Alphabet Is Reportedly Baking Gemini Directly Into Silicon
A new chip called "Frozen v2" would hardwire parts of Gemini's architecture into the hardware itself, betting efficiency gains are worth the loss of flexibility.
Alphabet shares rose 3% on Monday after The Information reported that Google is building a new AI server chip designed specifically to run its Gemini models, a project internally known as "Frozen v2."
The idea behind Frozen v2 is unusual: rather than building another general-purpose AI accelerator, Google would permanently etch parts of Gemini's architecture into the chip's silicon. That would cut down on the number of calculations the chip needs to perform and reduce how much data has to move around to answer a query, according to The Information's reporting.
The payoff, according to Google's own internal projections cited in the report, is significant. Engineers estimate the chip could handle six to ten times more tokens (the chunks of text AI models process) per unit of power compared with Google's current-generation TPUs, or tensor processing units, the custom chips that power much of its AI infrastructure today.
That kind of efficiency jump matters because power, not just chip supply, has become one of the biggest bottlenecks in AI infrastructure. Data centers running large models are increasingly constrained by how much electricity they can pull and how much heat they can dissipate, so a chip that does more work per watt is valuable even before you factor in raw compute gains.
A specialist, not a replacement
Frozen wouldn't replace Google's TPU lineup. Instead, The Information describes it as a more specialized branch sitting alongside Google's general-purpose chips, reserved for the specific job of running Gemini as efficiently as possible. Google is reportedly targeting 2028 for deployment, and the company currently views the project partly as a trial run rather than something it intends to manufacture at TPU-scale volumes.
That caution points to the chip's central limitation. Because Frozen v2 hardwires elements of Gemini's architecture directly into the chip, it would only remain useful if Google keeps that underlying architecture largely intact in future Gemini versions. AI model design moves fast, and architectures change. A chip this tightly coupled to one model family is a bet that Gemini's core structure will hold steady long enough to justify the specialization, according to The Information.
Why Google needs this now
The timing lines up with a compute crunch that has been building inside Google for a while. The report notes that internal shortages of AI computing capacity have created friction inside the company and reportedly pushed Google Cloud to turn away business from outside customers, an unusual move for a division competing directly with Amazon and Microsoft for enterprise cloud dollars.
Google has already been paying to close that gap in the near term. Just last month, the company agreed to pay SpaceX close to $1 billion a month for help meeting its enterprise compute commitments, according to the report. That kind of spending underscores how tight capacity has gotten, and why a chip promising several times the efficiency of current hardware would be attractive even with a 2028 timeline attached to it.
Alphabet did not immediately respond to a request for comment on the report. The stock move on Monday suggests investors read the news as a sign the company has a credible plan for easing its compute constraints and controlling costs on the infrastructure side of its AI buildout, rather than simply spending its way out of the shortage.
What happens next depends on how the tradeoff plays out. If Frozen v2 delivers anywhere near the efficiency gains Google's engineers are projecting, it could meaningfully lower the cost of running Gemini at scale. But locking hardware to a specific model architecture is a gamble, and Google's own apparent caution, treating this first version as a trial rather than a mass-production chip, suggests the company knows it.
