Zhipu AI (Z.ai)
Model-as-a-Service (MaaS) API platform + enterprise/on-prem model deployment + open-weight GLM models as ecosystem funnel; monetization shifting from on-prem projects to cloud API services (cloud segment 26.3% of FY2025 revenue, up from 15.5%)
Earnings, margins, COGS & capex
Classic frontier-lab P&L: FY2025 revenue RMB 724.33M (+131.9%, missing the ~RMB 756M estimate) against a RMB 4.72B total loss (+59.5% YoY) -- it loses roughly RMB 6.5 for every RMB 1 of revenue. R&D of RMB 3.18B dwarfs revenue 4.4x. The equity story rests not on current financials but on the MaaS pivot -- CEO Zhang Peng said on the earnings call that open-platform/API ARR reached ~RMB 1.7B (~$250M), up ~60x YoY (ARR, not recognized revenue), with MaaS gross margin lifted to 18.9% from 3.3% -- and on GLM-5's demonstration that Zhipu can train frontier models on 100% domestic silicon. Funding is equity-market-driven: ~$559M IPO (Jan 2026), ~$4B placement (Jul 2026), and a planned RMB 15B (~$2.2B) STAR Market A-share raise (tutoring accepted Jun 17, 2026).
Income statement — where each revenue dollar goes
% of revenueOf every $1 of revenue, ~81¢ is cost of goods and ~14¢ operating expense, leaving ~5¢ of operating profit (~-440¢ net).
Revenue trend
Margins
loss widening in absolute terms but narrowing relative to revenue growth
widening
improving on inference-side engineering optimizations
rising in absolute terms
COGS structure
Cost structure dominated by training/inference compute and talent, not disclosed line-by-line. Post-Entity-List (BIS, Jan 2025), frontier training is domestic: GLM-5 was trained on ~100,000 Huawei Ascend 910B chips with zero NVIDIA silicon. Domestic accelerators are likely cheaper per chip but lower performance-per-watt than NVIDIA, raising compute cost per useful FLOP; inference-side optimization lifted MaaS gross margin to 18.9%. Domestic inference support on non-Huawei silicon (Cambricon, Moore Threads) is reported but not company-confirmed in sourced reporting.
Capex
Not separately disclosed. Compute capacity buildout is the core use of proceeds: RMB 12B of the planned RMB 15B STAR raise is earmarked for general-purpose foundation-model R&D (RMB 2B for the MaaS platform, RMB 1B working capital); the Jul 2026 ~$4B placement is earmarked for R&D (hiring, computing capacity), business expansion, strategic investments and M&A, and working capital.
Latest earnings
Revenue miss: RMB 724.33M vs ~RMB 756M analyst estimate; total loss widened 59.5% YoY -- yet the stock closed +31.9% at HK$915 the next session (Apr 1), pushing market cap past HK$400B (~$51B) as AI momentum overrode the print
No formal numeric guidance; CEO Zhang Peng framed 2026 around the on-prem-to-cloud transition, noting clients who deployed open-source models locally are shifting toward the cloud API
- FY2025 revenue
- RMB 724.33M (+131.9%)
- FY2025 total loss
- RMB 4.72B (+59.5%)
- FY2025 adjusted net loss
- RMB 3.18B (+29.1%)
- FY2025 R&D expense
- RMB 3.18B (+44.9%)
- MaaS ARR (company-reported, unaudited)
- ~RMB 1.7B (~60x YoY); segment gross margin 18.9%
Growth drivers
- MaaS API platform scaling (~RMB 1.7B ARR in 2025 per the CEO's earnings-call disclosure, ~60x YoY) as the company pivots from on-prem project deployment to cloud services
- GLM-5 (Feb 11, 2026, 744B-param MoE, MIT-licensed) and GLM-5.2 ranking #1 among open-weight models on the Artificial Analysis Intelligence Index -- open-source funnel converting to paid API usage
- Sovereign/enterprise AI demand inside China and in non-US markets where a US-independent stack is a feature, not a bug
- Policy tailwind — Beijing's push for domestic AI champions; scarcity value as one of only two listed pure-play LLM labs (with MiniMax) on HKEX
- Planned STAR Market A-share listing (RMB 15B) adding mainland capital access and onshore investor demand
Bull & bear
Zhipu is the listed pure-play on China's entire sovereign-AI stack: if Chinese AI spend consolidates onto domestic, Entity-List-proof vendors, Zhipu is the category-defining asset with the models, the silicon independence, the policy backing, and now ~$4.5B+ of fresh capital.
