SenseTime Group: Generative AI Pivot and Valuation Reset

My qualitative label is: a company in transition. There are three reasons. First, the revenue engine has shifted; in 2025, generative AI replaced vision AI as the absolute core. Second, the income statement has reached an inflection point, but the cash flow and capex view has not caught up. Third, its competitors are AI utility-type players with deeper capital strength, such as Baidu and Alibaba, and vertical enterprise-intelligence players such as Fourth Paradigm and iFLYTEK, rather than traditional software companies. SenseTime's position is still unstable. It has left the old world, but it has not truly secured its seat in the new one.

History, Financial Review, and Share Price History

SenseTime emerged from the industrialization of a technical leap in Chinese academia. The Multimedia Laboratory at The Chinese University of Hong Kong had accumulated years of research in computer vision, and the academic paths of Tang Xiao'ou, Wang Xiaogang, Lin Dahua, and others directly shaped the company's early DNA. CUHK's commemorative article was direct: in 2014, the team made a breakthrough in facial recognition, believed "the time had come for technology to move toward industrialization," and founded SenseTime. The later board and core management team continued this academic entrepreneurship path: Xu Li handled strategy and operations, Wang Xiaogang led research, and Lin Dahua led scientific direction.

Early SenseTime solved the problem of putting advanced vision algorithms into real-world scenarios. Its first validated use cases were mobile beautification, financial identity verification, urban video understanding, transportation, and commercial analytics, rather than today's large models. The business model at this stage was essentially software licensing, project implementation, and algorithm integration. Gross margin was high, and the story was compelling. The problem was that smart-city and public-sector mega projects quickly pulled it into a structure with very slow collections. In 2021, revenue was RMB4.700 billion and gross margin was as high as 69.7%, but receivable turnover days had already stretched to 319 days. High gross margin did not automatically become high cash flow. This hidden risk later ran through almost the entire post-listing period.

SenseTime's development can be divided into four stages. The first was the algorithm commercialization period from 2014 to 2018: the company used leading vision AI models and scenario deployment capabilities to expand across mobile phones, finance, cities, and autos, and outsiders treated it as a Chinese AI unicorn. The second was the sanction and listing period from 2019 to 2021: the U.S. Commerce Department placed SenseTime on the Entity List in 2019, OFAC added the company to the NS-CMIC list in 2021, and the IPO was temporarily delayed, but it eventually listed on the Hong Kong Stock Exchange on 2021-12-30 at HK$3.85, raising about US$740 million. The third was the clearing and redirection period from 2022 to 2024: old smart-city and traditional vision businesses contracted, total revenue fell to RMB3.406 billion in 2023, and although it recovered to RMB3.772 billion in 2024, traditional AI still dropped 50.6% year on year in H1 2024. At the end of 2024, the company completed its "1+X" restructuring and formally made generative AI the core "1." The fourth is the validation period that began in 2025: revenue reaccelerated, H2 EBITDA turned positive, and the market started to take another look.

2023 was SenseTime's true turning point. On the surface, revenue fell 10.6% to RMB3.406 billion and net loss was still as high as RMB6.495 billion, which looked like a company whose old story had reached the end. Structurally, however, it was already actively cutting low-quality traditional business. In its 2023 results communication, the company explicitly wrote that traditional business as a share of total revenue fell from 81.9% in 2022 to 53.9% in 2023, with smart-city revenue already below 10%. This step was painful but necessary, because it pulled the company out of the old project-based logic of "many, scattered, slow-collection" deals and freed resources for the later generative AI push.

2024 was a year when organization mattered more than financials. Financially, revenue grew 10.8% year on year to RMB3.772 billion, generative AI revenue doubled year on year to RMB2.404 billion, vision AI revenue still fell 39.5% year on year to RMB1.112 billion, and full-year net loss was RMB4.307 billion. Beyond these numbers, the real key was the restructuring completed by management at the end of 2024: generative AI and vision AI were grouped as the core "1," while intelligent vehicles, smart healthcare, home robotics, smart retail, and other businesses were placed into "X," with relatively independent operating units. Reuters described it accurately at the time: SenseTime was forcing itself from a "vision AI company" into a "generative AI company."

2025 has so far been the year that most resembles the "new SenseTime." In April, the company released SenseNova V6 and SenseCore 2.0, emphasizing multimodal long chain-of-thought, global memory, reinforcement learning, and lower inference costs. The August interim report showed H1 2025 revenue rising to RMB2.358 billion, generative AI revenue of RMB1.815 billion, or 77% of total, and a significantly narrower operating cash outflow. By the time full-year results were disclosed in March 2026, annual revenue was RMB5.015 billion, generative AI revenue was RMB3.630 billion, vision AI had recovered to RMB1.08 billion, X businesses had fallen to RMB300 million, and both H2 EBITDA and H2 operating cash flow had turned positive. The message to the outside world was clear: the restructuring is visible in the income statement, not just in a slide deck.

