The Information report on Google’s internal “Frozen v2” chip, amplified by Reuters and Bloomberg Law. Alphabet shares rose as much as 3.7%. Google has not confirmed the project.
Frozen v2 would bake Gemini’s neural-network architecture into circuitry while keeping weights updateable. Claimed efficiency: 6–10× tokens per watt vs latest custom AI chips. Parallel to TPUs, not a replacement. Deployment targeted as early as 2028. Motivated by internal AI capacity crunch; Google Cloud reportedly turned away some external customers.
Google spokesperson: teams experiment with high-efficiency ideas; not every lab project reaches production. Rigidity risk if Gemini architecture drifts before 2028. Compared to startups like Taalas hardwiring models into silicon.