SemiAnalysis and Bernstein analysts say chip substrates are the primary bottleneck in AI infrastructure
SemiAnalysis CEO Dylan Patel said at CASPA's 35th annual conference that the biggest supply chain bottleneck in AI infrastructure is not power but chip substrates, and that cautious capacity expansion by Japanese suppliers makes it difficult to rapidly increase key materials. Bernstein senior analyst Stacy Rasgon expects semiconductor supply chain tightness to potentially persist until 2028. Patel also estimated that the cost of building 1 gigawatt of computing infrastructure has risen from about $40 billion in 2024 to more than $60 billion today, an increase of over 50% in two years; meanwhile, Nvidia's share of CoWoS and HBM resources has begun to decline and is expected to fall further in 2027.
Why read this
AI score 62
The two analysts, speaking at an industry conference, offered a specific judgment that the AI infrastructure bottleneck has shifted from power to chip substrates, along with cost and capacity data.
Topics
Body · AI translation
On October 28, news emerged that SemiAnalysis CEO Dylan Patel pointed out that chip substrates are the biggest bottleneck in artificial intelligence (AI) infrastructure, while Bernstein senior analyst Stacy Rasgon noted that semiconductor supply chain tightness could extend to 2028.
On the other hand, Google and Amazon are accelerating their in-house chip development, and Nvidia's share of HBM and CoWoS resources is declining, putting its future high profit margins to the test.
On October 23, Patel and Rasgon discussed whether AI investment can be sustained, semiconductor supply chain bottlenecks, and Nvidia's competitive advantages at the 35th annual CASPA conference held in Santa Clara, USA.
Patel pointed out that the biggest supply chain bottleneck currently facing AI infrastructure construction is not power, but chip substrates. Because relevant Japanese suppliers are cautious about expanding production, it is difficult to rapidly increase the supply of key materials.
Rasgon expects that the semiconductor supply chain tightness may continue until 2028, while the supply and demand situation in 2029 will depend on changes in market demand.
On the other hand, Nvidia's share of global consumption of CoWoS advanced packaging and HBM resources has begun to decline, and is expected to fall further in 2027.
As tech giants such as Google and Amazon accelerate their in-house chip development efforts, whether Nvidia's future market share and high profit margins can be sustained will also face new tests.
Patel pointed out that chip substrates are currently the most serious bottleneck in the AI infrastructure supply chain, with an impact even greater than that of land, power, and semiconductor manufacturing equipment.
Because chip substrates and related materials have suffered from long-term underinvestment, and Japanese suppliers are cautious about expanding production, capacity expansion for AI accelerators and CPUs has been constrained. In addition, memory chips and optical interconnect components are also facing tight supply.
By contrast, power shortages still have other solutions. SemiAnalysis estimates that by 2027, the scale of data centers under construction globally will approach 50 gigawatts (GW), of which the United States will account for about 35 gigawatts. Industry players can partially alleviate power supply constraints by deploying on-site gas turbines, fuel cells, and other equipment; however, expanding the semiconductor supply chain requires more time, making it difficult to rapidly increase supply.
Rising AI infrastructure construction costs are also putting pressure on industry expansion. Patel estimates that the total cost of building 1 gigawatt of computing infrastructure in 2024 was about US$40 billion, and has now exceeded US$60 billion, an increase of more than 50% in two years.