Alphabet Bets $185B on AI Infra

This level of capital expenditure more than doubles Alphabet's 2025 spending and is greater than the company's combined capex for the three previous years. The enormous outlay is intended to expand capacity for Google's internal AI development, including its Gemini models, and to meet what the company calls "significant cloud customer demand" for AI services. A key part of this investment is in Google's custom-designed Tensor Processing Units (TPUs), now in their seventh generation with "Ironwood". Unlike general-purpose GPUs from competitors like Nvidia, TPUs are ASICs (Application-Specific Integrated Circuits) built specifically for the matrix and vector math central to AI workloads, offering a different balance of performance and cost-efficiency, particularly for inference tasks. For large-scale users, TPUs can deliver a 20-50% lower total cost per useful FLOP compared to top-tier Nvidia GPUs. This custom silicon strategy extends to networking with Google's "Jupiter" data center fabric, which can scale to over 13 petabits per second. The architecture uses optical circuit switching (OCS) to dynamically reconfigure network connections, which has been shown to improve throughput by 30% and reduce power consumption by 40% compared to alternatives. This allows for massive clusters of TPUs, with a single pod connecting up to 9,216 chips, to operate efficiently. For platforms handling heavy video processing, this infrastructure build-out directly impacts the cost and scalability of AI-driven features like transcription, summarization, and text-to-video generation. The cost of AI video generation can range from $0.50 to $30 per minute, a significant reduction from traditional production costs. As newsrooms increasingly adopt these tools to create content for multiple platforms, the underlying infrastructure's efficiency becomes a critical competitive advantage.