In January 2026 Zhipu became the first foundation-model AI company to list anywhere — and its shares promptly surged more than 500%. Yet the same prospectus that made it the “global first LLM stock” disclosed a cumulative ~¥6.2 billion net loss on barely ¥685 million of cumulative revenue.

The genuinely open question is not whether Zhipu is capable — it is the #1 independent Chinese LLM maker, its open-weight GLM models top global open leaderboards, and FY2025 revenue more than doubled[10][36]. It is whether a Tsinghua-bred frontier lab that earns ~85% of its money from low-margin government and enterprise deployment, builds on US-restricted compute, and trades at a triple-digit sales multiple can convert technical standing into a durable, profitable business as DeepSeek and big tech bear down[16][47]. This study lays out both cases — and weighs them: the evidence leans toward Zhipu surviving the shakeout as the best-funded independent (medium confidence), and just as clearly toward a share price running well ahead of the disclosed economics (high confidence). The closing weighing in the Forward View states each lean, its confidence and the tripwires that would change it.

The decisive questions

Each links to the section that lays out the evidence on both sides.

The climb that frames the debate

Disclosed revenue by year (RMB millions). Growth is real and fast — roughly doubling annually to ¥724M in FY2025 — but the absolute base is small for a company that briefly carried a multi-hundred-billion-HK$ market cap.

How to read this

Nine sections, each built the same way: a neutral synthesis, framework visuals, a two-sided case-for / case-against ledger, dated and bilingual quotes, and the sources used. Start with the question that interests you, or read in order from Overview & Timeline.

Zhipu AI (智谱AI, internationally Z.ai) is a 2019 spinout of Tsinghua's Knowledge Engineering Group that builds the GLM family of foundation models, the 智谱清言/ChatGLM consumer app and AutoGLM agents. In January 2026 it became the first foundation-model company to IPO anywhere — a national-champion milestone — while still deeply loss-making.

What Zhipu actually does

Zhipu is a full-stack frontier lab: it pretrains its own GLM large language models rather than fine-tuning someone else's. The product surface spans the open-weight GLM series (GLM-4.5, 4.6 and the GLM-5 line), multimodal CogView/CogVideoX, the AutoGLM / Z.ai agents, and the consumer assistant 智谱清言 (ChatGLM), launched in August 2023[3]. Its roots are academic — chief scientist Tang Jie heads Tsinghua's KEG lab, and CEO Zhang Peng and co-founder Li Juanzi came from the same group[1][2].

How it got here

Zhipu raised more than ¥8.3 billion across eight rounds from a broad cap table — Tencent, Alibaba, Meituan, Ant, Xiaomi and Hillhouse alongside heavy state money (Beijing, Hangzhou, Zhuhai, Chengdu) — reaching a ~¥24.4B pre-IPO valuation[38][39]. The defining external shocks were the January 2025 US Entity List addition, which forced a domestic-compute pivot[47], and DeepSeek's 2025 breakout, which reset the whole field. In January 2026 it listed on the Hong Kong Stock Exchange — the “global first LLM stock” — and the shares then surged more than 500%[4][40].

The timeline

China launched ~305 large models in the '百模大战' (war of a hundred models), but the consensus is the market consolidates to 3–5 survivors[9]. A 2024 price war cut average LLM API prices ~92%, and DeepSeek's free open models reset the field in 2025[11]. Zhipu enters that shakeout as the #1 independent Chinese LLM maker (6.6% share, #2 overall) — the only non-big-tech firm in the top five[10].

The boom, the price war, the cull

After Baidu's ERNIE Bot and Alibaba's Qwen launched in 2023, roughly 305 large models appeared within a year, only ~40% of them completing regulatory filing[9]. In May 2024 DeepSeek-V2 triggered a brutal price war; Alibaba and ByteDance slashed API prices, and by late 2024 average China LLM prices had fallen ~92% versus May[11]. That structural margin compression is the backdrop to every Chinese model maker's economics — and to Zhipu's near-zero cloud-API gross margin.

Where Zhipu sits

By 2024 LLM-related revenue, Frost & Sullivan ranks Zhipu #1 among independent Chinese LLM developers and #2 overall at 6.6% share — the only independent in the top five, the rest being internet giants[10]. That framing is the bull case in one line: among companies whose only business is foundation models, Zhipu leads. The bear case is that the giants above and beside it (Alibaba, ByteDance) and the disruptor below it (DeepSeek) are exactly the players an independent must out-run.