This path has not been smooth. Being placed on the U.S. Entity List in 2019 means exports, re-exports, or transfers to listed entities of items subject to the EAR require licenses, and license exceptions are generally unavailable. Being placed on the NS-CMIC list in 2021 directly tightened the trading boundary for U.S. capital and U.S. investors. The company explicitly acknowledged these regulatory developments in its December 2021 supplemental disclosure. In other words, SenseTime has never been an ordinary Chinese technology stock from the first day of listing. Its financing, supply chain, and overseas expansion have always carried the shadow of a "restricted asset."

Founder Tang Xiao'ou died in December 2023 due to health reasons. This was both an emotional event and a governance event. The official obituary confirmed that Tang Xiao'ou passed away on the evening of 2023-12-15. After that, Xu Li became executive chairman and CEO, and the board and operating center moved further toward the executive team. For a company with a strong academic character, this means the company has moved from a stage driven by a founding research leader to one driven by a professional operating team. The tradeoff is clear: faster decisions, but the market will demand operating delivery more strictly.

The table below puts SenseTime's key financial inflection points in recent years side by side. Note that 2021 IFRS net loss was heavily affected by accounting items such as the fair value of preferred shares, so when judging operating quality, revenue mix, gross margin, adjusted loss, and cash flow are more meaningful.

The data already shows the problem. SenseTime is a story of "the revenue engine changed, while gross margin first moved lower," not a linear story where "revenue recovers and profit automatically follows." The reason is that generative AI brings computing power, data centers, server operations, AIDC depreciation, and cloud service costs together, rather than pure software gross margin. Gross margin fell to 41% in 2025, almost 29 percentage points below 69.7% in 2021. That is not bad news by itself, but it shows that the business model has shifted from asset-light algorithm licensing to a heavier infrastructure model that consumes more capex. If the market still views it through the 2021 "high-gross-margin AI platform" framework, it will misread the company.

The share price has moved along this line of cognitive correction. At listing, HK$3.85 reflected a scarcity premium. The 2022-2024 decline reflected the old-business downturn, continued losses, sanctions pressure, and disappointment that SenseTime was not among the top three in the generative AI race. After the 2025 interim and annual reports, the market again began to trade the "profitability inflection point" and "domestic computing power story." The average analyst target price recovered to HK$2.67, and CMBI also raised its 2026-2027 revenue forecasts by 3%-6% after FY25, lifting its target price to HK$2.50. Yet the share price was still only HK$1.48 as of 2026-06-12. This shows the market has not fully believed the story; it is merely giving SenseTime a renewed window to prove itself.

Business Model, Moat, and Governance

SenseTime's 2025 revenue mix is more persuasive than any narrative. Generative AI revenue was RMB3.630 billion, accounting for 72.4%; vision AI revenue was about RMB1.08 billion, or 21.6%; X businesses were about RMB300 million, or 6%. At the same time, IDC's 2024 data shows that SenseTime ranked third in China's large-model development platform market with a 12.2% share. In the vision AI market, it continued to rank first, and the company says it has led for ten consecutive years. In plainer terms: the old core business has not been completely lost, and the new story has already landed in revenue, but the new story is not a market that one company owns.

The issue is what exactly sits inside this RMB3.630 billion of generative AI revenue. The company emphasizes model training, fine-tuning, inference, Agent applications, and industry solutions. But from the characteristics of profit and cash flow, a meaningful portion is not pure software subscription. It looks closer to mixed revenue from "computing power leasing + model services + project integration." There are three pieces of evidence. First, 2025 gross margin was only 41.0%, well below typical high-gross-margin SaaS. Second, 2025 investing cash flow had a net outflow of RMB3.507 billion, mainly related to purchases of property, equipment, and intangible assets. Third, the company conducted consecutive placements, and the use of proceeds repeatedly mentioned "expanding the AI cloud platform," "expanding SenseCore," "improving domestic chip adaptation," and "Generative AI R&D and commercialization." It does not look like a light software machine. It looks more like an AI power plant still spending heavily on its base.

From an operating leverage perspective, its cost structure is also heavier than that of an ordinary software company. Fixed costs have two layers: one is R&D staff, model training, and algorithm engineering; the other is servers, racks, power, depreciation, cloud resources, and AIDC operations. Once revenue rises, margins can indeed improve, and this began to show in H2 2025. But if revenue growth slows, profit can be quickly consumed again, because computing power and R&D costs are both hard to compress immediately like sales expenses. This operating leverage is harder than traditional SaaS, and also more dangerous. It can pull H2 EBITDA positive, and it can also push EBITDA back into negative territory in the next price war.

