Source | Bohu Finance (bohuFN)

Author | All too well

Zhipu, the 'first large model stock,' is truly becoming legendary.

In April, people were still marveling at how a company with annual revenues of RMB 700 million could see its market value soar to HK$400 billion. Two months later, on June 22, its market cap briefly surpassed HK$1 trillion.

How did this happen? Or rather, why?

01 Everything AI-related is going wild

This year, everything connected to AI has gone wild.

According to a Crunchbase report, in the first quarter of this year, just four companies—OpenAI, Anthropic, xAI, and Waymo—accounted for 65% of global AI funding.

NVIDIA's market cap once surged to $5.7 trillion, with its stock price rising over 60% in the past year. Anthropic and OpenAI have also secretly filed IPO applications with the U.S. SEC, entering the final sprint toward listing, with both valuations nearing the $1 trillion mark.

The same is true in China.

According to rough statistics by Caijing Tianxia, around 12 Hong Kong-listed companies have seen their stock prices rise over 100% year-to-date with market caps exceeding HK$100 billion—all in the hard tech sector, largely benefiting from AI. Examples include MiniMax, Biren Technology (one of China's 'four little dragons' in GPUs), storage chipmaker GigaDevice, and semiconductor company Hua Hong Semiconductor.

The same trend is evident in China's A-shares. Companies like CXCT and SMIC have both surpassed RMB 1 trillion in market cap, with Cambricon soon to follow.

Meanwhile, the pace and scale of funding for emerging tech companies, including large model firms, have been unprecedented in China this year.

Take Moonshot AI, for example. After closing an $8 billion funding round in late February, it immediately launched a new round priced at $17 billion the same day. Two weeks later, its valuation was raised again to $18 billion, with $2 billion in new funding, bringing its post-money valuation to $20 billion. DeepSeek, which had relied on its parent company for funding, also began targeted invitation-only financing, estimated at around $40 billion, with many state-owned enterprises and internet giants actively engaging.

If debates last year centered on whether AI was a bubble, this year, while discussions about the AI bubble haven't stopped, the market seems more afraid of missing out.

This shift didn't happen overnight.

First, the gap in model capabilities is narrowing. The difference between top U.S. and Chinese models has shrunk from a generational lead to single-digit margins. According to Artificial Analysis, China's leading models were only 60% as capable as U.S. counterparts two years ago but now reach around 90%.

Second, China's cost advantage is becoming evident. UBS semiconductor team analysis shows that Chinese models' average API pricing is less than 20% of U.S. equivalents, with training costs below 10%, yet gross margins are comparable to Anthropic and OpenAI, ranging between 20% and 40%.

In other words, Chinese models aren't just catching up—they're showing commercial competitiveness. OpenRouter data shows that Chinese AI large models have ranked first globally in weekly call volume for eight consecutive weeks.

Third, demand-side changes are underway. Microsoft canceled internal Claude Code licensing partly due to cost pressures from token-based billing. Uber burned through its entire annual AI budget in the first four months of this year, forcing it to cut back on usage. Amazon even shut down an internal leaderboard encouraging employees to use AI more. Now, more companies are adopting multi-model strategies, choosing models based on need—giving Chinese large model firms more opportunities.

Of course, model capabilities remain crucial, but the market is no longer satisfied with ranking leadership; it wants to see real growth, revenue, and profits.

And in China, Zhipu has become the standout player in this transformation.

02 Why Zhipu?

With so many AI companies, why did Zhipu reach a trillion-dollar valuation first?

The reasons are twofold: sustained validation of its model capabilities and its increasing resemblance to a profitable company.

First, model capabilities. A key catalyst was GLM-5.2. On June 12, 2026, the U.S. Department of Commerce, citing national security, required Anthropic to restrict foreign access to its latest model, Fable 5. Anthropic subsequently suspended overseas services for two flagship models.

The next day, Zhipu announced full global access to GLM-5.2.

The timing was impeccable. For many developers, when access to top overseas models became uncertain, a capable and openly available alternative naturally saw its value reevaluated.

More importantly, GLM-5.2 itself proved highly competitive. It supports 1M lossless context windows, enhanced coding capabilities, Day 0 adaptation to multiple domestic computing platforms, and uses the MIT open-source license.

Elon Musk added fuel to the fire. On June 19, when asked on social media when Chinese large models would reach Anthropic's Fable level, Musk replied, 'Probably Q1 2027.'

Zhipu founder Tang Jie responded directly: 'Not that long.' Musk added that while benchmarks might catch up, Anthropic still held a significant edge in practical usability. Tang fired back: 'Focus is the only thing that matters.'

This confidence came from GLM-5.2. Public testing data shows that in the FrontierSWE programming benchmark, GLM-5.2 scored only about 1 percentage point lower than Anthropic's top closed-source model, Claude Opus 4.8, and outperformed OpenAI's GPT-5.5, ranking among the world's leading available models.

