Knowledge Atlas Technology (Zhipu AI) - 2513.HK

Research Report

Subject: Knowledge Atlas Technology Joint Stock Company Limited (智谱 / Zhipu AI / Z.ai), HKEX: 2513. Technology sector - artificial intelligence foundation models. Report date: 18 June 2026.

A note on reporting history before we begin. Zhipu listed on the Hong Kong Main Board on 8 January 2026. As a Hong Kong issuer it reports semi-annually, and it has published exactly one results set as a public company: full-year FY2025 (year ended 31 December 2025), released around 30 March - 1 April 2026, which sits comfortably inside the 90-day recency window. There are no six quarterly transcripts to draw on, because they do not exist yet. Where the standard report would lean on six concalls, this one leans on the single FY2025 results event, the management commentary around it, and the financial track record and forward statements disclosed in the IPO prospectus (which covered FY2022, FY2023, FY2024 and H1 2025). Sections 7 and 9 are written honestly against that constraint rather than padded with fabricated continuity. The next data point, H1 2026 interim results, is expected around August 2026.

Section 1: What the company does

Zhipu builds large language models - the same category of technology as OpenAI's GPT or Anthropic's Claude - and sells access to them to governments, large enterprises and developers, mostly in China and increasingly across Southeast Asia and the Middle East. Its model family is called GLM (General Language Model). The legal listed entity is "Knowledge Atlas Technology Joint Stock Company Limited," but everyone, including the company's own product branding, knows it as Zhipu (智谱) or, since a 2025 rebrand of its international platform, Z.ai.

There are two ways it makes money. The larger way is on-premise deployment: a Chinese ministry, state-owned enterprise, bank or energy company that cannot send its data to an external cloud buys a Zhipu model installed inside its own data centre, customised and fine-tuned to its needs, often bundled as a self-contained "AI-in-a-Box" server appliance. The smaller but faster-growing way is cloud deployment: developers and enterprises call Zhipu's models over the internet through an API and pay per use, or subscribe to packaged products such as a coding assistant or an enterprise agent.

The founding story matters because it explains the company's DNA. Zhipu was spun out of Tsinghua University in 2019, specifically the Knowledge Engineering Group, by professors Tang Jie and Li Juanzi. Tang Jie is one of the most cited academics in Chinese AI and had worked on BAAI's Wudao project, China's flagship early large-model effort. The company began life working on knowledge graphs (hence "Knowledge Atlas"), then pivoted hard into pre-trained language models. In 2021 it released the GLM framework, which it describes as China's first proprietary pre-trained large-model framework, and launched a Model-as-a-Service (MaaS) commercialisation platform the same year. That academic lineage gives it deep research talent and credibility with the Chinese state, but it also seeded a culture of publishing and open-sourcing models that, as we will see, is now central to its strategy.

The core problem Zhipu solves is this: a Chinese institution wants frontier-grade AI but cannot, for data-sovereignty, regulatory or security reasons, use a foreign model and often cannot use a public cloud at all. Zhipu gives them a domestically built, increasingly open model they can run behind their own firewall, with the company doing the integration work. For developers globally, the value proposition flipped in 2025-2026 to something sharper: Zhipu open-sources frontier-class models under permissive licenses, so a developer anywhere can download the weights and run a model that benchmarks near the best closed American models at a fraction of the cost.

What makes this hard to replicate is the combination of three things that rarely sit together: the research capability to train a genuinely frontier model, the systems engineering to make it run efficiently on domestic Chinese chips (it integrated FP8 and Int4 quantization on Cambricon silicon with GLM-4.6), and the trusted-vendor status to be allowed inside sensitive Chinese government systems. A pure research lab can do the first; a systems integrator can do the third; very few can do all three at once.

