How Zhipu AI Built a $128 Billion Open Source LLM Empire From Tsinghua

In January 2026, a little-known Chinese AI company called Zhipu AI went public on the Hong Kong Stock Exchange at a $6.5 billion valuation. Six months later, its stock has soared 1900%, and the company is worth over $128 billion. Its GLM-5.2 model, released under an MIT open-source license, competes with GPT-5 and Claude Opus on coding benchmarks. How did a Tsinghua University spinoff become one of the world's most valuable AI companies in just seven years?

The Zhipu AI story is one of the most dramatic in the AI industry. It's not just about the money—it's about what happens when a world-class university research team, a massive domestic market, and a deliberate open-source strategy all converge at exactly the right moment.

Zhipu (now rebranded as Z.ai internationally) is the first pure-play LLM company to list in Hong Kong, and it's already one of the most valuable. Its GLM (General Language Model) series has gone from a 130 billion parameter research model to a 744 billion parameter MoE (Mixture of Experts) system that beats most Western models on coding benchmarks—and gives it all away for free.

From Tsinghua Lab to Global Player

Zhipu AI was founded in 2019 as a spinoff from Tsinghua University's Department of Computer Science, specifically from the Knowledge Engineering Group (KEG)—one of China's most respected AI research labs. The team had been working on knowledge graphs and pre-trained language models for years before the generative AI boom made their work suddenly enormously valuable.

The founding story is classic Chinese academic entrepreneurship: a group of professors and PhD students who had been publishing research papers for years suddenly realized their work could become a product. What made Zhipu different from hundreds of other Chinese AI startups was its approach from day one: build everything on open foundations, target developers, and move faster than anyone else.

The GLM Model Family: From 130B to 744B

Understanding Zhipu's rise requires understanding the GLM model family's evolution. The progression from GLM-130B to GLM-5.2 tells the story of how Chinese LLM development caught up to and in some ways surpassed Western models in just a few short years.

Model	Total Params	Active Params	Context	Key Feature	Release
GLM-130B	130B	130B (dense)	2K	First serious Chinese LLM	Aug 2022
GLM-4	355B	32B	128K	MoE architecture	Jan 2025
GLM-4.7-Flash	—	—	203K	Free on API	Mid 2025
GLM-5	744B	40B	200K	256 experts, MLA/DSA, trained on Ascend	Feb 2026
GLM-5.1	744B	40B	200K	400 tokens/s output speed	Apr 2026
GLM-5.2	744B+	40B+	1M	MIT open-source, two reasoning levels	Jun 2026

Two details about GLM-5 deserve special attention because they reveal something important about Zhipu and China's AI ecosystem more broadly.

First: GLM-5 was trained entirely on Huawei Ascend AI processors using the MindSpore framework—not NVIDIA GPUs. This is genuinely significant. While most of the world's frontier models are trained on NVIDIA H100/H200 GPUs with CUDA, Zhipu proved you can build a world-class LLM on Chinese domestic hardware. The model was trained on 28.5 trillion tokens—more than some Western frontier models—and achieved competitive benchmarks despite the hardware constraints.

Second: The architecture uses 256 experts with 8 activated per token (a 5.9% sparsity rate), combined with DeepSeek-style Multi-head Latent Attention (MLA) and Dynamic Sparse Attention (DSA) for efficient long-context processing. Borrowing architectural ideas from competitors is common in the AI industry, but Zhipu's ability to implement and improve on those ideas at this speed is remarkable.

The Open Source Strategy: Why Give It Away?

Zhipu's decision to release GLM-5.2 under the permissive MIT license surprised many in the industry. Why would a company worth over $100 billion give away its crown jewels for free? The answer lies in understanding how Chinese AI companies think about market share, developer ecosystems, and long-term competition.

Capturing Developers, Not Revenue Per Query

Western AI companies like OpenAI and Anthropic focus on maximizing revenue per API call. Their models are closed, their APIs are expensive, and they compete on being the smartest model available.

Zhipu is playing a different game. By releasing a near-frontier model under MIT license, it's trying to do two things:

• Build developer mindshare: If developers build their applications on GLM because it's open, free, and good enough, they won't switch to a more expensive closed model later. This is the classic open-source playbook: commoditize the model layer and make money on services, tooling, and enterprise support.
• Create a Chinese standard: In China's domestic market, where geopolitical concerns make Western AI models risky for enterprise use, having an open-source domestic alternative that's competitive globally is enormously valuable. Every company that adopts GLM reinforces Zhipu's position as the default.

The Coinbase Effect

In early 2026, Coinbase—the $40 billion cryptocurrency exchange—made Chinese AI models its default for engineering use. The announcement sent shockwaves through the tech industry: here was a major American public company choosing Chinese open-source models over Western closed ones, and the reason was simple economics.

Coinbase's engineering team found that GLM models (alongside DeepSeek) delivered 91% of the performance of premium Western models at a fraction of the cost. For a company with hundreds of engineers generating millions of tokens daily, the savings were enormous—estimated at roughly 50% of their AI spending.

