As July approaches, Zhipu stands on the brink of a pivotal moment.

On July 8th, marking six months since Zhipu's IPO, 25.68 million restricted shares will be unlocked. This means that the number of new shares flooding the market overnight will be one-and-a-half times the current free float. An even larger block is yet to come, as shares held by employees and early investors, accounting for 40% of the total capital, will also be unlocked next January.

Anticipating this, Zhipu has already made its move: On June 1st, it announced plans to return to the A-share market and list on the STAR Market, aiming to raise no more than 15 billion yuan. Simply put, to sustain its development and internal "blood-making" mechanism, Zhipu needs to raise more funds from the market to fuel its progress.

So far, this is just a typical capital market story. But honestly, it's not the most compelling part. The real question is: Among so many large-scale AI companies, why has Zhipu been thrust into the spotlight?

Companies that bask in the brightest spotlight often possess two key strengths: the ability to deliver tangible results and the skill to weave those results into a compelling narrative that resonates with public sentiment. The former determines whether they deserve a place on the stage, while the latter dictates when they become the sole protagonists.

Zhipu's uniqueness lies in its mastery of both these strengths, which seamlessly complement each other.

However, it must be acknowledged that both these characteristics bear a striking resemblance to Zhipu's "benchmark" across the ocean.

—Introduction

01

Strikingly Similar Narratives

To understand Zhipu, one must first examine its "benchmark"—Anthropic.

Zhipu has openly acknowledged this benchmarking, even during its earnings calls.

An interesting detail: A year ago, Zhipu referred to itself as the "Chinese OpenAI." However, after releasing its first earnings report post-IPO, it shifted its rhetoric to become the "Chinese Anthropic." The superficial reason is a shift in business model benchmarking; but at its core, it's the desire to craft a more compelling story.

Speaking of Anthropic, it is the world's highest-valued AI company, with a valuation of $965 billion. Its large models, Mythos and Fable, are currently unparalleled in terms of power. But Anthropic is also known for something else within the industry—its ability to turn "security anxieties" into marketing hype and, undoubtedly, its flair for drama.

Its theatrics are non-stop.

In April of this year, Anthropic released a preview version of its next-generation model, Claude Mythos, but announced that it would not be publicly released. The reason? It was too powerful and too dangerous. Being so capable that it dares not let you use it—this is an excellent advertising slogan in commercial communication, not a disclaimer.

In early June, Anthropic dropped another heavyweight report, calling on global peers to "slow down or even pause cutting-edge AI research and development." However, this report was packed with self-serving propaganda: it claimed that over 80% of its code was written by its own Claude large model, that engineer productivity had increased eightfold, and that it had outperformed humans by 52 times in certain optimization tasks, and so on.

A manifesto urging the industry to slow down, yet proving "I'm the fastest" throughout. No wonder critics directly call it "a PPT designed to raise funds." The reason for this accusation is that facts rarely align so perfectly—just days before the long essay was released, Anthropic secretly submitted its prospectus.

On June 11th, its performance reached a new climax—its founder, Dario Amodei, released another long essay, asking the U.S. government to "regulate AI using the Federal Aviation Administration's (FAA) rule template and mandate third-party testing for models exceeding a certain computational threshold."

As a result, this time, it got what it asked for. The next day, the U.S. Department of Commerce locked its two most powerful models, Fable 5 and Mythos 5, in the same drawer as high-end chips.

But interestingly, many observers speculate that even this incident, including the ban, was scripted by Anthropic, because it repeatedly does one thing—converting every statement of "I'm too powerful" and every call for "regulation" into publicity, which is then converted into valuation—this is its favorite script.

To be honest, there are several AI companies globally with the strength to compete with Anthropic, but when it comes to capturing public sentiment and storytelling, Anthropic stands alone.

But interestingly, every punchline Anthropic throws on stage is steadily caught by Zhipu offstage, without fail.

In April, after Anthropic claimed that Mythos was "too advanced to show," the global AI community, especially developers, fell into a collective anxiety—everyone worried that frontier models would become "heard but untouchable" in the future.

As a result, only Zhipu seized this emotional window—because the very next day, Zhipu unveiled GLM-5.1, positioning it as the world's strongest open-source model, and it was open-sourced completely. This move not only reassured global developers but also elevated Zhipu's status, with immediate effects—its stock closed up over 11% that day.

In June, Zhipu's script reached an epic level—Anthropic's models were officially locked down by its own country, making the long-discussed topic in the AI industry—"closed-source models could be revoked at any time"—a reality for the first time. Yet again, the very next day, Zhipu "coincidentally" announced the full release of GLM-5.2, with million-level context and open-sourced under the most permissive MIT license.

To ensure everyone understood and to connect with Anthropic's drama, Zhipu added a not-so-subtle line this time—"Frontier intelligence should not belong to a few, nor should it be revoked at any time by a few rules."