I think SenseTime has only three moats that truly stand. The first is its long accumulation in vision AI. This is the combination of ten years of computer vision foundation, project delivery experience, and industry customer relationships. It is not a large-model startup that appeared only in 2023, which gives it an experience advantage in security, commercial analytics, smart terminals, smart vehicles, and other scenarios where it knows how customers actually deploy systems. The second is integrated infrastructure capability. SenseCore is not a slide-deck concept. It has become the base for selling services externally and running models internally, and IDC's placement of SenseTime among the top three large-model development platforms suggests this toolchain is not an empty shell at the customer end. The third is industry deployment rather than pure chatbot capability: Sensechat, office/knowledge/analytics Agents, cultural tourism, education, healthcare, and autos are all directions where it is trying to press foundation models into real scenarios.

There are also two claimed moats in market promotion that I do not fully buy. The first is "model capability itself is the moat." In today's Chinese AI market, this is hard to sustain, because Alibaba's Tongyi, Baidu's ERNIE, ByteDance's Doubao, and a group of open-source models are rapidly commoditizing pure model-layer capability differences. The second is "computing power scale is the moat." Computing power is important, of course, but if it mainly supports low-margin inference leasing or one-off delivery, it is more the cost side of the moat than pricing power. For SenseTime, the real determinant of future margin is how much revenue comes from reusable software and application layers, not how many PFLOPS it has.

Governance is where SenseTime deserves a discount. The company has a weighted voting rights structure. Xu Li, as co-founder, executive chairman, and CEO, is also a WVR beneficiary. This ensures control during the transition period, but naturally reduces ordinary shareholders' voice. In addition, the company has frequently used general mandate placements over the past two years: it raised about HK$2.5 billion in July 2025, another about HK$3.15 billion in December, and about HK$3.23 billion again in April 2026. It is difficult for the market to give a "governance premium" to a WVR company that is still loss-making and repeatedly needs external equity funding.

Investors also should not let the PR phrase "record collections" sweep away two financial-quality issues. First, customer concentration is rising. In 2025, one single customer contributed 19.0% of revenue, compared with 12.2% in 2024. Second, long-aged receivables remain the shadow left by the old era. The company disclosed in its 2025 interim report that, as of 2025-06-30, trade receivables aged more than 3 years still stood at RMB3.997 billion, and those aged 2-3 years were RMB1.008 billion. Management also acknowledged that some customers, especially those related to the public sector, were constrained by budgets, making long-aged collections still difficult. The company said at the end of 2025 that receivables aged over 2 years had improved. That deserves recognition, but it does not erase the risk.

Historical controversy should also remain on the desk. At the end of 2023, short seller Grizzly Research accused SenseTime of suspicious revenue round-tripping. Within the scope I could verify this time, I did not see subsequent public primary disclosures of regulatory penalties, auditor replacement, or financial restatement. Therefore, this cannot be written as an established fact. But it does remind us of one point: for AI companies that once had a high share of project-based revenue, revenue quality, related structures, and collection cadence will always matter more than "model ranking."

Industry, Cycle, and Horizontal Competitors

Placed back into the industry, SenseTime faces three overlapping markets, not a single track. The first layer is China's enterprise large-model platform market, where IDC's 2024 data ranked SenseTime third with a 12.2% share. The second layer is the vision AI software market, where SenseTime remains a leader. The third layer is the larger "AI cloud + inference infrastructure + industry applications" market, where it now faces full-stack giants such as Baidu and Alibaba rather than traditional CV companies. In other words, SenseTime is not competing inside a clean boundary. It is moving among multiple profit pools.

The industry is still in a growth phase, but the location of profit pools has begun to change. The model layer itself is increasingly being flattened by open source and low-price strategies from large companies. Profit is gradually concentrating at both ends: stronger and cheaper infrastructure and cloud resources on one side, and applications that can reach customer workflows, create reuse, and form subscriptions on the other. Baidu disclosed in Q4 2025 that AI Cloud Infra annual revenue was about RMB20 billion, up 34% year on year, with subscription revenue for AI accelerator infrastructure growing rapidly. In Q1 2026, AI Cloud Infra revenue further increased to RMB8.8 billion, up 79% year on year. Alibaba's Cloud Intelligence Group reached FY2025 revenue of RMB118.028 billion, grew 36% year on year in the December 2025 quarter, and saw AI-related product revenue grow at triple-digit rates for ten consecutive quarters. These two numbers reveal a harsh reality: China's AI battlefield is shifting from "who launched the model first" to "who can run models cheaper, sell them wider, and bind customers deeper."

