SenseTime's annual revenue is nine times that of MiniMax, and its staff has plummeted by 34%, but why is its market value still struggling to increase?

(Source: Intelligent Era AGI)

Back then, “AI Four Little Tigers”—SenseTime—was once thriving, eventually completing “the first IPO among AI Four Little Tigers on Hong Kong stocks.”

But today, SenseTime has no choice but to face an “upset” from an AI large-model company—MiniMax.

As everyone knows, before founding MiniMax, Yan Junjie (I/O) had worked at SenseTime Technology for more than 5 years, starting from an intern and rising all the way up; he previously served as SenseTime Group’s Vice President, the Deputy Director of the Research Institute, and CTO of the Smart City business group. This is also the “shining background” Yan Junjie presented to investors when MiniMax entered the AI large-model race.

And on the evening of March 24, SenseTime released its 2025 performance report, yet it still faced losses:

• Revenue reached 5.014 billion yuan, up 32.9%;

• Net loss was 1.782 billion yuan, narrowing 58.6% year over year;

• Adjusted net loss narrowed 54.3% year over year, down to 1.956 billion yuan.

If we follow the “profit-earning timeline points” announced by SenseTime Chairman Xu Li, then clearly, SenseTime did not achieve it in the end.

And to express that profits are “turning positive,” the first chart Xu Li showed at the earnings call: the second-half 2025 EBITDA data—376 million yuan, achieving a return to positive, as well as operating cash flow turning positive.

What is EBITDA? It’s an abbreviation for Earnings Before Interest, Taxes, Depreciation, and Amortization, which is usually translated as “profit before taxes, interest, depreciation, and amortization.” It refers to income before deducting interest, tax items, depreciation, and amortization—i.e., profit that adds back those four cost items: taxes, interest, depreciation, and amortization.

In general, hardly any companies or securities firms are willing to disclose EBITDA to the outside world as if it were true net profit—SenseTime is the exception.

So, although over the past 8 years, SenseTime’s revenue has kept growing and its losses have also been narrowing, at tonight’s earnings call, Xu Li and SenseTime CFO Wang Zhengpin repeatedly mentioned one word—“loss reduction”:

SenseTime’s operating resilience and results transformation in 2025 are reflected in high-quality growth and leap-forward loss reduction.

We have also established the trend of continuing to reduce losses significantly.

Overall, we expect that in 2026, at the level of adjusted capital investment, we will continue to maintain this high proportion of reduction trend, and looking at the full year, the adjusted net profit is very likely to achieve a return to positive for the year.

Worth noting is that, according to Intelligent Era AGI’s exclusive information, as of December 31, 2025, SenseTime had 2,472 employees. Compared with 3,756 the prior year, it net decreased by 1,284 people, representing a year-over-year decline of about 34.2%, and even employee benefits spending dropped significantly.

Evidently, although SenseTime’s 2025 annual revenue is 9.24 times that of MiniMax in the same period, its losses are also just as high. For capital markets, the new story in the AI sector is MiniMax and Zhipu—“AI 2.0.”

And because over the past 8 years SenseTime’s revenue and profits have fluctuated sharply, its stock price and market value seemingly have been unable to receive the same treatment and favor as MiniMax’s market cap of 320 billion.

I’ve tracked SenseTime for more than 7 years, and I still have very strong overall optimism about SenseTime. But with this kind of situation now, everyone is fairly anxious.

After stopping R&D growth, SenseTime is still going through a “winter”

SenseTime has been going through a winter for the past two years.

From large-scale layoffs to sharply reduced internal budgets, and even the so-called “X” that split each business to raise independent financing and operate at their own profits and losses—along with Xu Bing stepping down as an independent director, and the split-off of the Jiying business—everything points to one thing:

Xu Li wants SenseTime to become profitable.

This is easy to imagine. With the two Hong Kong-listed AI companies Fourth Paradigm and Wenwen (Chutalk) cutting investments significantly, their net losses are about to touch 0, and they have a good chance of turning profitable.

And with the “hottest” phase of “the lobster,” MiniMax and Zhipu—which have top large models—have seen their revenue and market value surge dramatically, and they very likely will enter a profit period this year.

