AI / LLMChina
← Sticker Wall
JOURNEY · DETAIL
Zhipu AI: From Tsinghua Academic Research to Enterprise Private Deployment LLM Platform
Founded: Zhang Peng, Tang Jie, Liu Debing · Beijing Zhipu Huazhang Technology Co., Ltd.
JOURNEY
Key Fields
FIELD STAMPSIndustryAI / LLM
RegionChina
ScaleGiant
ChannelOther
Origin
In 2019, Tang Jie, director of the Knowledge Engineering Group (KEG) at Tsinghua University's Department of Computer Science, led his team to commercialize years of academic research in knowledge graphs and LLM technology. The team initially focused on AMiner, an academic search tool providing paper retrieval services for researchers. While it gained significant user base and academic influence, it generated almost no commercial revenue, forcing the team to rely on horizontal projects to sustain operations. Before the ChatGPT explosion, Zhipu was already training the GLM series, which was initially more academic than commercial in nature.
Milestones
2019
Inception PMF
Zhipu AI was officially founded in June 2019, emerging from the Knowledge Engineering Group at Tsinghua University. Founders Zhang Peng and Tang Jie are both professors at Tsinghua. The initial product was the academic search platform AMiner. Despite its massive influence in academia, commercialization was difficult, and the team relied on contract projects to survive, with the company's accounts at one point becoming so tight that the founders had to use their own funds to cover expenses.
2020
In-house Model Development Turning Point
At the end of 2020, Zhipu released the GLM series of pre-trained models, pursuing an independent architecture rather than simply replicating GPT. The team invested significant computing power into training a 10-billion-parameter model focused on Chinese corpora. At the time, the domestic AI community generally did not favor the in-house LLM route, preferring to fine-tune open-source models. Zhipu faced repeated rejections in early fundraising, with its angel round only reaching the tens of millions of RMB.
2022
Product Pivot Transition
In August 2022, Zhipu launched and open-sourced the cognitive LLM GLM-130B. With 130 billion parameters, it achieved performance comparable to GPT-3 in bilingual (Chinese-English) benchmarks. The open-source strategy earned Zhipu recognition in the developer community, but commercialization remained unproven, with the company's main revenue still coming from a few government-enterprise POC projects, totaling less than 50 million RMB in annual revenue. ChatGPT had not yet been released, and investors showed little interest in the LLM track.
2023
ChatGPT Effect Dividend Growth
After ChatGPT ignited the domestic LLM track in 2023, Zhipu quickly became a capital favorite. In May, it completed a Series B financing round of several hundred million RMB, reaching a post-money valuation of approximately $1 billion and becoming one of the first domestic LLM unicorns. It launched the MaaS open platform, with API call volumes growing 50-fold within three months and monthly active developers exceeding 100,000. However, the high unit price of private deployment (millions of RMB) and long delivery cycles of 3-6 months became the biggest bottlenecks for scaling.
2024
Commercial Expansion PMF
In April 2024, Zhipu completed a new round of financing, bringing total funding to over 2.5 billion RMB, led by Middle Eastern oil funds, with a valuation of 10 billion RMB. That same year, private deployment orders surged in the financial sector, with over a dozen clients, including major state-owned banks and top brokerage firms, signing annual contracts worth tens of millions. Zhipu's annual revenue for 2024 exceeded 500 million RMB, though gross margins were limited to about 35% due to high computing and delivery costs.
2025
DeepSeek Impact and Self-Rescue Failure
Following the open-source release of DeepSeek-R1 in January 2025, its extreme cost-effectiveness disrupted the domestic LLM market. Zhipu was forced to cut its API prices from 5 RMB per million tokens to 0.5 RMB, a 90% reduction. In the third quarter, Zhipu's overall API revenue fell 40% quarter-over-quarter, with some small and medium-sized clients migrating to the DeepSeek ecosystem. Internal strategic disagreements arose: one side advocated for full open-source and community ecosystem development, while the other insisted on private deployment for top-tier clients. Ultimately, Zhang Peng decided to prioritize the private deployment route, leading to the layoff of approximately 15% of the API operations team.
2026
Deepening Private Deployment vs. DeepSeek Transition
In the first half of 2026, Zhipu significantly scaled back its public cloud API investment, shifting 70% of R&D resources to private deployment and industry-specific fine-tuning solutions. The GLM-5 series models were deeply optimized for finance and government sectors, and an out-of-the-box private deployment appliance solution was launched, compressing delivery cycles from 3 months to 2 weeks. In the first half of the year, new private deployment contracts exceeded 800 million RMB, with a 35% market share in the financial sector and successful bids for multiple provincial-level AI hub projects.
Turning Points
- The open-source release of GLM-130B in August 2022 allowed Zhipu to accumulate sufficient developer community recognition and technical reputation before the ChatGPT explosion; otherwise, the financing window would not have opened.
- The capital frenzy triggered by the ChatGPT effect in 2023 turned Zhipu from a niche academic startup into a unicorn overnight, but it also masked the core issue of insufficient commercial delivery capabilities.