This means Zhipu can now be considered a global first-tier player. But what's drawing even more attention is its commercialization speed.

For a long time, skepticism around large model companies has been similar: models keep improving, but can they make money? Zhipu's answer is that it's converting model advantages into revenue growth. Compared to localized deployment, MaaS (Model as a Service) has become Zhipu's most important growth engine.

Simply put, developers don't need to build their own computing power or download models—they can just use Zhipu's capabilities via API calls and pay per token. Currently, the core demand comes from code generation and Agent applications. Zhipu CEO Zhang Peng has also made clear that the company is focusing on coding and Agent directions at this stage.

In March 2026, Zhipu disclosed that its MaaS platform's API business had reached RMB 1.7 billion in annual recurring revenue (ARR), up 60x year-over-year, with over 242,000 paying developers and token calls growing 15x in six months.

More notable than the growth itself is its quality. After GLM-5's release in February, Zhipu raised prices for some services by 30% to 60%, with some API prices increasing by 67% to 100%. Yet call volume still grew over 400% after the hike (note: one reason for the 400% growth was the low base previously).

In other words, user growth wasn't built on low-price subsidies—the model's capabilities had gained some pricing power.

For an AI company, this ability is often more valuable than mere revenue growth. And the market's willingness to value Zhipu at nearly a trillion dollars is largely because this path already has a reference model:

Anthropic.

This year, Anthropic's commercialization has been astonishing. Its annualized revenue grew from $9 billion in January to $45 billion in May. Reuters also reported that the company expects to achieve an operating profit of about $559 million in Q2.

Anthropic seems to have proven that under frenetic growth, large model companies can still be profitable.

And Zhipu is increasingly seen as one of the closest Chinese counterparts to this story. From GLM-5 to GLM-5.1 to GLM-5.2, Zhipu has maintained a roughly two-month update cycle; meanwhile, developer counts, call volumes, and revenue have grown in tandem.

The market isn't betting on today's Zhipu—it's betting on whether it can become China's Anthropic.

03 What must Zhipu prove after its trillion-dollar valuation?

So, can it? The answer isn't optimistic.

Even Anthropic can't escape cost pressures. It admits that planned infrastructure spending may prevent it from maintaining profitability for the full year.

Models require continuous training, and as inference scales up, more clients lead to exponentially growing computing demands. Revenue is rising, but so is the burn rate.

The same is true for Zhipu. In 2025, Zhipu's net loss was RMB 4.718 billion, up 59.5% year-over-year, with R&D spending reaching RMB 3.18 billion, up 44.9%. This means the company spent RMB 4.4 on R&D for every RMB 1 earned.

Meanwhile, capital expenditures fell from RMB 460 million in 2024 to RMB 74.7 million, shifting from heavy asset computing investments to a 'leasing + servitization' model.

But this hasn't changed the core issue: improving model capabilities doesn't naturally reduce costs. Instead, as competition intensifies, investments become more rigid. The industry is stuck in a loop: without stronger models, growth stalls; but stronger models require higher costs.

Meanwhile, coding is the critical inflection point for AI commercialization. It transforms large models from generative tools into productive tools. Anthropic's rapid valuation surge is fundamentally due to Claude Code integrating into enterprise development workflows.

But the problem is that this market lacks stable moats.

OpenAI is striking back fast. Codex's weekly active users surged from 600,000 to 5 million in a short time.

Domestic competition is even denser. Zhipu launched its GLM Coding Plan in 2025 at about one-seventh of Claude's price, attracting over 150,000 paying developers in two months and exceeding 242,000 by year-end. Nine of China's top ten internet companies now deeply use GLM models.

But at the same time, 'everyone is doing coding.'

ByteDance validates through internal engineering before external rollout; Kimi bets on multi-Agent collaboration to enhance complex task handling; Alibaba and Tencent are also building proprietary programming model systems.

Zhipu's choice of an open-source path essentially tries to replicate Android's model: expand ecosystem share by opening model capabilities.

But the problem is that this path doesn't inherently create moats. The MIT open-source license lowers barriers and accelerates developer adoption but also means the technical approach can be quickly replicated. As model capability gaps shrink, competition will rapidly shift to more practical areas: customer acquisition costs, channel capabilities, and enterprise relationships—traditional strongholds of internet giants.

Most importantly, the gap between ARR of RMB 1.7 billion and a market cap of nearly HK$1 trillion is still astronomically wide.

Reference sources:

1. Caixin Weekly: Why is the market buying into Zhipu's trillion-dollar valuation?

2. Caijing Tianxia WEEKLY: Zhipu's 20x surge in six months stuns investors

3. Dingjiao One: RMB 700 million in revenue, HK$1 trillion in market cap—is Zhipu worth it?

4. Wall Street See: Why is the market buying into Zhipu's trillion-dollar valuation?

The cover image and illustrations belong to their respective copyright holders. If the copyright owners believe their works are unsuitable for public browsing or free use, please contact us promptly, and our platform will make immediate corrections.