Coinbase is far from alone. Airbnb uses Alibaba's Qwen. Lindy uses DeepSeek. Industry estimates suggest that 80% of American AI startups now use Chinese open-source models for at least some of their workloads. On platforms like OpenRouter, Chinese models' share has exploded from 2% to over 40% in less than a year.

Zhipu's Pricing Advantage

Even for its hosted API, Zhipu is dramatically cheaper than Western competitors. GLM-5 costs $1.00 per million input tokens and $3.20 per million output tokens—roughly 3x cheaper than Claude Sonnet for input and 5x cheaper for output.

GLM-4.7-Flash, Zhipu's fast lightweight model, is completely free on the API with a 203K context window. That's the strongest free model offering from any Chinese provider, and it's deliberately positioned to capture the long tail of developers who don't want to pay for API access but will build applications that eventually need premium features.

The Money Machine: How Zhipu Actually Makes Money

If Zhipu gives away its best models for free, how is it worth $128 billion? The answer tells you a lot about how the AI business model works in China versus the West.

Revenue: Growing Fast, But Still Small

Zhipu's revenue tells a story of explosive growth but modest absolute numbers. In 2022, the company made just 57.4 million yuan ($8M) in revenue. By 2025, that had grown to 724 million yuan ($100M)—a 12x increase in three years.

The API platform alone achieved an ARR (annualized revenue rate) of 1.7 billion yuan ($250M) in 2025, up 60x year-over-year. The platform has over 4 million registered users, and major Chinese tech companies including ByteDance, Alibaba, and Tencent all use GLM models in their products.

But here's the thing: $250M in ARR for a $128 billion valuation means Zhipu is trading at roughly 256x its annual revenue. For comparison, OpenAI trades at about 43x ARR, and Anthropic at about 107x. This is an extraordinarily rich valuation by any standard.

Company	Valuation (mid-2026)	Est. ARR	Valuation/ARR
OpenAI	$852B	~$200B	~43x
Anthropic	$965B	~$90B	~107x
ByteDance (AI)	$2200B+	~$200B	~11x (overall)
DeepSeek	$42B	~$1.5B	~28x
Zhipu (Z.ai)	$128B	~$500M	~256x
MiniMax	$30B	~$200M	~150x

Why Is the Valuation So High?

Several factors explain Zhipu's extraordinary valuation multiple:

• Scarcity premium: Zhipu is the only pure-play LLM company listed on the Hong Kong Stock Exchange. For Chinese investors who want exposure to AI but can't easily buy OpenAI or Anthropic, Zhipu is essentially the only game in town.
• Southbound capital flood: When Zhipu was added to the Hang Seng Tech Index and Stock Connect in June 2026, mainland Chinese investors could suddenly buy Hong Kong-listed AI stocks. The flood of southbound capital drove the price up dramatically.
• Domestic substitution narrative: In an environment of US export controls on AI chips, the "China must have its own AI models" narrative is extremely powerful. Zhipu is the most prominent publicly tradable bet on this outcome.
• GLM-5.2 technology premium: The model's competitive benchmark performance and 1M token context window signal to investors that Zhipu isn't just a cheap copy—it's genuinely at the frontier.

Whether this valuation is justified is a matter of intense debate. At 256x revenue, Zhipu is priced for perfection and then some. The company would need to grow revenue 20-30x just to get to a "reasonable" SaaS multiple. But in a market where scarcity and narrative drive prices, conventional valuation rules may not apply.

Funding and Investors

Before its IPO, Zhipu raised about $1.4 billion across 12 funding rounds. Its investor list reads like a who's who of Chinese tech: Alibaba, Tencent, Meituan, Xiaomi, and even Saudi Aramco's Prosperity7 Ventures. The January 2026 IPO raised approximately $558 million at a $6.5 billion valuation—meaning the stock has returned nearly 20x in six months for early public investors.

The Competitive Landscape in China

Zhipu isn't the only LLM company in China—not by a long shot. The Chinese AI market is intensely competitive, with multiple well-funded players all jockeying for position.

According to IDC data, in China's public cloud MaaS (Model-as-a-Service) market by revenue:

1. ByteDance's Volcengine (Doubao) — 40%+ market share
2. Alibaba Cloud (Qwen/Tongyi)
3. Baidu Smart Cloud (Ernie/Wenxin)
4. Zhipu AI (GLM)
5. China Mobile Cloud

So by raw market share, Zhipu is only fourth or fifth in China. But this ranking is misleading because it measures revenue, not influence. Zhipu's open-source strategy means its models are used far more widely than its revenue share suggests. Developers use the free open-source versions, companies self-host GLM models on their own infrastructure, and startups build products on GLM without paying Zhipu directly.

It's the classic open-source business model: give away the product to build the ecosystem, then monetize the enterprise customers who need support, customization, and hosted versions. Linux did this. Red Hat made billions from it. Whether Zhipu can pull off the same trick with LLMs remains to be seen—but the playbook is clear.