In short, with one sentence paired with a normal upgrade + open-sourcing, Zhipu elevated itself into the global AI order debate; it positioned its model as the "antidote" for developers worldwide who had long suffered under Anthropic; and it casually placed its product on the same table as the world's most cutting-edge models.

Three birds with one stone, all hits. Among so many large model companies globally, only Zhipu steadily caught the window created by others. This isn't just skill; it's talent.

Zhipu also understands one truth—stories should never be told by oneself; that's just self-talk. If you want to tell a story, do it through the most influential voices globally.

What illustrates this best is the deliberately selective amplification and precise editing of a conversation in the dissemination chain of this epic narrative.

The domestic version of the story stops at a two-sentence version—after releasing new versions, even Elon Musk started to view Zhipu favorably, which is indeed a fact—because on X, Musk predicted that Chinese models would catch up to Fable by the first quarter of 2027.

Tang Jie's response under Musk's post was simple—"It won't take that long."

Just two sentences encapsulated the endorsement of a big name, Zhipu's confident edge, and the dividends of China's catch-up, pushing public sentiment to its peak and sending the stock price soaring.

But the story doesn't end there—in fact, the complete conversation was a four-sentence version. According to netizens' translations of the public conversation record, the true sharpness of the dialogue lies in Musk's truncated next sentence.

The edited-out part is—after Tang Jie said, "It won't take that long," Musk immediately responded with a paragraph that, when translated, reads: Maybe you (Zhipu) are right on individual benchmarks, but when it comes to real-world practical value, even catching up by the first quarter of next year would be quite remarkable. Anthropic is on the right path—they focus all their energy on refining truly practical intelligence. This capability doesn't appear on leaderboards but directly translates into revenue.

Musk's face-to-face remark cut benchmarks and real skills that generate revenue into two separate things. Zhipu's dissemination conveniently downplayed the more critical latter half because it was something it couldn't handle—"Capability doesn't appear on leaderboards but directly translates into revenue" is precisely the opposite of Zhipu's current reality, as its actual situation is that its capabilities are strong and it consistently tops leaderboards, but its actual revenue is worlds apart.

So, Zhipu could only respond with a logically disjointed pretty phrase: "Focus is the only thing we need, especially on what the essence of intelligence truly is."

02

Different Soils, Different Growth

At this point, more detailed facts must be clarified, or this becomes mean-spirited.

First, Anthropic's drama works because it's built on a truly cash-printing business—its annualized revenue grew from $87 million in early 2024 to approximately $44 billion by May 2026; Q2 revenue is expected to double to $10.9 billion, marking its first quarterly operating profit of about $559 million; 85% of its revenue comes from enterprise clients renewing its programming packages, and as businesses become addicted to AI programming, this business grows rapidly without increasing sales costs, leveraging high growth.

Thus, security narratives are merely icing on the cake for Anthropic, not life-saving oxygen.

Similarly, Zhipu's skills are also exceptionally strong. Yesterday, I visited one of China's AGI "Four Little Dragons," and its executive said—Zhipu's capabilities in Cording are currently unparalleled in China.

More facts: GLM-5.2 scored 50 points on the Artificial Analysis comprehensive leaderboard, topping the open-source weight category and reaching the level of Opus 4.8. An even more impressive and remarkable feat lies beneath: the entire GLM-5 series was trained on 100,000 Huawei Ascend 910B chips, without using a single NVIDIA card. In today's world, where any country or company could face chip supply constraints, Zhipu has largely achieved supply chain security—this is real skill, not showmanship. It also means its claim that "frontier models shouldn't be revoked by a few rules" is indeed grounded in self-sufficient production capacity.

But after discussing achievements, Zhipu's dilemma must also be addressed.

Let's calculate the same metrics. Zhipu's 2025 revenue was 724 million yuan, with a net loss of 4.718 billion yuan—for every 1 yuan earned, it lost 6.5 yuan. For comparison, OpenAI still loses $1.22 for every $1 earned.

Meanwhile, Anthropic achieved its first quarterly operating profit of about $559 million in Q2 2026—one is starting to make money quarterly, while the other is still deep in the red. The gap is not small.

This gap is not primarily a technical issue.

The enterprise AI markets in China and the U.S. are fundamentally different. The U.S. has the world's most mature and comprehensive digital industry ecosystem, with SaaS penetration, cloud-native architectures, and API payment habits having been in place for over a decade. Anthropic selling Claude to Fortune 10 companies works because those clients have engineering teams capable of integrating it, making token-based billing a natural fit. In contrast, Chinese enterprises have uneven digital foundations, with many clients demanding data remain within their domain and models deployed on-premises.

Thus, "private deployment" is not a choice for Zhipu but a market demand. Under these circumstances, 70-80% of Zhipu's revenue comes from project-based work rather than subscriptions—this is a result of client demand, not a flawed business model.

Deeper still is the purchasing power gap. China's AI demand is undoubtedly the world's second-largest and may even become first in the future, but currently, Chinese enterprises' willingness and budget for software payments lag significantly behind the U.S. The price war that has driven Zhipu's cloud-based API gross margin down to 18.9% is not solely due to industry competition but also reflects the overall market's payment ceiling.

