AI基建三條線，誰先漲、誰最猛、誰還能追？
- 核心觀點：AI產業爆發呈現清晰的傳導鏈條：首先利好晶片（底層算力需求），其次暴露出能源瓶頸（資料中心高耗電），最後長期抬升儲存需求（AI系統持續吞吐資料）。這一邏輯在近一年的美股、A股相關資產漲幅中得到顯著驗證。
- 關鍵要素：
- 晶片層：AI需求最先轉化為GPU（NVIDIA）、HBM及先進製程訂單，NVIDIA 2026財年營收同比增長65%，設備端（ASML、Lam Research）開年漲幅達27%-30%。
- 能源層：AI資料中心單櫃功率提升至50-100千瓦（傳統為5-15千瓦），IEA預測2030年資料中心用電達945 TWh，翻倍增長，推動燃氣輪機（GE Vernova，年漲幅167%）、核電營運商等資產重估。
- 儲存層：AI帶來數據頻繁讀寫、實時調用和緩存壓力，儲存晶片原廠（SK海力士Q1營收增198%、利潤增406%）及HDD廠商（SanDisk年內漲350%）表現突出。
- 孫宇晨去年11月推文提及的晶片、能源、儲存三條線，若當時買入Micron、SanDisk等美股儲存概念股，至今漲幅達180%-552%。
- 國內資產方面，海光資訊2024年營收91.62億同比增長52.4%，國產算力晶片及儲存設計企業（兆易創新、普冉股份等）同步受益。
Original Author: Changan I Biteye Content Team
In November last year, Justin Sun posted a tweet:
If you treat this statement as an industry judgment rather than just an attention-grabbing quote, looking back you'll find:
These three tracks are almost the true profit path of the AI narrative.
If, after that tweet, you had bought US-listed storage concept stocks, what would the outcome be today?
• Micron: +214%
• Seagate: +180%
• Western Digital: +190%
• SanDisk: +552%
This article breaks down these three tracks:
Why does AI first benefit chips, then highlight energy bottlenecks, and finally drive long-term storage demand? Which assets have already outperformed in this structural cycle?
1. Chips: The First to Cash in on the AI Boom – Not Narratives, but Orders
What AI burns first is not the application layer, but the underlying computing power.
Whether it's training large models, daily inference, Agent calls, or multimodal processing, the first step is to get the computing running, and all these computations ultimately rely on GPUs, HBM, high-speed interconnects, and advanced manufacturing processes.
In other words, the growth in AI demand doesn't first propagate to the later stages. Instead, it translates directly and concretely:
More chips, more powerful chips, higher bandwidth chips
That's exactly why AI demand is first reflected in the chip sector.
Industry data has already made this clear. Based on fiscal year 2026 estimates, NVIDIA's revenue grew 65% year-over-year, indicating sustained demand for high-end computing chips.
🌟Assets in this Direction
Core Computing Layer: NVIDIA (NVDA), AMD, Broadcom (AVGO), TSMC (TSM)
Domestic Computing Layer: Haiguang Information (688041.SH), Cambricon (688256.SH), among others. Haiguang is one of China's representative x86 server CPU companies, with 2024 revenue of RMB 9.162 billion, up 52.4% year-over-year.
Semiconductor Equipment Layer: ASML, Applied Materials (AMAT), Lam Research (LRCX). ASML's US-listed ADR hit an all-time high at the start of 2026, surging over 8% on January 2nd, with a year-to-date gain of 27%. Lam Research is up 30% year-to-date, and Applied Materials is up 28% year-to-date. All three semiconductor equipment giants have significantly outperformed the S&P 500.
🌟Performance Over the Past Year
The chip track is the earliest and most significant gainer in this AI wave. NVIDIA, as the leader, has achieved cumulative gains of over 1,000% since early 2023. Equipment stocks continued to hit new highs in early 2026 and remain in a strong upward cycle. Citigroup released a research report predicting a "Phase 2 Bull Upcycle" for the global semiconductor equipment sector, with the main themes for chip stocks in 2026 clearly pointing to ASML, Lam Research, and Applied Materials.
2. Energy: As AI Scales, the Bottleneck Shifts from Chips to Electricity
No matter how many chips you have, they can't run without power.
