What happened
Two outlets reportedly describe Google as building a custom server chip called Frozen v2. The reported design would bake the Gemini architecture directly into silicon, with a 2028 target and claimed inference-efficiency gains of 6–10 times.
If accurate, the project would be part of a broader shift toward custom AI silicon. The supplied record also points to Microsoft’s move to AMD Helios, with Anthropic reportedly testing AMD, as evidence of pressure to reduce reliance on Nvidia’s pricing power.
Why it matters
Inference efficiency can shape the cost of operating AI systems. A Gemini-specific chip with the reported gains could alter those economics and add to the move toward more specialized hardware and less concentrated vendor dependence.
The most useful receipt to watch is not the reported target date alone. Watch for Google confirmation, technical specifications, benchmark methodology, or measured inference results that establish what Frozen v2 is designed to do and whether the stated gains are comparable.
Watch for a Google statement or technical release that confirms Frozen v2 and provides comparable, measured inference-efficiency results.
Upstream references
Digest dated 2026-07-21 · upstream model claude-sonnet-4-6. Source IDs are preserved for audit; the publishing host does not receive the upstream URL map.
- 1
9b0dbd4007ed550271bc5c09a39278b7ecf4321eReference from the upstream research server - 2
a1e3af7102d2e7d811a19098da994bc03f98cac0Reference from the upstream research server
This Research brief was generated by Terra from a dated upstream research digest. It has not received the source-by-source human review required for Reviewed analysis. Material limit: This brief relies on two reports and unconfirmed internal sourcing; the chip’s existence, 2028 schedule, architecture and claimed 6–10 times efficiency gains are not independently established in the supplied record.