Independent, Scarce, Penetrating

Times Create Heroes, and Heroes Shape the Times

Editor: Li Ran

Contributor: Xing Zhe

Source: Rao Cai - Rao Cai Research Institute

They say lock-up expiry is fiercer than a tiger, yet the 'first large model stock' defied the trend.

On July 8, Zhipu faced its first cornerstone lock-up expiry post-listing, corresponding to a market value exceeding HK$40 billion. Logically, such a scale could easily trigger sharp stock price fluctuations. However, the day's performance was surprising: after a slight dip at the open, the stock quickly rallied, closing up 13.35%. The next day, it rose another 11.3%, closing at HK$2,032. Even with subsequent corrections, the July 15 closing price of HK$1,707 remained higher than the pre-lock-up expiry price of HK$1,610. JP Morgan also raised its target price to HK$2,000, reiterating an 'overweight' rating.

How? First, cornerstone investors held firm. Prior to the lock-up expiry, multiple institutions publicly committed to maintaining their holdings. These included state-backed funds like the Beijing AI Industry Investment Fund and the Beijing Jingneng Green Energy M&A Investment Fund, as well as market-oriented institutions and early shareholders such as WT Asset Management, Optimas Capital Limited, and Luster LightTech. According to Securities Daily, these committed institutions held approximately 70% of the shares subject to lock-up expiry.

Second, positive business developments provided support. Shortly before the lock-up expiry, Zhipu released its new flagship model, GLM-5.2, which narrowed the gap with leading international models to just 1%-4% in processing million-scale long contexts.

Under this dual support, Zhipu passed an initial stress test. Ultimately, investing is about confidence and the future. A smooth transition validated the former, while future growth depends on the company's ability to forge a sustainable commercialization path.

1  Topping Evaluations: More Than Just a Technical Showcase

LAOCAI

What is certain is that GLM-5.2's timely release has opened up market imagination.

On June 17, Zhipu officially launched the GLM-5.2 large model and open-sourced it simultaneously. Clearly positioned, it is designed for long-duration, complex tasks, capable of engineering-ready processing of million-token-scale long contexts.

According to Zhipu, GLM-5.2 can handle over 880,000 tokens in a single continuous task, autonomously completing the entire software delivery process from development, joint debugging, testing to packaging and deployment. Previously, such tasks would require a team to collaborate for weeks, marking an industry revolution.

Behind this lies a profound shift in large models from basic dialogue to complex task execution. Beyond understanding longer information, models must maintain goals, invoke tools, and complete multi-step operations over tasks lasting hours or even longer, posing significant challenges to reasoning, memory, and execution capabilities.

Evaluation data speaks volumes. On the comprehensive capability rankings of international independent evaluator Artificial Intelligence, GLM-5.2 scored 51 points, ranking among the top open-source models and forming a 'tripartite governance' pattern alongside Anthropic and OpenAI.

Evaluations show that in Code Arena, a code assessment system with blind testing by over a million users, GLM-5.2 ranked first among globally available models. In authoritative evaluations like FrontierSWE and Terminal-Bench, it narrowed the gap with Claude Opus 4.8 to just 1%-4%.

Topping the rankings means Zhipu now has the strength to compete head-on with leading players. Industry analyst Li Xiaojing believes that GLM-5.2's improvement in evaluation scores is not the only highlight. Facing increasing external technical restrictions, the company did not take the old path of stacking massive computing power. Instead, through algorithmic architecture innovation and software-hardware Collaborative optimization (collaborative optimization), it continuously improved model training and inference efficiency. This pioneering approach is where the true strategic value lies.

Remarkably, after topping the evaluations, Zhipu set its sights even higher. On July 11, founder Tang Jie outlined a new post-Coding era direction in an internal letter titled 'The Giant Wave Has Arrived': while the industry generally accelerates commercialization, the company decided to push upward. Named the 'Touch High Plan', it strategically invests over the next two years without pursuing short-term application monetization, instead targeting the next frontier of AGI.

According to the internal letter, the 'Touch High Plan' focuses on four main directions:

First, long-duration tasks: enabling AI to evolve from 'instant Q&A' to 'grand projects', autonomously breaking down goals like 'designing a new cancer drug' into thousands of subtasks, learning, executing, and recording throughout the lifecycle.

