<div id="readability-page-1"><div><div><p><span>Author: Zhou Ailin, Tencent Finance</span></p><p><span>The stock prices of Chinese AI model companies have astonished investors—from being underestimated at the time of their IPO to reaching a market capitalization of nearly 1 trillion yuan at its peak, Zhipu AI's rise has been meteoric, earning it the label of &#34;China's Anthropic.&#34;</span></p><p><span>Undeniably, China is leading or has gained an edge in the &#34;open source ecosystem and price diffusion&#34; race. Especially for small teams and individual developers with limited budgets running non-core businesses, Chinese models offer a cost-effectiveness that US closed-source giants cannot match.</span></p><p><span>However, everything has a price.</span></p><p><span> Several investors and strategists told Tencent News's &#34;Deep Dive&#34; that Zhipu AI, specializing in the B2B programming track, is currently showing momentum that surpasses MiniMax, but the challenges it faces are also obvious. For example, there is a huge gap in computing power, with domestic computing power being only about one-twentieth of that overseas; the market price is not based on current profits, but on the &#34;faith valuation&#34; of AI native applications, with R&amp;D investment (3.18 billion yuan) equaling 4.4 times revenue (724 million yuan), making profitability a distant prospect; in addition, for this new high-priced stock with a small market capitalization, major shareholders and cornerstone investors will face a large-scale unlocking of shares around July. The extreme volatility of &#34;30% amplitude&#34; on May 29 was generally interpreted by the market as a pre-emptive stress test of the unlocking pressure—previous unrealized gains were more than ten times, creating a strong incentive to cash out. &#34;The token's growth rate is accelerating, and the number of Chinese users willing to pay for agents is also increasing, along with overseas users,&#34; Huang Hui, Chairman and Fund Manager of Xiaoyu Asset Management, told Tencent News's &#34;Deep Dive.&#34; &#34;The rise of technology inevitably brings bubbles, and many investors cannot predict the final outcome. As the saying goes, 'They wear the bubble on their sleeves, but you can't do anything about it.' This is a common characteristic globally, and funds can only choose to invest in AI.&#34; In the future, both Chinese and American large-scale models will occupy a place in the market, but competition will become increasingly fierce. Strong open-source models such as DeepSeek and Qwen continue to close the gap, leading to increasingly intense price wars. With more model stocks going public, the scarcity premium will also decrease. &#34;Market Dream Rate&#34; The coding field is a battleground for leading model vendors because users have a strong willingness to pay; for example, 37.9% of Claude's users utilize its coding capabilities. Zhipu, whose initial story seemed unappealing, has overtaken its Chinese counterparts precisely because of its vertical advantage in coding. Since the second half of 2025, market applications have become highly concentrated. Currently, over half of token consumption comes from coding scenarios, including many implicit coding demands generated by Agent applications. Thus, &#34;the Chinese version of Anthropic&#34; has become Zhipu's capital narrative. The success of Claude Code's technology and commercialization overseas has driven the entire US AI industry chain. However, intrinsic value remains a crucial factor to consider. In just five months, Zhipu AI's market capitalization soared from 51.1 billion to over 700 billion, a more than 13-fold increase—faster than any previous tech bubble in the Chinese stock market. Years ago, strategists used the term &#34;Price-to-Whatever Ratio&#34; to describe the extent of bubbles in certain stock markets. The reasonableness of a stock price is measured by earnings; therefore, the &#34;Price-to-Whatever Ratio&#34; is undoubtedly ironic—it's whatever the price is. Currently, listed AI model companies are not profitable, and it's unlikely they will be profitable in the next two years. Therefore, the benchmark for the capital market is &#34;Annual Recurring Revenue,&#34; calculated as: ARR = Monthly Revenue × 12. &#34;The market is currently focused on profitability. Although model companies aren't making a profit, as long as the ARR is rising, investors will temporarily buy in,&#34; a leading foreign mutual fund investment manager told Tencent News's &#34;Deep Dive.&#34; Both MiniMax and Zhipu have indeed seen their ARR more than double. Zhipu AI's annual ARR grew more than 60 times year-on-year – In 2025, Zhipu AI's total revenue reached RMB 724 million, a year-on-year increase of 131.9% – this is the financial report. Entering 2026, the pace accelerated sharply. According to a UBS research report, Zhipu AI's API platform ARR reached approximately $250 million, a year-on-year increase of about 60 times. The core engine driving this leap was the establishment of pricing power in the Coding sector – three price increases did not trigger user churn: overseas subscription pricing rose from $49/month at the beginning of the year, to $80 in February with the release of GLM-5, and then jumped again to $160 on April 11, a cumulative increase of 226%; among domestic paid users, the number of Coding Plan subscribers exceeded 240,000. However, looking at revenue data, Zhipu's R&amp;D investment (3.18 billion yuan) is 4.4 times its revenue (724 million yuan). For a direct comparison, iFlytek's revenue in 2025 was 27.1 billion yuan, net profit was 839 million yuan, and market capitalization was approximately 118.7 billion yuan; Zhipu's revenue was only 2.7% of iFlytek's, yet its market capitalization was several times higher. Zhipu's shareholders' equity decreased from -3.955 billion yuan (2024) to -8.111 billion yuan (2025), indicating a continued deterioration in its balance sheet. The market is not pricing in current profits, but rather in the &#34;belief valuation&#34; of AI-native applications—Huachuang Securities predicts revenue of approximately 3 billion yuan in 2026 (+317%), but the timeline for profitability remains far off, with HSBC expecting its first profit no earlier than 2029. Some have also made a prediction that, under a more optimistic scenario, Zhipu will successfully transform into an API-based platform and become profitable in 2029. This assumption includes: - Revenue CAGR maintained at 130% from 2026 to 2029 - API business share increasing from 27% to 65% - Net profit margin of 10% in 2029, revenue of approximately 35 billion yuan, and net profit of 3.5 billion yuan - A reasonable PE ratio of 60 times (premium for high-growth AI companies) at that time → a market capitalization of 210 billion yuan, still nearly 60% lower than the current market capitalization. At the same time, the uncertainty of the competitive landscape must also be considered. In the next three years, the model iteration speed will be extremely rapid, and today's &#34;number one in China&#34; could be overturned at any time by competitors such as DeepSeek or Qwen. &#34;The Chinese version of Anthropic&#34;? Another noteworthy issue is that China doesn't lack models; major model vendors have been strategically differentiating and making precise bets since the beginning of 2025. Zhipu GLM focuses on Web Coding, believing