SenseTime therefore naturally carries three cyclical attributes. It has a technology iteration cycle, a capex cycle, and a clear policy cycle. Technology upgrades determine model and inference efficiency. Capex determines whether SenseCore can keep up with demand. The policy cycle determines whether state-owned capital, public-sector customers, and large enterprises are willing to purchase local models and domestic computing power. Reuters' 2026 reporting on the rollout of "AI Plus" mentioned that China is pushing AI integration in key industries to reach 70% by 2027 and 90% by 2030. Another report mentioned that China was studying a large-scale national data center construction plan and strengthening the use of domestic supply chains. For a company like SenseTime, this is clearly a tailwind. But who gets paid for the policy wind ultimately depends on delivery capability, price, and financing capacity, not slogans.

Geopolitics is a long-term structural constraint, not a one-time transaction. The Entity List creates licensing barriers for items subject to the EAR. NS-CMIC restricts U.S. investors from trading relevant securities. For SenseTime, these two impacts have landed through three channels. In financing, U.S. dollar capital and some international institutions are naturally more cautious. In the supply chain, the company's public uses of proceeds repeatedly mention "improving domestic chip adaptation" and "building an AI cloud stack based on domestic chips." In overseas markets, the company has not been completely blocked, but it has made the Middle East and other non-U.S.-sensitive markets its breakthrough path, and in 2025 it deployed China's first overseas domestic-computing-power cluster in Saudi Arabia. The conclusion is simple: sanctions did not kill SenseTime, but they pushed it onto a path that is more capital intensive, more geographically selective, and more localized in supply chain.

Horizontally, SenseTime operates in a field with ample competitors. The truly relevant comparisons are four types of players, not old security-industry rivals. Baidu has become an AI utility company: it uses search cash flow to fund AI Cloud Infra and AI applications, can fight on price, and can endure over time. Alibaba is another, harder-to-handle opponent: cloud is the base, Tongyi is the model, e-commerce and office ecosystems provide distribution, and capex has almost no ceiling. Fourth Paradigm behaves more like an enterprise AI software company: its revenue scale is close to SenseTime's, but its revenue mix leans more toward platform software and industry solutions; hardware drags gross margin, but software characteristics are clearer. iFLYTEK is a "scenario-based national team" in China's government, education, and healthcare markets. Its large-model API/MaaS revenue was RMB385 million in 2025, up 263% year on year, but its larger base comes from education hardware and industry applications, not from purely competing with SenseTime in foundation models.

The real difference between SenseTime and these companies lies in where the money comes from. Baidu and Alibaba have mature cash cows, so they can build infrastructure first and wait for commercialization later. Fourth Paradigm's money comes more from enterprise software delivery, so the market focuses more on revenue quality and the speed of approaching break-even. iFLYTEK's money comes from education, government-enterprise, and integrated software-hardware ecosystems, so even if large-model API is still not large, it has a sufficiently large application base. SenseTime's problem is that it wants seats at all three tables: infrastructure, model platform, and industry applications. But it does not yet have a mature cash-cow table to feed the other two. That is why the capital market worries most about whether it will remain a "fast-growing heavy-asset AI business."

If I must assign SenseTime an ecological niche, I would define it as a transitional platform moving from a vision AI leader to a full-stack generative AI challenger. It is not a foundation-model leader in the industry's first-principles sense, not the price setter in cloud, and not a player with irreplaceable channels in a vertical scenario. Its opening is this: it understands delivery better than pure startups, pivoted to LLMs earlier than traditional CV companies, and sits closer to infrastructure than pure application companies. But if the industry enters a large-scale price war, or if customer procurement shifts toward cheaper general models plus integrators, this niche will weaken first rather than strengthen first.

Current Fundamentals, Bull-Bear Debate, and Valuation

SenseTime does not disclose quarterly reports, so the "last four quarters" can only be approximated using H1 2025 and H2 2025. In H1 2025, revenue was RMB2.358 billion, generative AI revenue was RMB1.815 billion and accounted for 77%, adjusted EBITDA was -RMB521 million, and operating cash flow remained a net outflow but narrowed significantly from earlier periods. Based on full-year results, H2 2025 revenue was about RMB2.656 billion, loss for the period was only about RMB293 million, EBITDA turned positive at RMB380 million, and operating cash flow also turned positive. The financial trajectory shows that SenseTime's current operating condition has "pulled out a recovery slope," rather than "remaining in free fall."

But this slope is still not enough to reassure investors, because once capex is added back, the story immediately reveals its true shape. Full-year 2025 operating cash flow had a net outflow of RMB301 million, while investing cash flow had a net outflow as high as RMB3.507 billion, mainly due to purchases of property, equipment, and intangible assets. In other words, the company has passed the ugliest section on the income statement, but on the cash flow statement it remains a business that consumes significant capital. This distinction is crucial: H2 EBITDA turning positive proves the business model has operating leverage, but it does not prove the business can already roll forward on its own.