A concrete example: a provincial government bureau wants an internal assistant that can read regulations, draft documents and answer staff queries without any data leaving the building. Zhipu ships a GLM model installed on hardware inside the bureau's data centre, fine-tunes it on the bureau's document corpus, builds the agent workflows on top (its CoCo enterprise agent or AutoGLM-style automation), and provides ongoing customisation. The bureau pays a large project fee. This is high-touch, project-based, labour-intensive work - closer to systems integration than to shrink-wrapped software - which is exactly why the economics, covered below, are the central debate about this company.

Section 2: Business segments

Zhipu reports two operating segments. The split is by delivery model, not by product - the same underlying GLM models flow through both.

On-premise Deployment (~74% of FY2025 revenue)

This is the legacy core and still the majority of revenue. The company installs and customises GLM models inside a customer's own infrastructure. Customers are large Chinese government agencies, state-owned enterprises, financial institutions and energy companies - organisations for which data cannot leave the premises. Within FY2025, the "enterprise general-purpose models" line (the closest proxy for classic on-premise project revenue) was around RMB 366 million.

The core capability here is not just the model; it is the ability to deliver a working, fine-tuned, secured system inside a sensitive institution and be trusted to do so. That trust took years to build and is reinforced by Zhipu's Tsinghua/state lineage. What is hard to replicate is the combination of frontier model plus the integration muscle plus the security clearances and reference accounts.

It exists as a distinct segment because its economics are fundamentally different from cloud. This is project-based, heavily customised, labour-intensive delivery with long payment cycles - the model that first-generation Chinese AI firms like SenseTime ran. As these projects scale, the non-standardised marginal cost rises, which is why the on-premise mix has been dragging blended gross margin down (more on this in Section 6 and 10). Competitively, within China this segment competes against the cloud giants' enterprise arms (Alibaba, Baidu, Tencent) and against systems integrators, but Zhipu's frontier model quality and neutrality (it is not a hyperscaler trying to lock customers into its cloud) are its edges. In the group, management treats this as the cash-generating, relationship-anchoring base - the cow that funds the bet - even as it openly wants the mix to shift toward cloud.

Cloud-based Deployment (~26% of FY2025 revenue and the growth engine)

This is the API and product business: developers and enterprises calling GLM over the internet, plus packaged products like the GLM Coding Plan and enterprise agents. It contains two of the fastest-growing lines disclosed for FY2025: open-platform/API revenue nearly quadrupled to around RMB 190 million, and the enterprise AI agent line grew about 249% to around RMB 166 million. Geographically the cloud business is where the overseas push lives.

The core capability is different here: it is about model quality on public benchmarks (which drives developer adoption), pricing, and the breadth of the product surface (coding, agents, vision, video). This segment is what makes Zhipu look like a software company rather than an integrator - it is the "develop once, replicate infinitely" dream. The reason it is a separate segment is purely economic: it scales with near-zero marginal cost per additional API call once the model is trained, and its margin trajectory is the opposite of on-premise. Cloud gross margin went from roughly 3% in 2024 to 19% in 2025 as volume scaled and Zhipu raised prices (a 30% increase on coding and a large increase on API access in early 2026), evidence of genuine pricing power. Within this segment the competitors are everyone - DeepSeek, Alibaba Qwen, Moonshot Kimi, MiniMax, ByteDance Doubao - in a brutal price-competitive API market, plus the global open-weight field.

Management's framing, and the market's, is clear: cloud and agents are the future of the financial model, and the strategic objective is to grow this mix until Zhipu's economics look like software rather than services.

Section 3: Products and business detail

The product catalogue is a ladder of GLM model generations plus a set of applications built on top.

