Phancy Group: Revenue Rose 35.6% to 7.135 Billion Yuan While Gross Margin Fell to 34.8% and Operating Cash Flow Stayed Negative for a Sixth Straight Year

The disputed secondary-source segment table in the research brief should be discarded. Official H1 2025 group revenue was RMB2.626 billion, so a purported interim breakdown summing to more than RMB4 billion cannot describe that six-month period. The filing-based FY2025 split is unambiguous: AI Platform revenue was RMB6.552 billion, API revenue was RMB79.9 million and Agentic AI revenue was RMB503.2 million. Their respective growth rates were 32.0%, 129.2% and 93.2%.

The historical platform series also requires care. Under the old FY2024 presentation, the Sage AI Platform generated RMB3.676 billion, grew 46.7% and represented 69.9% of group revenue. Under the new three-engine presentation, the FY2024 comparative for the broader “AI Platform” category was restated to RMB4.965 billion, or 94.4% of revenue. The apparent jump in platform share contains a major classification change. It is not evidence that the underlying mix moved from 69.9% to more than 90% in one year.

The market is trading three linked narratives. The first is an earnings inflection: Phancy expects H1 2026 profit attributable to owners of RMB90–130 million, compared with a loss in H1 2025. The second is an AI-usage inflection: Q1 token-call volume was almost six times the prior-year level, deployable computing resources increased by more than 200%, and Agentic AI orders on hand were 99% above year-end 2025. The third is a capital-markets inflection: Phancy is preparing a Shenzhen ChiNext listing that could bring an A-share valuation reference and RMB3.8 billion of planned new capital.

Each narrative carries a qualification. The H1 profit alert says other income increased substantially because of fair-value gains on investment projects. The token/API business represented only 1.1% of FY2025 revenue. The A-share offering could issue 62.0–186.1 million basic shares, with a potential 15% over-allotment, following an April 2026 H-share placement that had already enlarged the share count by 7.5%.

Profitability quality is the decisive near-term issue. FY2025 IFRS operating loss was RMB133.5 million. That operating loss is already struck after RMB96.2 million of other income. Equity-method results of RMB17.4 million and net finance income of RMB76.3 million then brought the pre-tax loss to RMB39.8 million. Government grants and tax refunds accounted for almost all FY2025 other income. The company recorded an adjusted net profit of RMB6.3 million only after adding back RMB44.1 million of share-based compensation; IFRS net loss remained RMB37.8 million.

Cash conversion was weaker. Operating cash outflow was RMB681 million in 2025, after RMB622 million in 2024 and RMB999 million in 2023. Purchases of property and equipment increased from RMB27.6 million to RMB266.9 million in 2025. A business that reports approximately break-even adjusted profit while consuming close to RMB1 billion after capital expenditure is not yet profitable on owner earnings.

The balance sheet provides time to improve. At December 2025, cash and cash equivalents were RMB1.997 billion, total liquid cash resources were approximately RMB3.774 billion, borrowings were only RMB202 million and the gearing ratio was 2.7%. The April 2026 placement subsequently added net proceeds of approximately HK$1.556 billion.

Receivables remain a warning signal rather than a liquidity crisis. Net trade receivables fell from RMB3.086 billion to RMB2.341 billion in 2025, reducing calculated average DSO from roughly 175 days to 139 days. Yet gross receivables more than one year overdue rose from RMB436 million to RMB914 million, even as the total balance fell, lifting them from 12.9% to 34.7% of gross receivables of RMB2.634 billion. The provision for credit-loss allowance was cut from RMB200.0 million to RMB17.4 million in the same year. That ageing deterioration matters because contractual credit terms are generally stated as no more than 90 days.

The central bull case says scale is finally turning a decade of R&D and enterprise relationships into profit. Revenue has compounded rapidly; selling and administrative expenses have been cut. API and Agentic AI are growing much faster than the core business, and domestic-computing requirements favour an independent orchestration layer. The A-share listing could finance product development without stressing the balance sheet and could expose the company to a deeper domestic shareholder base.

The central bear case says the company is increasing revenue by bundling more hardware, computing and outsourced delivery, while its software-like gross margin and cash economics deteriorate. Customer concentration, point-in-time recognition and ageing receivables weaken visibility. The first reported profitable half may depend substantially on investment revaluation rather than recurring operating profit. Open-weight models and hyperscalers can also commoditise the foundation-model layer faster than Phancy can build high-margin API revenue.

The share-price response to the July profit alert reflected this disagreement. The shares opened higher on July 22 but closed at HK$28.40, down 2.0% on the day. By July 31 they had recovered to HK$29.28. That price was about 47% below the HK$55.60 IPO price and 27% below the HK$40.36 April 2026 placement price. The trailing 52-week range of HK$23.70–70.00 shows how strongly the stock has traded on AI enthusiasm, fundraising and profit expectations rather than a stable earnings base.

Phancy’s qualitative portrait is a company in transition. It has moved beyond a speculative pre-revenue AI story: revenue exceeds RMB7 billion, customer relationships are real and losses have narrowed substantially. It has not yet crossed into high-quality compounding growth because recurring revenue, segment margins, renewal economics and positive free cash flow remain unproven.

Vertical history, financial review and price narrative

Phancy began in September 2014 as a technology start-up built around enterprise decision AI. Founder Dai Wenyuan had worked on machine-learning research at Baidu and served as a principal scientist at Huawei’s Noah’s Ark Lab. His formative problem was how to make machine learning usable by companies that lacked large teams of data scientists, particularly for risk control, marketing, fraud detection and operational decisions. Conversational AI was not the starting point. The company was incorporated in Shenzhen, later moved its registration to Beijing and converted into a joint-stock company in 2021.

