China’s pure-play AI champion has world-class model momentum, but its trillion-dollar valuation leaves little room for reality.
Knowledge Atlas Technology Joint Stock Company Limited (HKEX: 2513), branded internationally as Z.ai and domestically as Zhipu AI, is a leading Beijing-based pioneer in general-purpose large language models (LLMs) and artificial general intelligence (AGI) technologies.[1, 2, 3] Spun out of Tsinghua University’s Knowledge Engineering Group (KEG) in 2019, the company achieved a critical milestone on January 8, 2026, by listing on the Main Board of the Hong Kong Stock Exchange, becoming the world's first "pure-play" foundation model developer to successfully go public.[1, 2, 3, 4]
The company generates revenue through two core operating segments: On-premise Deployment and Cloud-based Deployment.[2, 5, 6] On-premise deployment delivers customized large model services deployed directly on the customer's private physical infrastructure, ensuring compliance with strict data security guidelines.[2, 5, 6] Cloud-based deployment provides access to its flagship General Language Model (GLM) family via the Zhipu MaaS (Model-as-a-Service) Platform, monetized through usage-based API calls, developer subscriptions, and software development kit (SDK) licensing.[2, 5, 7] Geographically, while the company’s core customer base is situated within the People's Republic of China, its programming and enterprise-level services have expanded globally to encompass developers in 218 countries and regions, supported by regional hubs in the Middle East, United Kingdom, Singapore, and Malaysia.[1, 5, 8]
The primary revenue-generating products and target customer segments of Knowledge Atlas are structured as follows:
| Product / Service | Segment Type | Primary Customer Profile | Monetization Structure | Provenance |
|---|---|---|---|---|
| GLM Enterprise Models | On-premise Deployment | State-Owned Enterprises, Banks, Government Agencies | Bespoke multi-million-yuan software licenses and on-site integration contracts | [2, 5, 6, 9] |
| Zhipu MaaS Platform (APIs) | Cloud-based Deployment | Private Enterprises, Internet Companies, Tech Developers | Usage-based pricing per million tokens consumed (proactively raised in 2026) | [2, 5, 7, 8] |
| GLM Coding Plan (CodeGeeX) | Cloud-based Subscription | Individual Software Developers, Enterprise IT Teams | Tiered monthly and annual recurring subscription fees | [2, 8, 10] |
| The Claw Plan (AutoClaw) | Hybrid Agentic AI | SMEs, Universities, Consumer End-Users | One-click localized model installation and software-as-a-service (SaaS) subscription | [5, 8, 11] |
Customers choose Knowledge Atlas over domestic and international alternatives due to its prestigious Tsinghua academic endorsement, full optimization with domestic Chinese chips (e.g., Huawei Ascend), and superior algorithmic efficiency, which enables high-order reasoning at a fraction of the cost of Western flagship models.[1, 7, 8, 9, 12]
The financial and operational performance of Knowledge Atlas is driven by three primary commercial engines: the expansion of developer token consumption, the rapid adoption of subscription-based developer products, and the ongoing structural migration of enterprise workloads from customized on-premise deployments to scalable cloud-based APIs.[8, 13]
At the core of the company's offering is the General Language Model (GLM) series.[1, 2] The company's flagship model, GLM-5.2 (released in June 2026), leverages a 744-billion-parameter Mixture of Experts (MoE) architecture with 40 billion active parameters, supporting a lossless context window of up to 1 million tokens.[12, 14] This model achieves top-tier performance on the global Code Arena blind benchmark, scoring 1,595 points—trailing Anthropic’s Claude Fable 5 by a minor margin while outperforming Google's Gemini models and OpenAI’s GPT-5.5 in complex software engineering and agentic execution tasks.[12, 14, 15]
Complementing its base models, the company sells specialized software applications.[1, 2] CodeGeeX (commercialized via the GLM Coding Plan) provides deep IDE integration for automated code generation, supporting over 242,000 paying developers globally as of early 2026.[2, 8] In the agentic AI market, the company offers AutoGLM, a smartphone agent capable of executing multi-step tasks like food delivery or purchasing from shopping history, and AutoClaw, which simplifies localized model installation under its highly successful "Claw Plan" subscription.[1, 2, 8] Additionally, the company monetizes multimodal video generation via its text-to-video model, Ying, and operates AMiner, a scientific data mining and academic search platform.[1, 2, 16]
The competitive moat protecting Knowledge Atlas is multi-layered, consisting of technical, structural, and regulatory advantages:
* High Customer Switching Costs: Knowledge Atlas has deeply embedded its GLM APIs into the core development pipelines of 9 of the top 10 internet companies in China.[8] Transitioning from GLM to an alternative model requires significant engineering resources to rewrite code interfaces, re-tune prompt architectures, and restructure database pipelines, resulting in strong retention.[3, 8]
* Hardware Adaptation and Domestic Substitution Lock-in: Due to its placement on the United States Commerce Department’s Entity List in January 2025, the company has completed "Day 0" hardware-software co-optimization with domestic Chinese semiconductor manufacturers.[1, 8] The GLM series is adapted to more than 40 domestic chips, featuring native FP8 and Int4 quantization on Cambricon processors, compatibility with Moore Threads and Kunlunxin GPUs, and full optimization on Huawei Ascend hardware.[1, 8, 9, 12] This comprehensive adaptation shields the company from Western export restrictions and makes it the default choice for Chinese state-affiliated enterprises.[1, 9, 17]
* Cost Advantage via MoE Architecture: The company's sparse MoE architecture yields a strong cost advantage.[7, 14] Under GLM-5.2, the company charges an output price of $4.13 per million tokens for long-context windows.[12] This is significantly lower than Western alternatives like GPT-5.5 ($45) and Claude Opus ($25), enabling rapid market share capture while maintaining positive unit economics.[7, 12]
* Ecosystem and Endorsement Advantage: The company's origins within Tsinghua University provide academic prestige, which is highly valued by government and municipal clients.[1, 9] With over 4 million registered platform users generating continuous training feedback, the company benefits from a data feedback loop that accelerates model optimization.[3, 8]
The addressable market for generative AI in China is expanding, driven by state-mandated digitalization and industrial automation.[3, 8] According to Visible Alpha consensus estimates published by S&P Global Market Intelligence, Z.ai's total revenue is projected to reach HK$3.2 billion ($409 million) in fiscal year 2026, representing a major step-up from 2025.[13] Over the long term, consensus models project that the addressable enterprise demand will allow the company’s top line to scale to HK$46.8 billion ($5.9 billion) by fiscal year 2030, representing an estimated revenue compound annual growth rate (CAGR) of 127% from 2025.[3, 13]
Knowledge Atlas competes primarily with other domestic AI giants, listed pure-play start-ups, and traditional cloud providers:
| Competitor | Stock Code / Type | Core Strategic Focus | Business Model & Revenue Mix | Competitive Position vs. Z.ai |
|---|---|---|---|---|
| MiniMax Group | HKEX: 0100 | Multimodal foundation models (text, voice, video, character interaction) | Dual-model; 71% Consumer B2C (Talkie app), 29% B2B Cloud APIs | Z.ai is gaining ground; Z.ai’s enterprise focus and coding leadership have driven its valuation above MiniMax |
| Haizhi Technology | HKEX: 02706 | Graph-model fusion; industry-specific AI agents (finance, regulatory) | Industry-level AI solutions; 100% enterprise-oriented | Z.ai maintains a broader horizontal foundation, whereas Haizhi dominates specialized regulatory niches |
| Alibaba Cloud (Qwen) | NYSE: BABA | Dense and MoE cloud-deployed general models | Consolidated cloud computing bundle; massive infrastructure | Z.ai holds an independent, cloud-agnostic position, though Alibaba holds a 3.86% equity stake in the company |
| Baidu (ERNIE) | HKEX: 9888 | Enterprise cloud APIs and consumer-facing search integration | Hybrid subscription and ad-supported enterprise cloud API access | Z.ai ranks as the preferred independent partner for government agencies seeking non-platform-aligned models |
Following its technical breakthroughs in June 2026, Knowledge Atlas has gained ground.[12, 17] The release of GLM-5.2 and the company's inclusion in the Hang Seng Tech Index have drawn capital flows toward Z.ai, establishing it as the largest pure-play listed AI model company in China by market capitalization.[10, 17]
Knowledge Atlas announced its audited full-year fiscal results for the period ended December 31, 2025, on March 31, 2026.[6, 8]
* Revenue Performance: Full-year revenue reached RMB 724.33 million, a year-on-year increase of 131.9% from RMB 312.4 million in fiscal year 2024.[6, 8] However, this result missed analyst consensus expectations of RMB 749.21 million by 3.32% due to severe compute resource constraints in late 2025.[10, 19, 20]
