Tech giants share the electricity costs! Is the power sector entering a golden era?

富途寰球私享匯

joined discussion ·

Alpha Call Spotlight | Reevaluation of CPU Value and Opportunities in the Power Chain under the AI Era Summary Report

I. CPU Sector: An Underestimated Shortage is Unfolding

Over the past two years, almost all the spotlight on AI semiconductors has been on GPUs and memory. However, in this issue’s discussion, analyst Tim Yang presented an easily overlooked assessment:CPUs are entering their own cycle of shortage-driven price increases, with Intel potentially being the stock with the highest upside.

Two forces are simultaneously driving up CPU demand.

The first force comes from the "old world's" wave of hardware upgrades.Cloud servers deployed on a large scale during the pandemic in 2020–2021 will reach the end of their five-year lifecycle around 2026, triggering a peak in replacement demand. At the same time, an easily overlooked fact in the market is that not all companies need GPUs to run AI. Intel’s fourth-generation Xeon CPUs already come with built-in AMX accelerators, which can efficiently perform small model inference—making a general-purpose server sufficient for small and medium-sized enterprises that only need customer service bots, internal applications, or have strong data privacy concerns, much more cost-effective than purchasing expensive GPUs. This logic mirrors what we previously discussed about storage, where SME demand for SSDs/HDDs outweighs that for HBM.

The second force comes from the paradigm shift towards Agentic AI—this represents the true incremental variable.Unlike traditional chatbots, Agents need to frequently invoke external tools such as database queries, Python script execution, and web scraping, and these operationsall run on CPUs.Data shows that tool processing latency accounts for over 90% of an Agent's total latency, and under large-batch scenarios, CPU energy consumption makes up 44%. In other words,In Agentic AI systems, CPUs are becoming bottlenecks earlier than GPUs.The compound annual growth rate (CAGR) of tasks executed by agents from 2025 to 2030 could exceed 500%. Combined with mainstream platforms like DouBao and ZhiPu fully transitioning to sandbox execution architectures (where CPU resource consumption is strongly correlated with user scale), the strategic importance of CPUs is being reevaluated.

Surging demand has collided with rigid supply.

Meta, Google, and Microsoft have raised their server demand forecasts for 2026 by 15%-20%. However, the issue is that production capacity cannot keep up. Intel's 10nm/7nm and Taiwan Semiconductor’s N5/N3 process capacities are strained, with Taiwan Semiconductor prioritizing GPU production, leaving very limited flexibility for server CPUs.

The results are already beginning to show:Intel has notified clients of an approximate 10% CPU price increase by the end of this month, while AMD server CPUs are expected to rise by 10%-15%.According to industry exchanges, this is likely not a one-time event — there may be another round this year and possibly again next year, with the supply-demand gap expected to be largely filled only by the first half of 2027.

Core target: Intel - the high-elasticity stock 'shunned' by the market

$Intel (INTC.US)$

Its product business remains the foundation, more stable than the market anticipates.CCG (PC processors) accounts for approximately 60% of revenue with a profit margin of 29%, making it a cash cow; DCAI (server CPUs + AI accelerators) makes up about 32% and represents the strategic focus. Q4 2025 earnings exceeded expectations, but weak guidance for Q1 2026 caused a sharp drop in share prices — the core reason being management underestimating the AI demand surge, resulting in finished goods inventory falling to 40% of its peak, creating a situation of "produce-as-you-sell."

However, this precisely means room for improvement. Given the substantial size difference between PC chips (~120mm²) and server chips (~500-600mm²),shifting just 10% of PC production capacity to servers could support a roughly 20% increase in server shipments.Moreover, because servers command much higher unit prices and profit margins compared to PCs, the overall boost to the company's profitability would be quite significant.

The foundry business is the true valuation option, and it is accelerating its realization.Here are several noteworthy developments:

>18A process(equivalent to Taiwan Semiconductor's N2) launched on March 12 with the third-generation Core Ultra. The yield rate has improved from the previous 50% to around 55%-60%, showing potential for upside surprises. A key context is that — Broadcom executives confirmed just yesterday that Taiwan Semiconductor’s overall capacity has hit its limit, making demand spillover an inevitable trend.

