Zhipu's "Aolong" is a bit hot to handle

Ask AI · Why is the continuous expansion of losses despite surging revenue?

Zhipu’s Longmen has only just crossed the halfway point.

Author | Jingxing

Editor | Gu Nian

From a big rise to a big fall, then back up again.

From April 2 to 8, Zhipu first surged 31.94% in three trading days after the earnings report, then plummeted 14.86%, and on April 8, it rose again, with the increase once approaching 20%. The value revaluation driven by Token economics is facing fierce debate between market bullish and bearish views.

The earnings report itself also presents a rather torn outcome. On March 31, Zhipu released its 2025 financial report showing that in 2025, the company achieved total revenue of 724 million yuan, a significant year-on-year increase of 131.9%, marking three consecutive years of doubling revenue.

What’s even more impressive is the cloud MaaS platform revenue. In 2024, Zhipu’s MaaS platform revenue accounted for only 15.5%, rising to 26.3% in 2025, with business scale reaching 190 million yuan, a year-on-year surge of 292.6%. The annual recurring revenue from the MaaS platform skyrocketed 60 times over the past 12 months, reaching 1.7 billion yuan. Against the backdrop of the domestic large-model industry mired in a price war, Zhipu completed a cumulative 83% API price increase in the first quarter of 2026. Meanwhile, the platform’s Token call volume not only did not decline but grew by 400% against the trend.

But contrasting sharply with the high revenue growth is Zhipu’s continuously expanding huge losses. The financial report disclosed that in 2025, Zhipu’s net loss for the year reached 4.72B yuan, with an adjusted net loss of 3.18B yuan, expanding by 29.1% year-on-year. Its annual R&D expenses hit 3.18 billion yuan, up 44.9%, more than four times its revenue scale.

This forms the basic fundamentals of the world’s first large-model stock: while revenue surges, losses remain high; local deployment business gross margin declines, but cloud deployment business volume and price both rise, with gross margin increasing from 3.3% to 18.9%, showing signs of a second growth curve. In response, Zhipu CEO Zhang Peng introduced a brand new narrative formula — Business value in the AGI era = Intelligent upper bound × Token consumption scale.

And Zhipu’s AGI growth story is inseparable from the concentrated explosion of high Token consumption scenarios since last year. From developers flooding into programming scenarios to Agent products releasing Token consumption potential, Zhipu has seized the rising beta of the large-model market. In other words, Zhipu’s revaluation is a victory on the demand side.

A lobster with two ways to eat?

In the business ecosystem, Zhipu positions itself as a substitute for Anthropic. Zhang Peng once publicly stated, “Our model is good enough, at a top level worldwide. We have huge advantages in price and cost. We joke that if Anthropic sells for $200, we sell for 200 RMB.”

This aligns with both parties’ business models. The GLM series models fully align with Anthropic’s Claude model, trying to expand the user base with lower Token prices, drive API call volume growth, and feed back into model iteration and Token fee increases. The financial report shows that Zhipu has become one of the domestic companies with the highest paid Token consumption, leading the domestic large-model market from paying for low prices to paying for results.

But regarding the attitude towards lobsters, Zhipu and its benchmark Claude developer Anthropic are diverging.

On April 4, Anthropic announced that Claude’s subscription quota no longer supports third-party tools like OpenClaw. Some users said OpenClaw is a bottomless Token pit, and subscription-based Tokens can’t withstand it: “When tasks start reading a file, a few rounds of calls are enough. The current MCP (Claude’s external communication protocol) is still single-round, each time only calling one tool, waiting for feedback, then calling the next. A single task can consume tens of millions of Tokens.”

According to platform plans, Claude’s subscription modes include Pro, Max (with 5x and 20x versions), and Team (standard and premium seats), priced at $20/month, $100/month, $200/month, and $25/user/month, $125/user/month respectively. In summary, the more money spent, the higher the usage quota, the stronger the model permissions, and the more advanced features. These are more suitable for developers who need high-frequency model calls and Token consumption.

In the era of intelligent agents, Anthropic is doing subtraction, while Zhipu tends toward addition.

On March 10, Zhipu released AutoClaw, supporting one-click installation to greatly simplify lobster usage; six days later, Zhipu released the deeply optimized base model GLM-5-Turbo for lobster scenarios, increased API prices by 20%, and launched lobster subscription packages at 39 yuan/month for 35 million Tokens, and 99 yuan/month for 100 million Tokens.

Anthropic, on the other hand, promotes its own lobster replacement tools, using official versions to meet users’ complex task needs and avoid unlimited overuse of computing power by subscribers.

Behind this are completely different business pursuits. While early movers chase independent lobster ecosystems, latecomers are still anxious about how to fill their stomachs.

Public data shows that Anthropic’s latest annualized revenue has reached $14 billion, compared to Zhipu’s revenue soaring to $700 million, which is not in the same order of magnitude.

In other words, facing Anthropic’s strong enterprise API ecosystem, Zhipu’s API platform is still in the early stage of ecosystem construction, and its user awareness remains at the level of overseas model alternatives at low prices, without a leading advantage. A comment from a Zhipu user explains this well: “(Cost performance) depends on who you compare with. You get what you pay for. Don’t expect too much.”

Token calls have no weak links

If “the intelligent upper bound determines pricing power” is Zhipu’s core narrative, then the lobster craze sweeping the AI circle this year provides tangible support for Zhipu’s story — replicating Anthropic’s business path, aiming for B to C.