Foundation models (the GLM family). The lineage runs GLM framework (2021), GLM-130B (2022), ChatGLM-6B (the open consumer-grade chat model that gave Zhipu enormous developer mindshare), then the GLM-4 series. The agentic era began with GLM-4.5 (July 2025), an open-source model built for agent applications, and GLM-4.6 (late September 2025), which expanded the context window from 128K to 200K tokens and was notable for integrating FP8 and Int4 quantization on Cambricon chips - a domestic-silicon milestone given the US export situation. GLM-5 launched 11 February 2026 as the flagship for coding and long-running agent tasks, a roughly 744-billion-parameter sparse Mixture-of-Experts model with about 40 billion parameters activated per token. The line then moved fast: GLM-5.1 (200K context) and, on 16-17 June 2026, GLM-5.2, a ~744-753B MoE with a 1-million-token context window, released fully open-source under the permissive MIT license. GLM-5.2 was benchmarked as beating GPT-5.5 on several long-horizon coding tasks at roughly one-sixth the cost, and its open release triggered a one-day share surge of about 33%.

Multimodal models. GLM-V / GLM-4.6V are the vision-language models. "Ying" (and the CogVideoX line) handle video generation. These broaden the model surface beyond text into the multimodal demands of agents and content.

Applications and agents. AutoGLM is a general-purpose mobile agent that takes voice commands and executes tasks inside a smartphone; AutoGLM Phone extends this into action-taking on devices. CoCo is the enterprise agent. The GLM Coding Plan is the packaged developer-coding subscription, which by 2026 had surpassed 242,000 paying developers and became a globally visible product. The Z.ai platform is the international developer brand and console, supporting cURL, official Python/Java SDKs and OpenAI-compatible API patterns.

The MaaS platform is the commercialisation layer underneath all of it: by 2025-2026 it served around 2.9 million users (about 15% paying) and roughly 12,000 enterprise clients. When GLM-5 launched, 9 of China's top 10 internet companies (including ByteDance, Alibaba and Tencent) integrated it, several within 24 hours - a striking signal of domestic distribution.

Manufacturing / delivery. There is no physical manufacturing in the classic sense, but there is a real "production process": training frontier models requires enormous compute (R&D was about 70% computing cost in 2024), the models must then be engineered to run efficiently on available - increasingly domestic - hardware, and for on-premise customers Zhipu must physically deploy, fine-tune and secure the system on the customer's premises, sometimes as a hardware appliance. The binding constraint is access to advanced AI chips, which the US Entity List directly threatens (Section 8).

Geographies and export. The home market is China, contributing the overwhelming majority of revenue (on-premise China clients were ~88% of on-premise revenue in H1 2025). The overseas push is the strategic frontier: "AI-in-a-Box" sovereign-AI deployments to governments and enterprises in Southeast Asia and the Middle East, with offices in the Middle East, the UK, Singapore and Malaysia and partnership programmes spanning roughly 10 ASEAN and 10 Belt-and-Road nations. Open-sourcing the frontier models (MIT-licensed GLM-5.2) is itself a distribution strategy - it puts Zhipu's model in front of every global developer and seeds downstream commercial pull.

A defining milestone: on 8 January 2026 Zhipu became the first pure foundation-model company in the world to go public, and the first of China's "Six Tigers" of AI to reach an IPO. On 8 June 2026 it was added to the Hang Seng Tech Index alongside MiniMax.

Section 4: Customers

The customer base splits cleanly along the two segments.

On-premise buyers are large Chinese institutions: central and provincial government bureaus, state-owned enterprises, banks and insurers, and energy giants. Inside these organisations the buying decision sits with IT/digitalisation leadership and procurement, often with sign-off escalated to senior management because these are large, sensitive projects. The criteria are: data never leaves the building, the vendor is domestic and trusted, the model is frontier-quality, and the vendor can actually deliver and maintain a customised system. Sales cycles are long - these are tendered enterprise/government projects with extended payment terms.

Why they choose Zhipu specifically: it pairs a genuinely frontier model with the integration capability and the trusted, Tsinghua-rooted, state-aligned brand. A government buyer is choosing a partner it can be seen to trust, and Zhipu's "first foundation-model IPO / national champion" positioning reinforces that. Switching costs are high once installed: the model is fine-tuned on the customer's data, embedded in workflows and agents, and re-procuring and re-integrating a rival is expensive and risky. That said, concentration risk is real - large lumpy government contracts mean revenue can be uneven, and a project-based book is less predictable than recurring subscriptions.

