Tech Giants Plan $680B Capex for AI Arms Race

- The planned 2026 capital expenditures for key players in the AI arms race include Amazon at around $200 billion, Alphabet at up to $185 billion, Microsoft projecting a pace of $144 billion, and Meta earmarking up to $135 billion. This level of spending by just four of the "Magnificent Seven" is comparable to 2.1% of the U.S. GDP, a larger share than the construction of the U.S. railroad system between 1850 and 1859. - Oracle's "Stargate" project, in partnership with OpenAI and SoftBank, is a multi-year, $500 billion initiative aiming to build out 10 gigawatts of AI data center capacity in the U.S. The recent 4.5-gigawatt expansion with Oracle is designed to house over 2 million AI chips and is estimated to create over 100,000 jobs in construction and operations. For scale, 4.5 gigawatts is enough to power approximately 3.4 million U.S. homes. - The massive power requirements for these AI data centers are a critical factor, with global demand projected to reach 68 gigawatts by 2027, nearly equivalent to the entire power grid of California. An AI-optimized server rack can consume 40-100+ kilowatts, a significant increase from the 5-15 kilowatts used by traditional server racks. - To reduce reliance on Nvidia, which holds a dominant market share, major cloud providers are investing in their own custom silicon. Microsoft's new Maia 200 AI accelerator, built on a 3nm process, is designed for inference and reportedly offers 30% better performance per dollar than their current hardware. Microsoft claims the Maia 200 has three times the FP4 performance of Amazon's third-generation Trainium chip and superior FP8 performance over Google's seventh-generation TPU. - Amazon's custom silicon efforts include the Trainium and Inferentia chips, developed by its Annapurna Labs subsidiary, which are designed to provide cost-effective alternatives for training and inference workloads on AWS. - Nvidia is accelerating its own roadmap to maintain its lead, moving to a one-year release cycle. Their upcoming "Rubin" architecture will succeed the "Blackwell" B200 and is expected to feature a new GPU, a new ARM-based CPU called "Vera", and HBM4 memory. The Rubin platform is also slated to use a 3nm process and will introduce NVLink 6 for faster interconnect speeds. - The Blackwell B200 GPU, a key component of the current spending boom, packs 208 billion transistors and offers up to 18 PFLOPS of sparse FP4 performance. The rack-scale NVL72 configuration, which combines 72 Blackwell GPUs, is liquid-cooled and can consume over 132 kilowatts per rack. - This massive capital outlay is impacting the financials of these tech giants, with free cash flow being closely watched by investors. For instance, Amazon is reinvesting nearly 90% of its operating cash flow back into capital expenditures, largely for AWS and data center expansion. This has led companies to tap into debt markets to finance their ambitious AI infrastructure projects.