- GLM-5 trained on ~100k Ascend 910B chips proves the US-independent stack works at the frontier -- the single most strategically valuable capability in Chinese tech, and Zhipu owns the flagship proof
- MaaS ARR of ~RMB 1.7B (company-reported), up ~60x in 2025 with segment gross margin already at 18.9%; if even half converts to recognized revenue in FY2026, revenue roughly triples again and the P/S de-risks rapidly
- Scarcity + flow: the only ways to own China's LLM race in public markets are Zhipu and MiniMax; the STAR dual listing adds onshore demand, index-inclusion potential, and a mainland bid
- Open-weight GLM-5.2 leading the Artificial Analysis open-weight rankings builds a global developer moat that survives the Entity List -- weights cross borders even when APIs cannot
- War chest (IPO + ~$4B placement + RMB 15B planned A-share raise) funds multi-year compute buildout and acquisitions while most private rivals scramble for capital
A ~$100B+ valuation on ~$105M of revenue and RMB 4.7B of widening losses is priced for flawless hyper-growth; the multiple rests on thin float, retail momentum, and scarcity that new listings will erode, while price competition and inferior silicon cap the path to profitability.
- ~900x trailing P/S (or ~400x the company-reported ~$250M ARR) with losses ~6.5x revenue -- the price implies multi-billion-dollar revenue at software margins within a few years, with no disclosed path
- Chinese model-API pricing is under sustained pressure from DeepSeek/Qwen/Doubao competition; Zhipu missed its first revenue estimate as a public company and its total loss grew ~60%
- Entity List (Jan 2025) forecloses US/Western enterprise revenue, confining Zhipu to the most price-competitive AI market on earth
- Float mechanics cut both ways: the same thin tradable share count that fueled a ~1,500% rally guarantees violent downside on sentiment reversal; the Jul 2026 placement priced ~13% below market and added only ~4.2% new shares
- Structural compute disadvantage: Ascend 910B on domestic fabrication trails NVIDIA frontier silicon; each frontier iteration costs Zhipu more per useful FLOP, and US tool controls could tighten the ceiling further
- Dilution treadmill: a ~$4B placement (~4.2% dilution) six months post-IPO plus a planned RMB 15B A-share raise signal ongoing heavy cash burn funded by shareholders -- and the stock fell on the STAR-raise announcement
What it is worth
Trailing/forward revenue multiples vs listed AI peers + reverse-DCF sanity check; no meaningful earnings-based method available (deeply loss-making)
~$10-25B (still ~40-100x the company-reported ~$250M ARR) on a China AI-sentiment reversal, an ARR-to-revenue conversion disappointment, tighter US tool controls hitting Ascend supply, or float expansion unwinding the squeeze -- the Jan 2026 IPO priced this same company at ~$6.6B
~$50-90B
if MaaS growth stays strong but decelerating, losses keep widening, and the scarcity premium erodes gradually as Moonshot and others list -- multiple compresses faster than revenue compounds
~$130B+ (revisits the late-Jun 2026 HK$1T peak) if FY2026 recognized revenue converges toward reported ARR (~$250M+) and keeps compounding, GLM retains open-weight leadership, the STAR listing lands with strong onshore demand, and Ascend next-gen silicon narrows the compute gap
At ~$100-115B market cap vs RMB 724.33M (~$105M) FY2025 revenue, Zhipu trades at roughly 900x+ trailing P/S, or ~400x the company-reported ~$250M MaaS ARR. Reverse-DCF framing: the price implies revenue compounding to a $5-10B+ run-rate within ~5 years at eventual software-like margins -- i.e., near-flawless execution in the world's most price-competitive AI market. The multiple is currently carried by scarcity (one of two listed LLM pure-plays), thin float (~4.2% added by the $4B placement), momentum, and the sovereign-AI narrative rather than by fundamentals. Vintage: Jul 2026. Not financial advice; no buy/own call is made on this or any mainland-China name.