The market is mainly trading three things now: "profitability inflection point + self-controlled Chinese AI infrastructure + scaled generative AI deployment." The first two have factual anchors: the April 2025 upgrades of SenseNova V6 and SenseCore 2.0, the three placement proceeds all revolving around infrastructure and Generative AI, and the Saudi overseas domestic-computing-power cluster already landed. The third has real but not fully proven commercialization signals: generative AI revenue's share has risen significantly, and Sensechat, office, and industry Agents are entering more daily-use scenarios. My judgment is that this narrative is not "overheated," because the share price remains far from its 52-week high. But it is also far from cold enough to ignore execution risk.

At the analyst level, sell-side consensus did recover after FY25 results. Aggregated data from Investing and Yahoo showed that the 8 analysts covering SenseTime had an average 12-month target price of about HK$2.67. After FY25, CMBI raised its 2026-2027 revenue forecasts by 3%-6%, with the core assumption that 2026 total revenue would rise to RMB6.39 billion, mainly driven by another 35% growth in generative AI, and lifted its target price to HK$2.50. This indicates the market basically recognizes the "direction," while the debate is concentrated on "quality" and "sustainability."

The strongest bull evidence has four points. First, the restructuring delivered faster than many expected, and generative AI has moved from a supplement to the main business. Second, vision AI did not die completely; it resumed growth in 2025, and the company remains a leader in IDC's vision AI market. Third, H2 EBITDA and H2 operating cash flow turned positive, at least proving that revenue growth is not all loss-making showmanship. Fourth, sanctions have not cut off its financing channels or overseas expansion; Hong Kong equity financing continues, and the Middle East market is progressing.

The strongest bear evidence also has four points. First, 2025 free cash flow remained clearly negative, showing the underlying business is still burning infrastructure capital. Second, customer concentration rose to 19%, and when combined with the historical receivables burden, cash-flow quality remains unstable. Third, being top three in large-model platforms does not equal pricing power, because Baidu and Alibaba have far stronger capital strength and cloud infrastructure. Fourth, equity financing has not stopped; three consecutive placement rounds in 2025-2026 will keep reminding the market that this story costs a lot of money.

On valuation, look through the numbers first, then discuss the story. Based on the 2026-06-12 closing price of HK$1.48 and an estimated fully diluted share count, SenseTime's equity market value is about HK$62.48 billion. Based on 2025 revenue of RMB5.015 billion, converted to about HK$5.803 billion, the price-to-sales ratio is about 10.8x. If year-end 2025 cash and term deposits totaling RMB13.17 billion are treated as cash equivalents, without precisely deducting borrowings, cash-adjusted EV/Sales is about 8.1x, and the true figure would be slightly higher. This is a valuation reflecting "the new story is already half established," not a valuation where "the market has completely given up."

In horizontal valuation, the most relevant comparison is Fourth Paradigm, not Baidu or Alibaba. Fourth Paradigm's current share price is about HK$28.74 and market capitalization about HK$15.4 billion. Based on its 2024 revenue of RMB5.261 billion, the current price-to-sales ratio is only about 2.5x, far below SenseTime's. Part of the difference is reasonable: SenseTime's generative AI revenue is growing faster, and the market gives it stronger imagination as an infrastructure platform. But the difference is large enough to remind investors that SenseTime's current valuation is not "undisputedly cheap among AI stocks." The market is giving SenseTime a future growth premium, not a current cash-flow premium.

Below are the three valuation scenarios I use. This is a worksheet that separates "growth, cash flow, and valuation multiple," not investment advice. More important than a precise decimal in the model is whether three assumptions can be satisfied at the same time: whether generative AI can still maintain high growth, whether cash flow can stop depending on placements, and whether SenseCore capex can slow down.

Working backward from this table to the current price gives a clear conclusion: HK$1.48 is not in an "extremely undervalued" zone. It is closer to the lower end of the base scenario. Therefore, the market's current implied expectation is not exaggerated, but it has already prepaid part of the "continued profitability inflection" and "marginal decline in financing need." If the next results show "revenue keeps rising, but capex and placement demand do not stop," the stock will be very fragile. Conversely, if it can maintain positive EBITDA and operating cash flow for two consecutive half-years, the market will quickly move it toward a higher multiple.