The early product thesis was that enterprise AI should be industrialised through an operating platform rather than rebuilt for every use case. Sage packaged model development, feature engineering, automated machine learning and production management into a common environment. Financial institutions were natural early customers because they had large structured datasets, measurable decision outcomes and clear economic incentives to improve approval, fraud and marketing models.

The first stage, from 2014 through roughly 2018, was product validation. The company built credibility in decision-oriented machine learning while China’s large enterprises were beginning to move from traditional business-intelligence rules toward predictive models. Its founders’ research backgrounds helped establish technical legitimacy, while venture funding allowed the company to spend ahead of revenue. The lasting asset from this phase was accumulated deployment knowledge inside regulated and data-sensitive organisations.

The second stage, from 2019 through 2022, was scale at the cost of cash. Phancy expanded from a platform licence into application development, packaged solutions and substantial implementation work. Revenue accelerated. Yet external technical-service spending, R&D and long customer payment cycles kept operating cash flow negative. Operating cash outflow was approximately RMB453 million in 2020, RMB770 million in 2021 and RMB780 million in 2022. Trade-receivable days rose from 86 in 2020 to 134 in 2022 as enterprise and public-sector business expanded.

This period created the company’s present ambiguity. Building a reusable platform required engineers to solve many customer-specific problems. Those deployments enlarged the product’s industry knowledge and references, but also trained customers to buy complete outcomes rather than standalone software. Phancy gained scale; it did not establish subscription economics.

The third stage began with the Hong Kong listing in September 2023. The IPO followed several earlier listing attempts and was completed after the U.S. Department of Commerce placed 4Paradigm on the Entity List in March 2023. The final offer price was HK$55.60, at the bottom of the marketed range, and the company sold approximately 18.4 million H shares, implying gross proceeds of about HK$1.02 billion before the over-allotment option and expenses. Cornerstone investors absorbed a large share of the offering, which helped complete the transaction but left a relatively constrained public float.

The IPO story described Phancy as China’s platform-centric decision-AI leader: a potential enterprise-software compounder with a broad industry runway. The market was asked to look through losses and cash burn toward the possibility that customer implementations would become standardised, platform-driven and repeatable. The share price did not sustain the IPO valuation because the evidence arrived unevenly. Revenue grew, but the company remained loss-making, cash flow stayed negative and capital markets repeatedly supplied new equity.

The fourth stage, beginning in 2024 and accelerating through 2026, is an attempt to turn project accumulation into a full-stack platform. In 2024 the legacy Sage platform grew strongly, while the company reduced selling and administrative expenses. In 2025 it reorganised its external reporting around three engines: AI Platform, API and Agentic AI. ModelHub XC was built to adapt open models to domestic accelerators; HAMi pooled GPU resources; PhanthyCloud exposed models, computing and tokens through a common service layer.

This strategic turn has a coherent logic. Open models reduce the value of owning one proprietary foundation model, but they increase the operational complexity of choosing, adapting, serving, governing and replacing many models. China’s domestic-chip ecosystem adds another compatibility layer. Phancy is positioning itself as the software and operations layer above that fragmentation. Its value would come from making heterogeneous models and chips usable inside enterprises, rather than from training the single best frontier model.

The financial record shows both progress and unfinished work.

Revenue and margin figures come from company filings; percentages may differ slightly because of rounding. H1 2026 remains preliminary guidance, not reported results.

Revenue growth has been real and broad enough to survive a slower Chinese enterprise-IT environment. The company expanded annual revenue from RMB3.1 billion in 2022 to RMB7.1 billion in 2025, a three-year compound rate of roughly 32%. Customer count, average contract value, hardware and computing pass-through, and application scope all contributed. The filing does not provide enough data to separate price, volume and customer-count growth precisely.

Margin movement tells the more important story. Gross margin declined by 12.3 percentage points between 2023 and 2025. The company attributes the 2025 decline mainly to product mix and hardware procurement. Finished-goods cost almost doubled from RMB2.064 billion in 2024 to RMB4.047 billion in 2025, far faster than revenue. Cloud and other technical-service fees increased to RMB724 million. Those figures indicate that a meaningful portion of incremental revenue came from compute, hardware or externally delivered work with lower margins than pure software.

Operating leverage appeared below gross profit. Selling expense fell 23.6% to RMB205 million and administrative expense fell 20.2% to RMB154 million. Employee count declined from 967 to 619. R&D expense rose only 7.7% to RMB2.337 billion, reducing the R&D ratio from 41.2% to 32.8%. The company narrowed its operating loss despite gross-margin compression.

The headcount reduction reads two ways. It could reflect software standardisation and organisational discipline. It could also reflect greater reliance on external technology-service suppliers, whose costs remain large. The company does not disclose delivery headcount, R&D headcount or the division of outsourced technical fees between product development and customer implementation. That prevents a clean test of whether labour intensity has structurally declined.

Cash generation has not followed the earnings improvement. Cumulative operating cash outflow from 2020 through 2025 was approximately RMB4.3 billion. In 2025, operating cash outflow increased to RMB681 million despite the adjusted profit turn, while property-and-equipment purchases rose to RMB267 million. The resulting simplified owner-cash outflow was close to RMB948 million before acquisitions, investment securities and financing flows.

Receivables explain part, but not all, of the gap. Calculated average trade-receivable days improved from around 175 in 2024 to 139 in 2025. Payables also fell sharply, from RMB2.183 billion to RMB1.172 billion, absorbing cash as suppliers were paid. Inventories rose from RMB172 million to RMB329 million, and prepayments and other receivables rose from RMB536 million to RMB1.091 billion. The company’s growth still requires working capital even when reported profit approaches break-even.