>14A processis expected to progress faster than 18A (the team stabilized after experiencing organizational turbulence during the 18A phase). Potential major clients point to two companies:Apple(entry-level M series processors, which could replicate the path taken by Taiwan Semiconductor and Apple to improve yields) andNVIDIA(The greatest strategic significance of Intel's inclusion lies in the expected gradual increase in Intel's foundry share for second- and third-generation GPUs). Notably, NVIDIA’s collaboration with Intel is not merely a 'gesture' to the US government — in scenarios pursuing higher general computing power and ecosystem maturity, the evolution speed of x86 may outpace ARM. Essentially, NVIDIA is placing bets on both tracks simultaneously.

Advanced packagingIt is the first link to be implemented and monetized. Intel's packaging costs are only 70%-80% of Taiwan Semiconductor's CoWoS, with a yield rate easily reaching 98% (back-end assembly interconnection, not involving EUV-level precision), much higher than front-end wafer manufacturing. About 68% of Taiwan Semiconductor’s capacity in 2026 has already been secured by NVIDIA.Google TPU v8 is expected to adopt Intel packaging by 2027, combined with automotive chips and NVIDIA’s back-end packaging cooperation, projected to contribute approximately $800 million in revenue by 2027 — a brand-new growth engine.

Valuation: SOTP breakdown points to $57, with potential upside of approximately 30%. Product business contributes about $37 (2026E revenue of $50 billion × EBIT margin of 25% × 15x EV/EBIT) + foundry business contributes about $18 (net assets of $80 billion × 1.2x PB, which aligns closely with capacity value estimated based on monthly capacity of 440,000 wafers, referencing Taiwan Semiconductor/GlobalFoundries) + equity assets such as Mobileye contribute about $2. The current PB is only 2x (compared to AMD at 5x), market expectations are low, showing significant room for recovery.

AMD vs Intel: How to choose? AMD offers stronger certainty, but much of its growth expectations have already been priced in; Intel, with purer CPU exposure, shows greater elasticity and broader recovery potential. Conservative investors should opt for AMD, while those seeking higher upside should prioritize Intel.$Advanced Micro Devices (AMD.US)$$Intel (INTC.US)$

II. AIDC Power Chain: America’s issue isn’t electricity, but 'timely electricity'

In the opening of Part II, analyst Freya highlighted a core assessment:China’s AI bottleneck lies in chips, while America’s AI bottleneck lies in electricity — and this is not a short-term transition, but rather a long-term structural contradiction.

Why is it said that the US electricity problem is almost unsolvable?

The data is very straightforward: American IDC power consumption as a proportion of total national consumption will soar from 4% in 2024 to 13.2% in 2030 (IEA forecasts a 3.4-fold increase), with a supply gap expected by 2028.12GW supply gapBut the real issue is not generating capacity, but the grid. A comparison between China and the US shows this: China adds over 400GW of new power generation capacity annually, has built a 50,000 km ultra-high voltage network for transmitting electricity from west to east, and coordinates supply and demand top-down through its 'East Data West Compute' policy — electricity is not an issue in China. In contrast, the situation in the US is almost a textbook case of what not to do:

> Aging grid: The average age of the North American grid is 35–40 years, with long-term lack of systematic upgrades;

> Regional fragmentation: Over 50% of data centers are concentrated in Texas, California, Northern Virginia, and similar areas, but there is poor interconnection between regional grids, leading to severe localized shortages;

> Non-compressible construction timelines: Approval and construction of new transmission lines requireMore than 8 years, while China's ultra-high voltage lines can be completed in 1-3 years;

>Net reduction in stable power sources: The peak of coal power retirement is approaching, wind and solar energy storage is unstable, and nuclear power construction has a long cycle.

One-sentence summary:Distant water cannot quench immediate thirst. Distributed power generation is not a temporary solution but a mainstream strategy for the long term.

