Phancy Group Co., Ltd. (06682): JPMorgan initiates Phancy at Overweight on heterogeneous-compute optimization and scalable API Token economics

Overview

This initiation report argues that Phancy is evolving from enterprise AI software and project deployment into an AI-infrastructure company that can monetize heterogeneous-compute optimization and Token production. JPMorgan’s constructive view rests on expanding AI-intelligence consumption, China-specific deployment complexity, fast API growth and a prospective shift toward more scalable economics.

Core views

JPMorgan initiates coverage of Phancy at Overweight with a Dec-2027 price target of HK$50, based on 15x 2028E EV/EBITDA. The report frames Phancy as an AI infrastructure and enterprise-deployment business rather than a traditional SaaS company: it combines enterprise deployments through AI Platform with heterogeneous-compute management, model adaptation and serving capabilities through HAMi/RISE and ModelHub. The target multiple gives Phancy a premium to mature infrastructure peers for faster growth and differentiated exposure to heterogeneous compute, but remains below a software-like 18-20x multiple because deployment and services remain capital- and service-intensive. The underlying demand argument is that intelligence consumption per task is rising. Reasoning models, agents, long-context workloads and multimodality consume more inference compute, while lower inference costs make more workloads economically viable. JPMorgan argues that efficiency improvements need not reduce total infrastructure demand because lower unit cost can stimulate incremental usage. China is especially relevant because enterprises may need to operate across Nvidia GPUs, Huawei Ascend and other domestic accelerators, as well as multiple foundation models including DeepSeek, Qwen, GLM, Kimi and MiniMax. This hardware and model fragmentation raises the operational value of resource scheduling, cross-chip optimization and model-aware serving. Phancy’s AI Platform is the existing demand base. It generated about Rmb3.1bn in 1H26, up 30% YoY and representing roughly 82% of group revenue. The business is largely project based, covering infrastructure deployment, maintenance and technical services, so its accounting profile is less recurring than conventional SaaS. JPMorgan nevertheless expects it to grow at a 26% CAGR in 2026-28E, supported by enterprise deployment demand and the company’s experience in heterogeneous environments. API is presented as the pivotal change in the business model because it connects Phancy’s infrastructure capabilities directly to usage-based Token economics. The company can deploy third-party models on its compute base, optimize the model-hardware pairing and sell inference capacity to enterprises without funding frontier-model training. API revenue reached about Rmb464mn in 1H26, up 861% YoY; JPMorgan forecasts Rmb1,507mn, Rmb3,013mn and Rmb5,424mn in 2026E, 2027E and 2028E, respectively. The report expects API to remain the fastest-growing segment and sees gross margin approaching about 30% over the next two years as utilization, procurement scale and serving efficiency improve. It distinguishes participation from value capture: Token growth alone shows demand exposure, while rising gross margin and gross profit per compute unit would show that Phancy retains a share of the efficiency value it creates. The report believes scarcity is migrating from access to raw compute capacity toward heterogeneous optimization and ultimately intelligence efficiency. A neutral platform can potentially manage a customer’s full compute pool across architectures and models, unlike vertically integrated chip vendors, model providers or cloud platforms that have advantages within their own ecosystems. That neutrality may be valuable for enterprises using mixed infrastructure, several models, private deployments or production-grade governance. However, JPMorgan stresses that this advantage is not automatic: Phancy must demonstrate measurable improvements in utilization, throughput, latency, memory efficiency, deployment speed or inference cost. Reuse of HAMi/RISE, ModelHub and related capabilities across customers would support more software-like economics; ongoing dependence on bespoke engineering would preserve project-like economics. JPMorgan forecasts 2026-28E revenue CAGR of 36%, revenue of Rmb10,628mn in 2026E, Rmb14,583mn in 2027E and Rmb19,798mn in 2028E, and adjusted net income of Rmb141mn, Rmb378mn and Rmb965mn, respectively. Its 2028E adjusted net-profit forecast is 37% above consensus, reflecting stronger API scaling, better unit economics and operating expenses growing more slowly than revenue. The report expects adjusted EBITDA to rise from Rmb139mn in 2026E to Rmb542mn in 2027E and Rmb1,311mn in 2028E. It also expects group profitability to recover as API economics improve, despite an initial mix effect: group gross margin fell from 37.7% in 1H25 to 32.8% in 1H26 while gross profit reached about Rmb1.2bn. Accounting earnings are not considered sufficient evidence because the business requires compute capacity, equipment and working capital. JPMorgan expects CFO to turn positive at Rmb523mn in 2026E, followed by Rmb376mn in 2027E and Rmb421mn in 2028E, but forecasts FCFF of Rmb125mn, negative Rmb170mn and negative Rmb246mn over those years as investment continues. The key financial test is therefore incremental capital productivity: the preferred outcome is for growing Token demand to be served through better utilization and throughput on the existing compute base, allowing gross profit and cash flow to outpace invested capital. Growth requiring proportional additions of owned or leased compute and working capital would restrain returns and multiple expansion. Agentic AI is a potential longer-term extension rather than a major valuation driver today. It generated about Rmb214mn in 1H26, up 5% YoY. JPMorgan wants to see revenue reacceleration, repeat adoption, cross-customer product reuse, retention and better unit economics before assigning it a more meaningful valuation contribution. The report identifies stronger API growth with higher utilization and margins, wider use and reuse of HAMi/RISE and ModelHub, and improvement in CFO, FCF and incremental ROIC as the clearest upside confirmations. A successful A-share listing could broaden the domestic funding base, although its valuation effect would depend on dilution and returns earned on new capital.

