← Back to feed

Alphabet is designing a new class of server chip that moves away from general-purpose AI accelerators toward model-specific silicon. The chip, internally called Frozen v2, permanently bakes portions of Gemini’s architecture directly into the hardware layer. The Information reported the project; TechCrunch confirmed the details. Alphabet stock rose on the report.

The claimed efficiency range: 6-10x better than the current generation of Gemini serving infrastructure. That is not a marginal improvement — at 10x, it changes the unit economics of inference at scale.

Why “Frozen” Is the Right Name

Standard AI accelerators, including Google’s own TPU line, are designed to run any model workload. The flexibility is a feature for training and experimentation but it is a cost centre for inference at scale, where the same architecture is executed billions of times per day.

Frozen v2 takes the opposite approach: the architecture of Gemini is not loaded onto the chip at runtime, it is embedded into the physical silicon. Certain layers, attention patterns, or routing structures that are fixed across Gemini inference calls are handled in hardware rather than computed dynamically. The chip becomes a specialized execution engine for one model family.

This is structurally similar to what Google attempted with earlier TPU generations but at a finer level of specificity. The closest public analogy is the work Etched was doing with Sohu — building a Transformer-only chip with no flexibility for other architectures — except Frozen v2 applies the same principle inside a hyperscaler at Google’s model specificity.

What This Means for the Gemini Competitive Position

Google entered 2026 trailing Anthropic on the Intelligence Index by 13 points and OpenAI by 8 on SWE-bench. It has closed that gap substantially with the Gemini 3.x generation, but the cost structure for serving a model as large as Gemini 3 Pro or a future Gemini 3.5 Pro at scale remains a constraint.

A 6-10x efficiency gain on dedicated serving hardware would:

• Cut the per-token cost of serving Gemini by that factor on the chips in question
• Reduce power consumption proportionally, which matters at the multi-gigawatt data centre scale Google is building toward
• Enable inference at margins that standard TPU or GPU infrastructure cannot match

The competitive implication is direct: if Frozen v2 delivers, Google can price Gemini substantially below rivals running on standard accelerators while maintaining better margins. That is the economic leverage that would let a $1.50/M Flash model remain profitable at volume while frontier labs running on H100s are squeezed.

The Tradeoff: Brittleness

Embedding architecture into silicon creates inflexibility. If the model architecture changes significantly — as it has between every Gemini generation — the chip becomes obsolete or requires costly revision. Google is betting that Gemini’s architectural core is stable enough to lock in hardware.

This also means Frozen v2 cannot serve other labs’ models. For a hyperscaler that rents compute to Meta and Anthropic, a model-specific chip narrows the use case. Google may be willing to accept that tradeoff to secure cost advantage for its own inference workloads.

No release timeline was disclosed. Alphabet did not comment.