* Net Loss and Profitability: The reported statutory net loss widened to RMB 4,718.2 million, compared to RMB 2,958.0 million in 2024.[6] This missed consensus estimates, which projected a narrower net loss of RMB 3,430 million, primarily due to higher-than-expected share-based compensation expenses (RMB 558.3 million) and intensive R&D spending.[6, 20] Adjusted non-IFRS net loss (excluding share-based compensation, listing expenses, and changes in the carrying value of financial instruments issued to pre-IPO investors) stood at RMB 3,181.97 million.[6]
* Segment Breakdown: The legacy On-premise Deployment segment generated RMB 534 million in revenue with a gross profit of RMB 261 million (48.8% gross margin).[13] The Cloud-based Deployment segment brought in RMB 190 million, with a gross profit of RMB 36 million (18.9% gross margin).[13]
* Loss per Share: Basic loss per share reached RMB 12.03, missing consensus estimates by 74% as widening losses diluted historical share values.[19]
For the first quarter of fiscal year 2026, the company reported strong operational metrics.[2, 11] Annual Recurring Revenue (ARR) expanded 6.4-fold in a single quarter, rising from $39 million in December 2025 to $250 million by March 2026, demonstrating commercial adoption for its MaaS products.[2, 11] Retained earnings of HKD 593.16 million were reported on the balance sheet for the quarter ending March 31, 2026.[21]
No explicit change to long-term GAAP guidance was issued during the March 31, 2026, announcement, but management outlined a structural shift in its product mix.[8] In entering fiscal year 2026, the company expects Cloud-based deployment to overtake On-premise deployment, projecting 2026 cloud revenue to surge to HK$2.0 billion (representing 63% of total forecasted revenue of HK$3.2 billion).[13] Additionally, to offset severe hardware shortages, the company proactively raised its API pricing by 83% and eliminated first-purchase promotional discounts in February 2026, shifting its strategic focus from volume acquisition to value pricing.[8]
On the earnings call, CEO Dr. Zhang Peng noted that the company is prioritising long-term token consumption and local chip adaptation over short-term profitability.[8] Management highlighted that the upcoming listing on Shanghai's STAR Market, aimed at raising RMB 15 billion ($2.22 billion) with RMB 12 billion dedicated to frontier general-purpose foundation model R&D, will ensure a sustainable operational runway through the late-2020s.[22, 23]
The FY2025 results initially prompted caution due to the wider-than-expected net loss.[20, 24] However, subsequent technical milestones and analyst upgrades shifted market sentiment [15, 17]:
* Stock Price Impact: On June 15, 2026, the stock surged over 32% following JP Morgan’s target price upgrade.[17] On June 22, 2026, a further 13% daily gain pushed the share price above HK$2,380, elevating the total market capitalization past HK$1 trillion.[10] The stock experienced a 10% technical pullback on June 23, 2026, to trade around HK$2,060 - HK$2,170.[2, 25, 26]
* Analyst Rating and Price Target Updates:
* JPMorgan: Maintained a "Buy" rating on June 17, 2026, and raised its price target from HK$1,400 to HK$1,800, citing pricing power and strong developer metrics.[17, 27]
* CLSA: Initiated coverage with a "Hold" rating on June 17, 2026, setting a target of HK$1,500.00, expressing near-term valuation caution.[27]
* Macquarie: Maintained an "Outperform" rating on June 15, 2026, with a target of HK$1,221.40.[17, 28]
* Jefferies: Maintained a "Hold" rating on April 14, 2026, with a target of HK$912.55.[27]
The company's market capitalization of approximately HK$933.6 billion implies a trailing Price-to-LTM-Sales multiple of 1,113.1x, representing a premium compared to peer averages of 9.0x and the technology sector average of 2.4x.[25, 29] This valuation is supported by a restricted free float.[10] Only 11.73 million H-shares (approximately 2.67% of total post-listing share capital) were unrestricted at listing, creating a supply-demand mismatch in the secondary market.[10] Valuation is also supported by expected passive inflows of HK$3.3 billion following the company's inclusion in the Hang Seng Tech Index, alongside potential eligibility for the Southbound Stock Connect.[11]
RISK SPECTRUM MATRIX
High Impact ┌───────────────────────────────┐┌───────────────────────────────┐
│ COMPUTE INFRASTRUCTURE ││ GEOPOLITICAL BARRIERS │
│ Severe hardware shortages ││ Placed on US Entity List; │
│ limit model API expansions ││ restrictions on top-tier GPUs │
└───────────────────────────────┘└───────────────────────────────┘
┌───────────────────────────────┐┌───────────────────────────────┐
│ VALUATION DE-RATING ││ FINANCIAL RUNWAY BURN │
│ Cornerstone lock-up expiry; ││ RMB 2.6 billion FCF burn; │
│ secondary float dilutions ││ high dependence on STAR raise│
└───────────────────────────────┘└───────────────────────────────┘
┌───────────────────────────────┐┌───────────────────────────────┐
│ COMPETITIVE PRICING ││ CUSTOMER CONCENTRATION │