SWOT
Strengths
- First-mover listed pure-play LLM lab globally (HKEX 2513, Jan 2026) with scarcity premium and ~$4.5B+ of fresh 2026 equity capital
- Proven frontier capability on 100% domestic silicon — GLM-5 (744B MoE) trained on ~100,000 Huawei Ascend 910B chips with zero NVIDIA dependency -- the strongest existence proof of a US-independent AI stack
- Open-weight GLM-5.2 ranks #1 among open-weight models on the Artificial Analysis Intelligence Index, driving global developer adoption and an ecosystem funnel
- Tsinghua-rooted research talent and a decade of KEG-lab lineage — broad strategic cap table (Alibaba, Tencent, Ant Group, Meituan, Xiaomi, Hillhouse, Qiming, Prosperity7/Aramco)
- Revenue growth inflecting — +132% FY2025 with company-reported MaaS ARR up ~60x and segment gross margin rising to 18.9%
Weaknesses
- Revenue (~$105M FY2025) is negligible against a ~$100B+ market cap (~900x trailing P/S); total loss is ~6.5x revenue and widening in absolute terms
- US Entity List (Jan 2025) blocks NVIDIA frontier-silicon access and most US enterprise customers/partnerships; overseas monetization structurally capped in Western markets
- Ascend 910B performance-per-chip trails NVIDIA frontier silicon — training efficiency and iteration speed depend on domestic fab (SMIC-class) supply and yields
- Thin tradable float amplifies volatility (the ~$4B Jul 2026 placement added only ~4.2% new shares); the ~1,500% rally is momentum- and scarcity-driven, not fundamentals-driven
- Revenue missed estimates in its first print as a public company — on-prem-to-cloud pivot creates near-term revenue-recognition turbulence
Opportunities
- China enterprise + government AI adoption with explicit policy preference for domestic, Entity-List-proof vendors
- Sovereign AI demand in the Middle East, Southeast Asia, and other non-aligned markets seeking a non-US stack
- STAR Market A-share dual listing (RMB 15B planned — tutoring accepted Jun 17, 2026, with Guotai Haitong and CICC) unlocking onshore capital and index inclusion
- MaaS ARR compounding if GLM-5/5.2 open-weight adoption converts to paid API usage at scale
- Potential consolidation — placement proceeds explicitly earmarked partly for strategic investments and M&A among China's fragmented AI tigers
Threats
- DeepSeek's efficient open models (V4 Pro, 2026) and Alibaba Qwen/ByteDance Doubao compress Chinese API pricing -- Zhipu already loses ~6.5x its revenue
- Moonshot AI's prospective HK IPO (targeting up to ~$18B valuation after a May 2026 $2B raise at $20B) and further AI listings dilute the scarcity premium supporting the multiple
- Tighter US export controls on chipmaking tools could constrain domestic-fab Ascend supply, throttling training capacity
- A China AI-sentiment reversal on HKEX would hit a thin-float, momentum-driven stock disproportionately hard (the stock fell on the STAR secondary-listing announcement itself)
- Frontier progress by OpenAI/Anthropic/Google widening the capability gap if domestic compute scaling stalls
Moats, dependencies & bottlenecks
Moats
Frontier capability on fully domestic silicon (GLM-5 trained NVIDIA-free on ~100k Ascend 910B) high within China; contingent on the Ascend/domestic-fab roadmap keeping pace The Entity List turned a liability into a moat inside China -- Zhipu is structurally the safest vendor for Chinese state and SOE buyers
Open-weight GLM ecosystem and developer mindshare (GLM-5.2 #1 on the Artificial Analysis open-weight index) DeepSeek (V4 Pro), Kimi K2.6, and Qwen fight for the same open-weight crown Open weights travel where the Entity List blocks APIs
Tencent, local-government funds, Prosperity7) high while Beijing prioritizes AI champions Preferential access to compute, government contracts, and the STAR listing channel
MiniMax listed the same week; Moonshot is weighing an HK IPO at up to ~$18B Cheap equity capital is itself a competitive weapon in a compute arms race
talent war with ByteDance/Alibaba/DeepSeek is intense
Dependencies
All frontier training now rides Ascend clusters; domestic fab yield/capacity and US tool controls are the hard ceiling
Burn is funded by serial equity raises; a sentiment reversal closes the funding window
regulatory/demand Policy is currently a tailwind but concentrates fate in one government's priorities
Alibaba (BABA) and Tencent (0700.HK) as strategic investors and cloud distribution channels distribution/strategic Both are also competitors via Qwen and Hunyuan
Already blocks NVIDIA silicon and US customers; further tightening (tools, HBM) would bite harder
Advantages
- Only battle-tested NVIDIA-free frontier training stack at scale (GLM-5 on ~100k Ascend chips) -- unique strategic asset inside China