Returns on capital remain negative on an operating-cash basis. Conventional ROE and ROIC are not useful while operating profit and free cash flow are negative and the balance sheet contains substantial equity raised in advance of future growth. The proper return test is whether API, agents and platform renewals can eventually produce positive cash after maintenance R&D and compute investment.

The balance sheet is strong in a solvency sense. Liquid cash resources of RMB3.774 billion exceeded borrowings of about RMB202 million at year-end 2025. Net current assets were RMB5.383 billion. The 2026 placement added more liquidity. Phancy has ample runway unless receivables deteriorate severely or planned infrastructure spending rises far beyond announced levels.

The capital-markets history is shorter than the business history but already has four distinct valuation regimes. At IPO, the company was valued as a scarce listed Chinese enterprise-AI platform. Subsequent weakness priced continued losses, geopolitical restrictions and doubts about project-heavy revenue. AI-agent and domestic-computing excitement later pushed the stock as high as HK$70 within the latest 52-week period. Profit-quality questions, repeated issuance and broader AI-stock volatility then brought it back toward the high twenties.

At HK$29.28, market capitalisation is approximately HK$16.31 billion, equivalent to RMB14.0 billion. That is about 2.0 times FY2025 revenue on a simple market-cap-to-sales basis. Deducting year-end liquid cash, adding debt and including April placement proceeds would produce a lower pro forma enterprise-value-to-sales ratio, although subsequent GPU purchases and other uses of cash make a precise post-placement enterprise value impossible before interim reporting.

The stock is optically inexpensive relative to many AI-labelled companies. The discount exists because Phancy’s revenue carries a 34.8% gross margin, negative free cash flow, substantial customer concentration and uncertain recurring economics. Its short listed history also makes a historical-percentile valuation claim unreliable. The relevant question is whether the company can graduate from roughly two times sales with weak cash conversion to a higher-quality software multiple, rather than whether the current multiple is low compared with theme-driven peaks.

Business model, moat, industry and peers

Phancy reports one IFRS operating segment, meaning management does not provide segment profit, segment assets or gross margin for the three commercial product families. It does, however, disclose revenue by business engine.

The figures are the company’s new reporting classification. They should not be spliced directly onto the old Sage, SHIFT and SageGPT series.

The AI Platform business is the economic core. It includes Sage software, model and chip adaptation, GPU virtualisation, cloud access, deployment and associated infrastructure. A customer may buy software licences, an integrated appliance, computing capacity, implementation services or some combination. That breadth helps Phancy win large transformation budgets, but it also obscures the value and margin of the reusable software component.

The API business sells token-based model and computing access. Its 129% growth rate and the sixfold increase in Q1 2026 call volume are strategically important, but FY2025 revenue was only RMB79.9 million. It would need to grow several-fold before it materially changes group gross margin or cash flow. The economic variables are revenue per token, inference cost per token, utilisation and customer retention, not token volume alone.

Agentic AI sells applications or business outcomes built on the common platform. Revenue almost doubled to RMB503 million in 2025, and orders on hand were up 99% at the end of Q1 2026. Agent work can deepen customer relationships and create high-value workflows. It can also recreate the economics of customised consulting if each deployment requires extensive integration, data preparation and human support.

The cost structure combines fixed and variable elements. Core platform R&D, model adaptation, product engineering and sales infrastructure are largely fixed over a reporting period. Hardware, rented computing, third-party cloud capacity, external technology services and implementation are more variable. The latter costs grew fast enough in 2025 to overwhelm operating leverage at the gross-profit line.

The evidence for scale economies is mixed. Selling and administrative ratios improved markedly, showing that corporate overhead need not grow with revenue. Gross margin fell, showing that the incremental contract mix was less profitable. Genuine platform leverage would eventually produce both outcomes at once: lower operating-expense ratios and stable or rising gross margin.

Phancy’s strongest moat is its enterprise deployment layer. Large regulated organisations have complex data permissions, legacy systems and security requirements. Once Sage is embedded in model development, governance and production workflows, replacement can be disruptive. The company’s experience across finance, energy, manufacturing, telecoms and retail gives it templates and operational knowledge that a new entrant must rebuild.

The second moat is heterogeneous infrastructure compatibility. ModelHub XC and HAMi address practical bottlenecks created by multiple open models and domestic accelerators. ModelHub had adapted more than 30,000 models by year-end 2025 and more than 70,000 by Q1 2026, while supporting chip platforms including Huawei Ascend, Cambricon, Kunlunxin and MetaX. HAMi allows fine-grained sharing of GPU capacity. These are useful engineering assets in a fragmented ecosystem.

The third moat is customer access. More than 1,000 contracted customers and a large order book provide reference cases and cross-selling opportunities. Yet concentration narrows the moat’s practical breadth: the largest customer supplied more than one-quarter of 2025 revenue. A broad logo count can coexist with dependence on a few large contracts.

The weaker “marketing moats” are model ownership and network effects. Phancy supports external model families such as Qwen, DeepSeek, Kimi, Hunyuan, GLM and MiniMax. This broad compatibility is useful, but it also shows that much of the intelligence layer is supplied by third parties. There is no disclosed user network effect comparable with a consumer platform, no public evidence of superior model benchmarks and no disclosed pricing premium.

Open-weight model commoditisation cuts in both directions. It reduces Phancy’s cost of accessing capable models and makes a model-neutral management layer more valuable. It also lets Alibaba, cloud providers, systems integrators and customers themselves assemble similar stacks. Alibaba’s Wukong enterprise-agent platform and continuing Qwen model releases show how quickly a hyperscaler can bundle models, cloud, agents and enterprise distribution.

