The explosive growth of AI compute demand is running headlong into four structural supply chain bottlenecks that will constrain the industry for years, according to a detailed analysis from Semiconductor Engineering.
The piece identifies advanced foundry and packaging capacity, memory supply (particularly HBM), data center power availability, and lasers for optical interconnect as the four critical choke points.
The scale of demand is staggering. Google Gemini alone processes more than 3.2 quadrillion tokens per month, up 7 times year-over-year. TSMC’s revenues are expected to quadruple between 2023 and 2028, and the global semiconductor market is forecast to reach US$1.5 trillion (approximately GBP 1.2 trillion) by 2030, with AI accounting for 55 percent of that.
“It will be a long time before we can meet customer demand,” TSMC CEO CC Wei told Bloomberg.
Bottleneck 1: Advanced foundry and packaging
TSMC holds an effective monopoly on leading-edge AI chip manufacturing thanks to its advanced process nodes and its CoWoS (Chip-on-Wafer-on-Substrate) packaging technology. CoWoS capacity has been growing at 80 percent compound annual growth (2022-2027), but AI accelerator wafer demand is growing more than 11 times over the same period, an arithmetic that guarantees persistent shortfall.
TSMC has outsourced some CoWoS work to ASE and Amkor and is developing a new CoPoS (Chip-on-Panel-on-Substrate) approach using glass core substrates designed to cut costs. But these are medium-term solutions.
Alternative foundries offer limited relief. Intel and Samsung have advanced nodes but lack comparable packaging maturity. SMIC is on trailing-edge nodes with poor power efficiency. GlobalFoundries tops out at 12 nanometers, insufficient for data center AI.
Meanwhile, ABF (Ajinomoto Build-up Film), a substrate material for which one company supplies more than 95 percent of the global market, saw prices rise 30 percent in 2026, with a supply gap of more than 20 percent projected for 2027.
Bottleneck 2: Memory, especially HBM
DRAM supply is dominated by three companies (SK hynix, Samsung, Micron), all now valued above US$1 trillion. High-Bandwidth Memory, which requires the complex stacking of 8 to 16 DRAM dies per module, is 80 percent manufactured in South Korea.
DRAM companies, long accustomed to boom-bust cycles, are now raising prices and signing long-term strategic supply agreements, such as the Micron-Anthropic deal. A new JEDEC HBM standard using standard packages and glass substrates to bypass the ABF shortage is under development but years away from production.
Chinese suppliers such as CXMT (US$8 billion revenue in 2025) and Yangtze Memory (building three new factories) are expanding but lag significantly on specifications.
Bottleneck 3: Power and data center infrastructure
Amazon’s CEO has named power as the number one constraint on AI data center buildout. Electrical grids in most regions cannot add capacity fast enough. Community resistance is growing: a Texas Tribune poll found most Texans oppose new data centers in their areas.
Hyperscalers are pursuing creative workarounds. Tesla, Sunrun, and Renew Home (a Google spin-off) plan to tap residential home batteries to supply 17 large data centers during peak demand. Microsoft signed a 20-year deal with Chevron via Joulent for a 2.7-gigawatt gas plant in the Permian Basin.
Other infrastructure is tight too: transformers and high-voltage breakers are in short supply. GE Vernova, the leading gas turbine manufacturer, is sold out through 2029. Solar-plus-battery configurations remain uneconomical without natural gas backup due to winter worst-case scenarios.
Bottleneck 4: Lasers for optical interconnect
The least immediate bottleneck but the fastest-growing. Lasers are required for scale-out links (pluggable transceivers) and future scale-up links (co-packaged optics), with the latter requiring 10 to 100 times more lasers per system.
The top three laser suppliers, Coherent, Lumentum, and Sumitomo, hold a combined 68 percent market share. All are sold out and require upfront cash for capacity expansion. Nvidia invested US$2 billion in each of Coherent and Lumentum in March 2026. Coherent has doubled its indium phosphide output in 2026 and plans to more than double it again in 2027.
This bottleneck is considered the most manageable because indium phosphide fabrication requires less capital and shorter lead times than leading-edge logic or memory fabs.
Geopolitical overlay
Two of the four bottlenecks have a geographic concentration risk. TSMC’s leading-edge capacity is concentrated in Taiwan, within proximity to China. South Korea produces 80 percent of the world’s HBM, and all of its fabs and packaging facilities are within range of North Korean missiles.
“The supply chain is responding, but these bottlenecks are likely to be with us for years,” Tate concluded.
Sources: Data Center AI Growth Faces Challenging Bottlenecks (Semiconductor Engineering, July 2026)