For a long time, Zhipu’s revenue structure has been clear: “two legs.” One is localized deployment, emphasizing a major share in government and enterprise markets, indirectly reaching enterprise development teams, supported by high ticket prices and high gross margins, forming the early revenue base; the other is MaaS cloud deployment, serving enterprise and individual developers through API calls and subscription services, billed by Token consumption or monthly packages, with stronger scalability. Zhipu has also released AI code editors and code analysis tools to increase developer stickiness and bind to Zhipu’s development ecosystem.

Historically, Zhipu relied mainly on localized deployment revenue, focusing on government and enterprise markets, indirectly reaching development teams. Its C-end business has long been marginalized. A notable phenomenon is the reduction in marketing efforts toward the C side.

During the Spring Festival of 2026, Baidu, Alibaba, ByteDance, and Tencent launched a red envelope war, making AI-led red envelopes a new year’s user mindset. Meanwhile, Zhipu remained silent in promotion, with C-end activities basically halted. The next time ordinary users pay attention to Zhipu is on March 10, when the official released the AutoClaw promotional article “Today, install Lobster on every computer.” By March 16, Zhipu quickly updated the world’s first deeply optimized base model for lobster scenarios, GLM-5-Turbo, specifically optimized for lobster needs, seeking the next growth point in Token consumption.

This is the reason why the market overestimates Zhipu. During the earnings call, when asked whether “API growth is sustainable,” Zhang Peng confidently said that the industry has been looking for a simple, economical, and powerful business model to boost growth in recent years. Today’s AI capabilities have moved from being usable and fun to truly solving increasingly complex and important problems, turning Token API calls and consumption into real economic value. OpenClaw will make future API and Token consumption grow exponentially, forming a deterministic, long-term trend.

But in the lobster wave, Zhipu is not the biggest winner in the domestic market.

OpenRouter OpenClaw model data shows that Zhipu’s GLM-5 once topped the popularity charts, but for most of the time, models like Kimi, MiniMax, Step (Stair Star), MiMo (Xiaomi), and Qwen (Alibaba) led the OpenClaw call volume charts.

The OpenClaw model rankings show a relay-style rotation of top spots. Kimi K2.5, due to its cost advantage, was set as the official free main model of OpenClaw and quickly topped the charts; MiniMax and Step, with their multi-modal versatility and base model advantages, strongly reversed the trend.

The reason is that in Token bottomless scenarios like OpenClaw, the main consumption comes from routine automated calls by larger user groups, while models like GLM, known for complex production tasks and higher call costs, face scale resistance in this Token war.

Data shows that in 2025, Zhipu’s cloud deployment gross margin surged, and with the release of GLM-5 and GLM-5-Turbo models, prices were increased twice, gaining market recognition. But looking at the revenue structure, Zhipu’s localized deployment revenue reached 535 million yuan, accounting for 73.7% of total revenue, still its main income source.

Although Zhipu has optimized GLM-5-Turbo for lobster scenarios, emphasizing better performance in long-chain tasks, it has not overtaken the C-end market due to global brand inertia. Instead, it quickly gained access to internet giants like ByteDance, Alibaba, and Tencent after release, becoming an auxiliary role in their intelligent transformation.

Large models are a marathon race

From Zhipu’s 2025 financial performance, it seems to have achieved a perfect comeback, but behind the numbers, the cost paid is substantial.

In 2025, Zhipu’s sales costs increased over 200% year-on-year to 427 million yuan. The financial report shows this surge is due to increased computing service expenses from business expansion; additionally, R&D expenses rose 44.9% to 3.18 billion yuan, driven by higher employee costs and payments to third-party cloud providers.

Compared to internet giants with heavy assets and autonomous computing power, Zhipu, as a technology-oriented supplier, prefers a light-asset model. Its computing power in 2025 heavily depends on leasing from third-party providers, with downstream clients bearing the cost. This means Zhipu must pay high procurement costs for computing power to support explosive Token call volume, without building its own infrastructure.

This route has both pros and cons. Compared to BAT’s pursuit of autonomous, controllable infrastructure, Zhipu emphasizes building barriers at the technology and application layers, investing all available funds into core competitiveness. The downside is that the company must maintain extremely strong model competitiveness to offset rising computing costs. From gross margin levels, Zhipu’s overall gross margin dropped from 56.3% in 2024 to 41% in 2025, partly due to high computing costs squeezing profits.

On the other hand, the large-model industry is essentially an endless marathon. From DeepSeek to Seedance, from Manus to OpenClaw, no one knows when the next windfall will arrive.

At the AGI-Next frontier summit earlier this year, Zhipu founder and chief scientist Tang Jie said:

“Large models are now more about racing speed and time. Maybe our code is correct, and we can go further in this aspect, but maybe after failure, it’s just half a year, and then it’s gone.”

Every new product can redefine industry perceptions of AI capabilities. In this race, if a generation of models cannot keep up, all previous advantages may be lost in a short time.

Some industry insiders close to large models say that the product iteration of large models is extremely rapid. Even if a single model can achieve a leading capability, it’s hard to fundamentally change the game. Long-term resilience and not falling behind are more important:

“For model companies, it’s very hard to stand out. Outsiders don’t perceive it clearly. Application companies might achieve positive ROI or high gross margins, build barriers along the tech route, and even release their own models, which can attract more overseas attention than flagship models from top model companies.”

This sketches the survival dilemma of model companies: without deep infrastructure and application ecosystems, it’s hard to control market pricing power. Zhipu has been fortunate to seize the lobster-driven intelligent agent wave, but how long this dividend can last and where the next challengers will emerge remains to be seen.

**

< PAST · Recommended Reads >

**

＞霸王茶姬 is far from the “second-tier” moment

**

**

＞梁柱难撑腾讯音乐的买买买

**

**