Buying chips is just the beginning. Long-term operation of large models, data centers, and inference services requires continuous power supply and extra load for cooling and heat dissipation. Traditional data center cabinets typically have a power density of 5 to 15 kW, but AI data centers have significantly raised this to 50 to 100 kW. The power consumption and cooling pressure are on a completely different scale. An IEA analysis this year mentioned that data center electricity consumption could reach around 945 TWh by 2030, roughly double current levels, with AI as the primary driver. The US Department of Energy has also explicitly stated that the growth in data center power demand is putting significant pressure on regional power grids.
🌟Assets in this Direction
Gas Turbines: GE Vernova (GEV): Gas turbine orders are booming, with total orders reaching $59 billion in 2025 and the backlog growing to $150 billion. Management raised the 2026 revenue guidance to $44-$45 billion.
Independent Power Producers: Constellation Energy (CEG): The largest zero-carbon power operator in the US, with nuclear assets securing long-term power purchase agreements directly with tech giants. Vistra (VST): Holds both nuclear and gas assets, with the midpoint of its 2026 EBITDA guidance approximately 30% higher than 2025.
Uranium Resources: Cameco (CCJ): The world's largest publicly traded uranium miner, a beneficiary of the nuclear renaissance upstream.
🌟Performance Over the Past Year
GE Vernova's stock price rose 167% over the past year. The 52-week low was $408, and it hit a high of $1,181, nearly doubling from its low. Constellation Energy hit an all-time high in 2025 but subsequently pulled back about 28% from its peak due to regulatory policy uncertainties, currently trading at a relatively low level. Vistra remains strong overall, with long-term power supply contracts with data centers continuously materializing. The energy sector has overall been repriced from a traditional defensive position to a core beneficiary of AI infrastructure.
3. Storage: The Most Easily Overlooked, but a Long-Term Beneficiary
The core logic for storage is simple: AI is not a one-off call. It is fundamentally a system that continuously processes, stores, and calls data.
Training requires reading massive data, saving checkpoints during the process, inference requires loading models and caches, and RAG and Agents constantly access knowledge bases, logs, and memories.
Consequently, AI brings not just "more data," but:
• More frequent data reads and writes
• More real-time data access
• More complex data management
• Greater pressure on data migration and caching
Looking deeper, the more expensive the GPU, the less it can sit idle. Therefore, the industry will increasingly focus on how to deliver data to the computing power end faster and more reliably.
In other words, the more AI develops, the more storage becomes not just a "warehouse for data," but the data foundation that ensures the entire AI system can run continuously.
🌟Assets in this Direction
Memory Chip Manufacturers: SK Hynix (000660.KS), Samsung Electronics (005930.KS), Micron Technology (MU)
NAND / SSD / HDD Manufacturers: SanDisk (SNDK), Seagate (STX), Western Digital (WDC)
Domestic Storage Design: GigaDevice, Puya Semiconductor, Dongxin Co., Beijing SinoWealth, Montage Technology, along with storage module manufacturers such as Demingli, Shanneng Xinchuang, and Longsys.
🌟Performance Over the Past Year
Since the beginning of 2026, the storage sector has been one of the strongest segments in the AI industry chain. On the US stock side, driven by AI infrastructure investment and high-capacity storage demand, Seagate, SanDisk, and Western Digital have all surged significantly this year. Reuters reported in late April that Seagate and Western Digital have more than doubled year-to-date, while SanDisk has risen around 350% year-to-date. Memory chip manufacturers have also strengthened in tandem. Micron has risen sharply this year, and SK Hynix continues to benefit from HBM shortages and capacity grabs by major clients, reporting Q1 revenue up 198% year-over-year and operating profit up 406% year-over-year, further solidifying its profitability.
Final Thoughts: Chips Go First, Energy Catches Up, Storage Benefits Last
The first phase of AI realization is chips; the second bottleneck is energy; the third long-term beneficiary is storage.
Correct logic does not equal a comfortable entry point. Structural opportunities exist, but it's not about blindly chasing highs.
What's truly valuable isn't the hype itself, but where you stand in the industry chain.
Disclaimer: The above is merely an industry chain review and does not constitute investment advice. In particular, some targets have already seen exaggerated gains in 2026. Correct logic does not guarantee a comfortable entry point.
