Report Interpretation

Overview

This initiation report argues that Phancy is becoming a full-stack AI infrastructure provider rather than remaining solely an enterprise AI software vendor. Macquarie expects growth in compute management, model adaptation, API/token services and Agentic AI to expand earnings and support a valuation re-rating, while flagging execution, capital intensity and cash-conversion risks.

Core views

Macquarie initiates coverage of Phancy Group at Outperform, contending that the market underestimates its transition from legacy enterprise AI software into a broader AI infrastructure platform. The company’s stack spans HAMi vGPU compute management, ModelHub XC model adaptation, AI-platform deployment, API/token inference services and Agentic AI applications. The central thesis is that Chinese enterprises increasingly need a neutral layer to manage a mix of NVIDIA and domestic accelerators amid supply constraints and fragmented hardware ecosystems. Phancy’s vendor-neutral orchestration is therefore positioned at several points in the AI value chain rather than relying only on software licensing or hardware resale. The infrastructure thesis rests on GPU efficiency and adaptation. HAMi pools heterogeneous GPUs into a virtualised layer, supports compute partitioning down to 1% and MB-level memory allocation, and, according to the company, can improve GPU utilisation by 5-10x and reduce hardware costs by up to 80%. Reported examples include Ke Holdings increasing utilisation from 13% to 37%, NIO raising CI-pipeline utilisation from about 5% to 30-50% while reducing simulation GPU-hours by 30%, and China Merchants Bank doubling hardware utilisation. ModelHub XC had more than 200,000 adapted and certified models by mid-2026, targets more than one million by 2027, and is intended to reduce the “CUDA wall” between mainstream models and domestic chips. Macquarie views these capabilities as increasingly valuable to enterprises adopting Huawei Ascend, Cambricon, Kunlunxin and other alternatives. The report identifies API/token services as the most important long-term growth engine. Phancy owns or leases compute infrastructure and monetises API calls and token consumption; 2025 API revenue was Rmb79.9m, or 1.1% of group revenue, but grew 129.2% YoY, while management stated that 1Q26 token revenue exceeded the full-year 2025 total. Macquarie forecasts API revenue of about Rmb1.3bn in 2027E and Rmb3.0bn in 2028E, or roughly 16% of group revenue. Demand indicators include a US$200m three-year prepaid Huanxi Media agreement and a Rmb1bn high-performance-computing services contract with China Mobile Ningxia. The company uses cost-plus API pricing targeting a 20-25% gross margin based on projected GPU utilisation. Macquarie nevertheless stresses that Token Factory must prove it is incremental consumption revenue rather than relabelled compute or cloud revenue. Agentic AI is smaller but strategically differentiated because its Result-as-a-Service model links payment to client value creation such as cost savings, efficiency gains or incremental revenue. The business is concentrated in new-energy power, with film and entertainment added in 1H26. Macquarie argues that vertical deployment data may form a flywheel: stronger models improve outcomes, reinforce client lock-in and support replication. Agentic AI generated Rmb503m in 2025, up 93.2% YoY, and Rmb214m in 1H26, up 4.9% YoY; management guided to 50% growth for FY26, while Macquarie models 20% growth through 2028E. Key tests are whether revenue-sharing contracts can scale beyond energy and whether accumulated vertical data produces defensible advantages. Financially, Phancy’s revenue rose from Rmb4.20bn in 2023 to Rmb5.26bn in 2024 and Rmb7.14bn in 2025. Macquarie forecasts revenue of Rmb10.95bn in 2026E, Rmb15.20bn in 2027E and Rmb19.79bn in 2028E, a 34% CAGR over 2026-28E. The AI Platform remained 91.8% of 2025 revenue at Rmb6.55bn, while orders on hand were Rmb8.9bn in March 2026 and Rmb7bn in August, normally converting to revenue within 12 months. The shift is still early: AI Platform accounted for more than 90% of revenue in 2025 and API plus Agentic AI only 8.2%. The report expects operating leverage but not a pure-software margin profile. Cost of sales rose 54.2% in 2025, faster than revenue growth of 35.6%, as computing power and hardware procurement expanded. Gross margin fell from 42.7% in 2024 to 34.8% in 2025 and is forecast to moderate toward about 29% by 2028E as compute-related revenue grows. Nevertheless, operating losses narrowed from Rmb545m in 2023 to Rmb133m in 2025; adjusted net profit reached Rmb17.8m versus an adjusted Rmb265.0m loss in 2024. Macquarie forecasts adjusted profit of Rmb132.5m in 2026E, Rmb385.7m in 2027E and Rmb648.3m in 2028E, assuming revenue rises faster than R&D and employee costs. R&D intensity fell to 32.8% from 41.2%, headcount fell to 619 at end-2025 from 967 at end-2024, and revenue per period-end employee rose to about Rmb11.5m from Rmb5.4m. The report cautions that some labour has shifted into third-party service and computing costs. Cash flow and capital intensity remain material constraints. Receivables declined to Rmb2.34bn at end-2025 from Rmb3.09bn despite revenue growth, improving estimated receivable days to about 139 from 225, and credit-loss provisions fell to Rmb17m from Rmb200m. Yet operating cash flow remained negative because lower receivables were offset by a Rmb1.01bn decline in payables and higher inventory and prepayments. Macquarie estimates free cash flow of negative Rmb963m in 2025 and forecasts negative Rmb759.5m, negative Rmb457.3m and negative Rmb55.4m for 2026E-28E. It estimates annual GPU depreciation of about Rmb660m across FY26-FY28E, although prepaid or deposit-funded cloud contracts and the prior net-cash position partly support the build-out. Macquarie values Phancy through a sum-of-the-parts approach because its software-oriented AI Platform/Agentic AI operations and capital-intensive API business have different economics. It applies 3x 2027E EV/revenue to AI Platform plus Agentic AI after removing hardware pass-through revenue, and 3x 2027E EV/revenue to API, slightly below the stated global neocloud peer average of 3.3x due to chip-supply uncertainty, competitive inference conditions and the early-stage Token Factory. The resulting HK$59 12-month target implies about 21x 2029E earnings. Potential catalysts are A-share listing progress, potentially in 1H27 although management gave no timeline, public token-factory listings that provide valuation references, stronger API revenue guidance and new tender wins.

