Alphabet is designing a new server chip, internally called Frozen v2, that aims to run Gemini models between six and 10 times more efficiently than Google's current AI silicon, measured in tokens generated per unit of power. The chip is slated for a 2028 release, according to a report from The Information citing anonymous sources. Alphabet's stock climbed roughly 3% on Monday morning after the news broke, ahead of the company's earnings report later this week.
The efficiency target matters because Google has told investors it plans to spend between $180 billion and $190 billion this year on AI infrastructure, a figure that has repeatedly rattled the market. A chip that generates six to 10 times more output per watt would meaningfully change the unit economics of serving Gemini at scale, both for consumer products and for Google Cloud customers renting inference capacity.
Google did not confirm the report to TechCrunch, but it did not deny it either. The company issued a statement pointing to its ongoing hardware-software co-design work without addressing Frozen v2 by name.
Key facts
- 01Frozen v2, Google's next in-house AI server chip, is slated for a 2028 release.
- 02The chip is designed to be 6 to 10 times more efficient than Google's current AI silicon, measured in tokens per unit of power.
- 03Alphabet's stock climbed roughly 3% on Monday morning after the report surfaced.
- 04Google has committed $180B to $190B in AI-related capital expenditures this year.
- 05OpenAI unveiled its first custom inference chip, Jalapeño, in June; Anthropic is in talks with Samsung on a chipmaking partnership.
"Our teams are constantly researching and experimenting with new innovations to deliver maximum performance and efficiency for our users and customers," Google said. "While not every project moves into production, this rigorous exploration is central to our full stack approach. By co-designing our hardware and software from the ground up, we ensure our systems are integrated and highly optimized for real-world workloads."
“Our teams are constantly researching and experimenting with new innovations to deliver maximum performance and efficiency for our users and customers.”— Google, company statement
Frozen v2 would extend Google's decade-plus lead in custom AI silicon. The company has been shipping tensor processing units internally since 2015 and has used them to train and serve Gemini alongside Nvidia GPUs. The new chip is the latest attempt to push more workloads off Nvidia hardware, whose margins currently sit on the balance sheet of every major AI buyer.
The rest of the industry is chasing the same goal. In June, OpenAI announced its first custom chip, an inference processor called Jalapeño. Earlier this month, Anthropic was reported to be discussing a new chipmaking partnership with Samsung. Every frontier lab now sees vertical integration into silicon as a requirement, not an option — a shift from the 2023 era when renting Nvidia H100s was the entire strategy.
The competitive pressure is compounded by supply. Global AI compute capacity remains constrained, and firms that own their own chip designs get earlier and cheaper access to fab capacity than those bidding on the open market. For Google, which is trying to serve Gemini across search, Workspace, YouTube, and Cloud simultaneously, in-house silicon is the only way the math works at the volumes it needs.
Efficiency claims made three years out are worth reading with some skepticism. Frozen v2 is a 2028 product, and the six-to-ten-times figure is measured against Google's current chips rather than whatever Nvidia or AMD will be selling by then. Chip roadmaps slip; competitors' roadmaps also advance. The report is sourced to anonymous individuals, and Google's statement stops well short of confirming the specifications or the timeline.
Still, the market reaction — a 3% jump on Monday morning — shows how tightly Alphabet's stock is now bound to the perception that its $180B-to-$190B capex will translate into durable cost advantages. Investors have been asking for a clear return on that spend since it was announced earlier this year, and a credible efficiency roadmap is one of the few answers Google can give without waiting for Gemini revenue to catch up to the bill. Frozen v2, if it ships on time and hits its numbers, would let Google serve models at a fraction of the per-query cost of a Nvidia-hosted competitor — the kind of structural advantage that reshapes who can afford to run frontier AI at consumer scale.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
