A US district judge gave final approval Monday to Anthropic's landmark $1.5 billion settlement of a class-action copyright lawsuit, closing one chapter in the AI industry's long-running legal battle over training data. But the bigger story this week isn't a courtroom — it's a chip lab in Mountain View.
Google parent Alphabet is designing a new server chip called "Frozen v2" that could make its Gemini models six to 10 times more energy-efficient than current hardware, as first reported by The Information. The chip, slated for a 2028 release, is part of a broader push by Google to reduce the astronomical power costs of running large language models at scale.
"We are constantly researching and experimenting with new innovations to deliver maximum performance and efficiency," Google told TechCrunch. The company emphasized its "full stack approach" of co-designing hardware and software from the ground up.
The numbers explain the urgency. Google has said it plans to spend between $180 billion and $190 billion on its AI buildout this year. With that much capital at stake, investors have been watching closely for signs that those dollars will translate into real returns. News of Frozen v2's efficiency targets sent Alphabet's stock up roughly 3% on Monday ahead of the company's earnings report later this week.
The Efficiency Arms Race
Frozen v2 isn't Google's first custom silicon for AI. The company already ships Tensor Processing Units (TPUs) now in their fifth generation, purpose-built accelerators that power both Gemini training and inference. But the new chip targets a specific pain point: the cost-per-token of serving AI models to billions of users.
"Everybody is chasing the same physics problem," said Dylan Patel, chief analyst at SemiAnalysis, a semiconductor research firm. "Training gets all the headlines, but inference — actually running these models — that's where the real money gets burned. If Google can serve Gemini at a fraction of the current energy cost, they flip the economics of the entire AI business."
The efficiency push comes as Meta, OpenAI, and Anthropic compete to offer the cheapest API pricing. Anthropic recently cut Claude API prices across several tiers, and OpenAI's GPT-5.6 family introduced tiered pricing aimed at undercutting rivals. Hardware efficiency directly determines how low these companies can go without bleeding cash.
Beyond Google: Who Else Is Building AI Chips
Google's custom silicon strategy puts it in an unusual position relative to other AI labs. OpenAI relies heavily on Microsoft's Azure cloud and Nvidia GPUs. Anthropic has partnered with Amazon and Google Cloud. Meta designs its own accelerators but still depends on Nvidia for the bulk of its compute.
"Nvidia still dominates, but the challengers are getting serious," said Patel. "Google has the unique advantage of controlling both the model and the silicon. That integration lets them optimize in ways that Nvidia's customers can't."
The chip could be between six and 10 times more efficient than Google's existing AI chips, measured by tokens generated per watt. If those numbers hold up in production, it would represent a generational leap in efficiency. Most chip generations improve by 30-50% in perf-per-watt.
Google isn't alone in chasing better AI economics. Amazon's Trainium and Inferentia chips power parts of its AWS AI services. Microsoft is rumored to be developing its own AI accelerators under the "Athena" project. Meta has deployed its MTIA chips for recommendation systems and is working on a new generation for generative AI workloads.
What Frozen v2 Means for Gemini Users
For the end user, better chip efficiency could mean lower prices, faster responses, or both. Google has already integrated Gemini into Search, Workspace, Android, and Cloud — making it one of the most widely deployed AI models in the world. Every millisecond and every milliwatt matters at that scale.
"Google's position is actually quite strong here," said Chirag Dekate, an analyst at Gartner. "They have the search revenue to fund massive capital expenditure, they have the engineering talent to design custom silicon, and they have the distribution to put AI in front of billions of users. Frozen v2, if it delivers, reinforces that flywheel."
The chip is expected to tape out in 2027 with production silicon arriving in 2028. That timeline puts it in the same window as Nvidia's next-generation "Rubin" architecture and Intel's Falcon Shores successor. The AI chip race is far from over — it's only getting started.
The Copyright Ruling That Still Matters
While the chip news dominated Monday, the Anthropic settlement approval also carries weight for the AI industry. Judge William Alsup of the U.S. District Court for the Northern District of California gave final approval to the $1.5 billion deal, which resolves claims that Anthropic illegally downloaded millions of copyrighted books.
Alsup had previously ruled that training AI on copyrighted text qualifies as fair use — widely seen as a win for the industry. But Anthropic's decision to settle rather than litigate means the fair use question will never reach a federal appeals court. Reuters reported that the settlement closes this case without creating binding precedent, leaving the broader legal questions unanswered.
The Bigger Picture
The twin stories of Google's Frozen v2 chip and the Anthropic settlement illustrate two sides of the same coin. The AI industry is simultaneously racing to build more efficient hardware while navigating legal and regulatory frameworks that have yet to catch up with the technology's rapid pace.
For Alphabet, the bet on custom silicon represents a vote of confidence in its ability to compete with Nvidia on its own terms. Unlike most of its rivals, Google can design, manufacture, and deploy its own accelerators, giving it end-to-end control over the AI stack. If Frozen v2 delivers on its efficiency promises, it could give Google a meaningful cost advantage in the race to serve AI at global scale.
The broader trend is clear: the AI industry is moving from a phase of "build anything" to "build efficiently." The companies that master the economics of inference — not just the quality of their models — will be the ones that survive the shakeout. With Frozen v2, Google is placing a very expensive bet that hardware integration is the path forward. And for now, investors seem to agree.
For more on the latest AI developments, visit our AI category page.