Analysis framework

JPMorgan begins with AI-infrastructure demand and China’s fragmented hardware and model environment, then assesses how Phancy’s AI Platform and API businesses monetize that complexity. It tests the thesis through segment growth, utilization, API margin, operating leverage, cash conversion and returns on incremental capital, and values the company on 2028E EV/EBITDA to reflect its infrastructure-heavy operating model.

Methodology notes

• Valuation methodsEV/EBITDA valuation

  15x 2028E EV/EBITDA target-multiple valuation

  JPMorgan uses EBITDA rather than P/E because depreciation, deployment costs and investment can materially affect reported net profit in Phancy’s capital- and service-intensive business. The 15x multiple balances faster growth and differentiated optimization exposure against a still-large deployment and services component.

• Industry AnalysisSupply-demand framework

  AI intelligence-consumption demand and infrastructure-efficiency supply analysis

  The report links rising reasoning, agentic, long-context and multimodal workloads to demand for inference capacity, then evaluates whether heterogeneous optimization can improve utilization and reduce serving cost as capacity becomes more available.

• Corporate Fundamentals and FinanceFree cash flow analysis

  CFO, FCF and incremental ROIC assessment

  JPMorgan treats cash conversion and returns on added compute and working capital as a separate test from accounting earnings, because strong revenue can coexist with weak free cash flow in an infrastructure-heavy model.

Asset mapping & comparison

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

• Phancy Group Co., Ltd. (06682.HK)

  Primary covered company; positioned to benefit from heterogeneous-compute optimization and enterprise AI-infrastructure demand.

  Strengths

  AI Platform provides an enterprise deployment base; API offers direct exposure to Token consumption; neutral cross-platform optimization can be valuable in mixed hardware and model environments.

  Weaknesses

  Revenue remains partly project based and capital intensive; current economics still depend materially on deployment, services, compute and working capital.

  Comparison

  JPMorgan views its growth and heterogeneous-compute exposure as deserving a premium to mature infrastructure peers, but not a software-like 18-20x EV/EBITDA multiple.

  Risks

  Slower AI demand, pressure on inference economics, commoditization of optimization capabilities, weak utilization or continued capital intensity could limit margins and returns.

Key data

• Rating and target priceOverweight; HK$50.00 Dec-27 targetTarget is based on 15x 2028E EV/EBITDA; price was HK$27.60 on 23 Sep 2026.
• 2026-28E revenue CAGR36%JPMorgan forecast.
• 2028E adjusted net profitRmb965mnApproximately Rmb0.9bn; JPMorgan’s forecast is 37% above consensus.
• AI Platform revenuec.Rmb3.1bn in 1H26Up 30% YoY and about 82% of group revenue.
• API revenueRmb464mn in 1H26Up 861% YoY; forecast at Rmb1,507mn/Rmb3,013mn/Rmb5,424mn in 2026E/2027E/2028E.
• API gross marginc.30%Expected over the next two years as utilization, procurement scale and serving efficiency improve.
• Group gross margin32.8% in 1H26Down from 37.7% in 1H25 as revenue mix changed.
• 2028E adjusted EBITDARmb1,311mnVersus Rmb139mn in 2026E and Rmb542mn in 2027E.
• Cash flow from operationsRmb523mn/Rmb376mn/Rmb421mnJPMorgan forecasts for 2026E/2027E/2028E.
• FCFFRmb125mn/(Rmb170mn)/(Rmb246mn)JPMorgan forecasts for 2026E/2027E/2028E, highlighting continuing investment needs.

Impact & implications

JPMorgan argues that Phancy’s valuation upside depends on proving that heterogeneous optimization and API serving can turn rapid Token growth into reusable product revenue, higher margins and cash returns. Sustained revenue growth without improving utilization, gross margin and returns on capital would offer weaker support for multiple expansion.

Risks

• AI compute demand may grow more slowly than forecast if efficiency gains outpace the creation of new economically viable workloads, reducing Token demand, utilization and API capacity absorption.
• Independent inference economics may weaken if model developers cut official API prices, cloud providers subsidize inference, open-model competition or commercial licensing changes reduce third-party endpoint value, or capacity shortages raise procurement costs.
• The optimization layer may commoditize as chip vendors improve software, cloud providers internalize optimization or open-source tools standardize deployment.
• Capital intensity may remain high: GPU purchases, leases and working-capital needs could leave operating cash flow and FCF weak even amid rapid revenue growth.
• Productization may progress slowly, leaving Phancy reliant on customized deployment and limiting margin expansion and operating leverage.

What to watch

• API revenue growth alongside utilization and gross-margin improvement.
• Gross profit per compute unit and whether API gross margin approaches about 30% over the next two years.
• Broader adoption and cross-customer, cross-hardware reuse of HAMi, RISE and ModelHub.
• Revenue growth relative to R&D and infrastructure spending, plus CFO, FCF and incremental ROIC.
• Evidence of repeat customer adoption, retention and improved unit economics in Agentic AI.
• The dilution and capital-return outcome of any successful A-share listing.