Google Eyes Computing Power, Likely Impact On Ad Serving 11/21/2025

At an all-hands, Google VP Amin Vahdat said the company must double AI computing power every six months — targeting roughly 1,000x capacity in four-to-five years — to meet demand from agentic, compute-intensive advertising systems and called AI infrastructure the most critical and costly front in the AI race. That buildout is driven by the need to run complex ML models and real-time bidding/optimization that adjust bids and placements in milliseconds to improve ROAS and lower CPA, fundamentally changing how ads are created, timed and served. Google plans to pair more efficient models with custom silicon — highlighted by the public launch of its seventh-generation Ironwood TPU, which it says is nearly 30x more power-efficient than its 2018 Cloud TPU and can support models like Gemini 3 — and presented the plan shortly after Alphabet’s better-than-expected Q3 results.

Analysis

At an all-hands meeting Google VP Amin Vahdat said the company must double AI computing power every six months and presented a slide targeting roughly 1,000x capacity growth over four to five years to meet demand from agentic, compute-intensive advertising systems; the presentation followed Alphabet’s better-than-expected third-quarter results. The explicit linkage between compute scale and advertising — faster ad creation, sub-millisecond bidding adjustments, and agentic purchasing — frames this buildout as strategic to monetization through improved ROAS and lower CPA.

Vahdat called AI infrastructure "the most critical and also the most expensive" front in the AI race and emphasized custom silicon and more efficient models as levers. Google’s public launch of the seventh‑generation Ironwood TPU, described as nearly 30x more power-efficient than its 2018 Cloud TPU and able to support models such as Gemini 3, illustrates the company’s path to efficiency gains.

Operationally, delivering real-time bidding and optimization at scale requires both sustained capital investment and engineering execution; successful scale could widen Google’s ad-tech moat by improving performance and lowering per-conversion costs. The plan’s feasibility effectively replaces a traditional Moore’s Law cadence with an aggressive compute-scaling target, so investors should treat execution, capex trajectory, and measured ROAS improvements as primary signals of commercial success.

Key risks include the pace of infrastructure deployment, potential margin pressure from elevated capex, and the challenge of translating raw compute into sustained advertiser ROI; absent clear metrics on TPU deployment and ad-performance gains, the timeline and commercial payoff remain uncertain. Monitor published capex guidance, TPU adoption, and empirical ROAS/CPA trends to validate the thesis.

Key Decisions for Investors

• Consider a selective overweight in Alphabet (GOOGL/GOOG) if you believe Ironwood and custom silicon will sustainably reduce model costs and materially improve ad ROAS, but size positions to reflect significant execution risk
• Monitor Alphabet’s capex guidance, TPU deployment cadence, and public metrics on model latency, ROAS and CPA as near-term indicators; trim exposure if capex rises without commensurate monetization gains
• Watch procurement and demand signals from semiconductor suppliers (including Intel) for increased AI-infrastructure bookings as a secondary way to play rising compute demand, but avoid assuming benefits are confined to any single vendor until contracts or procurement details are disclosed
• Mitigate risk through position sizing or hedges given the plan’s capital intensity and aggressive six-month doubling cadence; reassess as empirical deployment and ad-performance data arrive