The margin-of-safety review is colder than the scenarios. First, the current price is roughly at a premium to the bearish-scenario intrinsic value, so the margin of safety is zero. Second, the most fragile assumption across the three scenarios is that "the share of software/application value-add in generative AI revenue will continue to rise," rather than revenue growth. If revenue is mainly driven by low-margin inference and computing power services, the valuation multiple in the base scenario should be cut. Third, if there is zero growth over the next 3 years and the market prices SenseTime as an "AI project company that is still not self-funding," reasonable value is only around HK$1.25, and the annualized return from buying at the current price would be about -5.5%, clearly below the Hong Kong 10-year government bond yield of 3.32%. Fourth, this is a typical case where "the story has improved, but the price still has not provided enough margin of safety." My conclusion: there is no margin of safety.

Risks, Catalysts, Tracking Indicators, and Research Uncertainty

The first risk most likely to create permanent capital loss for SenseTime is misjudging revenue quality, with medium probability and high impact. Worse than generative AI not growing is generative AI continuing to grow, but mainly through computing power leasing, model-training outsourcing, and project delivery, leaving gross margin, collections, and capex looking nothing like a software platform. The observable indicators are clear: whether gross margin falls below 35%, whether operating cash flow turns consecutively negative again, and whether placements continue. If these three appear at the same time, the market will reprice SenseTime from an "AI platform" into a "heavy-capital project stock." Revenue may still grow, but valuation will step down.

The second risk is customer concentration and long-aged receivables, with medium-high probability and high impact. In 2025, a single customer accounted for 19% of revenue, while in the 2025 interim report, receivables aged more than 3 years were still close to RMB4.0 billion. Management improved collections in H2 2025, but that looks more like cleanup of historical baggage than disappearance of the risk. If the macro environment or local budgets tighten again, the gap between revenue recognition and cash realization will widen again. Shareholders, not accounting profit, will be hurt first, because the company may again choose equity financing to fill the cash gap.

The third risk is a large-company price war, with high probability and high impact. Baidu's AI Cloud Infra had already reached about RMB20 billion for full-year 2025 and RMB8.8 billion in 2026Q1. Alibaba Cloud grew 36% year on year in the December 2025 quarter, and AI-related product revenue grew at triple-digit rates for ten consecutive quarters. They do not need to be smarter than SenseTime in every niche scenario. As long as they are cheaper in model calls, inference resources, and delivery pricing, they can compress SenseTime's profit room. The most dangerous thing in AI is someone else being willing to use a mature cash cow to subsidize 20 points of price, not someone else's model being 5 points better.

The fourth risk is geopolitical and supply-chain constraint, with medium probability and high impact. The legal texts of the Entity List and NS-CMIC do not change every day, but their impact on financing, supply chain, and customer psychology is continuous. The company's public uses of proceeds repeatedly emphasize domestic chip adaptation, which shows this is not an abstract risk. Its choice to place overseas computing clusters in Saudi Arabia also shows that international expansion is geographically selective. If the United States further tightens relevant export, investment, or third-country transshipment rules, SenseTime's ceiling in high-end computing access and overseas commercial cooperation will be lower.

The fifth risk is dilution, with high probability and medium-high impact. Many investors are much more sensitive to "narrowing losses" than to "share-base expansion." But for a company like SenseTime that is still heavily investing in its base, the latter is just as damaging. The three placements in July 2025, December 2025, and April 2026 have already told the market that as long as revenue growth is racing against capex, the company has a strong incentive to refinance when the share price recovers slightly. For ordinary shareholders, this means that even if the business moves in the right direction, per-share value may not be realized in sync.

Positive catalysts are also clear. The strongest would be delivering all three at the same time for two consecutive half-years: positive EBITDA, positive operating cash flow, and no new large placement. That matters more than releasing another model version. Second, if the company can prove that Agents, office, knowledge, and industry applications continue to take a rising share of generative AI revenue, rather than relying only on inference calls to drive revenue, the market will be more willing to assign a higher multiple. Finally, if vision AI continues to recover under overseas and high-quality customer strategies, it will become an underestimated cash-flow buffer.

The tracking table below lists the indicators I think truly matter. Its purpose is to judge whether SenseTime is becoming a company "more like a platform" or "more like a computing power contractor," not to predict every quarter's share price.

Among these indicators, the three most important are operating cash flow, long-aged receivables, and refinancing. They sit closest to the source of permanent capital loss. Model upgrades, ranking lists, and product launches can all be lively, but if cash flow does not improve, old receivables are not collected, and the share base keeps being diluted, shareholder returns will ultimately be consumed. The tracking sources are simple: company annual and interim reports, placement announcements, results-call language, and the financial reports of major competitors.

This report still has four research blind spots. First, the company does not disclose quarterly financials, so judgments about the "latest four quarters" can only be approximated through half-year splits. Second, generative AI revenue is not publicly split into "model/application value-add" and "inference/computing power services," so revenue quality can only be judged indirectly. Third, precise splits among borrowings, maintenance capex, and expansionary capex were incomplete in the primary fragments retrievable this time. Fourth, key private competitors such as ByteDance, Zhipu, MiniMax, and Moonshot lack audited data on the same basis, so horizontal comparison is naturally less solid than with listed companies.

