Anatomy of Zhipu's Open-Source Gamble: $110M Revenue, Volatile Valuation

On November 9, blogger ccerrato147 published a long thread systematically breaking down Zhipu's (Z.ai internationally) history and its "build the product, then give it away free" business model: the company booked $110M revenue last year with R&D spending over twice that, and in 2025 reported $110M revenue against $700M in losses, after being added to the US Entity List in January. Whether open-source prestige can convert into contracts others can't win is the thread's central open question.

Confirmed

• Tsinghua origins: Zhang Peng joined Tsinghua University's Knowledge Engineering Group (KEG) in 2002, with the creed "Paper to Product"; the first product was academic search engine AMiner, which became Zhipu in 2019. When GPT-3 launched in 2020, Zhang's advisor Zhang Bo judged it "spouts nonsense and doesn't know what it knows," yet Zhang Peng insisted on rebuilding methods from the papers in-house.
• Early open-source reputation: Starting January 2022, the team trained the 130-billion-parameter GLM-130B on 96 A100s, fitting on a single A100 server, repeatedly hitting OOM, gradient, and hardware failures along the way; the final successful training took about two months and roughly $600K. In November 2022, GLM-130B became the only Chinese model among Stanford's 30-model benchmark—scores below closed-source models, but it put them on the world's radar.
• Closed-source monetization failed: GLM-4 launched January 2024 with 200 billion parameters, benchmarking at roughly 90% of GPT-4, sold closed-source via API only; but just four months later DeepSeek V2 slashed prices and the whole industry followed, killing that revenue route.
• To-C flop: Zhipu's consumer app never broke 11M MAU, versus ByteDance's Doubao at roughly 324M; Zhipu then pivoted to B2B enterprise business.
• Open-source peak: In July 2025, Zhipu open-sourced GLM-4.5 weights under MIT license, topping Hugging Face within 48 hours; afterward, companies outside China began integrating it into their own products.
• IPO and valuation swings: Zhipu listed on the Hong Kong exchange in January at roughly a $7B valuation; within five months its market cap briefly hit $170B (nearly half of Alibaba's), but by end of September about $130B had evaporated.
• Competitive landscape: Three weeks after ChatGPT's launch in late 2022, Zhipu already had a comparable model in training and rushed out a chatbot over Chinese New Year; investors had once demanded Zhipu halve its valuation, then capital flooded in and the same race saw over 130 competitors.

Why it matters

• Zhipu represents an extreme business path: build the model on paid R&D, then give the weights away free, betting open-source prestige converts into enterprise contracts. There are two monetization routes—cloud API access and private deployment—the logic being that even with a freely downloadable model, customers still need GPUs, operational assurance, security review, and responsive support; as ccerrato147 puts it, "that gap is the business, but it hasn't been filled yet."
• The author also flags the model's inherent contradiction: every downloader could become a competitor rather than a customer. With R&D spending double revenue and wild market-cap swings, whether this model can work remains the case study the whole industry is watching.

2026-11-09 ~ 2026-11-09 · 13 related posts

• Zhipu surged to $170B within five months of its Hong Kong IPO, then lost ~$130B — ccerrato147 · 2026-11-09
• Zhipu earned $110M but spent twice that on R&D — can open-weight models pay the bills? — ccerrato147 · 2026-11-09
• [source] Zhipu Earned $110M Last Year and Spent Twice That on Research — ccerrato147 · 2026-11-09
• From Tsinghua's KEG Lab to Zhipu: Zhang Peng Joined in 2002 Under 'Paper to Product' — ccerrato147 · 2026-11-09
• Zhang Bo Called GPT-3 a Model That 'Talked Nonsense' — Zhang Peng Built One Anyway — ccerrato147 · 2026-11-09
• GLM-130B Training Cost About $600K: Two Months on 96 A100s — ccerrato147 · 2026-11-09
• GLM-130B Was the Only Chinese Model in Stanford's 30-Model Benchmark in Nov 2022 — ccerrato147 · 2026-11-09
• After ChatGPT Shipped, Zhipu Already Had a Comparable Model Training — 130 Rivals Followed — ccerrato147 · 2026-11-09
• Zhipu's Chat App Peaked at 11M MAU vs Doubao's 324M — It Pivoted to Enterprise — ccerrato147 · 2026-11-09
• GLM-4's Closed API Model Lasted Just Four Months Before DeepSeek V2 Price Cuts — ccerrato147 · 2026-11-09
• GLM-4.5 Hit #1 on Hugging Face Within 48 Hours of MIT-Licensed Release — ccerrato147 · 2026-11-09
• [source] Zhipu: $110M Revenue, $700M Loss — What Exactly Is the Open-Weights Business Selling? — ccerrato147 · 2026-11-09
• [source] Zhipu's bet: open-model reputation converting into contracts nobody else can serve — ccerrato147 · 2026-11-09