一、几则值得并置观察的行业信号
8月23日，阿里巴巴宣布港股上市以来首次新股配售，募资规模达800亿港元，并明确将100%投向AI基础设施建设。其中有一个容易被忽略的结构性细节——阿里账面并不缺钱（截至6月底现金储备约700亿美元），却仍选择以稀释股权的方式再融资。这背后的信号是：AI算力投入的资本开支强度，已经开始突破经营现金流可自我覆盖的边界，必须借助外部股权融资来补足资金缺口。
从投入强度看，阿里单季资本开支达676.78亿元人民币，同比增长75%——折算下来，相当于每天烧掉约7.5亿元。腾讯同步加码，二季度资本开支攀升至528亿元，约为上年同期的近三倍。仅阿里、腾讯两家，单季资本开支合计已超1200亿元。
加杠杆最激进的是字节跳动。据彭博社报道，字节正讨论2026年最高700亿美元的AI数据中心投入计划（约合人民币4700亿元），这一体量据称相当于腾讯、阿里、百度三家的AI投入总和。
海外同样进入高强度投入周期。谷歌将资本开支指引上调至1950亿—2050亿美元；Meta把2026年资本开支指引下限从1250亿美元上调至1300亿美元；OpenAI将截至2030年的算力支出上调至7500亿美元。据不完全统计，2026年全球九大云服务商总资本开支预计突破8867亿美元。
二、回报率已越过融资成本线
巨头为什么敢如此密集地砸钱？答案落在算力投资的回报模型上。
按阿里CEO吴泳铭在财报会上的口径：AI算力资本支出预计三年内回本，随着AI产品毛利率持续提升，回收周期有望缩短至2.5年甚至两年左右。
海外模型的回报更陡峭。海外云厂商算力租赁的投资资本回报率（ROIC）约28.7%，国内云厂商在中性假设下约13.5%，二者均高于各自的边际融资成本。换言之，这一轮资本投入在账面上已经具备"自我造血"能力。
马斯克旗下xAI的超级计算集群是更极致的样本：Claude、谷歌均是其大客户，按现有租金水平测算，两年即可回收全部资本开支。
正因如此，算力租赁目前看确实是一门好生意。前段时间有位此前做不动产投资的客户问我，什么生意好做且适合他，我给的建议是：造一座水电站，下面建一个算力中心。
三、供需临界点与代际风险
那么，算力租赁这么火热，会不会有隐忧？会。资本疯狂涌入的必然结果，是供给越过需求，只不过这个需求的临界点究竟在哪里，各家测算口径不一；更何况它本身还是一个高速演化中的变量。
从当下的运行效率看，问题已经显现：GPU利用率偏低，国内智算中心的平均利用率不足30%，即便在剔除约20%的冗余配置后，仍有约50%的算力处于闲置状态。与此同时，这个赛道还叠加了一层技术风险——一旦出现代际性技术突破，已建成的集群资产就会在租赁端承压，届时价格是否会出现雪崩式下行，目前无人能够给出确定性答案。
四、头部厂商为何仍敢重注
既然存在明显隐忧，为什么头部大厂依然敢下重注？逻辑很简单：押注AI赛道几乎是必须动作，不上车的代价就是出局。而且它们都自带保障——因为它们本身就是最大的算力消耗方。
阿里的通义、腾讯的混元、字节的豆包，每天在进行推理和模型训练，消耗的算力本身就是天文数字。对它们而言，自建算力即便对外租不出去，内部消化也划算；对外租赁反而只是边际收益的补充，而不是盈利的前提。
五、给投资方的建议
如果是单纯要投资算力基础设施，我的建议是：把思维从"建设思维"切换到"运营思维"——唯有如此，当行业整体承压时，才能守住细分市场。
所谓运营思维，首先是不要把智算局限在硬件堆叠上，而要优先考虑运营效率的提升与生态体系的协同演化。比如，如何高效榨取有限资源，最大化单位GPU、单位功耗所能产出的Token数量。这是第一个维度——指向更优质的服务与更强的价格韧性。
第二个维度，是深入到应用场景中去，不只提供裸算力，而要叠加应用落地：
如果今天要投算力，别只做"算力二房东"。裸算力没有护城河，谁便宜用户就用谁。真正的壁垒，是"算力＋场景"。谁离用户最近，谁就赚走最大的利润。
I. A Set of Industry Signals Worth Examining Together
On August 23, Alibaba announced its first new share placement since its Hong Kong IPO, raising HK$80 billion, with 100% of proceeds earmarked for AI infrastructure. One easily overlooked structural detail: Alibaba is not short of cash on its books (cash reserves stood at roughly US$70 billion as of end-June), yet it opted to raise capital via equity dilution. The signal behind this move is clear: the capital expenditure intensity of AI compute investment has begun to outpace what operating cash flow can self-fund, requiring external equity financing to fill funding gaps.