Governance is founder-controlled. Dai Wenyuan is the ultimate controlling shareholder and remains chairman and chief executive. Founder control supports long product-development horizons, but minority shareholders bear the usual risk that expansion, strategic investments or fundraising take priority over near-term per-share returns.

Capital allocation has favoured growth and liquidity rather than buybacks or dividends. Phancy raised approximately HK$1.31 billion through a 2025 placing at HK$50.50 and another HK$1.56 billion through an April 2026 placing at HK$40.36. It then proposed a Shenzhen IPO and a RMB400 million GPU-server purchase. The sequence is rational if incremental capital produces high-return platform infrastructure; it is dilutive if the company continues to fund low-margin hardware and project revenue.

Share-based compensation is another dilution channel. Approximately 23.25 million options were outstanding at June 30, 2026, equal to about 4.2% of registered shares before any A-share issue.

The external industry backdrop remains favourable. IDC projects rapid expansion in Asia-Pacific AI-platform and AI spending as enterprises move from experimentation into production. IDC has estimated Asia-Pacific AI spending could reach US$175 billion in 2028, representing a 33.6% compound growth rate from 2023, while its AI-platform forecast implies even faster growth in the platform layer.

China’s specific demand drivers are domestic substitution, data security, public-sector procurement, model commoditisation and the need to use AI on private enterprise data. IDC expects many large Chinese enterprises to prioritise AI-sovereignty requirements, including non-public hosting, open technologies and regional partners. These preferences support Phancy’s private-deployment and domestic-chip positioning.

The industry’s profit pool is less settled. Foundation-model creators seek returns through APIs and cloud consumption. Chip and computing providers earn from scarce infrastructure. Enterprise-software vendors monetise workflows and data. Integrators earn implementation fees. Phancy sits across all four layers, but its 2025 gross margin suggests that infrastructure and implementation currently absorb much of the revenue pool.

The company is exposed to a technology-iteration cycle, a corporate-capex cycle and a policy cycle. New model generations can make existing applications more capable and stimulate demand, but they can also render proprietary features obsolete. Large customers’ budget and acceptance schedules determine revenue timing. Domestic-computing policy favours local compatibility, while export controls constrain access to U.S.-origin technology.

The U.S. Entity List is a structural constraint rather than a temporary headline. The 2023 rule places 4Paradigm under licensing requirements for items subject to the Export Administration Regulations, with a presumption of denial for most applications. The company has responded by supporting domestic accelerators, but the restriction can still affect access to high-end GPUs, software tools, suppliers and international customers.

The horizontal comparison reveals what Phancy still needs to prove.

Peer figures use each company’s latest publicly available period and are not perfectly period-matched.

Kingdee represents the closest proof standard for Chinese enterprise software. Its growth is slower, but subscription ARR, contract liabilities, a 67.1% gross margin and positive operating cash flow show what standardisation and recurring billing look like. Customers choose Kingdee for core ERP, finance and business-management workflows. Phancy is more AI-centric and may grow faster, but it has not shown equivalent revenue visibility or cash conversion.

SenseTime became a model, computer-vision and AI-infrastructure company. Its generative-AI revenue grew rapidly, but computing costs and continuing losses make it a capital-intensive upstream comparison. Phancy’s relative advantage is narrower enterprise workflow and decision deployment; SenseTime’s advantage is deeper model and vision research. Both remain exposed to falling inference prices and domestic compute constraints.

Palantir is the global conceptual benchmark, not a valuation peer. It proves that a deployment-intensive enterprise-AI company can eventually achieve software economics when its ontology, data platform and applications become reusable. Palantir reported 85% year-on-year revenue growth in Q1 2026 and guided to roughly US$1.8 billion of Q2 revenue with more than US$1.06 billion of adjusted operating income. That margin profile is far beyond Phancy’s. The comparison establishes the standard Phancy must reach; it does not justify applying Palantir’s multiple to Phancy today.

Alibaba is the largest ecological threat rather than a pure peer. It can combine Qwen models, cloud computing, enterprise agents, developer tools and existing customer distribution. A customer may still choose Phancy for model neutrality, private deployment and cross-chip orchestration, but Alibaba can subsidise one layer to win consumption in another.

Phancy’s niche is an independent enterprise-AI control and deployment layer for complex Chinese organisations. It takes budget from traditional systems integrators, standalone machine-learning tools and bespoke AI development. Hyperscaler bundles, ERP vendors adding native AI and internal customer engineering teams can take the same budget back.

Current fundamentals and capital markets

The latest completed annual period showed accelerating revenue and sharply narrower losses, but weaker gross margin and cash conversion. FY2025 revenue rose 35.6% to RMB7.135 billion. Adjusted net profit was RMB6.3 million, while IFRS net loss was RMB37.8 million. Operating cash outflow increased to RMB681 million.

The quarterly progression indicates that demand remained strong. Revenue reached RMB2.626 billion in H1 2025, RMB4.402 billion for the first nine months and RMB7.135 billion for the full year. The implied second half was much larger than the first, reflecting normal enterprise procurement seasonality and contract acceptance. Q1 2026 revenue rose 35.4% to RMB1.458 billion, with a 35.1% gross margin.

H1 2026 reported results were not yet available on the research base date. The company had only issued a preliminary profit alert and said interim results were expected by the end of August. The request to establish actual H1 figures therefore cannot be satisfied without using information published after 2026-08-03.