Second, autonomous agent systems: evolving from 'intelligent assistants' to 'digital employees', building a collaborative society of thousands of professional agents to achieve 'self-driving' digital productivity.

Third, fully self-trained: using synthetic data factories and AI self-play to generate knowledge from nothing, and allowing systems to self-modify code within secure sandboxes, freeing evolution speed from human physical limitations.

Finally, extreme safety governance: embedding ethics and regulations as foundational axioms into models, investing billions to tackle 'mechanical interpretability', transforming black boxes into transparent systems, and participating in international AI governance to prevent technological abuse.

Industry insiders told Rao Cai that Zhipu's 'Touch High' strategy demonstrates significant openness and breakthrough potential. The fact that GLM-5.2 achieved full adaptation with domestic computing power platforms like Huawei Ascend, Pingtouge, and Moore Threads on its first day is a positive signal. As the logic of 'open-source drainage (driving traffic) and API monetization' continues to prove effective, Zhipu's growth potential will further unfold.

2  Soaring Volume and Prices, Back-to-Back Enterprise Deals

LAOCAI

Indeed, with the right path and direction, success depends on execution and steady accumulation. The effectiveness of this combination is evident in past financial data:

In 2025, Zhipu's revenue reached RMB 724 million, up 131.9% year-on-year, doubling for three consecutive years and making it China's largest large model company by revenue.

Breaking down the revenue structure: the main segments are local deployment and cloud deployment (MaaS). The former contributed RMB 534 million, up 102.3% year-on-year, though its share of total revenue dropped from 84.5% to 73.7%. This is because cloud deployment services grew faster, with revenue reaching RMB 190.4 million, up 292.6% year-on-year, and its share jumping from 15.5% to 26.3%.

In 2026, Zhipu implemented a series of price adjustments. On February 12, it launched the new flagship model GLM-5, followed by price hikes for the GLM Coding Plan package, with overall increases starting at 30%. On March 16, it released GLM-5-Turbo and again raised API prices by 20%. After several adjustments, Zhipu's large model product prices have cumulatively increased by over 60%.

Benefiting from strong market demand, these price hikes did not deter customers. In Q1 2026, API pricing increased by a cumulative 83%, while usage volume surged 400% year-on-year. By March 2026, Zhipu's MaaS platform had over 4 million registered users across 218 countries and regions, with over 242,000 paying developers. With both volume and prices rising, the cloud business accelerated into the monetization phase.

Why are customers willing to pay more despite price hikes? The stickiness formed by a good user experience is a key factor. As Zhipu CEO Zhang Peng explained, using international peer Anthropic as an example, 'When a model is strong enough, the API itself becomes the best business model. This logic can be summed up as: the smarter the model, the more confident it is in raising prices; the more indispensable it is to users, the greater the token consumption, and the higher the revenue ceiling.'

Real money from government and enterprise markets also validates this logic. In 2026, Zhipu secured several major deals:

On June 16, Shenzhen Smart City Technology Development Group Co., Ltd. awarded Beijing Zhipu Huazhang Technology Co., Ltd. a RMB 93.9 million contract for a global science and technology innovation analysis service project based on domestic large models.

On June 2, China Unicom (Guangdong) Industrial Internet Co., Ltd. awarded Zhipu a RMB 9.79 million contract for the 2026 Guangdong Industrial Internet Pilot Base Shenzhen AI + Glasses New Construction Project.

On February 11, the China Meteorological Administration's website showed that Zhipu won a RMB 10.9 million contract for the 'Smart Short-Term Meteorological Flash Warning Information Release Platform Construction Project - Smart Short-Term Warning Meteorological Service Large Language Model' initiated by the CMA Public Meteorological Service Center.

From industrial internet to government digitalization, Zhipu's enterprise client base is expanding rapidly. According to the company, as of July 2026, its enterprise clients have chosen private deployments covering key industries such as finance, energy, manufacturing, healthcare, and government services. Beyond data security and compliance, the company provides not just model capabilities but also deliverable, controllable, and iterable industry solutions.

