era: when a company's valuation anchor point shifts from "how much money it makes today" to "whether it can define an entire era in the future," the most important competitive edge, besides technology itself, is the ability to tell this story and the credit record of past "counterintuitive" choices being validated.

This time, Zhipu is betting not only on the technological direction but also on whether this narrative can hold until the next real technological breakthrough arrives.

Tang Jie sent this letter with a core message: do not pursue short-term monetization, continue to invest in AGI.

Four directions:

• Long-term tasks—AI evolves from "answering questions in seconds" to being able to work continuously for months.

• Autonomous intelligent agents—from "one-person companies" to "no-person companies," with hundreds or thousands of AIs collaborating and dividing labor.

• Self-training—AI writes code, generates data, and trains itself.

• Extreme safety governance—investing hundreds of billions in "interpretability" to make black-box models transparent.