The guidance implies H1 2025 revenue of roughly RMB2.63 billion at either end of the growth range, exactly consistent with the filed figure. The wide H1 2026 range therefore arises from uncertainty over the current period, not an inconsistent comparison base. A RMB800 million revenue interval equals 25% of the low end. That is unusually broad for a half that had already closed.

Several explanations are possible. The accounts may have been at an early stage of closing. Contract acceptance or principal-versus-agent assessments may affect which projects are recognised and at what gross amount. Hardware and computing deliveries may also have been awaiting documentation. These are inferences rather than confirmed causes. The range itself indicates lower visibility than investors normally expect from a recurring software model.

Profit quality matters more than whether the projected H1 figure falls at RMB90 million or RMB130 million. Management cited three drivers: core-business growth, a lower selling-expense ratio and substantially higher other income, mainly from fair-value increases in investment projects. The third item is non-operating and potentially volatile.

FY2025 offers a useful bridge. Operating loss was RMB133.5 million. That figure is already struck after RMB96.2 million of other income, of which government grants and tax refunds supplied RMB95.9 million. Equity-method results of RMB17.4 million and net finance income of RMB76.3 million then brought the pre-tax loss to RMB39.8 million. The business had not reached recurring operating profitability even though adjusted net income was slightly positive.

The H1 alert may represent a genuine further step in operating leverage. Selling expense has already fallen substantially and revenue is growing faster than R&D. Yet the filing does not quantify fair-value gains, operating profit or cash flow. Until those figures appear, the evidence supports a turn in reported profit, not a confirmed turn in recurring owner earnings.

The operating businesses show different momentum. AI Platform remains the principal revenue engine. API has the highest usage growth, with Q1 token volume nearly six times the prior-year period and above the entire FY2025 volume by approximately 40%. Agentic AI’s order pipeline almost doubled from year-end. ModelHub adapted-model count rose beyond 70,000.

The US$200 million Huanxi Media cooperation should not be treated as booked revenue. The May 2026 announcement sets a three-year target of no less than US$200 million in token fees, a floor rather than a cap, of which a first binding two-year API contract of US$20 million has been signed and is prepaid. On the same day Phancy agreed to subscribe about HK$201 million for roughly 14.30% of Huanxi Media, which reported FY2025 revenue of HK$342.1 million and a net loss of HK$498.3 million, so part of that demand is vendor-financed. Revenue recognition will depend on actual consumption, contract terms and collectability.

The market is currently trading the combination of a first profitable half, token growth and a potential A-share re-rating. It is not yet trading proven free cash flow. The profit alert’s muted same-day close suggests investors distinguished between the headline profit and its non-operating components.

The Shenzhen plan has advanced beyond an informal aspiration but remains at an early regulatory stage. Pre-listing tutoring was accepted by the Beijing CSRC office on March 25, with Huatai United Securities appointed. The board later proposed a ChiNext offering of 62.0–186.1 million new A shares before any greenshoe. The company still requires Shenzhen Stock Exchange review and CSRC registration; no exchange acceptance or registration had been announced by the base date.

The planned RMB3.8 billion proceeds are substantial relative to the existing balance sheet. RMB2.55 billion is allocated to full-stack AI-platform R&D and industrialisation, RMB500 million to domestic AI IT-application innovation and RMB750 million to working capital.

Dilution depends on pricing and issue size. The basic A-share issue would equal roughly 11%–33% of the existing registered share count. With the maximum greenshoe, new shares could reach approximately 214 million, or 38% of the current count and 27.7% of post-offering shares. The cash raised accompanies the dilution, so per-share value depends on whether the A shares are issued above or below the existing business’s intrinsic value and whether the proceeds earn adequate returns.

The A-share listing could narrow the valuation discount if domestic investors assign a higher multiple to an enterprise-AI pure play. It may also create an H/A price relationship, more analyst coverage and a new acquisition currency. Conversely, a large issue, conversion of domestic shares into tradable securities and continuing option grants increase the supply of shares.

The bull and bear disagreement can be reduced to five measurable disputes.

First, bulls see the gross-margin decline as temporary mix pressure from rapid infrastructure deployment; bears see it as evidence that the “platform” is economically a hardware-and-services bundle. The next two reports must show whether gross margin stabilises as token and agent revenue grows.

Second, bulls see the H1 profit as operating leverage; bears see investment revaluation and government-linked income supporting weak operations. The decisive figure is recurring operating profit before fair-value gains, grants and share compensation.

Third, bulls view the order book as revenue visibility; bears note that 99.8% point-in-time recognition and customer concentration make backlog timing unpredictable. Cash collections and contract liabilities matter more than headline orders.

Fourth, bulls believe open models enlarge demand for Phancy’s orchestration layer; bears believe Qwen, DeepSeek and other models make the technology stack easier for cloud vendors and customers to assemble without an independent platform.

Fifth, bulls expect the Shenzhen listing to unlock value; bears see another round of capital raising before the existing business has generated free cash flow.

Valuation, risks and tracking

Phancy cannot be valued credibly on headline P/E. FY2025 IFRS earnings were negative, adjusted earnings were close to zero and H1 2026 profit includes an undisclosed fair-value component. Annualising the H1 guidance would produce RMB180–260 million of accounting profit, equivalent to a superficial P/E of roughly 54–78 times at the current RMB-equivalent market capitalisation. Recurring profit would be lower if fair-value gains are removed.

Cash-flow passthrough is poor. A five-year operating-cash-flow-to-net-income ratio is not economically meaningful because both measures have been negative for most of the period. More usefully, operating cash flow remained negative every year from 2020 through 2025 and cumulative outflow was about RMB4.3 billion. Adjusted FY2025 profit of RMB6 million corresponded with RMB681 million of operating cash outflow.