Analysis framework

Macquarie first frames Phancy’s strategic repositioning within China’s heterogeneous AI-compute market, then evaluates product capabilities and customer examples. It assesses monetisation through the API/token and Agentic AI businesses, models segment revenue and operating leverage, reviews working capital and compute-capex needs, and values the distinct business models using a sum-of-the-parts EV/revenue framework.

Methodology notes

• Competition & strategyValue chain analysis

  Full-stack AI value-chain positioning

  The report assesses Phancy’s exposure from GPU orchestration and model adaptation through deployment, routing and inference consumption, rather than treating it as a single software business.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  Domestic GPU fragmentation and enterprise AI deployment

  Macquarie links constrained access to NVIDIA hardware and diverse domestic accelerators to demand for neutral orchestration and model-adaptation tools.

• Valuation methodsSOTP (Sum-of-the-Parts) Valuation

  Sum-of-the-parts EV/revenue valuation

  The report assigns separate 3x 2027E EV/revenue multiples to the software-oriented AI Platform/Agentic AI businesses and the API business because their growth, margins and capital intensity differ.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Phancy Group (06682.HK)

  Primary covered company and beneficiary of China’s AI infrastructure build-out.

  Strengths

  Vendor-neutral HAMi vGPU orchestration, ModelHub XC adaptation ecosystem, enterprise and SOE customer base, growing API/token business and emerging operating leverage.

  Weaknesses

  API business is early-stage, AI Platform still contributes more than 90% of revenue, and gross margins are diluted by compute and hardware content.

  Comparison

  Macquarie applies a 3x 2027E EV/revenue multiple to the API business versus a stated global neocloud peer average of 3.3x, reflecting a discount for uncertainty.

  Risks

  GPU supply and cost volatility, token pricing pressure, capital intensity, dilution, cash conversion and competition.

Key data

• 12-month target priceHK$59.00Outperform initiation; 118.4% 12-month TSR versus HK$27.02 at 3 Sep 2026.
• Revenue forecastRmb10.95bn in 2026E; Rmb15.20bn in 2027E; Rmb19.79bn in 2028EMacquarie forecasts a 34% revenue CAGR over 2026-28E.
• API revenueRmb79.9m in 2025; about Rmb1.3bn in 2027E; Rmb3.0bn in 2028E2025 revenue grew 129.2% YoY and represented 1.1% of group revenue.
• 2025 gross margin34.8%Down from 42.7% in 2024 as hardware and computing-power content increased; forecast toward about 29% by 2028E.
• Adjusted net profitRmb17.8m in 2025; Rmb132.5m in 2026E; Rmb385.7m in 2027E; Rmb648.3m in 2028E2025 was the first full year of adjusted profitability.
• Orders on handRmb8.9bn in March 2026; Rmb7bn in August 2026The report says these orders normally convert into revenue within 12 months.
• Free cash flowNegative Rmb963.2m in 2025; negative Rmb759.5m in 2026E; negative Rmb457.3m in 2027E; negative Rmb55.4m in 2028ECompute build-out and working-capital needs keep cash generation negative in the forecast period.

Impact & implications

Macquarie believes Phancy’s infrastructure and API businesses can alter its growth mix and market valuation if GPU utilisation, token demand and deployment execution scale as forecast. The report also makes clear that the transition requires substantial compute investment and leaves earnings, cash flow and multiples sensitive to API monetisation, chip access and competitive pricing.

Risks

• A-share listing execution, regulatory approval or timing could disappoint and remove a potential re-rating catalyst.
• Export controls, GPU shortages or higher procurement costs could constrain deployments, delay compute-asset payback and pressure Token Factory profitability.
• Greater inference capacity from cloud providers, foundation-model vendors and MaaS operators could lower token pricing and restrict margin expansion.
• Further equity, convertible or other fundraising for compute expansion could dilute shareholders if API monetisation scales slowly.
• Historical placements and any future concerns about fundraising, counterparties, disclosures or governance could weigh on valuation.
• Long payment cycles, collections deterioration or higher receivables could constrain cash generation.
• Hyperscalers, cloud providers or open-source alternatives could weaken Phancy’s growth outlook and pricing power.

What to watch

• Progress and timing of a potential A-share listing, which Macquarie considers possible in 1H27 but without management guidance.
• API-business revenue guidance, token demand, GPU utilisation and the ability to establish incremental Token Factory revenue.
• New tender wins and the disclosed pipeline conversion from orders on hand into revenue.
• Public listings of token factories such as SiliconFlow and Infinigence AI as potential valuation reference points.
• Whether Agentic AI revenue-sharing contracts become repeatable beyond energy and whether vertical data creates defensible advantages.
• Cash conversion, contract liabilities or customer prepayments, compute depreciation and the sustainability of operating leverage.