Industry analyst Wang Tingyan believes that in exploring commercialization paths for large models, focusing on the B-end is proving to be a shorter and more stable route to scale. Multi-million-dollar government contracts not only bring revenue but also valuable brand endorsements, helping to break through to consumer markets for large models. With each market trust vote, Zhipu is evolving from a technology leader into an infrastructure partner for industrial intelligence.

3  From Technical Leadership to Commercial Closure: Several Hurdles Remain

LAOCAI

Looking back, Zhipu's rise stems from its own efforts but also benefits from the era's winds. A new global wave of AI infrastructure development is unstoppable.

However, a large market also means fierce competition. The track (track) is still emerging, and everyone is a dark horse. Winning now does not guarantee always winning. For Zhipu to Continuously taking the lead (continue to lead) and maintain its position, it must stay vigilant and continuously address gaps. Compared to global giants like Anthropic, it faces several key challenges:

First, rigid computing power costs and uncertain profitability demand further scale effects.

In 2025, Zhipu's revenue hit a record high, but net losses reached RMB 4.718 billion, or RMB 3.182 billion after adjustments, primarily due to R&D spending of RMB 3.18 billion, equivalent to 4.4 times revenue.

A significant portion of this went to purchasing third-party computing power services. Meanwhile, Anthropic achieved annualized revenue of USD 30 billion while spending just one-fourth of OpenAI's model training budget.

Industry analyst Sun Yewen notes that this does not reflect Zhipu's inefficiency. With Google raising its 2026 capital expenditure forecast to USD 180-190 billion, the intensity of the global computing power arms race is clear. Rigid growth in computing power costs hangs like the Sword of Damocles over Zhipu's path to profitability.

Second, intense price wars from competitors squeeze pricing power and raise doubts about sustained volume and price growth.

On April 26, DeepSeek announced API price adjustments, slashing input cache hit prices across its V4 series to one-tenth of the initial prices. Its official API pricing page showed that the reductions applied to all V4 models, focusing on input cache hit scenarios.

After the adjustment, DeepSeek-V4-Flash cost RMB 0.02 per million tokens for input cache hits. The enterprise-grade DeepSeek-V4-Pro saw an even steeper drop, from RMB 1 to RMB 0.1 per million tokens for cached inputs, and with limited-time discounts, the actual price fell to RMB 0.025 per million tokens.

Internet giants soon followed, with some even 'selling cloud services at a loss', sparking short-term concerns about MaaS market development—the more users, the greater the losses. Without a clear product performance gap, price wars became the primary means of market capture. Despite rising demand for AI deployment driving up industrial chain (supply chain) prices, scale expansion has not reduced marginal costs enough to offset rising computing power costs.

Fortunately, Zhipu maintains relatively strong pricing power thanks to its technical barriers in coding and private deployment demand from government and enterprise clients. However, sustaining excess profits in this environment requires maintaining a far larger technical lead over competitors—no easy feat.

Third, questionable Token demand growth needs further consolidation.

The current explosive growth in Token consumption highly depends on the penetration of AI agent applications. If agents fail to move beyond developers to ordinary users or if scenario rollouts lag, the growth curve of Token demand could inflection point (turn) at any time.

For Zhipu, the confidence to raise prices comes from its temporary lead in model intelligence. The problem is that competitors are not idle. With rapid market iteration, giants like ByteDance, Alibaba, and Tencent are investing heavily in R&D. Once the technological gap between large models is closed, the risk of eroding the demand base looms.

Finally, balancing business transformation with investment and monetization.

Zhipu faces a 'transformation paradox': the profitable local deployment business has a limited ceiling, while the API business, with its room for imagination (imagination space), currently has a gross margin of just 18.9% and faces low-price competition from internet giants and DeepSeek.

According to Gelonghui, Zhipu has set a goal to increase API business revenue to 50% of total revenue. While appealing, achieving this is no easy task. API competitiveness ultimately depends on model performance, response speed, and stability—all requiring sustained technological iteration.