The maintenance-versus-growth capex split is not disclosed. FY2025 depreciation on property and equipment was only RMB12 million, while cash purchases were RMB267 million and property and equipment rose from RMB35 million to RMB291 million. That indicates most 2025 capex was probably growth-related. The additional proposed RMB400 million GPU-server purchase reinforces that inference.

A reasonable maintenance-capex assumption is RMB15–30 million, close to recent depreciation and basic replacement needs. Even on that lenient basis, FY2025 owner earnings were approximately negative RMB696–711 million when operating cash flow is reduced by maintenance capex. Deducting all capital expenditure produces negative RMB948 million. The gap versus adjusted profit is far greater than 30%, so owner-earnings and enterprise-value-to-sales measures should dominate valuation.

Historical valuation offers little anchor. The stock has traded for less than three years, was loss-making throughout most of that time and moved through a 52-week range of HK$23.70–70.00. Multiple changes have reflected AI sentiment, liquidity and listing expectations as much as fundamental earnings.

Peer valuation also needs restraint. Kingdee deserves a higher sales multiple because its gross margin, ARR, contract liabilities and cash flow show subscription economics. Palantir deserves a much higher multiple because it has exceptional growth and operating margins. SenseTime’s capital intensity and losses show that an AI label does not guarantee premium economics. Phancy should not receive a SaaS multiple until its disclosure and cash conversion resemble SaaS.

The absolute valuation below uses 2027 revenue, normalised owner-free-cash-flow margins and EV/Sales. It explicitly incorporates additional shares from options and a possible A-share issue. Net cash includes assumed fundraising proceeds less infrastructure deployment.

The conservative case assumes that platform growth slows, hardware remains a large share of revenue, gross margin stays close to current levels and cash flow only reaches break-even. Its permanent-loss trigger is continued negative operating cash flow combined with customer or receivable deterioration.

The base case assumes revenue remains above the industry IT-spending rate, gross margin recovers modestly and API plus Agentic AI become material without eliminating implementation work. A 4%–6% owner-FCF margin would still be well below mature software peers. The main catalyst is evidence that recurring operating profit and collections improve before the A-share offering.

The optimistic case requires substantial standardisation. API and agents must increase their combined revenue contribution, gross margin must recover toward 40%, customer concentration must fall and owner-FCF margin must reach high single digits. It also assumes capital raised in Shenzhen earns adequate returns rather than funding pass-through compute revenue.

This is valuation-scenario analysis within a research framework, not investment advice.

At the current price, the market appears to discount approximately the conservative value while granting partial credit for a profitable transition. It does not appear to price the full base case. Four lines in the interim report will decide the expectation gap: gross margin, operating profit excluding investment revaluation, operating cash flow and the ageing of receivables.

The most fragile base-case assumption is a 4%–6% owner-FCF margin by 2027. Cutting the midpoint from 5% to 3.5%, approximately 70% of the original assumption, would justify reducing the EV/Sales multiple from 1.8 times to around 1.5 times. On the same revenue, net-cash and dilution assumptions, base value falls to roughly HK$34 per share.

There is no discount to conservative value at HK$29.28. The current price sits within the conservative fair-value range rather than 20% below it. The margin of safety is therefore zero.

A flat-earnings exercise leads to the same conclusion. Annualised H1 accounting profit of RMB180–260 million would yield only 1.3%–1.9% on the current RMB14.0 billion market capitalisation, before removing fair-value gains. That is below the 3.51% Hong Kong 10-year government-bond yield as of July 31 and around the 1.71% China 10-year yield, despite carrying much greater business risk. There is no margin of safety at this buy price.

Margin-of-safety sufficiency verdict: none.

The principal permanent-loss risks are specific.

The highest-probability risk is that revenue remains project and hardware heavy. Its probability is medium to high and its impact is high. If AI Platform revenue continues to grow at 25%–35% while group gross margin falls below 32%, the market will stop treating that growth as software growth. Lower gross profit would constrain R&D, recurring operating profit and the valuation multiple simultaneously. The observable indicators are cost of finished goods, point-in-time revenue, segment mix and group gross margin.

Profit-quality risk has a high probability over the next report and medium-to-high impact. If most H1 profit comes from fair-value gains, grants or finance income, the expected earnings inflection will be delayed. The transmission path is immediate: consensus earnings fall, P/E becomes unusable again and the A-share re-rating story weakens.

Receivable and customer-concentration risk has a medium probability and high impact. Gross receivables older than one year already doubled to RMB914 million, while the largest customer contributes 26.9% of revenue. Delayed payment or reduced procurement from that customer could create impairment, cash outflow and a revenue miss in the same period.

Dilution risk has a high probability and medium impact. Options, H-share placements and the proposed A-share issue could increase the fully diluted share count materially. The direct effect is lower ownership per share. The indirect effect is more damaging if investors conclude that the business requires repeated equity financing to fund growth.

Technology and price-competition risk has a medium probability and high long-term impact. Alibaba and other cloud platforms can bundle open models, agents and computing. If API prices fall faster than inference costs, Phancy’s token volume can grow while API gross profit remains small. The observable indicators are API revenue versus token volume, revenue per token and cloud-service costs.

Export-control risk has a medium probability and high tail impact. Further restrictions on advanced accelerators or software could increase compute cost and slow model adaptation. Domestic-chip compatibility mitigates the risk but may not reproduce the performance, software ecosystem or availability of unrestricted global hardware.

The tracking dashboard focuses on data that can overturn the thesis.

Underlying figures come from company filings and calculations described in this report.