The 2026 listing wave

The market's verdict turned euphoric in early 2026: Zhipu IPO'd on 8 January and MiniMax doubled on its HKEX debut a day later, opening a wave of Chinese AI listings[12]. Yet domestic commentators increasingly call the “六小虎/六小龙” (six-tiger) framing a previous-generation label, arguing the story has shifted from telling a tech narrative to proving commercial value — and that valuations compress as model supply grows[13].

About 84.5% of 2024 revenue came from local/on-premise deployment for government, finance and energy clients (gross margin ~66%), versus just 15.5% from cloud/API — which ran a gross margin of 3.4%, even slightly negative in H1 2025[15]. The whole investment case turns on whether Zhipu can shift that mix toward the higher-margin API/MaaS line, which the CEO targets at 50% of revenue[18].

How the money is actually made

Despite being a frontier model lab, Zhipu's revenue today looks closer to a systems integrator's: in H1 2025 over 80% came from localized deployment and customized AI solutions for large government, finance and energy customers[16]. Critics liken this to the “AI 1.0” project model of SenseTime — high-touch, custom, and hard to scale — with top-five customers a sustained 45–61% of revenue[15]. The cloud/API (MaaS) business, the part that should scale like software, is still small and was barely profitable.

Source: Zhipu prospectus. On-prem GM ~66%; cloud/API GM 3.4% (2024), -0.4% (H1 2025). [15]

The margin caveat

Zhipu's headline gross margin (56.3% in 2024) looks software-like next to rival MiniMax's 12.2% — but analysts note the gap is partly an accounting choice: Zhipu books most model-training spend as R&D rather than cost of sales, flattering reported margin[17]. R&D is overwhelmingly compute — salaries are only 14.8% of R&D, with over 70% going to compute service fees — and in FY2025 R&D ran about 4.4× revenue[49][29].

The pricing turn

Having cut prices in the 2024 war, Zhipu reversed in 2026: as GLM models improved it raised API prices — reportedly ~+83% for GLM-5 versus GLM-4.7, then a further ~+10% with GLM-5.1 — while call volume rose ~400% and cloud gross margin climbed from 3.4% toward ~18.9%[19][20]. That is the bull's proof that a strong model confers pricing power; the bear's reply is that the base is tiny and the margin still thin.

Running the CEO's 50% target through FY2025 margins

A back-of-envelope worth doing (our arithmetic, illustrative): hold FY2025 segment gross margins fixed — on-prem 48.8%, cloud/API 18.9%[15] — and shift the mix to the CEO's 50/50 target[18]. Blended gross margin computes to 0.5 × 48.8% + 0.5 × 18.9% ≈ ~34% — about seven points below the 41% Zhipu actually reported for FY2025[17]. For the target mix merely to hold blended margin at 41%, cloud/API gross margin would have to reach ~33% — nearly doubling again from 18.9%. The arithmetic cuts against both camps: the “lighter” model is not automatically a richer one at today's unit economics, which makes the cloud line's margin trajectory — not its revenue share — the single figure to watch.

Zhipu's rivals attack from every direction: DeepSeek's free open models reset mindshare and pricing, Alibaba (Qwen) and ByteDance (Doubao, ~260M MAU) bundle AI at scale, and the other “tigers” jostle for the same enterprise budgets[23]. Kai-Fu Lee argues only DeepSeek, Alibaba and ByteDance ultimately dominate — the sharpest statement of the bear case[21].

Five Forces

Industry attractiveness for a Chinese independent foundation-model maker, each force rated with a sourced basis. Click a force to see the evidence.

Who Zhipu competes with

DeepSeek is the disruptor — its free, MIT-licensed R1 sharply lowered the cost of deploying a top-tier model locally, directly hitting the 2B/2G order books of the tigers and making their fundraising harder[22]. Alibaba's Qwen leads global open-source download share (~60%) and rides Alibaba Cloud's distribution; ByteDance's Doubao (~260M MAU) and Alibaba's Quark (~180M MAU) dwarf any startup app[26][23]. Among the tigers, Moonshot/Kimi and MiniMax lean consumer, while Baichuan pivoted to healthcare and 01.AI stepped back from frontier pretraining[45].