In terms of investment scale, Alibaba’s quarterly capital expenditure hit RMB 67.678 billion, a 75% year-on-year increase. Converted, this equals roughly RMB 750 million burned every single day. Tencent ramped up investment in tandem; its Q2 capital expenditure climbed to RMB 52.8 billion, nearly triple the figure for the same period a year prior. Combined, Alibaba and Tencent alone posted quarterly capital expenditure exceeding RMB 120 billion.
ByteDance has taken the most aggressive leverage approach. According to Bloomberg, Byte is discussing AI data center investment plans of up to US$70 billion for 2026 (approximately RMB 470 billion), a sum reportedly equivalent to the combined AI spending of Tencent, Alibaba and Baidu.
Overseas players have also entered a high-intensity investment cycle. Google raised its capital expenditure guidance to US$195–205 billion; Meta lifted the lower bound of its 2026 capex guidance from US$125 billion to US$130 billion; OpenAI revised its projected compute spending through 2030 upward to US$750 billion. Based on incomplete statistics, total capital expenditure by the world’s nine major cloud vendors is projected to surpass US$886.7 billion in 2026.
II. Return Rates Have Crossed the Cost of Capital Threshold
Why are giants pouring capital in such concentrated fashion? The answer lies in the return model for compute investment.
Per Alibaba CEO Wu Yongming’s remarks at the earnings call: AI compute capital expenditure is expected to recoup costs within three years, and as gross margins on AI products keep rising, the payback period may shorten to 2.5 years or even around two years.
Return profiles for overseas models are even steeper. The return on invested capital (ROIC) for compute leasing among overseas cloud vendors stands at roughly 28.7%, while domestic cloud vendors reach approximately 13.5% under neutral assumptions. Both figures exceed their respective marginal cost of capital. In other words, this round of capital deployment already carries the capacity for organic cash generation on paper.
X’s supercomputing cluster under Elon Musk serves as an even more extreme case: Claude and Google are among its major clients. Measured against current rental rates, full capital expenditure can be recovered in two years.
For this reason, compute leasing appears to be a compelling business for now. Not long ago, a former real estate investment client asked me what business would suit him well. My recommendation: build a hydropower station, and construct a compute center underneath it.
III. Supply-Demand Inflection Point and Generational Technology Risk
Still, for all the hype around compute leasing, are there hidden risks? Yes. The inevitable outcome of frenzied capital inflow is supply outstripping demand, though estimates for where this demand tipping point lies vary widely across firms. Worse still, demand itself is a rapidly evolving variable.
Signs of strain are already visible in current operational performance: GPU utilization remains low. Average utilization rates at domestic intelligent computing centers sit below 30%. Even after stripping out roughly 20% allocated for redundancy, nearly half of available compute capacity remains idle. Meanwhile, the sector faces an additional layer of technological risk. Should a generational technology breakthrough emerge, existing built clusters will face pressure on leasing revenue. No one can yet offer a definitive answer on whether rental prices will collapse in such a scenario.
IV. Why Top Vendors Still Bet Heavily
Given these clear risks, why do leading large tech firms continue to make massive commitments? The logic is straightforward: betting on the AI track is almost mandatory. Failure to participate means being edged out. Moreover, they have built-in safeguards — they are themselves the largest consumers of compute resources.
Alibaba’s Tongyi, Tencent’s Hunyuan, and Byte’s Doubao consume astronomical volumes of compute daily for model training and inference. For them, self-built compute capacity remains economically viable even if no surplus capacity is leased externally. Leasing to third parties merely delivers incremental revenue, rather than serving as a prerequisite for profitability.
V. Recommendations for Investors
If investing purely in compute infrastructure, my advice is to shift mindset from a construction mindset to an operations mindset. Only in this way can market positioning be preserved amid industry-wide stress.
The operations mindset first means treating intelligent computing as more than hardware stacking. Priority should go to lifting operational efficiency and the coordinated evolution of ecosystem partners. For instance, how to extract maximum value from limited resources and maximize Token output per GPU and per unit power consumption. This first dimension underpins superior service and stronger pricing resilience.
The second dimension involves deep integration into real-world use cases. Instead of supplying bare compute capacity alone, layer in application implementation:
For compute investment today, avoid being merely a “compute middleman”. Bare compute capacity has no moat; users migrate to whichever provider offers lower pricing. True barriers lie in compute plus use cases. Those closest to end users capture the largest profit pools.