Zhipu's R&D spending is twice its revenue; how long can free open-source models sustain growth?
Zhipu releases all its latest models free and open-source, but R&D spending is already twice revenue—the real question, hanging unanswered at the piece's close, is how long this business of burning cash for market attention can last.
The argument · tap a timestamp to hear it
Paper-to-product culture turned knowledge graphs into revenue
Zhipu traces back to Tsinghua's Knowledge Engineering Group, founded in 1996 and focused on knowledge graphs and information extraction. The lab's internal motto was P2P—Paper to Product—reflecting the conviction that AI required industry engagement. In 2006, KEG spun out AMiner, an academic search engine. Starting in 2015, it built commercial partnerships with Microsoft, Elsevier, Google, ExxonMobil, and IBM, generating revenue that convinced the team to incorporate AMiner as a standalone company.
GPT-3's scale forced Tsinghua to bet on building its own large models
GPT-3's release in 2020, with 175 billion parameters and generation quality far exceeding Google's BERT and other mainstream models at the time, made Zhang Peng realize that classical NLP's complex syntactic pipeline was unnecessary. His advisor Zhang Bo agreed the large language model represented a new paradigm but warned that models ‘spout utter nonsense’and don't truly understand what they say. Zhang Peng pushed ahead anyway, rallying the team to train their own billions-parameter Chinese-English bilingual model. With OpenAI's details unpublished, they had to discover everything themselves.
GLM proved that blending BERT-style fill-in-the-blank with GPT-style generation works at scale
In the WuDao project of 2021, Tsinghua proposed GLM (General Language Model), a pretraining framework that used bidirectional context to fill blanks like BERT, but generated token-by-token like GPT rather than predicting entire masked spans at once. Beyond 100 billion parameters, the model began showing signs of intent understanding—for example, on arithmetic questions about a newborn's head length, it divided rather than simply multiplying 50 by 1/4, suggesting it grasped the problem's underlying logic rather than pattern-matching formulas.
— Zhang PengGLM-130B went open-source because monetization looked impossible at the time
In 2022, Zhipu trained GLM-130B, a 130-billion-parameter dense model (avoiding the sparse mixture-of-experts used in WuDao 2.0) chosen so it could run on a single A100 server. Training compute cost roughly 4 million yuan, with total end-to-end cost around 10 million yuan. After successful training, the team chose to open-source it—Zhang Peng's reasoning was blunt: even closed-source, it wouldn't make money anyway, so trade it for reputation. GLM-130B became the only Chinese model to appear on Stanford's November 2022 large model benchmark report.
— Zhang PengConsumer chat lost to Baidu and ByteDance; Zhipu pivoted entirely to enterprise
Zhipu's consumer chatbot, Zhipu Qingyan, launched in August 2023 as one of China's earliest but couldn't compete against Baidu's Wenxin and ByteDance-backed Doubao. Doubao eventually reached 324 million monthly active users; Qingyan peaked below 11 million. Paid subscriptions saw no consumer appetite. Zhipu turned to model-as-a-service for enterprises—Zhang Peng's logic: enterprise customers grow suspicious of prices too low; the enterprise market's price wars are far less vicious than consumer.
— Zhang PengDeepSeek proved MoE architecture could reach GPT-4 caliber; Zhipu adapted
The industry had doubted whether mixture-of-experts architecture could deliver GPT-4-level performance until DeepSeek's R1 validated the path. Zhipu studied DeepSeek's model architecture and built its own variant, releasing its first mixture-of-experts flagship model, GLM-4.5, on July 28, 2025—and open-sourced it for the first time. Expecting modest response, it topped Hugging Face's leaderboard within 48 hours; even U.S. companies began integrating it into their products. Zhang Peng realized new model launches no longer required heavy-handed marketing campaigns.
— Zhang PengStock soared tenfold then crashed back; R&D burn now exceeds revenue by twofold
Zhipu went public on Hong Kong's exchange in January 2026; the stock price soon surged roughly tenfold, reaching a peak market cap around $170 billion in June, surpassing Xiaomi and BYD and equivalent to more than half Alibaba's market value. It later fell back to around $50 billion, with the company issuing new stock to institutional investors in July to shore up cash. First-half revenue grew 400% year-over-year to $140 million; annualized revenue reached roughly $1.6 billion. The critical problem: R&D spending now exceeds revenue by more than twofold—whether the free open-source model strategy can sustain this burn rate remains an open question.
In their own words · checked verbatim
Unless you take the leap and make the pursuit, you’ll never know just how fast or how far you can run.
Zhang Peng5:13
A newborn baby’s head is one-quarter of its body length. If a baby is 50 centimeters long at birth, how long is its head?
Zhang Peng8:21
I think those companies struggling to make real impact dampened investor interest in future startups bearing "AI breakthroughs".
Zhang explained that while excessive price-cutting does happen in enterprise, it happens less than in consumer. Because in the end, the client wants the product and is suspicious of a bid that is priced too low.
Soon after the listing, the stock price soared some ten times.
GLM-5.3 and its Flash version in particular seems to have gathered a lot of attention - though DeepSeek remains the de facto number one Chinese AI lab.
How long can this immensely dilutive business model go on? We shall see.
Figures
| GPT-3 parameter count | 175 billion | 4:13 |
| GLM-130B end-to-end cost | 10 million yuan | 11:27 |
| ByteDance Doubao monthly active users | 324 million | 16:47 |
| Zhipu Qingyan peak monthly active users | Under 11 million | 16:47 |
| GLM-4 vs. GPT-4 benchmark level | ~90% | 17:48 |
| Zhipu Hong Kong IPO peak market cap | $170 billion | 21:51 |
| Zhipu Hong Kong IPO market cap after pullback | $50 billion | 21:51 |
| Annualized revenue (ARR) | $1.6 billion | 21:51 |
Glossary
- GLM (General Language Model)
- Tsinghua's pretraining framework that fuses BERT-style fill-in-the-blank with GPT-style token-by-token generation.
- MoE (Mixture-of-Experts)
- A sparse model architecture that activates only part of its parameters at inference, more compute-efficient than dense models.
- Model-as-a-Service
- A business model delivering large model capabilities to enterprise customers via cloud API or private deployment.
- KEG (Knowledge Engineering Group)
- Tsinghua's lab that spawned Zhipu, focused on knowledge graphs and information extraction research.
- P2P (Paper to Product)
- KEG's internal motto encouraging research findings to become commercial products.
How to listen
Investors and practitioners tracking China's AI landscape, Series funding valuations, or the business models around open-source large models.
14:37–15:40, a listing of startup names in the hundred-models race; skip if you follow China's AI industry closely.