Technological iteration, in turn, demands high R&D investment, whether for computing power procurement, algorithm optimization, or talent acquisition. If external financing tightens or shareholders demand earlier profitability, the trade-off between 'investment and monetization' will intensify, affecting progress toward the target.

Listing these challenges does not imply pessimism but highlights a fact: Zhipu's current market value needs greater certainty in closing the commercial loop. The distance between technical premiums and commercial premiums is the key yardstick for measuring the company's long-term value.

4  Cutting Costs and Balancing Short-Long Term: Achieving the Crucial Leap

LAOCAI

This brings us back to the opening point: lock-up expiry is just the beginning; profitability is the ultimate goal.

In 2025, Zhipu's revenue exceeded RMB 720 million, but net losses surpassed RMB 4.7 billion. Operating activities resulted in a net cash outflow of approximately RMB 2.246 billion for the year, with cumulative losses from 2022 to 2025 nearing RMB 8.5 billion.

Faced with continuous operational financial losses, Zhipu needs to continuously raise external funds. It will be listed on the Hong Kong Stock Exchange in January 2026 and subsequently exercise the over-allotment option to raise approximately HK$4.9 billion. On July 9th, it announced another placement, raising a total of approximately HK$31.4 billion, primarily for research and development investment, business expansion, and capital structure optimization.

Once the A+H listing is completed, the timeline for transitioning from 'burning money for technology' to 'trading technology for cash' will also be put on the agenda. After all, the more financing and burdens there are, the more investors expect returns.

Meanwhile, market competitors are not waiting. For example, Anthropic is expected to achieve positive cash flow as early as 2027. For Zhipu, the key to escaping losses in the fiercely competitive domestic price war environment is to stabilize the basic profit from localized deployments while rapidly increasing the proportion of API business volume. In other words, both cost-cutting and revenue-generating measures are essential.

On the business front, Zhipu's gross profit margin for cloud deployment services rose from 3.3% in 2024 to 18.9% in 2025, mainly due to improved model inference efficiency and declining marginal costs from expanded computing power. However, the gross profit margin for localized deployment services slipped from 66% to 48.8%, primarily due to increased delivery resource investments driven by rising customer demand. After offsetting these effects, the company's overall gross profit margin fell from 56.3% in 2024 to 41.0%.

It should be noted that before going public, Zhipu had already initiated a series of 'cost-cutting and efficiency-enhancing' adjustments, integrating resources from government (G-end) and business (B-end) sectors, streamlining peripheral operations, optimizing personnel structure, and focusing resources on the core direction of the MaaS platform. However, the effectiveness of organizational streamlining and strategic focus ultimately needs to be reflected in financial metrics. The doubling of revenue and halving of net profit in 2025 indicate that this adjustment still has a long way to go.

Looking further ahead, globalization shortcomings also constrain revenue growth. Claude serves enterprise clients worldwide, whereas Zhipu's revenue relies heavily on the domestic market, with overseas commercialization still in its infancy. After reaching a certain level of penetration in the domestic government and enterprise market, expanding incremental space will require overcoming barriers such as geopolitics, channels, and brand influence.

Fundamentally, the investment logic for large model companies is essentially a game between long-termism and short-term efficiency. The former requires companies to invest in research and development, refine models, and cultivate ecosystems regardless of cost, aiming to secure a pricing power high ground in the ultimate Token economy. The latter demands that money be spent wisely, questioning the destination and effectiveness of every dollar: Can short-term losses build core barriers? Can burning money lead to repeat purchases? Can growth cover financial losses and subsequently outpace competitors?

From this perspective, the lock-up expiration is merely a small liquidity test. Zhipu's real challenge lies in proving that it can not only spend money but also make money when the market loses patience for grand AGI visions. By delivering a commercially sustainable and profitable answer, it can complete the crucial leap from 'burning money for technology' to 'trading technology for cash.'

As the saying goes, investing is about confidence and the future—the two are interdependent and jointly influence capital flows. It is commendable that the share price did not decline but increased on the lock-up expiration day. However, Zhipu must still race against time, itself, and the market. After all, with the lock-up lifted, shareholders have the right to reduce their holdings at any time.

This article is originally written by Rho Finance.

Please leave a message if you need to reprint it.