Positioning

Mapping the field on openness (closed/proprietary → open-weight) against go-to-market (consumer-app-led → enterprise/infrastructure-led). Hover a point for the basis. Zhipu sits in the open-weight, enterprise-heavy quadrant — close to Alibaba's Qwen and apart from the consumer-led tigers and closed OpenAI.

The stated strategy is to stay at the open-weight frontier: MIT-licensed GLM models that top global open agentic/coding leaderboards, distributed to win developers worldwide[14][24]. The moat is meant to be developer gravity + enterprise/government relationships + state backing + a sovereign-compute pivot. The counter: open weights let rivals reuse GLM for free, and the compute the strategy depends on is US-restricted[32][31].

The open-weight bet

Zhipu open-sources its flagship models under the permissive MIT license, positioning GLM as “ARC” (Agentic, Reasoning, Coding) foundation models optimized for agent frameworks like Claude Code and Cline[27]. The traction is real: GLM-4.5 passed 10M Hugging Face downloads in 30 days with 230+ derivative models, and GLM-5 ranked #4 globally / #1 open on Artificial Analysis, with GLM-5.1 reported as the strongest open model on some coding benchmarks[24][14][33]. The logic, per CEO Zhang Peng, is that a strong-enough model confers pricing power — an Anthropic-style thesis.

The bear caveat: the absolute scale gap is vast — Anthropic's ARR is roughly 76× Zhipu's — and Zhipu's 2025 R&D (¥3.18bn) was about 4.4× its full-year revenue, so the analogy flatters a far smaller, far more loss-making business[29].

Sovereign compute — moat or constraint?

After the Entity List cut off Nvidia, Zhipu trained its GLM-Image model entirely on Huawei Ascend chips with the MindSpore framework — a fully domestic-compute training run, and a hedge that aligns with state priorities[30]. But it disclosed neither the chip count nor the efficiency versus equivalent Nvidia systems, and skeptics argue the episode shows export controls working: domestic chips lag a generation, SMIC yields ~50%, and ~85%+ of Chinese compute is still Western[31].

SWOT

Applied even-handedly — strengths and threats given equal weight.

Revenue roughly doubled every year to ¥724M in FY2025 (+132%), but the FY2025 net loss was ¥4.72bn — and cumulative losses reached ~¥6.25bn on ~¥685M of cumulative revenue[36][35]. Pre-IPO cash (~¥2.55bn) was thin against the burn rate, which is why the January-2026 HKEX raise — and a planned ¥15bn STAR Market secondary — matter so much[50][7].

The headline numbers

Source: Zhipu HKEX prospectus and FY2025 results (RMB). Figures reconciled to 亿-denominated originals. [34][36][35][49]

Losses outran revenue

The defining fact of the prospectus is that the loss line grew faster than the revenue line: the loss-to-revenue ratio climbed from ~250% (2022) to over 1,200% in H1 2025[49]. Net loss by year ran ¥143M → ¥788M → ¥2.96B, then ¥4.72B for FY2025[35][36]. R&D — overwhelmingly compute, with salaries just 14.8% of it — was the engine of the burn[49].

Funding and the listing

Pre-IPO, Zhipu raised more than ¥8.3bn across eight rounds to a ~¥24.4bn valuation, with a cap table spanning Tencent, Alibaba, Meituan, Ant and Xiaomi plus heavy state funds[38][39]. The January-2026 HKEX IPO priced at HK$116.20 and raised ~HK$4.35bn (~US$558M); 70% of proceeds were earmarked for 2026–2028 foundation-model R&D[5][37]. A planned STAR Market A-share secondary would add ¥15bn — ¥12bn of it for GLM-6 and compute[7].

The valuation gap

The market's reaction dwarfed the fundamentals: shares rose 500%+ within weeks (to HK$725 by 20 Feb), and by late May the stock had touched ~HK$1,995, briefly pushing market cap past HK$881bn[40][53]. Critics put the implied price-to-sales north of 100× (one analyst pegged fair value at HK$223 on a 100× multiple); the first post-IPO results — revenue doubling, but a ¥4.72bn loss — still sent shares up ~30%+, capturing the tension precisely[53][36].

