Alphabet Inc Class CDeveloping Frozen v2 chip to hardwire Gemini architecture, targeting 6-10x tokens per watt, easing compute constraints and unlocking cloud backlog.
Alphabet is developing a new AI chip called Frozen v2 that hardwires the architecture of its Gemini model directly into silicon, targeting six to 10 times more tokens per watt than current TPUs by 2028. The specialized server chip is designed to ease compute constraints that have forced Google Cloud to turn away capacity requests from major customers, including Meta, despite a backlog that has swelled to $462 billion—more than 10 times its 2025 full-year cloud revenue of $43.2 billion. Google Cloud's revenue jumped 67% year-over-year to roughly $22.8 billion in Q2 2026, but CEO Sundar Pichai noted that revenue would have been higher without compute limitations. Alphabet has raised its full-year 2026 capital expenditure guidance to between $180 billion and $190 billion to expand infrastructure, while the cloud division's operating margin has already improved from 17.8% to 32.9% as utilization rises. Frozen v2 is expected to deploy in 2028 as a complement to existing TPUs, potentially unlocking meaningful additional capacity without proportionally massive new builds.
Alphabet Inc Class CDeveloping Frozen v2 chip to hardwire Gemini architecture, targeting 6-10x tokens per watt, easing compute constraints and unlocking cloud backlog.
NVIDIA CorporationGoogle's custom Frozen v2 chip could reduce reliance on NVIDIA GPUs for AI inference, posing competitive threat.
Meta Platforms Inc.