Core primary materials mainly come from SenseTime's 2025 annual report, 2025 interim report, annual and interim results announcements, placement announcements, and official governance pages. Industry and competitor comparisons mainly use official financial reports/announcements from Baidu, Alibaba, Fourth Paradigm, and iFLYTEK. Regulatory references mainly use BIS, Federal Register, and OFAC. Market data and rates use public data from Reuters, Yahoo, XE, and TradingEconomics as of 2026-06-12.

Zen Horizon Synthesis

Across its history, SenseTime has truly proved two capabilities. The first is the ability to turn frontier algorithms into paid products. This was proved in the vision AI era and has been proved again in the revenue surge from generative AI. The second is the ability to shift gears at the worst moment. The smart-city downturn, sanctions pressure, founder's death, and long-term trading below the IPO price could easily have killed the story at any Chinese AI company, but SenseTime at least moved its business engine from old vision AI to new generative AI, and it did so faster than many expected. Its past success was not only an era dividend; it also included real technical accumulation and execution.

But only half of the factors that once made it successful still work today. The academic background and algorithm accumulation remain. The vision AI customer base remains. The organization's willingness to endure pain and rebuild the business also remains. What has weakened is the commercial environment. SenseTime in 2017-2021 operated in a market where AI solution supply was scarce. SenseTime in 2026 operates in a market where model capability is being democratized faster, cloud vendors are exploding capex, and customers are becoming better at negotiating prices. The factors behind past success have not disappeared, but their scarcity has declined.

Horizontally, SenseTime's most real advantage versus peers is that underlying technology, infrastructure, and scenario delivery exist inside one company at the same time. This makes it easier to deliver than pure model startups in many enterprise scenarios, and easier to tell an AI story than traditional integrators. But its weaknesses are also clear: it does not have Baidu's and Alibaba's mature cash flow, Fourth Paradigm's more software-like revenue mix, or iFLYTEK's deep traffic entrance in education, government, and healthcare. This weakness is not temporary. It is the natural squeeze point of SenseTime's business model.

The current valuation is prepaying for two future events that must happen at the same time, rather than rewarding past success. First, generative AI revenue must continue to grow rapidly. Second, the cash conversion of that revenue must improve significantly. This is where the market is most likely to misjudge the company. Many people see H2 EBITDA turn positive and assume operating quality has already transformed. The reality is closer to "the income statement has improved first, while the cash flow statement has not graduated." If the market discovers in 2026 that SenseTime still needs frequent financing to support SenseCore and model R&D, today's price will be hard to call well protected.

The most important variable over the next year is whether the company can achieve positive EBITDA, positive operating cash flow, and no further large placement for two consecutive half-years, not model version updates. Over the next three years, the key variable is whether the share of application-layer and software-layer revenue within generative AI can rise, and whether gross margin can stay above 38%. Over the next five years, the key variable is whether SenseTime can grow from a participant in domestic AI infrastructure construction into an AI platform with truly reusable revenue. If it completes only the first layer, it will be a high-volatility theme stock. If it completes the first two, it deserves a more stable growth-stock valuation. If it completes all three, only then does it deserve to be discussed as a long-term compounder.

Bull and Bear Cases

The bull case can be condensed into four sentences. First, 2025 generative AI revenue reached RMB3.630 billion and accounted for 72.4%, showing that the transition has been written into the revenue mix rather than staying on the org chart. Second, H2 2025 EBITDA and operating cash flow both turned positive, proving that operating leverage is beginning to appear. Third, its position in the vision AI market remains, providing customer entry and a cash-flow buffer for new businesses. Fourth, despite sanctions, the company can still raise funds in Hong Kong and advance in the Middle East, showing that financing channels and non-U.S. overseas channels have not been blocked.

The bear case is equally hard. First, full-year 2025 operating cash flow remained negative, and investing cash flow had a net outflow of RMB3.507 billion, showing that free cash flow is far from turning positive. Second, single-customer revenue share reached 19%, and long-aged receivables remain a major burden. Third, Baidu and Alibaba are pulling AI competition from model capability toward infrastructure and price wars, and SenseTime does not have cash cows of the same scale to support long-term subsidies. Fourth, three consecutive placement rounds in 2025-2026 show that shareholder dilution risk is real, not theoretical.