│ DeepSeek price cutting on ││ Top 5 clients account for │
│ premium developer models ││ 26.5% of total revenue stream │
Low Impact └───────────────────────────────┘└───────────────────────────────┘
Low Probability High Probability
The business model is highly sensitive to rapid model obsolescence.[12, 14] If future model iterations (e.g., GLM-6) fail to secure top-tier benchmarks, or if developers find that the model requires excessive trial-and-error to complete tasks, developer churn will increase.[12, 14] Furthermore, actual developer testing shows that GLM-5.2 Max consumes more time and tokens to complete multi-step reasoning tasks than Western competitors, indicating potential computational inefficiencies in complex real-world workflows.[14]
The domestic Chinese generative AI market is characterized by price competition.[8, 30] Competitors like DeepSeek and MiniMax continue to release open-source models, which pressure Z.ai’s premium-priced API and subscription models.[20, 30] While Knowledge Atlas successfully raised prices in early 2026, a sustained price war from capitalized players (e.g., Tencent, Alibaba) could compress margins.[8, 30]
The company exhibits customer concentration, with its top five clients accounting for 26.51% of FY2025 revenue.[6] This concentration is compounded by structural capacity constraints.[1, 10] In early 2026, computing power shortages forced the company to cut its GLM Coding Plan daily registration quotas by 80%, resulting in refund requests and limiting near-term customer acquisition.[1, 10]
As a developer of frontier LLMs, the company is subject to regulatory oversight.[1, 9] It operates under strict guidelines established by the Cyberspace Administration of China (CAC) regarding algorithm registration, generative AI content filters, and state-aligned safety benchmarks.[9, 22] Any compliance failure could result in platform suspension, fines, or brand damage.[9]
Prior to its IPO, the company faced capital constraints, with an annual free cash flow of -RMB 2.69 billion and less than one year of cash runway.[19, 31] While the IPO raised HK$4.9 billion, the company continues to run a high R&D burn rate.[6, 11] If the proposed secondary listing on Shanghai's STAR Market is delayed or rejected, the company may face cash constraints to fund its model training roadmaps.[22, 23]
The Chinese AI industry is constrained by geopolitical restrictions.[1, 8] US embargoes on advanced semiconductor shipments limit direct access to top-tier training chips, leaving the company reliant on domestic fabricators whose production yields are vulnerable to supply-chain disruptions.[1, 8] On a macroeconomic level, technology valuations remain sensitive to global capital flows and Federal Reserve interest rate policy, with shifts in global liquidity impacting high-multiple assets.[28]
To analyze the total return potential for Knowledge Atlas Technology over a five-year holding period (FY2026 to FY2030), a quantitative model was constructed.[3, 13, 29] The current share price of HK$2,060 is utilized as the baseline valuation point.[25]
$EPS_{Year5} = \frac{\text{Net Income}}{\text{Shares Outstanding}} = \frac{\text{HK\$6,552,000,000}}{\text{520,000,000}} = \text{HK\$12.60}$
$EPS_{Year5} = \frac{\text{HK\$11,700,000,000}}{\text{500,000,000}} = \text{HK\$23.40}$
$EPS_{Year5} = \frac{\text{HK\$1,100,000,000}}{\text{550,000,000}} = \text{HK\$2.00}$
| Scenario | Revenue in Year 5 (HKD) | Margin / Earnings Assumption | Valuation Multiple Assumption | Current Share Price (HKD) | Implied Future Share Price (HKD) | 5-Year Total Return | Annualized Return (CAGR) | Probability |
|---|---|---|---|---|---|---|---|---|
| High Case | 65.0 Billion | 18% Net Margin / Net Income HK$11.70B | 50.0x Forward P/E | 2,060.00 | 1,170.00 | -43.2% | -10.6% | 20% |
| Base Case | 46.8 Billion | 14% Net Margin / Net Income HK$6.55B | 35.0x Forward P/E | 2,060.00 | 441.00 | -78.6% | -26.9% | 50% |
| Low Case | 22.0 Billion | 5% Net Margin / Net Income HK$1.10B | 15.0x Forward P/E | 2,060.00 | 30.00 | -98.5% | -56.8% | 30% |
VALUATION GRAVITY IMMINENT
Insiders and co-founders maintain significant share ownership.[18, 32] Co-founder Tang Jie owns a direct economic interest valued at approximately HK$66.6 billion, and Chairman Liu Debing controls shares worth approximately HK$70 billion.[9] The company has implemented employee shareholding platforms (Huihui and Zhideng) representing 16.55% of total equity.[9, 18] This structure aligns the interests of nearly half of the workforce (451 employees and consultants) with long-term enterprise value.[9]
The company is transitioning from on-premise custom integration contracts, which historically had long payment collection cycles, to highly scalable cloud-based APIs and recurring software subscriptions.[9, 13] This is evidenced by ARR expanding sequentially to $250 million by March 2026.[2, 11] However, near-term revenue quality is affected by the low gross margin profile (0-5%) of public API calls, which are pressured by compute subsidies.[7]