it to be the path to AGI applications; Kimi bets on a paradigm similar to LangChain, exploring ultimate Agent applications; Doubao focuses on C-end consumer applications and daily life experiences; Tongyi Qianwen adopts a relatively balanced strategy, committed to the comprehensive development of open source, multi-size models, and multimodal capabilities. Thus, &#34;The Chinese version of Anthropic&#34; has become Zhipu's capital narrative. However, capital narratives are not an accurate description of business reality. The similarities between the two lie in their technological paths and business philosophies (strong model → API → Coding track); the differences lie in their market environment, monetization logic, and revenue scale, and these differences are often the core. Anthropic's success grew within the unique paid culture, developer ecosystem, and enterprise SaaS procurement system of the United States. Zhipu faces a price-sensitive market with a fragmented competitive landscape and strong intervention from cloud vendors—therefore, industry insiders believe that its more likely ultimate goal is to become a core model supplier for China's AI infrastructure, rather than replicating Anthropic's independent flywheel business model. However, some argue that the prospects for &#34;token going global&#34; are undeniable, meaning the revenue ceiling for model companies is significantly increased, thus fueling the expansion of this capital narrative. Multiple mainstream benchmark tests show that Chinese models have largely caught up with the US in basic capabilities, but still lag behind by 5% to 8% when dealing with complex problems. For example, according to the advanced version of SWE-bench Pro, whose problems are more complex and closer to the real development environment, the test results show that Claude 4.7 achieved a score of 64.3%, Kimi K2.6 achieved 58.6% (this is the first open-source model to surpass GPT-5.4 in this test), and DeepSeek V4 achieved 55.4%. It is evident that the US model leads by 5% to 8%, which means that in actual development, a complex bug that the US model might solve in one attempt might require 3 to 5 attempts for the Chinese model to barely pass. Undeniably, the Chinese model, with its extremely high algorithmic efficiency, such as Hybrid Expert Architecture (MoE) and reinforcement learning, achieves extremely high cost-effectiveness (costing nearly 100 times cheaper), making it particularly suitable for users with limited budgets. However, the biggest challenge facing Zhipu today is the bottleneck in computing power and the uncertain prospect of profitability. Zhipu CEO Zhang Peng disclosed at the 2025 earnings conference: &#34;In the first quarter, API call pricing increased by 83%, and even so, call volume increased by 400%, still resulting in supply falling short of demand.&#34; Zhipu's previous growth was mainly due to model optimization, specifically employing two technical methods: model pruning and dynamic scheduling strategies based on peak and off-peak business periods. Profitability uncertainty remains prominent. Experts from the independent third-party research institution Acecamp mentioned that while Zhipu's programming model performed well in B-end order volume in the first quarter and April of 2026, there is controversy regarding its adoption of the ARR metric. This is because orders from large listed company clients are uncertain; even if penalties are required, cooperation may terminate due to corporate strategic adjustments. Therefore, including such large, monthly-paid orders in ARR may not be entirely applicable, even though financial reports show that its programming business's ARR has reached over one billion yuan. Increased competition among models may also impact profits. In terms of C-end commercialization, Zhipu currently has relatively few highlights, with programming packages being one of the few product lines performing well. In contrast, Kimi's C-end agent matrix covers multiple vertical scenarios, and each line has achieved effective monetization. If Kimi launches new models with parameter scales of 2.5T or even 3.0T within the next two to three months, Zhipu's market presence in the programming track will likely face a more direct impact. Furthermore, while the competitive landscape of models is much clearer than it was a year or two ago, the landscape itself is still evolving rapidly, especially regarding whether programming has become a mature area for large-scale model commercialization. Some argue that, unlike Anthropic, domestic companies face several challenges in this field: First, the advantages of the data flywheel are fragmented, with each company having its own customer base; second, most companies are capable of developing programming models to a high level of fundamental excellence, and Zhipu has not yet formed a solid technological barrier; third, unlike Anthropic, Zhipu lacks rich monetization scenarios and user reputation through programming capabilities, instead focusing primarily on selling programming model capabilities. The upcoming lock-up period is another significant hurdle for every IPO, after the initial sweet period of soaring stock prices. Many companies' stock prices have never returned to their peak. In the past, discounted IPOs were common in Hong Kong, but post-listing gains were relatively limited—because the pricing already fully reflected the fundamentals. Without the expectation of &#34;doubling in value upon listing,&#34; the appeal of cornerstone investors exchanging shares is greatly diminished. A primary market participant using a &#34;cornerstone anchoring&#34; strategy told Tencent News's &#34;Deep Dive&#34; that the logic has changed in the last two years. AI technology IPOs generally have a &#34;scarcity premium&#34;—Anthropic and OpenAI are not yet listed, and Zhipu and MiniMax are the only two globally tradable large-scale pure-play stocks, with the secondary market willing to offer valuation premiums far exceeding their fundamentals. In this situation, cornerstone investors locking up shares for six months offers extremely high expected returns. Furthermore, Hong Kong stocks didn't have this &#34;index inclusion game&#34; in the past, but now, MSCI inclusion, Hang Seng TECH inclusion, and the opening of the Southbound Trading Connect—each of these events triggers forced buying by passive funds. Speculative funds positioned themselves in advance and cashed out after the initial investment. &#34;VC/PE funds hold shares, with a fund lifespan of 5-7 years, nearing their exit window. Cornerstone investors themselves, with their six-month lock-up period expiring, are required by fiduciary duty to realize profits, and they never initially intended to hold long-term,&#34; the aforementioned person stated. Hong Kong stocks will see a historic peak in share unlocking in the second half of 2026, with cornerstone investors unlocking over HK$100 billion from June to August. It is foreseeable that in the next stage, Chinese models will remain globally competitive, but stock price volatility may intensify. After the selling pressure from unlocking, companies with strong fundamentals will experience long-term steady growth, but the key lies in whether the capital narrative can continue, which requires sustained digestion of sky-high valuations through ARR and profit growth.</span></p></div></div></div>