Cloud / developer buyers are a very different population: 12,000-plus enterprise clients and millions of developers using the API and the GLM Coding Plan, plus the 9-of-10 top Chinese internet firms that integrated GLM-5. Here the decision-maker is a developer or an engineering lead, the criterion is price-performance on benchmarks (especially coding and agentic tasks), and the sales cycle is essentially self-serve and instant. Switching costs are low in principle - APIs are designed to be OpenAI-compatible, so swapping is easy - which is precisely why model quality and price matter so much and why Zhipu's ability to raise prices 30%+ in early 2026 without losing demand is meaningful evidence of stickiness through quality rather than lock-in.

Contract structure mix. On-premise is project-based and milestone-driven (lumpy, lower predictability, long cycles). Cloud is usage-based API plus subscription products (the coding plan, enterprise agent seats), which is where recurring revenue is building - cloud annualised recurring revenue was cited at roughly RMB 1.7 billion. The strategic direction is to grow the predictable recurring cloud book relative to the lumpy project book.

Section 5: Competitive landscape

This is one of the most crowded competitive arenas in technology. Zhipu competes on three fronts at once: against the other "Six Tigers" startups, against China's cloud and internet giants, and against the global frontier labs.

The startups ("Six Tigers"). MiniMax (also newly HKEX-listed, 0100.HK) is the closest peer by trajectory but a different strategy - consumer-first (the Talkie AI companion) with a large overseas revenue share. Moonshot AI (Kimi) is a strong coding/long-context rival. Baichuan, StepFun and 01.AI round out the group. DeepSeek, though not formally a "tiger," is the most important competitor of all on research credibility: backed by quant firm High-Flyer, it pursued frontier research without near-term commercial pressure and produced the V3/R1 architectural breakthroughs that reset the field.

The giants. Alibaba's Qwen is arguably the single strongest open-weight competitor globally and has deep cloud distribution. ByteDance (Doubao), Baidu (Ernie) and Tencent (Hunyuan) all field competitive models with vast user bases and infrastructure. Notably, Xiaomi emerged in 2026 as a surprise volume leader on some token-throughput measures.

The global frontier. OpenAI and Anthropic remain the closed-source quality benchmark; the entire Chinese open-weight cohort, Zhipu included, competes by getting close to that frontier and then giving the weights away.

Where Zhipu wins: it is, by 2026, a genuine global frontier participant in coding and agents (GLM-5.2 benchmarking near the best closed models), it has the best-in-class trusted on-premise/government franchise in China, and its open-source-plus-sovereign-AI playbook is a coherent distribution strategy. Where it is exposed: the API market is viciously price-competitive, model leadership is fragile and can be leapfrogged in a single release cycle, and against the giants it lacks their balance-sheet depth, cloud distribution and consumer install base.

Barriers to entry are paradoxical. Training a frontier model is extraordinarily expensive (Zhipu's R&D ran to roughly RMB 3.18 billion in 2025, ~44% of revenue) and requires scarce talent and chips - high barriers. But open-sourcing collapses part of the moat: once GLM-5.2 weights are public under MIT, anyone can run them, and rivals can distil from open models. The durable moats are therefore not the model itself but the trusted-vendor on-premise franchise, the domestic-chip engineering, and brand/distribution.

Market caps are rough peer-size references only, as of June 2026, and move significantly; private companies marked "—".

Section 6: Industry

Zhipu sits in the foundation-model layer of the AI stack - the companies that train large general-purpose models and sell access to them. Demand is driven by three forces: enterprise and government digitalisation (the desire to automate knowledge work), the agentic-AI wave (models that don't just chat but execute multi-step tasks and write code), and, in China specifically, a national-policy push for technological self-sufficiency that channels state and SOE spending toward domestic AI champions.