What the price assumes

Date-stamping the marks: the IPO valued Zhipu at ~HK$51.8bn (8 Jan 2026)[5]; the peak market cap was ~HK$888bn (29 May 2026)[53] — a ~17× span in under five months — with implied price-to-sales running from ~244× in mid-January to ~890× at the peak and above 750× after the pullback to ~HK$630bn[55][54]. Peers frame the extremity: Moonshot's May-2026 round priced it at ~US$20bn on ~US$200M ARR — roughly 100× — and DeepSeek's reported target is ~US$59bn[42][41]. The implied read (our arithmetic, illustrative): for a 750× trailing multiple to compress to even a rich-software ~25× with no fall in the share price, revenue must grow ~30× from ¥724M — about five straight years of doubling (2⁵ ≈ 32×), or four years at the current +132% pace[54][36]. That is the bar both bull and bear are measured against: the bull must show the doubling persists; the bear needs only one year in which it does not.

Zhipu is the first of the group to be public and the #1 independent Chinese LLM maker — but it is bracketed by a free open disruptor (DeepSeek), consumer-led tigers (Moonshot, MiniMax), a big-tech incumbent (Alibaba/Qwen) and the global benchmark (OpenAI)[10]. Its differentiation is open-weight models plus enterprise/government depth; its challenge is that nearly everyone else is either cheaper, bigger or better-distributed.

The benchmark table

Private-company valuations (DeepSeek, Moonshot, OpenAI) are press estimates; Zhipu's and MiniMax's are public but volatile. Figures as of early–mid 2026. [44]

How to read it

On independence and openness, Zhipu and Alibaba's Qwen anchor the open-weight enterprise end, while DeepSeek combines openness with a consumer/developer pull Zhipu lacks[26][22]. On traction, the consumer-led peers (Moonshot's ~$200M ARR, MiniMax's apps) and big tech's hundreds of millions of MAU expose Zhipu's thin consumer presence[42][23]. On capital, being public is a genuine edge over the still-private DeepSeek and Moonshot — though DeepSeek's reported ~$50bn+ target valuation shows where investor enthusiasm concentrates[41].

What the comparison settles — and doesn't

It settles that Zhipu is a genuine top-tier independent with a real capital and listing advantage. It does not settle direction: a strong-model-equals-pricing-power path could lift it toward the survivors, while free open rivals and big-tech distribution could grind its margins for years. On the table's own evidence the lean is asymmetric: on capital and standing the edge is Zhipu's — the #1 independent and the first to list[10][4] — while on mindshare and distribution it belongs to DeepSeek and the giants[23][26]. This study therefore reads the peer set as supporting survival at medium confidence, not dominance.

The serious risks stack up: the US Entity List caps access to leading compute; the business loses billions and burns cash; DeepSeek and big tech squeeze both mindshare and margins; the stock trades at a triple-digit sales multiple on a 2.67% float with 2026–27 lockup cliffs; and heavy state ownership cuts both ways[47][35][55]. The counter: revenue doubling, returning pricing power, GLM's open-source standing, and durable state backing.

Compute & export controls

On 16 January 2025 the US added Beijing Zhipu Huazhang and affiliates to the Entity List, finding they “advance the PRC's military modernization” through advanced AI; a license is now required with a presumption of denial, cutting Zhipu off from leading Nvidia GPUs[47]. Zhipu called the move factually baseless with “no substantial impact” and pivoted to domestic chips, even training GLM-Image entirely on Huawei Ascend[48][30]. Skeptics counter that the episode shows controls working — domestic chips lag a generation, yields are low, and most Chinese compute is still Western[31].

Cash burn & commercialization

The loss-to-revenue ratio exceeded 1,200% in H1 2025, and at 30 June 2025 cash of ~¥2.55bn sat against ~¥7.09bn of net current liabilities — a thin runway that made the IPO existential rather than optional[49][50]. Caixin frames Zhipu bluntly as “cash-hungry,” with ~85% of sales in low-margin on-prem deployment[51]. A concrete commercialization warning sign: Zhipu's government large-model bid-wins slipped to #5 in H1 2025 from #3 a year earlier[52].

Valuation & lockups

The post-IPO move was extraordinary: ~16× to a peak near HK$1,995 and a market cap briefly above HK$881bn, an implied price-to-sales critics put in the hundreds — above 750× even after a pullback[53][54]. Only ~2.67% of shares truly float; cornerstones (~70% of the offer) unlock around July 2026 and 60%+ of early-holder shares in January 2027 — a wall of potential supply against a tiny float[55].

