Google's 'Frozen v2' chip targets a 6-10x efficiency leap for Gemini, just days before earnings

Google's "Frozen v2" Chip Targets 6-10x Efficiency Leap for Gemini, Days Before Earnings

A leaked internal Google project called Frozen v2 could make the company's AI servers six to ten times more efficient by a measure that matters more than raw speed: tokens generated per watt of power. Slated for 2028, the chip was first reported by The Information on Monday 1.

If Google delivers even the low end of that range, the cost of running Gemini on Google's own infrastructure could fall by roughly the same factor. Inference, the computing work that happens every time a user prompts a model, is where ongoing AI costs accumulate. A sixfold efficiency improvement at the silicon level flows straight into product margins, pricing, and the unit economics of everything from API calls to consumer subscriptions. For context on why investors cared immediately: Alphabet's stock rose about 3% on Monday morning after the report 1.

The timing is no accident. Google's earnings report lands later this week 1. Investors have been anxious about the company's spending; earlier this year, Google said it plans to spend between $180 billion and $190 billion building out its AI strategy 1. A chip that promises to make that investment stretch further is exactly the signal the market wanted.

Google declined to confirm or deny the report. Its statement emphasized continuous research and "co-designing our hardware and software from the ground up" 1. That non-denial is telling. In a cycle where every frontier lab is racing to design custom silicon, confirming the project would hand competitors a roadmap.

Frozen v2 is not happening in isolation. The same reporting cycle captured OpenAI announcing its first custom inference chip, called Jalapeño, in June 1. Earlier this month, Anthropic was reported to be in discussions with Samsung about a chipmaking partnership 1. Every major frontier lab is now designing or commissioning its own silicon. The goal has shifted from catching Nvidia to reducing dependence on Nvidia entirely.

Google already builds its own AI chips. The TPU line, which powers Gemini today, gives Google a head start over labs still shopping from Nvidia's catalog 1. Frozen v2 would extend that lead, and a 6-10x efficiency jump over Google's current generation would put real distance between Google's internal compute costs and what any lab paying Nvidia's margins can achieve.

For Nvidia, the question is whether its data center revenue trajectory already prices in this decoupling. Custom silicon does not eliminate GPU demand overnight; training the next generation of models still requires the massive parallel computing power that Nvidia's architecture handles well. But inference, where most serving costs pile up, is exactly the workload that custom chips are built to run more cheaply. If three of the largest AI labs are simultaneously building inference alternatives, the volume of Nvidia GPUs purchased for serving models could plateau well before the training side does.

The downstream effect on AI products themselves could outpace the effect on chip stocks. If inference costs drop six to ten times, every assumption about what an AI feature costs to operate gets rewritten. Subscription prices, API rates, free-tier economics, and the break-even point for AI features inside non-AI products all shift. Companies that have been holding back features because per-user compute costs were too high would find the math suddenly workable. The efficiency gain does not just improve Google's margins; it changes what is economically viable to build across the industry.

Frozen v2 is a 2028 target, and two years is a long time in a field where model architectures and hardware requirements evolve every few months. The Information's report cites anonymous sources, and Google's own statement noted that "not every project moves into production" 1. The 6-10x figure is a design aspiration, not a benchmarked result.

But the market's reaction, the timing, and the competitive context all point the same way. Custom silicon is no longer a hedge; it is the primary compute strategy for every frontier lab. The labs that build the most efficient inference chips will run the cheapest models, serve the most users, and hold the strongest margins. Google just told the market it intends to be that lab. Whether it can execute by 2028 will be one of the defining questions of this compute cycle.