TLDR
Zhipu AI's stock surged 25x in five months to a HK$1.33T market cap, then collapsed 80% in four months despite strong fundamentals. The episode dissects four layers of valuation collapse: scarcity premium evaporation, technical lead half-life, deteriorating unit economics, and a trust crisis. It argues the market is shifting from rewarding model leadership to demanding profitable unit economics and sustainable compute efficiency.
For you
PRO members can sign in and request more information — or their own version of this digest, tuned to what they care about.
Sign in to unlockSurprising Insights
Stock crashed 80% while revenue grew 4x, models improved, and >HK$75B was raised.
The first layer to collapse was scarcity premium, not fundamentals — liquidity increase and competitor IPOs erased it.
A model's leadership window can be weeks; Theta decay means time value erodes before next release.
Revenue growth accelerated cash burn because each API call consumes real compute; capacity became the revenue ceiling.
Primary market still values peers at 50-75x ARR while public market re-rates on quarterly unit economics.
Key takeaways
Scarcity premium is a valuation artifact, not a moat — it vanishes when competitors and liquidity increase.
Technical leadership has a short half-life (Theta decay); market now pays for sustained capability (org, data flywheel, compute autonomy), not one-time benchmark wins.
Revenue growth ≠ value growth when marginal costs are high; unit economics (revenue per compute dollar, gross margin) are the new focus.
Business model ambiguity (cloud-native vs IT services) caps valuation multiples; API margins are squeezed by compute costs and price wars.
Trust is revenue infrastructure for API companies; data privacy incidents directly threaten renewal rates.
Primary/secondary market valuation divergence (50-75x ARR vs public market) signals a pricing regime shift.
Long-term winners must sell intelligence at higher prices while lowering compute costs — profit determines the valuation anchor.
Notable quotes
“技术决定一家公司能否入场,商业模式决定它能够走多远,而利润决定它的中枢。”
▸ Notable quotes · 1:07“稀缺是一种估值,而不是一条护城河。护城河的定义是即便竞争者涌入它依然有效,而稀缺的定义恰恰是竞争者一旦涌入它就不复存在。”
▸ Notable quotes · 2:38“技术领先是一种会随时间自动贬值的资产……大模型是典型的相对性能资产,今天你的模型排在全球前三,不意味着三个月之后还在这个位置。”
▸ Notable quotes · 4:33“从卖项目到卖消耗,这是质的转变……每一元收入仍然对应着大约2.1元的经调整亏损。”
▸ Notable quotes · 6:16“性价比这个标签就从差异化优势变成了行业的准入资格。”
▸ Notable quotes · 8:04“过去三年大模型行业最重要的问题是谁的模型更强,而接下来的几年这个问题会逐渐变成谁能够把智能卖得更贵,同时把算力成本压得更低。”
▸ Notable quotes · 14:50Chapters
0:00 — Stock Rollercoaster & Scarcity Premium Collapse: Zhipu listed at HK$116.2, peaked at HK$2,980 (HK$1.33T market cap), then fell to HK$501.5 (HK$244.5B). The initial surge was driven by extreme scarcity (<3% free float, only 11.74M tradable shares) and status as China's only listed pure-play LLM giant. Scarcity premium evaporated after cornerstone unlock (25.68M shares) and a 20M-share placement at HK$1,588, which expanded float and coincided with Moonshot's Kimi K3 launch.
2:55 — Technical Lead Half-Life & Theta Decay: Moonshot's Kimi K3 (2.8T params) topped Arena, causing Zhipu to drop 28% then 19% in two days. JPMorgan cut target prices and long-term P/E, citing rapid rotation of model leadership. Zhipu's GLM-4-5.3 (750B params, post-training gains) was deemed a routine iteration, not a flagship leap. Technical lead is a decaying asset (Theta decay); leadership windows are weeks, not years, and competitors' release cycles are uncontrollable.
5:32 — Unit Economics & Business Model Pivot: H1 revenue HK$954M (+400% YoY), API share jumped from 26% to 86.5%, token volume up 40x, API gross margin improved from -0.4% to 24.6%. But blended gross margin halved to 26.4%, adjusted net loss widened to HK$1.96B. R&D spend 2.2x revenue. Each API call consumes compute — revenue ceiling = data center capacity. GLM-4 launch caused 10x demand spike, exhausted compute, forced Coding Plan suspension until July financing enabled expansion.
7:58 — Price War, Capital Pressure & Financing: DeepSeek V4.1 Flash, Moonshot, OpenAI, Anthropic all cut API prices. Zhipu raised Coding Plan prices, revealing weak pricing power. IPO proceeds (HK$4.9B) nearly depleted by June; July placement (HK$31.4B) and September 'small equity, large debt' deal (US$2B equity at HK$714, US$3B zero-coupon CB) were urgent. Total raised >HK$75B, but still thin vs big tech capex (Alibaba HK$190B in H1). Market now scrutinizes revenue per compute dollar, margin trajectory, cash conversion.
10:18 — Trust Crisis: Zcode Data Upload Incident: Developers found Zcode silently uploaded entire workspaces (git history, ops) encrypted to cloud; keys only server-side. Zhipu apologized, blamed default indexing, promised destruction. Taiyuan Chengming Tech alleged uploads continued post-apology. ~400 developers formed rights group. Zhipu open-sourced Zcode, invited audits, added no-retention mode. Stock dropped sharply; for an 86.5% API-revenue company, developer trust is revenue foundation.
11:17 — Four Lessons & Market Regime Shift: 1) Scarcity premium vanishes with competition/liquidity — test moats by asking if advantage survives three equal rivals. 2) Market pays for sustained leadership capability (org, engineering, data flywheel, compute autonomy, commercialization flywheel), not one-time benchmark wins. 3) Growth without margins burns capital; key metric shifts to revenue per compute dollar and gross margin. 4) Primary/secondary valuation split (Moonshot ~50x ARR, DeepSeek ~75x ARR vs public market quarterly re-rating) signals pricing inflection. Historical analog: Amazon vs Cisco post-dotcom.
14:13 — Future Hurdles & Outlook: Jan 2027: ~30% original shares unlock. Zcode trust repair ongoing. A-share STAR Market filing approved (up to HK$15B). Zhipu holds: open-source tier-1 model capability, domestic-cluster compute autonomy, API-led subscription path. Market no longer denies the business but refuses to pre-pay for future. Core industry rule shifting from 'whose model is stronger' to 'who monetizes intelligence at higher price while lowering compute cost'.
Speakers
大飞 (Da Fei) — Host / Analyst