The Bilingual Advantage

One of GLM's underrated strengths is its bilingual performance. GLM-5 natively supports English, Chinese, and 15+ additional languages. But the real advantage is in Chinese-English bilingual capability—something Western models have historically struggled with.

Independent evaluations show that GLM-5 matches or exceeds GPT-4's performance on Chinese language understanding and generation tasks. For businesses that work across Chinese and English markets, this is a killer feature. A marketing agency serving both China and the West can use one model for both languages instead of stitching together GPT-4 for English and something else for Chinese.

This bilingual advantage is part of why Zhipu is attracting attention from global companies. It's not just cheaper—it's actually better at certain tasks that involve both languages.

Challenges and Risks

For all its success, Zhipu faces significant challenges that could derail its trajectory:

1. The Valuation Question

At 256x revenue, Zhipu is priced for absolute domination of the global AI market. If growth slows, if competition intensifies, or if the broader AI sector cools off, the stock could fall dramatically. The 1900% rally in six months has all the hallmarks of a classic tech bubble—though predicting when bubbles pop is famously difficult.

2. Hardware Constraints

That GLM-5 was trained on Huawei Ascend chips is both an achievement and a limitation. Huawei's chips are widely seen as being 1-2 generations behind NVIDIA's latest. If export controls tighten further and Chinese chip development can't keep pace, Zhipu may face hardware constraints that Western competitors don't.

3. DeepSeek and Qwen Are Also Winning

Zhipu isn't the only Chinese open-source LLM maker. DeepSeek has become the darling of Western developers, and Alibaba's Qwen series has massive resources behind it. The open-source LLM market could become commoditized, with Zhipu's models becoming just one option among many rather than the clear leader.

4. Commercialization Uncertainty

Zhipu gives away its best models for free. The company's stated monetization strategy is enterprise services and premium features, but this is unproven at scale. The open-source playbook worked for software companies like Red Hat and MongoDB, but whether it works for frontier AI models is still an open question.

5. Regulatory and Geopolitical Risk

As a Chinese AI company with global ambitions, Zhipu faces unique geopolitical risks. Could the US government restrict use of GLM models by American companies? Could China restrict export of its most advanced AI technology? These are real risks that could impact Zhipu's international expansion plans.

What's Next for Zhipu?

Zhipu's trajectory over the next 2-3 years will be one of the most interesting stories in AI. Here's what we're watching:

AutoGLM: The Agent Play

In July 2026, Zhipu raised an additional $500 million and announced AutoGLM 2.0—an agent framework positioned as a direct competitor to GPT-5's Agent Mode. If Zhipu can build an open-source agent platform that matches closed alternatives, it could dramatically accelerate adoption among developers who want to build AI applications without lock-in.

International Expansion

Zhipu has rebranded internationally as "Z.ai"—a clear signal that it wants to compete globally. The GLM-5.2 open-source release, the Coinbase integration, and the company's growing presence on developer platforms all point to an aggressive international push. Can a Chinese AI brand succeed in Western markets? Zhipu is about to find out.

Coding as the Beachhead

Zhipu's strongest benchmark performance is in coding. GLM-5 scores 77.8% on SWE-bench Verified, within three points of Claude Opus. The GLM Coding plans (Lite, Pro, Max, Team) are specifically targeted at developers. Zhipu's high-speed version (GLM-5.1) outputs 400 tokens per second—faster than any other major LLM API. For coding use cases where speed matters, Zhipu is genuinely competitive.

This makes sense as a strategy. Developers are the most influential users in AI—they build the applications that everyone else uses. Win the developers, and you win the platform. That's how Microsoft won the PC era, how AWS won the cloud era, and how Zhipu hopes to win the AI era.

Conclusion: The Open Source Disruptor

Zhipu AI's journey from a Tsinghua research lab to a $128 billion public company in seven years is remarkable by any standard. But what makes Zhipu truly important isn't its valuation—it's what it represents: the rise of Chinese AI companies as genuine global competitors, not just cheap copies.

With GLM-5.2 under MIT license, Zhipu is betting that open source will win in AI the same way it won in software. The logic is compelling: if a model that's 90% as good as GPT-5 is free and open, why would developers pay for the closed version? For many use cases—especially coding, summarization, and enterprise applications—they won't.

Zhipu's success also highlights a fundamental shift in the AI industry. Five years ago, everyone assumed that AI would be dominated by American companies. Today, Chinese companies like Zhipu, DeepSeek, and ByteDance are not just competitive—they're often the innovators, setting the pace on open-source releases, context window sizes, and pricing models.

The question isn't whether Chinese AI companies will continue to gain global influence. They already have. The real question is whether the AI market will end up looking like the smartphone market (a few closed ecosystems dominating) or like the software market (open source winning most of the value). If it's the latter, Zhipu's $128 billion valuation might turn out to be cheap.