- First-listed status = cheapest capital among China's AI tigers exactly when compute capex decides the race
- Open-weight leadership (GLM-5.2, MIT-licensed) sustains global relevance despite the Entity List
- Backing from nearly every Chinese tech major plus Aramco's Prosperity7 -- distribution and political cover
- Revenue growing >130% with an ARR pivot that, if audited numbers confirm it, would re-rate the fundamental story
Weaknesses
- Revenue (~$105M FY2025) is negligible against a ~$100B+ market cap (~900x trailing P/S); total loss is ~6.5x revenue and widening in absolute terms
- US Entity List (Jan 2025) blocks NVIDIA frontier-silicon access and most US enterprise customers/partnerships; overseas monetization structurally capped in Western markets
- Ascend 910B performance-per-chip trails NVIDIA frontier silicon — training efficiency and iteration speed depend on domestic fab (SMIC-class) supply and yields
- Thin tradable float amplifies volatility (the ~$4B Jul 2026 placement added only ~4.2% new shares); the ~1,500% rally is momentum- and scarcity-driven, not fundamentals-driven
- Revenue missed estimates in its first print as a public company — on-prem-to-cloud pivot creates near-term revenue-recognition turbulence
Bottlenecks
- Domestic accelerator supply — Ascend 910B allocation and domestic-fab yields gate training-run scale and iteration cadence
- Monetization conversion — turning open-weight adoption and company-reported MaaS ARR into recognized, margin-bearing revenue
- Performance-per-watt gap vs NVIDIA-class silicon raises cost per frontier training run
- Thin tradable float limits institutional position-building and amplifies volatility
- China-centric enterprise TAM at fiercely competitive API prices caps near-term revenue per token
Top signals & trends
Top signals
Onshore capital + mainland/index demand; also a dilution event -- the stock fell on the initial announcement
The single most important fundamental checkpoint for the multiple
Leaderboard slippage would undercut the ecosystem-funnel thesis
Better domestic silicon directly raises Zhipu's frontier ceiling
More float = less scarcity squeeze; the Jul 2026 placement priced ~13% below market
Each new listing dilutes Zhipu's scarcity premium and offers a cheaper relative entry
Zhipu is the poster child of the targeted category
Trends
Zhipu is the flagship beneficiary and existence proof
Funding window wide open as of Jul 2026 -- but windows close
Compresses API pricing; value migrates to compute, distribution, and agents
Caps Zhipu's global TAM while entrenching its domestic position
Zhipu's own pivot, confirmed by the CEO at FY2025 results; higher-quality recurring revenue if execution holds
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
Primary training silicon -- GLM-5 trained on ~100,000 Ascend chips; private company
China's leading foundry for domestic AI accelerators at 7nm-class nodes; the physical bottleneck of the Ascend supply chain (mainland listing 688981.SS also exists; analysis context only)
Domestic AI accelerator maker; GLM inference support reported (analysis context only, not a recommendation)
Domestic GPU maker; GLM inference support reported
STAR Market IPO tutoring institutions / sponsors
On-prem deployment legacy base migrating to MaaS; government/SOE procurement favors Entity-List-proof domestic vendors
Developers via open-weight GLM + MaaS API platform ~RMB 1.7B 2025 API ARR (company-reported on earnings call, unaudited; ARR, not recognized revenue)
Southeast Asia) US-independent stack marketed to non-aligned governments
Private (High-Flyer-backed); hyper-efficient MIT-licensed open models (V4 Pro) set the price floor for Chinese AI APIs
Qwen open-weight family + Alibaba Cloud distribution; also a Zhipu investor
Private; largest consumer AI distribution in China and massive compute budget
Second listed Chinese LLM lab (HK IPO Jan 2026, $619M raise, +109% debut from HK$165 to HK$345); closest listed comp
Private; raised $2B at $20B valuation (May 2026), weighing an HK IPO at up to ~$18B; Kimi ARR passed ~$200M in Apr 2026
Incumbent Chinese search/AI platform with its own model line and cloud
Private; global frontier leader -- capability benchmark Zhipu is measured against, though not a direct market rival post-Entity-List
Private; frontier US lab, enterprise/agentic leader; competes for global open-market mindshare
Frontier models + TPU vertical integration; the template for the full-stack play Zhipu attempts domestically
Competes directly for the open-weight developer ecosystem GLM-5.2 currently leads