Governance & state ties

Leadership keeps control — the founder group plus employee platforms hold ~33% — alongside big-tech holders (Meituan ~5.5%, Tencent ~2.2%, Alibaba ~2.2%, Ant ~1.7%) and a deep bench of state funds[56]. Cumulative government support has been estimated at ~¥4.4bn[59]. That backing is a durable advantage at home — and, with Entity-List status, precisely what bars Zhipu from Western markets and raises independence questions. Internally, the org is Tsinghua-heavy (657 of 883 staff in R&D) and saw several commercialization-executive departures around mid-2025[57].

Forward view — three scenarios to weigh

Not predictions — conditions to watch. As of June 2026 the evidence sits closest to the base case: growth and the API turn are real[20], but project revenue still dominates[16] and the stock is priced beyond even the bull path[54].

The weighing

On whether the first listed LLM maker earns its valuation: the evidence leans no at current marks (high confidence on the gap, not on its timing). The controlling evidence is an implied price-to-sales critics put above 750× even after the pullback[54] and a true float of ~2.67% facing a cornerstone unlock around 7–8 July 2026 and a 60%+ early-holder unlock in January 2027[55], which outweighs the +131.9% revenue growth and ~60× MaaS-ARR jump[36][20] because even five more years of uninterrupted doubling would only bring the multiple back to rich-software levels with zero share-price appreciation. The strongest surviving counter-argument: scarcity as the only listed frontier lab plus ~¥4.4bn of cumulative state support can hold a narrative premium far longer than fundamentals justify — the shares jumped 30%+ on a ¥4.72bn-loss results day[59][36]. What would flip this reading: the July-2026 cornerstone unlock absorbed without a sustained drawdown, and FY2026 results (due ~March 2027) holding revenue growth above ~100% while the adjusted loss narrows. Pre-mortem: if this looks wrong in two years, the most likely reason is that MaaS ARR kept compounding fast enough for the fundamentals to chase the price — or, on the other side, that the lockup cliffs hit the thin float and the re-rating came harder and sooner than even bears modeled.

On frontier lab vs project-services shop: the evidence leans toward a real but unfinished transition (medium confidence). The controlling evidence is cloud/API gross margin rising from 3.3% to 18.9% while MaaS ARR grew ~60× to ~¥1.7bn[15][20], which outweighs the “SenseTime-like AI 1.0” charge because the scalable line is now simultaneously growing and gross-profitable — the one thing the bear reading said it could not do. The strongest surviving counter-argument: ~85% of revenue is still custom on-prem work whose own margin fell to 48.8%, top-five customers run 45–61% of revenue, and government bid-wins slipped to #5[16][15][52]. What would flip this reading: cloud/API below ~30% of revenue, or its gross margin back under ~15%, at the FY2026 results; conversely, reaching the CEO's 50% target early[18] would settle it for the bulls. Pre-mortem: if this looks wrong in two years, the most likely reason is a DeepSeek-class free model re-commoditizing the API just as the mix shifted — or, on the other side, 信创 on-prem demand proving a deeper, stickier margin pool than the integrator label implied.

On open-weighting GLM — moat or giveaway: genuinely contested, and the deadlock is specific. The distribution evidence is real — 10M Hugging Face downloads in 30 days with 230+ derivatives, GLM-5 ranked #1 open[24][14] — and so is the first monetization proof: API prices up ~83% while call volume rose ~400%[19]. But rivals reuse the weights free[32] and Qwen, not GLM, holds ~60% of global open-source download share[26]. What deadlocks it: the monetized API ARR (~¥1.7bn) is still smaller than the ¥3.18bn of R&D that produces the models[20][29]. What would flip it: API pricing holding through the next flagship release with ARR clearing ~¥3bn (covering R&D) — or a visible loss of the #1-open leaderboard slot to Qwen/DeepSeek. Pre-mortem: if this looks wrong in two years, the most likely reason is that developer gravity converted to paid usage faster than the giveaway eroded it — or, on the other side, that free rivals captured the derivatives and Zhipu paid to train everyone else's models.