The market is large and growing fast but hard to size cleanly because it spans cloud API spend, enterprise AI services and sovereign deployments. The directional signals are vivid: Zhipu's own revenue grew 132% in 2025 to RMB 724 million and the company (and covering analysts) point to a projected trajectory toward roughly RMB 3.0 billion in 2026, RMB 7.8 billion in 2027 and RMB 21.0 billion in 2028 - aggressive figures that imply the addressable market is expanding far faster than typical software. On global token-throughput measures (e.g. OpenRouter data), Chinese open models collectively took 45%+ of weekly volume by April 2026, a structural shift toward Chinese open-weight models in global inference.

In the global supply chain, Zhipu is at the model layer, dependent below it on compute (chips and data centres - its critical constraint) and feeding above it into applications and agents. China's distinctive industry feature is a deployment-led playbook: where Western labs chase consumer scale and the largest possible models, Chinese labs including Zhipu emphasise efficient models that can be deployed into enterprises and onto domestic hardware.

Import-substitution dynamics are central and run two directions. On chips, China is being forced toward domestic substitutes (Cambricon, Huawei) because US export controls restrict advanced foreign silicon - Zhipu's FP8/Int4 work on Cambricon is exactly this. On models, Chinese open-weight models are now substituting for American models in much of the developing world and in cost-sensitive global developer use, a reverse import-substitution where China is exporting the substitute.

Regulation shapes everything. In China, generative-AI services require regulatory approval and content controls, which favours trusted domestic incumbents. Externally, the US Entity List designation (Section 8) is the defining regulatory fact for Zhipu. Cyclicality is less an economic cycle than a technology cycle: leadership turns over with each model generation, and capital availability for loss-making AI labs is sensitive to sentiment.

Industry tailwinds: relentless enterprise AI adoption, the agentic shift expanding usage, national self-sufficiency spending, and global demand for non-US open models. Industry headwinds: brutal API price wars compressing margins, the chip-access ceiling on Chinese labs, and the structural reality that open-sourcing erodes the value capture of the very models being given away.

Section 7: Growth triggers

These are drawn from the FY2025 results commentary (released around 30 March - 1 April 2026), management's IPO-era statements, and dated product announcements. Because Zhipu has only one public results event, several "triggers" are product-launch and strategy statements rather than recurring-concall guidance; each is dated.

• GLM-5 series capability ramp driving API volume and pricing. Management and covering analysts frame enhanced GLM-5 capabilities as the engine for an exponential increase in API call volumes and the ability to raise prices. (FY2025 results commentary, ~Apr 2026.) Evidence already on the table: Zhipu raised coding-plan prices ~30% and API prices sharply in February 2026 while demand held.

• Enterprise AI agent business scaling (the "Claw/Clip Plan" agent push). The agent line grew ~249% in FY2025 and management points to agent products as potentially central to the financial model going forward. (FY2025 results commentary, ~Apr 2026.)

• Sovereign-AI expansion overseas. Expansion of sovereign/on-premise models into overseas markets - Southeast Asia and the Middle East, via "AI-in-a-Box" - is named as a primary 2026-2028 growth driver, supported by new offices in the Middle East, UK, Singapore and Malaysia. (FY2025 results commentary, ~Apr 2026.)

• Open-sourcing frontier models as a distribution flywheel. The MIT-licensed open release of GLM-5.2 (16-17 June 2026) with a 1M-token context window is the most recent and concrete trigger - it expands the global developer base that downstream commercial cloud revenue is built on. (Product announcement, 16-17 Jun 2026.)

• Domestic-chip engineering reducing dependence on restricted hardware. Continued FP8/Int4 quantization work on Cambricon silicon (begun with GLM-4.6, Sept 2025) is positioned as enabling Zhipu to keep scaling despite Entity List constraints. (Product announcement, Sept 2025; reiterated through 2026.)

• Cloud margin inflection. Cloud gross margin moved from ~3% (2024) to ~19% (2025); management's strategy is to keep growing the cloud mix so blended economics improve over time. (FY2025 results commentary, ~Apr 2026.)

