Alphabet raised its 2025 capex guidance to $91-93 billion while Amazon increased its outlook to $125 billion, creating combined infrastructure spending exceeding $216 billion. Both companies direct the majority toward AI data centers and custom chip procurement.
Anthropic agreed to deploy 1 million Trainium2 chips from Amazon's custom silicon program. OpenAI contracted $250 billion in Azure cloud services from Microsoft, requiring massive GPU and accelerator deployments. These commitments translate to direct semiconductor demand across training and inference workloads.
Micron Technology responded with a $24 billion Singapore investment to expand memory-chip capacity. High-bandwidth memory (HBM) demand from AI accelerators drives the expansion, as each training cluster requires terabytes of fast memory.
TSMC, NVIDIA, and AMD face supply pressure as hyperscaler orders stack up. TSMC's 3nm and 2nm process nodes run at capacity, with lead times extending to 6-9 months for advanced packaging. NVIDIA's Blackwell architecture and AMD's MI300 series compete for the same foundry allocation.
Custom silicon programs complicate the supply picture. Google's TPU v5, Amazon's Trainium and Inferentia, and Microsoft's Maia chips reduce dependence on NVIDIA but require separate wafer allocation. Each hyperscaler builds parallel supply chains, multiplying total semiconductor demand.
Memory makers Samsung, SK Hynix, and Micron expand HBM production lines to meet 2026 delivery schedules. HBM3E capacity sold out through Q3 2025, with hyperscalers securing long-term supply agreements at premium pricing.
The investment thesis centers on sustained demand through 2026-2027. Hyperscaler capex plans extend multi-year, creating visibility for semiconductor manufacturers. TSMC announced capacity additions in Arizona and Taiwan, targeting AI chip production. Samsung allocated $228 billion through 2030 for foundry and memory expansion.
Investor focus shifts to execution risk. Delivery delays, yield issues, or packaging bottlenecks could constrain hyperscaler buildouts. Advanced packaging capacity remains the critical constraint, as CoWoS and similar technologies require specialized equipment with 18-month lead times.
Q1-Q2 2026 earnings calls will reveal actual deployment rates versus announced commitments. Semiconductor equipment orders from ASML, Applied Materials, and Lam Research provide leading indicators of sustained capacity expansion.