99% of Investors Miss The AI Backbone
99%的投资者错过了AI基础设施
$7 trillion is pouring into data center infrastructure. 60% goes to compute, 40% to facilities. Everyone watches Nvidia's stock price — almost nobody pays attention to the buildings going up behind it. The money isn't in the hype. It's in the hardware, the power, and the concrete.
7万亿美元涌入数据中心基础设施。60%流向算力，40%流向设施。所有人盯着英伟达股价——几乎没人注意背后拔地而起的建筑。钱不在炒作里，在让AI成为可能的硬件、电力和混凝土中。
The $7 Trillion Split
7万亿美元分配
McKinsey breaks the $7T into two buckets. The two layers behave nothing alike — different growth, different risk, different winners.
麦肯锡将7万亿分为两个桶。两层完全不同——不同的增长、风险和赢家。
Key Insight
核心洞察
Compute is betting on who wins the AI race. Facilities is betting that the race happens at all. Someone has to build the track.
算力层押注谁赢得AI竞赛。设施层押注竞赛本身会发生。总得有人建赛道。
Compute Layer — $4.2T
算力层 — $4.2万亿
Where most of the revenue growth is happening. Chips, memory, servers, networking, and the semiconductor supply chain behind them all.
大部分营收增长发生在这里。芯片、内存、服务器、网络以及背后的半导体供应链。
AI Chips
AI芯片
NVDA still way out front. AMD has real momentum with MI300. AVGO and MRVL inside almost every major AI system. INTC pushing hard to get back in. Nvidia's newest chips pull up to 3× more power than last gen — driving the entire power/cooling chain.
NVDA遥遥领先。AMD MI300势头强劲。AVGO和MRVL几乎在所有AI系统中。INTC努力回归。英伟达最新芯片功耗比上一代高3倍——驱动整个电力/散热链。
HBM Sold Out for Years
HBM多年售罄
MU, Samsung, SK Hynix — all sold out in high bandwidth memory for years. Systems can't run without massive memory. Memory demand is exploding right alongside GPU demand, but most people overlook it.
MU、三星、SK海力士——高带宽内存多年售罄。系统没有大量内存无法运行。内存需求与GPU需求同步爆发，但大多数人忽略了。
Turning Silicon Into Racks
从硅片到机架
SMCI scaling almost faster than anyone. DELL and HPE anchor the enterprise market. LCAP (Lenovo) huge across Asia. Growing group of GPU cloud providers: Applied Digital, Okami, Digital Ocean, Iris Energy — leasing AI compute as fast as hardware arrives.
SMCI扩张几乎最快。DELL和HPE锚定企业市场。LCAP（联想）在亚洲很大。GPU云提供商群体增长：Applied Digital、Okami、Digital Ocean、Iris Energy——硬件到手即出租。
The Choke Points
瓶颈点
TSM manufactures almost every advanced AI chip. ASML is the choke point for the tools everyone needs. LRCX, KLA, AMAT handle the rest of the equipment. Without these four, no AI chip exists.
TSM制造几乎所有先进AI芯片。ASML是所有人需要的工具的瓶颈。LRCX、KLA、AMAT处理其余设备。没有这四家，没有AI芯片存在。
Moving Data Between Racks and Regions
在机架和区域间传输数据
ANET leads cloud-scale switching. CSCO drives enterprise traffic. CIEN, LUMN, COHR move data across long distances between buildings, regions, and countries.
ANET领航云规模交换。CSCO驱动企业流量。CIEN、LUMN、COHR在建筑、区域和国家间长距传输数据。
Hyperscalers — Their Own Bucket
超大规模——单独一桶
They don't just buy data center capacity — they build entire regions, negotiate multi-GW power deals, design their own chips, and run the clouds where every enterprise AI workload lands.
他们不只买数据中心容量——他们建整个区域、谈多GW电力交易、设计自己的芯片、运营每个企业AI工作负载所在的云。
Amazon · Microsoft · Google
Amazon · Microsoft · Google
The only players touching every layer of the stack. AWS, Azure, Google Cloud. Designing their own silicon: Amazon Trainium/Graviton, Google TPUs, Microsoft Maya. Amazon just rolled out in-house liquid cooling because traditional suppliers are backlogged for years — they won't wait until 2027.
唯一触碰每个层级的玩家。AWS、Azure、Google Cloud。设计自有芯片：Amazon Trainium/Graviton、Google TPU、Microsoft Maya。Amazon刚推出自研液冷——传统供应商积压数年——他们不愿等到2027。
Oracle · Alibaba · Tencent · IBM
Oracle · 阿里 · 腾讯 · IBM
Oracle accelerating with AI/HPC leasing. Alibaba and Tencent run massive AI regions across Asia. IBM carving out regulated industry niches. But the big three still sit in a category of their own — they're the only ones touching every layer.
Oracle加速AI/HPC租赁。阿里和腾讯在亚洲运营大规模AI区域。IBM切入监管行业细分。但三大仍有独立类别——他们是唯一触碰每个层级的。
Facilities Layer — $2.8T
设施层 — $2.8万亿
The part you never see, but nothing works without it. Power, cooling, electrical gear, real estate — paid every time a new data center flips the lights on.
你看不到的部分，但没有它一切都不行。电力、散热、电气设备、地产——每次新数据中心开灯就付费。
Keeping the Racks Alive
让机架活下去
VST tied directly to AI data center rise. ETN and SBGSY (Schneider Electric) handle electrical distribution and switchgear. JCI, TT (Trane), DJK (Daikin) manage thermal. MOD (Modine) and IVT (Invent) growing fast as racks shift to high-density cooling.
VST与AI数据中心崛起直接相关。ETN和SBGSY（施耐德）处理配电和开关设备。JCI、TT（特灵）、DJK（大金）管理散热。MOD和IVT随高密度散热快速增长。
When the Grid Can't Keep Up
当电网跟不上
AI pushing power demand higher than the grid was ever designed for. SIE (Siemens), ABB, PWR (Quanta) seeing real tailwinds. BE (Bloom Energy) and CMI (Cummins) for on-site generation and backup when facilities need more stability than the grid can provide.
AI推高电力需求超电网设计极限。SIE（西门子）、ABB、PWR（Quanta）迎顺风。BE（Bloom Energy）和CMI（康明斯）提供现场发电和备用，当设施需要比电网更稳定时。
The Shell That Keeps Everything Alive
让一切存活的壳
EQIX (Equinix), DLR (Digital Realty), IRM (Iron Mountain) build and lease the shells. They provide the space, interconnects, and reliability that let everyone else plug in. Steady income, benefits no matter which chip or cloud wins.
EQIX（Equinix）、DLR（Digital Realty）、IRM（Iron Mountain）建设并出租壳体。提供空间、互联和可靠性，让其他人接入。稳定收入，无论哪个芯片或云赢都受益。
$100 Allocation
100美元配置
Not every part grows at the same speed. Match the growth, risk, and timing with what the data tells us.
并非每个部分增速相同。根据数据匹配增长、风险和时机。
5-Year Projection
5年预测
That $100 grows into roughly $200–$235. No guessing, no moonshots — just clean exposure across the parts of the data center stack doing the real work.
100美元增长至约$200-$235。没有猜测、没有赌注——只是对数据中心栈真正干活的部位的干净敞口。