Thus, Zhipu's comprehensive gross margin fell from 56.3% in 2025 to 41%, with three price hikes and margin pressure reflecting both hard constraints from computational costs and soft constraints from the macro market environment.

Another factor can only be hinted at here—Chinese AI offers global developers too many choices. China is already the world's largest supplier of open-source models, providing nearly 30% of the global high-quality tokens—with so many options, the pricing power for closed-source and paid APIs naturally narrows.

So here's the paradox: The models are so capable, yet making money is so hard!

03

Market Capitalization Propped Up by Narratives Will Be Reclaimed by Facts

Whether a company is overvalued or undervalued is ultimately its own business. Zhipu's problem is that this market capitalization may bring many benefits but also many objective issues—and these issues are no longer just Zhipu's problems; the entire Chinese AI industry will have to foot the bill.

The first and shallowest layer involves retail investors. The July share unlock and an even larger wave next January will hurt those who bought in late; the ones holding nearly zero-cost shares, ready to cash out anytime, are cornerstone investors and employees. This layer is painful but still the lightest, as it only affects those who voluntarily reached out and signed risk disclosures.

The second layer is more severe: the distortion of industry pricing anchors.

If "over 1200x price-to-sales ratio" becomes a widely accepted market benchmark (though almost impossible), then when the next AI company raises funds, investors' calculations change: "If Zhipu is valued that high, your 100x price-to-sales ratio is too cheap."

This is not a flattering portrayal of Zhipu; rather, it highlights the inflation of a bubble across the entire industry. What are the possible consequences? Good companies are coerced into accepting inflated valuations, only to be crushed under the weight of inflated expectations. Meanwhile, bad companies exploit the situation to raise funds, continue to burn through cash, and keep spinning tales.

The more insidious harm is that the "inability to tell a compelling story" becomes a competitive disadvantage. If a CEO admits during a roadshow, "Our profit margins are thin, and customer retention is challenging,"—under the looming shadow of Zhipu's trillion-yuan market capitalization, this admission is perceived as a sign of incompetence. Consequently, everyone feigns confidence; no one dares to speak the truth. The ripple effect is catastrophic: the quality of real information deteriorates, pricing logic crumbles, and the entire industry becomes a contest of who can tell a story most reminiscent of Anthropic.

The third layer of harm is the most fatal: resource misallocation.

This is the most easily overlooked yet longest-lasting form of damage.

When a non-replicable company becomes the benchmark, mainstream resources increasingly tilt toward "similar stories"—not necessarily by copying Zhipu's model (which is difficult to replicate) but by selecting projects based on "market capitalization narratives" rather than "revenue health." Money flows to the most skilled storytellers, not the most effective problem-solvers.

Zhipu has already established itself as one of the leading companies in China's AI sector. If even it relies primarily on 'expectations' rather than 'revenue' to sustain a trillion-dollar market capitalization, what message does this send to latecomers?

The worst-case scenario for China's AI sector is not about falling behind technologically, but about fostering a generation of companies that 'lead the world in benchmarks, rival industry leaders in valuation, yet consistently fail to generate the profits they should.'

While Anthropic across the ocean has already demonstrated, with an annualized value of $44 billion, that 'AI can generate substantial profits,' China's most valuable AI companies are still struggling to turn a profit. This is not the intended trajectory for the industry.

Actually, there is a fourth layer of concern.

As we discussed in the previous section, China has provided both local and global markets with abundant, cost-effective open-source tokens, establishing itself as the country with the most vibrant open-source ecosystem in the global AI sector and the 'world factory' for highly cost-effective tokens.

However, every token comes with a cost. Even if the cost is minimal, multiplying it by the vast usage volume results in an astronomical figure. But if this prosperity is not sustained by a self-supporting closed loop and is instead reliant on financing to subsidize profit margins, then the real concern is not whether Zhipu is overvalued. Instead, it is whether, when the closed loop eventually collapses due to escalating costs, these companies that dare to invest heavily in R&D and fully leverage domestic computing power will collectively suffer. Will it lead to the collapse of the research and development environment that China's large models have painstakingly built up at its most critical juncture?

From this perspective, rather than allowing these companies to rely on a never-ending cycle of narratives and investors filling in the profit gaps to sustain themselves, it would be preferable to let them earn more through their own capabilities in a market willing to pay for the true value of AI.

What truly needs to be protected is not the market capitalization inflated by sentiment, but the best Chinese AI teams represented by Zhipu and their products, as well as the nascent yet essential environment—a market willing to pay for good technology.

In July, the curtain of lock-up expiration will soon rise. Regardless of how market capitalization fluctuates, for China's AI industry, the real test lies not in how well the stories are told, but in whether what remains after the sentiment fades are tickets for sustainable growth or merely tickets for having watched a trillion-dollar spectacle.