科技播客Dwarkesh Patel近日发文警示，若Anthropic等AI实验室保持当前增长势头，明年底年化收入可能达到1000亿至1500亿美元水平，但行业算力供应增速远未匹配这一需求[1]。Patel指出，领先AI实验室的物理算力扩张每年仅增长约3倍，远低于收入增速，这种失衡可能导致计算力价格上涨10倍以上[1]。
目前算力定价差异巨大。单张H100显卡现货租赁年化价格仅1.6万美元，而谷歌约11万张GPU集群每月租金达9亿美元，单张GPU每小时租金约为现货租赁单价的两倍[1]。相比之下，硅谷软件工程师平均年薪25万美元以上，与H100显卡年化租赁价格相差15倍[1]。
算力供应面临多重制约。到2027年底，AI芯片将占据台积电N3产能的86%，从目前的60%上升至此水平，达到物理意义上的绝对饱和[1]。Patel分析认为，摩尔定律、新晶圆厂建设、抢占先进制程产能这三个技术与产业乘数均已逼近天花板[1]。
Tech podcaster Dwarkesh Patel has warned that compute prices could surge tenfold or more if artificial intelligence labs like Anthropic maintain their current growth trajectories while hardware supply fails to keep pace [1]. According to Patel's analysis, if leading AI laboratories continue expanding at their recent rates, Anthropic's annualized revenue could reach between $100 billion and $150 billion by year-end, yet the industry's physical compute capacity is growing at only approximately 3x annually [1]. This fundamental mismatch between demand acceleration and supply constraints threatens to create a severe computational bottleneck for the sector.
The strain on GPU availability is already evident in current pricing disparities. A single H100 graphics card rents for approximately $16,000 per year on the spot market, while Silicon Valley software engineers earn average annual salaries exceeding $250,000 [1]. Google's cluster of roughly 110,000 GPUs costs approximately $900 million monthly to lease, with per-chip hourly rates reaching roughly double the spot rental price [1]. Beyond immediate pricing pressures, the semiconductor industry faces structural constraints that threaten to exacerbate shortages further. By the end of 2027, AI chips are projected to consume 86 percent of Taiwan Semiconductor Manufacturing Company's N3 process capacity—up from the current 60 percent—representing near-complete physical saturation [1]. With Moore's Law, new fab construction, and access to advanced manufacturing nodes all approaching their practical limits, the industry lacks sufficient multiplier effects to bridge the growing compute supply-demand gap [1].