Therefore, this is extremely awkward for SenseTime—originally it was supposed to be the first IPO, but the younger brothers all became profitable; while SenseTime did so much technology work and also rode the GPT boom, and even shifted all its financial losses onto its businesses, yet the group still hasn’t turned profitable. This is hard to face investors’ expectations.

So, SenseTime’s first area it moved to “cut” is R&D.

According to the prospectus, SenseTime’s R&D expenses in 2025 reached 3.775 billion yuan, down from 4.131 billion yuan the previous year, a decline of 8.6%.

“Mainly due to reduced employee benefit expenses, partially offset by increased server operations and cloud service costs.”

In other words, originally AI companies need to invest heavily in R&D technology, hire PhDs, and nurture stacks of AI talent. But now employee benefits are gone, and they still need to invest in compute power operations and buy compute cards.

Regarding the doubts from securities firms, what Xu Li said at the earnings call was this:

Everyone, remember that among the three expense items in operating costs in 2025, we saw declines, with a total annual decline of 11%. So when it comes to our cost reduction and efficiency improvement, we actually have many specific measures. For example, the close integration of large devices and large models itself creates extremely high cost-effectiveness. For example, building a talent base in Wuhan also helps us further lower labor costs.

In addition, we continuously optimize job positions and office rents domestically and internationally—so in fact, overall, from multiple perspectives and in many small ways, we keep controlling costs.

Then, we also keep optimizing our revenue structure. Our provision for accounts receivable’s negative impact on the income statement is actually being repeatedly reduced. Remember, it was -780 million yuan in 2024 and fell to only -290 million yuan in 2025. This is also one of the big factors.

Finally, there is the 1+X strategy and deepening. It has led to some X enterprises successfully obtaining external financing and achieving off-balance-sheet status, which also creates some additional room for the group’s overall profit improvement.

How to balance R&D investment and loss reduction?

Actually, we have always treated R&D as the core competitiveness of the company. In 2025, we indeed further increased our compute power investment for R&D. But because our R&D productivity improvement is very significant, and because we also have some strongly R&D-driven X businesses that successively became off-balance-sheet, our total R&D costs ended up declining. Moreover, we believe these two things—efficiency improvement and stronger R&D results—are not contradictory.

I won’t judge whether this passage is right or wrong, but I think the logic in his statement is actually hard to be internally consistent—scenarios, data, software, hardware, and talent are all indispensable. Only a powerful ecosystem can truly unlock AI’s limitless potential.

This is a PPT released at the Siemens AI Conference. I’m quite in agreement with it.

At the earnings call, Lin Dahua (DH) mentioned that in the fourth quarter of 2025, a new generation of native multimodal model architecture, NEO, was officially launched and open-sourced. It completely abandons the traditional design of visual encoders and the patchwork-style assembly of backbone models. Instead, it builds a unified language and vision end-to-end “human-like life state” architecture from the ground up. This new architecture brings a significant improvement in model learning efficiency.

Lin Dahua revealed that this year’s second quarter, SenseTime will update two model technologies:

1. RiRiXin will release version 7.0, expected to be released in the second quarter of this year;

2. SenseTime will release a brand-new open-source model based on the second-generation NEO architecture. This model has already been validated first in the industry for understanding and generating unified new scale laws (Scaling Law) under a native multimodal architecture, using new “associations” to understand generation under a unified architecture, with an independent learning curve—significantly surpassing the previous open-source benchmark model for unified understanding. It even takes ByteDance as the benchmark and is expected to achieve a multiple-level leap in both performance and cost-effectiveness.

Lin Dahua said the significance is that it will open up more downstream application space across the entire AI ecosystem, broadly empowering agentic AI application scenarios.

As for SenseTime’s AI application situation…

I won’t comment—I can only say the competition is fierce. Back then, the first-mover code and Office “Little Wanxi” were ultimately PK’d by Zhipu Byte’s Vibe Coding; and even the ranking in self-media—can that be considered performance?

I don’t understand.

But the chart from SenseTime’s earnings call can give everyone some food for thought—how AI can be applied and deployed in practice, and even how it might be replaced: AI + becomes, AI + search, AI + office software, AI + research, embodied intelligence, AI + marketing, and so on.