This is the key judgment that Freya repeatedly emphasized during this sharing session. There are three reasons:First, time is money—literally.The potential annual revenue of AIDC could reach tens of billions of dollars. The commercial value of going live 3-5 years earlier is enough to cover the entire CapEx and fuel costs of building one’s own power plant. Waiting for the grid means financial losses, and the AI race can't afford to wait.Second, distributed solutions are inherently superior to connecting to the public grid.Directly connecting natural gas pipelines or nuclear power to an internal data center power plant is far more reliable than connecting to an old, fragmented public grid.Third, the policy has paved the way.The 2025 DATA Act allows AIDC to bypass federal power regulation through off-grid energy facilities, and since December of last year, large data centers have been permitted to directly connect with power plants behind the meter. The effective power capacity metrics of the US grid have been declining --Even if the grid continues to improve, backup power is still necessary, and distributed generation will coexist with grid-connected power in the long term.

Priority in equipment selection: Gas turbines as the main option, diesel generators for emergencies, and nuclear power for the long term

The current landscape is that global new orders for gas turbines have exceeded 85GW, with effective production capacity at only 50GW, and orders are scheduled until 2028-2030. Gas internal combustion engines are booked until 2029.Diesel generators have thus been 'elevated' to primary power use—they represent the fastest installation solution available today.

Diesel generator sector

Weichai Power (Hong Kong shares)$WEICHAI POWER (02338.HK)$- Dominating delivery timelines, StarGate certification, overseas pricing advantages

Weichai’s strategy is clear: through its wholly-owned subsidiary Baudouin (a French brand), it has formed an exclusive engine supply relationship with Generac, the leading US generator manufacturer. The key advantage is lead time—while Caterpillar and Cummins have delivery cycles of up to 100 weeks, Generac, supported by Weichai's production capacity, requires only 36–60 weeks, a gap significant enough to influence customer purchasing decisions.

In terms of brand certification, the bidding documents for the OpenAI Stargate project in Dezhou this February showed that Broadwell was listed alongside Caterpillar and Cumminsas jointly awarded the contract—this is an endorsement from a top-tier North American data center for the brand. On the profitability side, high-power diesel generators in North America are priced at $600,000 to $800,000, while similar domestic products cost only ¥1.5 million, making the profit margin for overseas business much higher than domestically.

Caterpillar (US stock)$Caterpillar (CAT.US)$- The king of market share, with valuation being redefined

Holding a 48% share of the North American high-power diesel generator market, demonstrating absolute dominance. Sales volume is expected to grow by about 80% from 2025 to 2030. More importantly, the valuation logic is changing—orders are backlogged until 2029, production expansion is limited, and the market is beginning to price this traditional cyclical stock as a growth stock.

Cummins (US stock)$Cummins (CMI.US)$Its 95-liter high-power diesel generator has industry barriers, but its product lineup does not match Caterpillar’s (lacking gas turbine layout), leaving relatively limited room for imagination.

Gas turbine/natural gas power generation track—the long-term main battlefield for primary power supply

Weichai Power (Hong Kong shares)$WEICHAI POWER (02338.HK)$- PSI is the overlooked trump card

Weichai holds a 46.5% stake in its US subsidiary, Power Solutions International (PSI), which has obtainedcertification as a qualified supplier from the four major US cloud vendors, with 1MW-class natural gas engines and full product lines certified by UL. Weichai is developing both diesel and natural gas solutions simultaneously, with North American order visibility continuously improving. Recent pullbacks due to tensions in the Middle East have not changed fundamentals — electricity shortages persist, overseas supply chains remain tight, and the situation between the US and Iran has significantly eased, leading tostrong rebound potential。

Dongfang Electric (Hong Kong shares)$DONGFANG ELEC (01072.HK)$- Dual drivers of overseas breakthroughs and domestic outperformance

Overseas: The core product, G50 gas turbine, has a production cycle of about one year (compared to over two years for Siemens, GE Vernova, and Mitsubishi), with an annual capacity of 30 units, ensuring relatively sufficient supply. Recently, it secured an order for 20 × 50MW gas turbines in Canada (total value ¥40 billion, gross margin 40%-50%) — representing both profit growth and high-end market brand recognition, opening doors for future entry into North America’s AIDC market. Domestically: State Grid and Southern Grid’s CapEx on power equipment for 2026 exceeded expectations. Combined with increased demand for hydropower equipment from the Yarlung Zangbo River Hydropower Station project and rising energy security concerns following the Hormuz Strait incident (China's coal and nuclear power self-sufficiency rates are high, potentially leading to increased baseload power investments), Dongfang Electric, as one of the largest holders of domestic baseload power equipment market share, stands to benefit with certainty.