Pre-mortem

If this investment loses 50% three years from now, the first scenario I worry about would happen in 2027: Alibaba Cloud and Baidu AI Cloud Infra keep expanding, open-source models flatten enterprise customers' perception of model differences, and inference and GPU cloud prices fall 30%-40% within one year. SenseTime is forced to exchange lower prices for growth, generative AI revenue growth drops from 51% in 2025 to 10%-15%, gross margin falls from 41% to 33%-35%, and the H2 EBITDA-positive state disappears again. The market compresses its valuation from a little above 8x EV/Sales to 4-5x, and the share price has a chance to return to the HK$0.70-HK$0.80 range. This scenario does not require the company to "fail." It only requires it to lose growth-stock pricing eligibility in a price war.

The second scenario is more like a slow grind. In 2026-2027, public-sector and quasi-public-sector customer budgets remain tight, and improvement in long-aged receivables stops. To maintain SenseCore expansion and model R&D, the company conducts two more placement rounds of 4%-5% each. Revenue may still grow, and EBITDA may even look acceptable, but per-share value is continuously diluted, and the market ultimately stops believing that "this financing round is the last one." Under this scenario, the stock does not need to fall by half in one day. It can lose 40%-50% bit by bit over 18 months. For shareholders, the result is no different.

Final Research Conclusion

SenseTime is worth researching now, but it is not worth rushing to own at any price. The hardest stage is already half over: the most severe part of the old-business downturn has been cleared, organizational focus has shifted to generative AI, and revenue growth and loss narrowing are improving. But the other half is harder: it must prove that it can turn models, inference, and applications into sustainable cash flow, rather than piling up "AI high growth" through constant balance-sheet expansion and placements. For a company still carrying sanction constraints, WVR governance discounts, heavy-capital base investment, and long-aged receivables from the past, this proof process will not be short.

Therefore, my judgment is to be strict on price. This is neither pessimism nor chase-price optimism. The current HK$1.48 share price is no longer as desperate as the worst moment in 2024, but it also does not give investors enough room for error. It is more a stock that "can be tracked, but should not lead investors to relax discipline just because H2 turned positive." The most effective evidence that would change my view is consecutive positive operating cash flow, a significant decline in long-aged receivables, and an end to dependence on high-frequency equity financing, not another generation of models. As long as two of these three remain unfulfilled, I will not upgrade it from "Watch" to "Cautious Buy."

【Company Profile Scorecard】

• Fundamental quality: Medium

• Growth: High

• Moat: Medium

• Financial resilience: Medium

• Management credibility: Medium

• Valuation attractiveness: Low

• Risk level: High

• Suitable investor type: High-risk speculation

【Investment Rating】

• Rating: Watch

• One-sentence investment thesis: Generative AI growth is real, but cash generation and dilution constraints are still not fully reflected in the share price.

• Acceptable hold price: 1.42-1.92 HKD

• Clearly overvalued price: 2.40-2.75 HKD

• Current price category: Acceptable to hold

• Worth waiting for a better price: Yes. The more ideal trigger would be the share price returning below HK$1.00, or, without a price decline, the company achieving both positive operating cash flow and positive EBITDA for two consecutive half-years with no new large placement. The opportunity cost of waiting is potentially missing valuation repair driven by policy and sentiment.

• Target holding period: 1-3 years

• Expected annualized return: bearish about -19%; base about +13%; bullish about +59%

• Maximum downside risk: about 45%-55%; triggers would be a generative AI price war, gross margin falling below 35%, operating cash flow turning clearly negative again, and a new equity financing round.

• Signals that trigger reassessment: two consecutive half-years of positive operating cash flow; gross margin below 35% for two consecutive reporting periods; single-customer revenue share above 20%; receivables aged over two years stop improving; another placement with >4% dilution within 12 months.

【Ideal/Fair Buy Price】0.85-0.95 HKD Rationale: This corresponds to the bearish-scenario intrinsic value of about HK$1.10-HK$1.20 with at least a 20% margin of safety reserved. Only in this range is the price sufficient to cover the triple uncertainty of weaker-than-expected revenue quality, high capex, and refinancing.

【Valuation Range】

• current: 1.48 (as of the 2026-06-12 close)

• bear (conservative · ideal buy zone): [0.85, 0.95]

• base (reasonable · acceptable hold zone): [1.42, 1.92]

• bull (optimistic · above clear overvaluation line): [2.40, 2.75]

Other Securities Mentioned in the Report

• 09888.HK - Baidu, used as the direct comparison for enterprise large-model platforms and AI cloud infrastructure.

• 09988.HK - Alibaba, used to compare Tongyi and Alibaba Cloud's ceiling in capex, model distribution, and price-war capability.

• 06682.HK - Fourth Paradigm, used to compare enterprise AI software companies' revenue quality, path toward break-even, and current valuation.

• 002230.SHE - iFLYTEK, used to compare domestic computing power adaptation, deployment in government/education/healthcare scenarios, and large-model commercialization paths.

This report is based on public information and does not constitute investment advice. Markets carry risk; invest with caution.