Section 8: Key risks

US Entity List designation (high probability, structurally severe). In January 2025 the US Commerce Department's BIS added Beijing Zhipu Huazhang Technology and subsidiaries to the Entity List, with the aggravating "Footnote 4" designation, which extends license requirements not only to US-origin items but to foreign-produced items containing US technology, software or components. Mechanism: this restricts Zhipu's access to advanced AI chips and tooling, the single most important input for training frontier models. The company publicly stated the inclusion "lacks a factual basis" and would not have a substantial impact, and it has responded by engineering for domestic chips (Cambricon). But the risk is that if Chinese domestic compute cannot keep pace with the frontier, Zhipu's models fall behind in a field where one generation of lag is fatal. This is not a tail risk; it is the live, defining constraint.

Persistent and widening losses funded by external capital. FY2025 adjusted net loss was about RMB 3.18 billion, widening ~29% year-on-year, with R&D at ~RMB 3.18 billion (44% of revenue). Mechanism: the business burns cash at a scale far above its revenue, and that is sustainable only while capital markets and strategic investors keep funding it. A sentiment shift, a failed model generation, or a financing window closing would force a painful retrenchment in the R&D spend that is the entire competitive premise. Management itself attributes the widening loss explicitly to "substantial R&D spending."

Margin erosion from the on-premise mix and API price wars. Blended gross margin fell from ~64.6% (2023) to ~56.3% (2024) to ~41% (2025). Mechanism: on-premise project work carries rising non-standardised marginal cost as it scales, and the cloud API market is a price war. If the hoped-for cloud margin inflection stalls or the on-premise book stays dominant, the company could grow revenue while gross profit lags - precisely the FY2025 pattern (revenue +132%, gross profit only +69%).

Open-source self-cannibalisation. Zhipu's distribution strategy is to give away frontier weights under MIT. Mechanism: the more valuable the open model, the more it undercuts the company's own paid API and lets competitors distil from it. The strategic bet is that open weights pull commercial deployment and brand; if that pull is weaker than hoped, the company will have armed its rivals for free.

Model-leadership fragility. Competitive position rests on being at or near the frontier in coding and agents. Mechanism: a single strong release from DeepSeek, Qwen, Moonshot or an American lab can reset the benchmark hierarchy overnight, eroding both developer mindshare and pricing power. The field already turns over every few months.

Lockup expiry and shareholder overhang (near-term, sentiment). Around 8 July 2026, roughly 11.6% of shares unlock, with the largest unlocking holder being a state-backed cornerstone, and early pre-IPO investors (Alibaba, Tencent, Meituan, HongShan, multiple government funds) sitting on paper gains reportedly exceeding 15x after a ~6-month run. Mechanism: strong incentive to take profit can pressure the stock; comparable high-valuation HK new-issue unlocks have historically been followed by 20-40% pullbacks. This is a share-price risk, not a business risk, but it is imminent.

Section 9: Walk the talk

This section is constrained by reality: Zhipu has been public for barely five months and has reported once. There is no six-concall track record to audit. What can be done is to test the company's IPO-era commitments and its single FY2025 results against subsequent facts, and to read its multi-year operating record from the prospectus. I will not manufacture a pattern that the limited public history cannot support.

The pre-IPO commitment that can be checked is the strategic pivot Tang Jie articulated after the DeepSeek shock of early 2025: a stated ambition to achieve global leadership in large models by 2026, via new model architectures, reinforcement-learning research and the GLM-5 line. Measured against this, the company has, so far, delivered concretely. GLM-5 shipped on schedule in February 2026, and GLM-5.2 (June 2026) was independently benchmarked as beating GPT-5.5 on several long-horizon coding tasks at a fraction of the cost - a credible claim to frontier participation in at least the coding and agent domains. On the narrow question "did they ship the models they said they would, on time, at the quality they implied," the answer through mid-2026 is yes.