Snowflake's AI accounts surged from 9,100 to 13,600 in a single quarter, sending the stock up 37%; Agentforce ARR topped $1.2bn, +205% YoY; the four hyperscalers' combined 2026 capex exceeds $600bn, all funneled into AI infrastructure. The 'SaaS is dead' thesis is unraveling fast — do you think the application-layer monetization inflection point has finally arrived?

<p>Snowflake's AI accounts surged from 9,100 to 13,600 in a single quarter, sending the stock up 37%; Agentforce ARR topped $1.2bn, +205% YoY; the four hyperscalers' combined 2026 capex exceeds $600bn, all funneled into AI infrastructure. The 'SaaS is dead' thesis is unraveling fast — do you think the application-layer monetization inflection point has finally arrived?</p>

Trillion-Dollar Compute Buildout Sets the Stage — SaaS Harvest Season Is Here!

02513

<p>Zhipu AI's market cap surged over 13-fold, driven by high ARR growth in the coding sector. However, challenges persist: domestic computing power lags significantly, R&D costs far exceed revenue, and profitability remains distant. Major shareholders face share unlocking pressures around July, potentially causing volatility. While open-source models offer cost advantages, intense price wars and decreasing scarcity premiums pose risks to this 'faith valuation' bubble.</p>

How long can the "market dream rate" of AI-modeled stocks last?