Caterpillar (US stock)$Caterpillar (CAT.US)$- The most comprehensive product portfolio

The wholly-owned subsidiary Solar Turbines produces gas turbines and also provides reciprocating generators. Together with the diesel generator business, they form a three-pronged product portfolio covering all AIDC scenarios. They have signed orders for 2GW with American Electric Power and $800 million/1.4GW with Atlas, with a backlog of gas turbine orders exceeding 5GW. By 2030, they plan to increase gas turbine production capacity by 150% and gas generator capacity by 100% compared to 2024, corresponding to a total power generation of 50GW. The delivery speed is slightly faster than GE Vernova, with more aggressive capacity expansion.

GE Vernova (US stock)$GE Vernova (GEV.US)$- Absolute leader, but valuation has already moved ahead

Holds the largest share of global gas turbine installations (about double that of competitors), with the highest technological barriers and essentially pricing power — unit prices have risen from $1,200–1,400/kW to $2,400–2,500/kW. Data center orders will increase from 10%–15% to about one-third, with customers proactively switching to long-term supply agreements (SRA), which is conducive to valuation restructuring. However, precisely because it is the core target,current valuations have largely priced in AIDC orders and pricing power expectations, if AI momentum pulls back, there is significant risk of valuation decline.

Comprehensive preferred choice

III. Q&A Highlights: Three questions investors care about most

Q1: Could Intel’s received packaging/flow test orders just be a 'formality' or a gesture of goodwill?

Unlikely. Packaging belongs to the backend assembly interconnect process and does not involve EUV lithography-level precision; its yield rate is naturally ≥98% — as long as the process runs smoothly, mass production faces almost no technical obstacles. Additionally, packaging defects are usually detectable and isolatable single-point issues, unlike frontend flaws where one defect can ruin an entire wafer. Intel has clearly split its front-end and back-end operations and adopted a more open stance, while Taiwan Semiconductor's capacity is fully utilized, making it highly probable that demand overflow will convert into substantial orders.

Q2: What is the likelihood for domestic companies like Dongfang Electric and Weichai Power to secure long-term large overseas orders?

Under normal circumstances, brand validation for medium and heavy-duty gas turbines requires 50,000 to 60,000 hours, or up to seven years. However, with the current extreme power shortage in the AIDC sector, the risk appetite of overseas clients has marginally shifted – some are choosing a "trial by practice" approach, opting for domestic equipment to meet short-term delivery needs. Weichai's generator sets have already received certification for the Gateway Project, while Power Solutions International (PSI) has been qualified as a supplier by the four major cloud vendors. Dongfang Electric's acquisition of a high-end order from Canada (with a gross margin of 40%–50%) serves as a substantial signal of brand breakthroughs. Meanwhile, Dongfang Electric is exporting to the Middle East, Central Asia, and Southeast Asia; although the gross margin is relatively low, this demonstrates accumulating validation in product quality and delivery capabilities. Both brand recognition and product quality are improving, indicating sustainability.

Q3: Will improvements in chip energy efficiency, grid upgrades, and expanded production by equipment manufacturers quickly reverse the power shortage?

Three counterarguments: ① The increase in chip energy efficiency cannot offset the total growth in electricity demand driven by AI; ② The physical construction cycle for the power grid takes over eight years and is non-compressible – policy is one thing, but actual construction is another; ③ Effective capacity indicators for the US power grid are already declining, meaning even if there are improvements, backup power solutions will still be necessary.The core tracking indicator for the power equipment chain is the growth rate of AI capital expenditures (CapEx) by cloud vendors. In terms of investment advice, priority should be given to targets that are less priced in and have weaker correlation with AI momentum – these have limited downside but greater elasticity.

393K Views