- After DeepSeek-R1's open-source, free-to-use strategy disrupted the API market in 2025, Zhipu was forced to slash API pricing by 90%, compelling the company to shift entirely from the public cloud route to private deployment, which ultimately led them to their true competitive moat.
- The launch of the private deployment appliance in 2026, which compressed delivery cycles from 3 months to 2 weeks, increased the efficiency of signing large standard contracts by 4 times, pushing financial sector market share past 35% and establishing them as a local competitor capable of standing against DeepSeek.
Failures & Pitfalls
- From 2019 to 2021, Zhipu operated with an academic mindset; AMiner had academic influence but almost no commercial revenue, leaving the company so cash-strapped that founders had to pay salaries out of pocket.
- After the ChatGPT explosion in 2023, Zhipu rushed to launch a MaaS platform to capture the API market but failed to build an effective customer success system. Many small and medium-sized enterprise clients did not convert to paid plans after free trials, and API revenue could never cover computing costs.
- By 2024, while private deployment had secured large orders, delivery relied heavily on manual labor, with each project requiring 5-10 people on-site for 3-6 months. Delivery costs accounted for over 40% of contract value, severely dragging down gross margins.
- After the release of DeepSeek-R1 in 2025, Zhipu's response was sluggish; API price adjustments were delayed by nearly two months, during which about 30% of small and medium-sized clients were lost to competitors, and internal strategic disagreements led to the departure of three key members of the core engineering team.
关键成功要素
- The Tsinghua academic pedigree is the underlying logic for Zhipu's government and enterprise trust; Professor Tang Jie's background in the Knowledge Engineering Group makes financial and government clients naturally trust their data security capabilities.
- The true barrier to private deployment lies not in the model itself, but in the delivery system. Zhipu's ability to compress the delivery cycle to 2 weeks in 2026 is the core reason for its competitive success.
- The disruption caused by DeepSeek is essentially a positive development, forcing Zhipu to abandon the unprofitable and hard-to-scale API route and go all-in on the path of private deployment, which possesses a genuine moat.
- The financial industry is the golden scenario for private deployment; compliance requirements prohibit data from being moved to the public cloud. Zhipu's ability to achieve a 35% market share among banks is a critical survival benchmark.
Lessons
- The greatest danger for an academic startup is not a lack of technical strength, but the absence of cash flow to survive the long cold-start period before capital takes notice. Using contract projects to sustain oneself is a pragmatic survival strategy.
- Building community recognition through open-source models is crucial. Zhipu's open-sourcing of GLM-130B in 2022 laid the foundation of trust needed to quickly gain capital attention after the ChatGPT explosion.
- API price wars are low-dimensional competition. The real money is always in the ability to solve complex delivery problems for top-tier clients. The barrier for private deployment comes from engineering, not model parameters.
- When a more powerful open-source competitor emerges, do not engage in a price war. Immediately pivot to private deployment scenarios that competitors cannot cover, and build a moat through delivery capabilities rather than model capabilities.
Core Data
- Cumulative Financing：Over 2.5 billion RMB (based on public information, independent verification not performed)
- Valuation：Approximately 10 billion RMB (post-Series B+ in April 2024) (based on public information, independent verification not performed)
- 2024 Revenue：Exceeded 500 million RMB (based on public information, independent verification not performed)
- Private Deployment Gross Margin：Approximately 35% (2024) (based on public information, independent verification not performed)
- Team Size：Approximately 800 people (mid-2025) (based on public information, independent verification not performed)
- R&D Resource Allocation：70% shifted to private deployment (2026) (based on public information, independent verification not performed)
- New Private Deployment Contracts：Over 800 million RMB (H1 2026) (based on public information, independent verification not performed)
- Financial Industry Market Share：35% (2026) (based on public information, independent verification not performed)
- API Price Reduction：From 5 RMB per million tokens to 0.5 RMB, a 90% reduction (2025) (based on public information, independent verification not performed)
Competitors / Peers
Zhipu's main competitors in the private deployment track include Baidu's Ernie Bot Enterprise Edition and Alibaba's Tongyi Qianwen private deployment solution. Baidu has an advantage in the central enterprise market due to its cloud service ecosystem, but its model capabilities in Chinese scenarios are considered inferior to Zhipu's. Alibaba's Tongyi Qianwen uses group cloud resources to break through with lower prices but lacks the Tsinghua-backed trust endorsement in the government-enterprise compliance sector. In 2026, the real threat comes from DeepSeek, which attracts a large number of small and medium-sized integrators to build their own solutions using open-source, free models, impacting Zhipu's pricing space among mid-tier clients. Additionally, Moonshot AI and Wall-E Intelligence are also competing for the enterprise service market; Moonshot AI enters legal compliance scenarios with long-text capabilities, while Wall-E Intelligence targets retail and manufacturing scenarios with edge-side lightweight deployment. Zhipu's differentiation lies in its years of accumulated government-enterprise delivery experience and the Tsinghua brand endorsement, which are difficult for pure internet-native competitors to replicate in the short term.