On commercial guidance, the IPO and FY2025 framing set a steep trajectory (revenue toward ~RMB 3.0 billion in 2026 against RMB 724 million delivered in 2025). FY2025 revenue grew 132%, a strong number, but it missed the consensus analyst estimate of roughly RMB 760 million - a small miss, but a miss on the very first scorecard as a public company, which is worth noting against the bullish guidance. More importantly, the FY2025 results show the company delivering on the parts of the story it controls (model quality, API and agent growth of ~4x and ~249% respectively, price increases that stuck) while the loss widened and gross margin compressed - so management is delivering growth and capability but has not yet demonstrated the path to the improving economics it has promised. That demonstration is still entirely forward-looking.

The honest read: this is a young public company with a strong, verifiable track record of shipping frontier models on time and a credible-but-unproven story on economics. The credibility that exists is engineering credibility (they build what they say they will build). The financial credibility - turning frontier models into improving margins and eventually profit - cannot yet be assessed because there is no public history to assess it against. The next true test is the H1 2026 interim result in August 2026.

Section 10: Shareholder friendliness index

Dividends. Zhipu pays no dividend and should not be expected to. It is a deeply loss-making, growth-stage AI lab (FY2025 adjusted net loss ~RMB 3.18 billion) that earmarked roughly 70% of its IPO proceeds for R&D. There is no dividend history to report across the last three years and no realistic prospect of one while the company is burning cash to fund frontier model development.

Buybacks and dilution. No buyback programme has been announced or executed - again, entirely expected for a company that listed only in January 2026 and is consuming, not returning, capital. On dilution: the relevant dynamic is the opposite of buybacks. Zhipu raised roughly US$1.5 billion across many private rounds before listing and then raised a further ~HK$4.3 billion (US$560 million) in the January 2026 IPO by issuing new H-shares, so the share count has been growing as the company funds itself. The near-term share-supply event is not a buyback but the ~8 July 2026 lockup expiry releasing ~11.6% of shares. No MoatMap insider/buyback data block was supplied with this brief, and a manual review of HKEX disclosures finds no repurchase activity since listing; for completeness, the absence covers both the recent (post-listing) window and the company's entire short public life - there is simply no buyback to report from any window because the company has only existed publicly for ~5 months.

Verdict: Hoards Capital (by design). Zhipu returns nothing to shareholders and is in a heavy capital-consumption phase - appropriate for a frontier AI lab, but shareholders are buying growth optionality, not capital return, and should expect continued dilution and zero distributions for the foreseeable future.

Section 11: Insider activities

The listing venue is Hong Kong, so the primary source is HKEX Disclosure of Interests (DI) filings (Forms 3A/3B for directors and chief executives, and substantial-shareholder notices) on the HKEX news portal. No MoatMap insider/buyback data block was injected into this brief, so the assessment relies on HKEX DI and the IPO/lockup framework rather than a feed.

Recent transactions. Zhipu has been listed only since 8 January 2026, and its founders, controlling shareholders and the eleven cornerstone investors are subject to standard post-IPO lockups (the cornerstone/pre-IPO lockup running to on/around 8 July 2026). As a direct consequence, there have been no material open-market director or controlling-shareholder dealings in the period since listing - the shares are contractually locked. This is the expected and benign state for a company in its first five months public; it is not a signal in either direction.

Buys - read the signal. There have been no open-market insider purchases to interpret. None would be expected during lockup, and insiders already hold very large pre-IPO positions, so there is no conviction-buying signal available to read at this stage.

Sells - work out the why. There have been no completed insider sells, again because of the lockup. The forward-looking item to watch, rather than a past transaction, is the ~8 July 2026 lockup expiry: roughly 11.6% of shares unlock, with a state-backed cornerstone the largest unlocking holder, and early backers (Alibaba, Tencent, Meituan, Xiaomi, HongShan and multiple government funds) sitting on paper gains reported above 15x. The incentive to trim is strong, but any such sale would be a portfolio/profit-taking decision by financial investors after a ~6-month run, not necessarily a verdict on the business - the reason, if selling occurs, would most plausibly be diversification of an enormous unrealised gain rather than a fundamental concern. As of this report no such sales have been disclosed.

