目录
I. After Strong Demand, the Market Needs a More Concrete Earnings Model
II. Improving Chinese Demand Creates the First Growth Opportunity for Existing Power Products
III. 800V Remains the Direction of Travel, but Commercialization Requires Multiple Rounds of Customer Validation
IV. Gas Turbine Supply-Chain Bottlenecks: Equipment, Processes, and Effective Output
V. The Difference Between KSTAR and Jianghai Capacitor Lies Between Qualification and Expectations
VI. Valuations Already Look Far Ahead, Magnifying the Risk of Earnings Delays
VII. What Developments Could Genuinely Change the Investment View Over the Next Four Quarters?
本内容基于公开资料和研报数据整理，不构成任何投资建议，不代表任何个人观点，仅供学习参考，请理性阅读
Orders remain robust for AI computing-center power suppliers, but earnings conversion is increasingly diverging. Gas-turbine capacity expansion is constrained by equipment availability and production yields, 800V DC power still requires validation, and customer qualification determines which buyers new capacity can serve. The next phase of competition will hinge on delivery schedules, product margins, and collection quality.
I. After Strong Demand, the Market Needs a More Concrete Earnings Model
The most consequential shift in Goldman Sachs’ latest research is its focus on what happens after orders are placed. Its September 4, 2026 report, China Industrials Technology: AI Computing-Center Power Supply, summarizes company discussions held during the Asia Leaders Conference. The conference ran from August 31 to September 2 and hosted approximately 40 companies; the power-supply analysis covers the relevant participants. This scope matters: the report provides operating signals from the supply chain, not a statistical survey of the entire industry.
The companies interviewed generally did not report weakening demand. Jereh Group’s gas-turbine ancillary business, KSTAR’s data-center electrical equipment, and Megmeet’s server power supplies all have order or demand support. Goldman Sachs has also become more constructive on Chinese demand in 2027–2028. At the same time, the report cautions that some new power architectures are commercializing more slowly than the market expects, while the ability to bring capacity online as planned is becoming increasingly important. Rising demand and below-expectation earnings can occur simultaneously.
This changes the sequence for evaluating a company. Early discussions of AI power supply often began by estimating the market size and power components required per rack, then assigning supplier shares. Once the industry enters the delivery phase, the questions become: Which customers have completed qualification? Which products are already in mass production? How much can the factory ship this quarter? How much profit will remain after delivery? Long-term market potential still matters, but these operating steps must bridge the gap between that opportunity and the next earnings report.
The same demand can appear successively in customer plans, project pipelines, framework agreements, purchase orders, and supplier revenue. Adding these figures together would overstate the market; treating them as equally reliable would overstate earnings visibility. A project pipeline represents potential demand, purchase orders are generally closer to execution, and revenue recognition still depends on delivery and contract terms. All company operating figures discussed here are interpreted according to the timing and attribution of the underlying research discussions.
Goldman Sachs rates KSTAR Buy, Megmeet Neutral, and Jianghai Capacitor Sell, illustrating this divergence. All three companies are exposed to data-center power supply, yet their ratings differ sharply. Understanding the distinction requires assessing both what each company can deliver and what its valuation already assumes. Even a company with rapid business growth can come under pressure if market expectations are higher still.
Earnings conversion faces at least three constraints. Customer qualification determines the addressable sales base; capacity and production yields determine actual delivery volumes; and product pricing and costs determine the return on each unit of revenue. Cash flow faces the additional constraint of collections. During capacity expansion, upfront spending on equipment, raw materials, and work in progress may increase, so profit growth and operating cash-flow improvement may not occur simultaneously.
The analysis also works in reverse. Delays in commercializing new technologies do not negate demand for existing products, while ongoing factory expansion does not imply that a company lacks earnings power. Each product must be assessed separately: How much does the existing business contribute? When will new businesses begin contributing? Is the market applying the same growth rate to both? Divergence in the power-supply value chain begins at the product and customer levels before appearing in company financials.
II. Improving Chinese Demand Creates the First Growth Opportunity for Existing Power Products
Incremental Chinese demand may initially emerge through higher power ratings and increased shipment volumes for existing server power supplies. Goldman Sachs cites improving domestic availability of GPUs and ASICs as one reason for its more constructive view of AI computing-center demand in 2027–2028. GPUs are general-purpose parallel-processing chips, while ASICs are specialized chips designed for specific tasks. Better chip availability removes one constraint on system deployment, although power-supply orders must ultimately align with server procurement and data-center construction schedules.
The report notes that server power supplies are moving from 2–4 kW to 5.5 kW. This upgrade requires suppliers to rebalance efficiency, thermal management, reliability, and space constraints. Customers are purchasing power products capable of stable long-term operation. Higher power ratings can increase the value per unit, but pricing and profit still depend on design complexity, competition, material costs, and price reductions after volume production begins.
The percentage increase in power output therefore cannot be equated directly with revenue growth. Revenue also depends on server volumes, the number of power supplies required per server, supplier share, and unit pricing. Redundancy design further affects procurement volumes. Even as power per unit increases, the number of modules used in a system may change. The report does not provide a complete system-configuration matrix covering these variables, so it cannot support a reverse calculation of industry-wide revenue.
Megmeet provides a more concrete reference point. According to the report, the company has secured volume orders for 5.5 kW server power supplies, completed qualification with several domestic cloud service providers or contract-manufacturing customers, and is also seeing increased orders from US cloud providers. Cloud service providers are commonly abbreviated as CSPs. The combination of qualification and volume orders is closer to revenue generation than a product-development announcement alone, but execution cadence and product margins still require monitoring.
Honor Electronic’s progress reflects a different timetable. According to the conference discussions, small-volume shipments of its 5 kW project for Google are expected in the fourth quarter of 2026, followed by a production ramp in the first half of 2027. Two production lines in Vietnam have designed annual capacity of 400,000–500,000 units of 5–5.5 kW power supplies. These represent three distinct stages: expected small-volume shipments, an anticipated future ramp, and designed capacity. Designed capacity becomes sales volume only if customers continue placing orders and the factories reach the corresponding utilization rates.
The value of such projects depends on how long the customer relationship lasts. Entering a supply system for the first time generally requires R&D; and qualification investment; those upfront costs become easier to amortize if subsequent product generations continue to win orders. Conversely, customers may reallocate supplier shares when moving to the next-generation platform, diminishing the value of qualification secured for the previous generation. Research cannot stop at the statement that a supplier has entered a major customer’s supply chain; it must continue tracking mass-production share and continuity across product generations.
Overseas demand stability also remains conditional. Goldman Sachs reports that Jereh Group’s existing US projects feature off-grid operations, local production, and core US components, while KSTAR primarily supplies North America through an ODM model. ODM refers to products designed and manufactured to customer specifications and generally sold under the customer’s brand. These operating models may explain why the companies interviewed have not yet seen material disruption, but they do not establish that all future projects will remain unaffected by policy.
Off-grid or behind-the-meter power addresses specific constraints in electricity supply and project deployment. Behind-the-meter power generally refers to energy resources installed on the customer side of the meter; whether an individual project is fully off-grid must still be confirmed separately. Project ownership, equipment sourcing, local production arrangements, and future rules may each affect implementation. Revenue forecasts should therefore track actual customers and performance obligations project by project, rather than extrapolating the characteristics of one company’s current projects to the entire industry.