The interim report should explain why the guidance range was so wide and separate operating earnings from fair-value gains. Over subsequent periods, API revenue must be disclosed alongside token volume, because volume without unit economics can mislead. Receivable ageing, customer concentration and operating cash flow should be tracked from annual and interim filings. A-share progress should be followed through CSRC and Shenzhen exchange disclosures rather than press speculation.

Positive catalysts include H1 revenue near the upper end of guidance, gross margin above 35%, operating profit before investment gains, positive operating cash flow, lower customer concentration, meaningful API monetisation and an A-share issue at a valuation accretive to existing shareholders.

Negative catalysts include H1 revenue near or below the low end, fair-value income exceeding operating profit, gross margin below 32%, another increase in overdue receivables, a large-customer decline, an aggressively dilutive A-share structure or evidence that token prices are falling faster than compute costs.

Cross-synthesis, uncertainties and conclusion

Looking vertically, Phancy has proven one durable capability: it can bring machine learning and generative AI into large Chinese enterprises with difficult data, security and infrastructure constraints. The company survived the transition from predictive decision models to foundation models because it did not tie its entire product to one algorithmic generation. Sage became a broader deployment layer, and the company added model adaptation, GPU orchestration, cloud access and agents as customer requirements changed.

That adaptability came from management and accumulated engineering more than from luck. China’s enterprise-digitisation cycle, venture funding and policy support supplied favourable conditions, but many AI start-ups exposed to the same tailwinds failed to reach RMB7 billion of revenue or more than 1,000 customers. Phancy’s customer access, technical workforce and willingness to solve difficult deployment problems created a real business.

The same history produced its structural weakness. Each time customers asked for more complete outcomes, Phancy expanded the bundle. Software became mixed with implementation, third-party technology services, hardware and computing. Revenue grew faster, but the economics moved away from subscription software. Gross margin fell, revenue remained almost entirely point-in-time and cash flow stayed negative.

The company’s fate now turns on whether accumulated custom work can be productised. ModelHub and HAMi are plausible examples of reusable intellectual property created from deployment pain. A model-adaptation service becomes valuable when every customer otherwise repeats the same chip and model work. A GPU-sharing layer becomes valuable when scarce computing is underutilised. Agents can package industry knowledge into repeatable workflows. These products can create software-like leverage if the next customer requires much less incremental labour and hardware than the previous one.

The financial record has not yet confirmed that transition. Employee count and expense ratios fell, which is constructive. Gross margin and operating cash flow moved in the opposite direction. The strongest interpretation is that internal operating efficiency improved while revenue mix became more hardware- and compute-heavy. A full platform inflection requires both dimensions to improve.

Horizontally, Phancy’s advantage is independence and implementation depth. Alibaba can provide an integrated cloud and model stack, but some customers want model neutrality, private deployment and cross-vendor infrastructure. Kingdee owns deeper ERP workflows, though its AI layer is newer and tied to its enterprise-management products. SenseTime has broader model and vision research, but Phancy is more tightly oriented around enterprise decisions and operations. Palantir has superior ontology and software economics, but carries a very different geopolitical, customer and capital-market context.

Phancy’s weakness is structural until evidence changes. Its gross margin is roughly half Kingdee’s. It does not disclose ARR or retention. Its largest customer supplies more than one-quarter of revenue. Overdue receivables have aged. Operating cash remains negative. Those facts cannot be dismissed as temporary simply because revenue and token calls are rising.

The current valuation rewards some future success but not an extreme outcome. At roughly two times FY2025 sales and about 47% below the IPO price, the shares do not embody a Palantir-like business. A conservative operating outcome is already reflected around the current price. The market is also giving value to the possibility that API, agents and the A-share listing improve the quality and reference multiple.

The market’s likely misjudgement lies in separating revenue growth from revenue economics. Some investors may underestimate the value of Phancy’s domestic compatibility and customer installation base. Others may mistake every yuan labelled AI Platform for high-margin platform software. The filings support neither extreme. The company has meaningful technology and customer assets embedded in a delivery-heavy financial model.

Over the next year, the critical variables are H1 profit composition, gross margin, operating cash flow and A-share timing. The investment case improves quickly if the company reports operating profit before fair-value gains, maintains at least a 35% gross margin and reduces operating cash outflow. It deteriorates if reported profit is primarily investment income while receivables or hardware costs rise.

Over three years, the critical variable is productisation. API and Agentic AI need to become a material share of revenue, with disclosed unit economics or at least visible group-margin improvement. Customer concentration should decline, and recurring or over-time revenue should rise from its negligible base. A gross margin around 36% with positive mid-single-digit owner cash flow would validate the base valuation case.

Over five years, Phancy must establish itself as the control plane for enterprise AI rather than a reseller of computing and customised engineering. The durable value would lie in model governance, infrastructure orchestration, enterprise ontology, workflow integration and customer switching costs. Open models would then enlarge its addressable market. Without that control-plane position, open models and cloud bundling will compress pricing and leave Phancy competing for project margins.

Core bull reasons

• Revenue increased from RMB3.1 billion in 2022 to RMB7.1 billion in 2025, while Q1 2026 maintained 35.4% growth.
• Selling and administrative expenses fell in absolute terms during 2025, showing genuine overhead leverage.
• API revenue grew 129.2%, Agentic AI revenue grew 93.2%, and Q1 token calls were almost six times the prior-year level.
• Model and domestic-chip compatibility addresses a real procurement and deployment problem for regulated Chinese enterprises.
• Liquid resources, the April placement and the proposed A-share proceeds provide substantial funding for productisation.