Net assessment. Insider activity is effectively nil due to the lockup, so there is no buy/sell signal to extract from the last twelve months. This is neutral - the genuine read is simply "locked up, nothing to see yet." The meaningful event is ahead, not behind: how founders, the state-backed cornerstones and the strategic investors behave at and after the July 2026 unlock will be the first real insider signal this company produces, and it is worth watching closely. Per the disclosure standard: Hong Kong HKEX DI data is publicly accessible, and the finding is that no material insider transactions exist to report because of the post-IPO lockup, not because data was unavailable.

Section 12: Scenarios

Bull case. Zhipu's pivot to frontier leadership compounds. GLM-5.2 and its successors stay genuinely competitive with the best American models in coding and agents, and the open-source-under-MIT strategy works exactly as intended: the free weights become the default substrate for developers across the non-US world, and that mindshare pulls a fast-growing, high-margin commercial cloud and agent business behind it. The price increases of early 2026 prove to be the start of durable pricing power. Sovereign-AI "AI-in-a-Box" deals close across Southeast Asia and the Middle East, giving Zhipu a second growth leg that is less exposed to the Chinese domestic price war. Cambricon and other domestic chips keep pace well enough that the Entity List becomes an inconvenience rather than a ceiling. The cloud mix climbs, blended margins inflect upward, and the company walks the path from RMB 724 million toward the multi-billion trajectory it has sketched. In this world Zhipu is one of a small handful of globally relevant foundation-model companies and the clear flag-carrier of Chinese open AI.

Base case. Zhipu remains a credible top-tier Chinese lab and a real global participant in coding and agents, but in a permanently crowded, fast-moving field where it trades the benchmark lead back and forth with DeepSeek, Qwen, Moonshot and the American frontier. Revenue keeps growing quickly off a small base, led by API and agents, while the on-premise government franchise provides a steady, lumpy floor. The cloud margin keeps improving but blended margins stay pressured by the on-premise mix and price competition, so losses narrow only gradually and the company keeps consuming capital. Overseas sovereign deals materialise but more slowly and in smaller size than the bull case. The Entity List forces ongoing engineering compromises that cost some frontier ground but not catastrophically. The stock, having run ~15x in six months, lives or dies in the near term on the July lockup and on whether each model generation lands - high volatility around a genuinely growing business.

Bear case. A model generation disappoints, or rivals leapfrog decisively, and Zhipu's claim to frontier participation erodes - in which case both developer mindshare and pricing power evaporate, because in an open-weight world there is nothing to retain users who can switch APIs in an afternoon. The Entity List bites harder than expected: domestic chips cannot keep pace, training the next frontier model becomes impossible at competitive cost, and Zhipu falls a generation behind in a field where that is fatal. The open-source strategy turns out to have armed competitors more than it built a commercial funnel. Meanwhile the RMB 3 billion annual cash burn collides with a colder capital-markets sentiment, forcing a cut in the very R&D that justifies the company. The July lockup unleashes heavy selling by 15x-up early backers, the share price re-rates down hard, and a company valued as a frontier champion is repriced as a loss-making regional integrator with a commoditising product. The on-premise government book survives, but the equity story does not.

Section 13: Further reading

• AI Economics in the East - MBI Deep Dives, 2026-04-13 [paid]
• AI Economics in the East: Part 2 - MBI Deep Dives, 2026 [paid]
• AMD MI355X GLM-5 Inference: Up to 40% Cheaper per Million Tokens than B200 on SGLang FP8 - SemiAnalysis, 2026 [free] - Technical benchmark of Zhipu's GLM-5 inference economics on AMD vs NVIDIA silicon, relevant to the model's cost-competitiveness and chip-access constraints.