Knowledge Atlas is the leading independent general-purpose LLM developer in China, commanding a 6.6% market share by revenue.[2] Its model portfolio is integrated into major domestic industries, including the state power grid, PetroChina, and major commercial banking systems.[9] Its inclusion in the Hang Seng Tech Index further solidifies its domestic positioning.[10, 17]
The growth outlook remains strong.[3, 13] The transition to Agentic Engineering, the scale of developer subscription platforms (such as the GLM Coding Plan and the Claw Plan), and the projected revenue expansion to HK$46.8 billion by 2030 represent a strong growth profile.[8, 13]
The company’s financial health is constrained by high cash burn, with a statutory net loss of RMB 4.72 billion in FY2025 and a free cash flow trend of -RMB 2.6 billion.[6, 19, 31] While its cash position of RMB 2.26 billion was bolstered by the January IPO, the company remains dependent on secondary market equity raises (such as the STAR Market application) to sustain its intensive R&D program.[6, 11, 22]
The durability of the business is supported by its academic heritage, state-backed customer relationships, and proprietary patents.[1, 7, 9] The primary risk to viability is compute access under US technology restrictions.[1, 8] However, its partnership with Huawei Ascend suggests that the core model architecture is adapted to survive hardware restrictions.[1, 8, 17]
Management is allocating capital directly into frontier R&D and compute infrastructure, which is necessary to maintain competitiveness.[6, 24] This reinvestment prioritizes market share expansion and model capacity over near-term shareholder return or capital distributions, which may lead to ongoing dilution.[6, 11, 22]
Sell-side sentiment is positive, with 13 "Buy" and 3 "Hold" recommendations.[33] Consensus price targets have risen, led by JPMorgan's target upgrade to HK$1,800, though the current trading price remains above the average consensus target of HK$1,306.16.[27, 33]
The company is unprofitable, with statutory net losses widening to RMB 4.72 billion and a net margin of -264.86% in FY2025.[5, 6] Loss ratios exceed 10:1 relative to revenue in certain quarters, and structural breakeven is not expected to occur before fiscal year 2029.[3]
While the company has a strong record of technological milestones (creating China's first 100-billion-scale model and pioneering agentic software), its public market trading history is limited to less than six months and is characterized by high price volatility.[11, 25, 34]
HIGH-GROWTH CASH BURNER
Knowledge Atlas occupies a strong position in China's artificial intelligence and large language model sector.[2, 3] Its academic heritage, state-backed integration across infrastructure projects, and native adaptation to domestic hardware establish it as a key beneficiary of China's sovereign technology strategy.[1, 9] This competitive positioning is supported by strong operational momentum, as shown by its ARR expanding sequentially to $250 million by March 2026 and the positive reception of its flagship GLM-5.2 model.[2, 11, 17]
However, the investment thesis is constrained by valuation risk.[29] Trading at a trailing Price-to-Sales multiple of 1,113x, the company's valuation has decoupled from its fundamental growth path due to a limited initial free float (2.67%) and speculative retail inflows.[10, 29]
While consensus projections expect revenue to reach HK$46.8 billion by FY2030, multiple normalization is expected to limit total returns as the company's float expands.[13, 15, 29] The upcoming cornerstone lock-up expiration on July 8, 2026, and the secondary listing dilution on the STAR Market represent key catalysts that may lead to a valuation correction.[15, 22]
VALUATION GRAVITY AWAITS
Knowledge Atlas is in a strong medium-term upward trend, supported by the technical breakout of GLM-5.2 and JPMorgan's target price upgrades.[15, 17, 35] The share price of HK$2,060 is trading above its 50-day moving average of HK$1,495.14 and its 200-day moving average of HK$1,261.06, indicating positive technical momentum.[25, 35]
However, the daily 14-day Relative Strength Index (RSI) stands at 78.59, placing the stock in overbought territory and indicating short-term technical exhaustion.[29, 35] The near-term outlook is cautious, with potential downside pressure expected as the July 8, 2026, cornerstone lock-up expiration approaches, which will introduce new shares into the float.[15]
OVERBOUGHT CORRECTION RISK
View Knowledge Atlas Tech Joint (2513.HK) stock page
Loading the interactive version of this report…