Finally, let’s talk about business revenue:

• Revenue from generative AI was 3.630 billion yuan, up 51% year over year;

• Revenue from visual AI was 1.083 billion yuan, up 3.4% year over year. Of this, in the second half of last year, the revenue growth rate for this segment reached 20.9% year over year.

• Revenue from X innovation was 302 million yuan, down from 321 million yuan the previous year; the share of total revenue dropped from 8.6% to 6.0%.

SenseTime said that the decline in X business revenue was mainly because the intelligent driving business detached from the consolidated financial statements in August 2025.

SenseTime said that as time goes on, the composition of the X innovation business is expected to evolve—either by incubating more X innovation businesses, or by having existing X innovation businesses attract external investors and detach from the consolidated financial statements. Therefore, going forward, the year-over-year data for this revenue will become less meaningful.

I understand that SenseTime may incorporate all profitability projects of all Guo Xiang Capital into SenseTime’s X business—for example, Qiangnao Technology, which is about to IPO.

SenseTime has also hinted that SenseTime Healthcare might change the “Xiangting” brand, and that Xiwang might become another profitable segment within SenseTime’s X business.

Large devices: high inference demand; heterogeneous training; it has already achieved a performance catch-up against overseas top chips on domestically made hardware that is constrained by process.

SenseTime’s large devices are actually an important business that I’m fairly bullish on. After all, SenseTime has cards and compute power; selling cards, selling compute power, and selling cloud services are already major profit-making channels for current big manufacturers and AI cloud providers.

SenseTime disclosed that as of the date of this performance announcement, the total operating compute power scale of SenseTime’s large devices has increased to 40,400 PetaFLOPS (FP16).

In terms of cases, with the support of CATL, SenseTime built the world’s first AI intelligent system that fully connects the entire chain of “compute power management—IDC operations—energy storage system.” The system can accurately predict power load via large models and dynamically generate the optimal energy dispatch strategy. “We expect this system can achieve a 7% savings on electricity costs and reduce more than 4,000 tons of carbon dioxide emissions.”

Yang Fan said at the earnings call that while ensuring continuous “breaking through limits” in model R&D, they also ensure efficient, low-cost operation of upper-layer applications. Taking the LightX2V world model inference system as an example—leveraging frontier technologies such as step distillation and extreme quantization—we not only were first to achieve real-time video generation of simulations of the physical world, but also, under constraints on domestic hardware process technology, achieved an overperformance compared with top overseas chips. With outstanding compute efficiency, this system received official recommendations from multiple leading vendors including Alibaba and Tencent. Its open-source model has surpassed 10 million cumulative downloads on HuggingFace and ranks consistently in the global top ten (comparable to OpenAI’s gpt-oss).

What’s more, SenseTime has even achieved million-card heterogeneous mixed training. Of course, this also relies on support from compute card partners such as Moore, Biren, Muxi, Huawei, and others.

Yang Fan also revealed that SenseTime’s overall compute power demand is changing: domestic inference has strong demand, but heterogeneous compute is mostly for training, using different chips to achieve optimal cost-effectiveness.

When it comes to domestic products, what we see now is that at the very beginning, its market primarily came from government and enterprise customers, including major central and state-owned enterprises, as well as many domestic research institutions. In fact, they have some clear needs for domestic solutions, and based on the past two years, the scale has been continuously growing.

But in the past few months, there has been a clear new trend: more and more internet companies and mid-to-early stage science and technology innovation firms are increasingly taking an open and embracing attitude toward domestic chips.

So from this perspective, it’s actually very good. It means that after all these years, SenseTime’s coordination with domestic hardware vendors—along with all the exploratory work we’ve done, accumulating such knowledge and technical reserves—means we also hope we can seize the opportunity of the current market trend of domestic compute acceleration and capture it well.

Looking ahead to 2026, SenseTime says it will continue to deeply cultivate the native multimodal architecture, establish SenseTime’s global leading position in integrating native multimodal large models and spatial intelligence. At the same time, it will seize first-mover opportunities in the native AI agent market to achieve two-way explosive growth in both traffic scale and commercial value. It will also fully push domestic chip adaptation, continuously lowering the inference costs of large models.

© This article is original content by Intelligent Era AGI (weixin6060000)

For massive information and precise analysis, all in the Sina Finance APP