Core bear reasons

• Gross margin fell from 47.1% in 2023 to 34.8% in 2025 as finished-goods and computing-related costs increased.
• Almost 99.8% of FY2025 revenue was recognised at a point in time, while no ARR, retention or renewal data are disclosed.
• FY2025 adjusted profit of RMB6 million coincided with RMB681 million of operating cash outflow and RMB267 million of property-and-equipment purchases.
• The largest customer contributed 26.9% of revenue, and gross receivables older than one year rose to RMB914 million.
• H1 2026 profit guidance explicitly depends partly on investment fair-value gains, while further equity issuance could materially dilute existing holders.

Pre-mortem

One plausible three-year loss script begins with product mix. During 2027, Alibaba and other cloud vendors bundle enterprise agents and open models at prices 30%–40% below independent providers. Phancy responds by increasing bundled computing and hardware to protect customer volume. Revenue still grows 10%–15%, but gross margin falls from 35% to 28% and API revenue per token declines. Operating cash flow remains negative, the market cuts EV/Sales from approximately 2 times to 0.8–1.0 times, and additional A-share dilution lifts the share count above 750 million. The share price could fall into the HK$12–16 range.

A second script begins with concentration and accounting quality. A major customer delays acceptance and payment in 2027, causing revenue to miss expectations by RMB1 billion and requiring a large receivable impairment. Investors then discover that the 2026 profit turn was mostly fair-value income rather than recurring operations. Gross margin remains around 33%, operating cash outflow exceeds RMB800 million and the Shenzhen listing is delayed. A valuation near net cash plus a low multiple on project revenue could reduce the shares by more than 50%.

Research uncertainties

The first blind spot is the absence of reported H1 2026 results. The exact revenue, gross margin, operating profit, fair-value gains and cash flow were unavailable at the base date.

The second is segment economics. Phancy reports only one IFRS segment and does not disclose gross margin, profit, cash flow or capital employed separately for AI Platform, API and Agentic AI.

The third is revenue recurrence. Contract duration, renewal rate, net retention, revenue per token and backlog cancellation provisions are not disclosed.

The fourth is customer detail. Concentration is disclosed, but the identities, industries, contract terms and payment status of the largest customers are not.

The fifth is A-share pricing. The issue size and intended proceeds are known, but timing, issue price, final dilution and the value created from the proceeds cannot yet be estimated reliably.

The company has credible enterprise-AI technology, strong revenue growth and a balance sheet capable of funding the transition. Its investment quality is constrained by weaker gross margin, negative owner earnings, concentrated customers and insufficient disclosure of recurrence. At HK$29.28, the shares offer exposure to a potentially important Chinese enterprise-AI platform, but no clear margin of safety against the possibility that the business remains a capital-intensive project shop.

I would require two pieces of evidence before treating Phancy as a compounding software asset: recurring operating profit excluding investment revaluation, and positive operating cash flow while gross margin remains at least in the mid-thirties. The present price discounts some risk but does not compensate adequately for both unproven cash economics and prospective dilution.

【Company-profile scores】

• Fundamental quality: medium
• Growth: high
• Moat: medium
• Financial soundness: strong
• Management credibility: medium
• Valuation attractiveness: medium
• Risk level: high
• Suitable investor type: event-driven or high-risk growth investors able to monitor filings and dilution closely

【Investment rating】

• Rating: Watch
• One-line thesis: Rapid revenue growth and real deployment assets are offset by negative owner earnings, falling gross margin and non-operating profit support.
• Current-price classification: outside the three bands
• Whether to wait for a better price: yes; wait for HK$21–24 and evidence of operating profit excluding fair-value gains, gross margin of at least 35% and improving operating cash flow
• Opportunity cost of waiting: the shares could re-rate before cash-flow confirmation if H1 results reach the top of guidance or the A-share process advances
• Target holding horizon: three to five years after operating-economics confirmation
• Expected annualised return: conservative approximately −2% to 1%; base approximately 6% to 13%; optimistic approximately 18% to 24%
• Max-loss risk: approximately 50%–60%, triggered by sub-30% gross margin, sustained negative cash flow, a major-customer or receivable problem and heavy dilution
• Reassessment-trigger signals: gross margin below 32%; two consecutive reporting periods of negative operating cash flow after reported profit; largest-customer share above 30%; overdue receivables above 35% of gross receivables; or fully diluted shares above 750 million without a corresponding increase in owner earnings

【Ideal Buy Price】21–24 HKD

The range is about 20% below the midpoint of the conservative value of HK$27–30 and requires operating-quality conditions, rather than price alone, to be satisfied.

• Acceptable hold price: HK$31–44
• Clearly overvalued price: HK$62 or above

【Valuation Range】

• current: 29.28 HKD (close as of 2026-07-31)
• bear (conservative · ideal buy zone): [21, 24]
• base (fair · acceptable hold zone): [31, 44]
• bull (optimistic · above the clearly-overvalued line): [62, 68]

Other tickers mentioned

• 0268.HK: Chinese enterprise-software benchmark with subscription ARR, higher gross margin and positive cash conversion
• 0020.HK: China AI-model and infrastructure peer with greater foundation-model exposure and computing intensity
• PLTR.US: global decision-AI platform benchmark showing the margins possible after deployment work becomes reusable software
• 9988.HK: hyperscaler and Qwen ecosystem owner whose cloud-and-agent bundle is a major competitive threat
• 002230.SHE: Chinese speech and enterprise-AI vendor relevant to government and industry-solution competition
• 688111.SHG: Chinese office-software vendor illustrating how AI can be embedded into established enterprise workflows
• 0700.HK: Chinese cloud and model ecosystem participant with distribution and infrastructure advantages
• BIDU.US: Chinese foundation-model and cloud provider competing for enterprise AI workloads

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