On surviving DeepSeek and big tech: the evidence leans toward survival, not yet dominance (medium confidence). The controlling evidence is capital and standing — ~HK$4.35bn raised at IPO plus a board-approved ¥15bn STAR Market raise[5][7], as the #1 independent at 6.6% share and #2 overall[10] — which outweighs Kai-Fu Lee's three-winner forecast[21] because the consensus end-state of 3–5 survivors[9] leaves at least one seat for a funded, frontier-ranked independent, and Zhipu holds the best-evidenced claim to it. The strongest surviving counter-argument: distribution — Doubao's ~260M MAU and Qwen's open-source share mean the giants can out-bundle anything an independent ships[23][26]. What would flip this reading: the ¥15bn STAR raise failing or shrinking materially, or the next Frost & Sullivan ranking dropping Zhipu out of the overall top three; a further slide in government bid-win rank past #5 would also count[52]. Pre-mortem: if this looks wrong in two years, the most likely reason is that “survival” quietly became a Baichuan-style niche retreat — or, on the other side, that state-anchored procurement made the independent seat more profitable than the squeeze narrative allowed.

Method

Research proceeded by fan-out web search and direct fetching of primary and reputable secondary sources, in both English and Chinese. Every URL cited was opened and read during the run; each claim was transcribed into a structured manifest tagging it with a source tier, a confidence level and a stance (supporting / critical / neutral). Because Zhipu is a Chinese company, the bulk of the candid, current coverage — the prospectus breakdowns, the price-war and 'six-tigers' debate, the layoff and bid-win reporting — is in Chinese; roughly 60% of the sources here are Chinese-language, each carrying the original text alongside an English translation. The load-bearing figures — FY2025 revenue ¥724M (+132%) against a ¥4.72bn net loss, the cumulative ~¥6.25bn loss, the ~84.5% on-prem revenue mix, and the Jan-2026 HKEX listing — rest on Zhipu's prospectus and first post-IPO results as reported by Chinese financial press and confirmed in English coverage[36][35][15][4].

Frameworks used

The analysis applies the Pyramid Principle for the answer-first summary; Porter's Five Forces for the competitive landscape, each force rated with a sourced basis; a peer-comparables benchmark against DeepSeek, Moonshot/Kimi, MiniMax, Alibaba's Qwen and OpenAI; a 2×2 positioning map on open-vs-closed and consumer-vs-enterprise; a SWOT applied even-handedly; and a bull/base/bear scenario set that closes with an explicit weighing — a stated lean, a confidence level and concrete tripwires for each decisive question. BCG and Ansoff matrices were skipped — Zhipu's lines do not yet divide into separately-sized, independently-growing businesses, and an empty framework is worse than none.

Disclosed vs. estimated

Disclosed figures come from Zhipu's HKEX prospectus and its first post-IPO annual results — revenue, net loss, gross margin, the revenue mix, cash and the funding history are reported numbers, not estimates[34][38]. Market-share (the Frost & Sullivan 6.6% / #2 ranking) is a third-party estimate on a specific revenue definition[10]. Valuations and market caps are point-in-time and unusually volatile here — the stock swung from a ~HK$52bn IPO cap to a peak above HK$881bn — so any single number is a snapshot, and the implied multiples (e.g. P/S ~244x) are critics' calculations[53]. Peer figures for the private rivals (DeepSeek, Moonshot, OpenAI) are press estimates, not audited disclosures. Currency is RMB unless a source reports HK$ or US$; 亿 (100M) and 万 (10,000) units were reconciled against each source's own digits.

Neutrality & independence

This study compiles, then weighs. Every section pairs the case for and against with sourced evidence; the Executive Summary states where the evidence leans and at what confidence, and the Forward View closes with an explicit weighing — leans, confidence levels and tripwires — while stopping short of any buy/sell call. We researched in Chinese as well as English specifically to avoid an outsider/Western framing and to surface the domestic debate — including domestic criticism. The Teardown is independent and not affiliated with, or endorsed by, Zhipu AI, and this is not investment advice — no rating, price target, or recommendation to buy or sell any security. The achieved evidence mix (see the Sources) is balanced between supporting, critical and neutral citations, with a majority of home-language sources, by design.

Cross-checked at build time by an automated link checker; a few primary sources may be paywalled or bot-walled and were verified manually. See Methodology & Limits.