Google is developing a new artificial intelligence chip designed to run its Gemini models far more efficiently, according to a report first published by tech publication The Information. The effort — which The Information said is internally known by the codename "Frozen" — sent Alphabet shares noticeably higher as investors bet that cheaper, faster inference could protect Google's margins as AI computing costs climb.
The report was rapidly confirmed and amplified across financial media. Reuters reported that "Google plans new chip to run Gemini models more efficiently," citing The Information, while CNBC wrote that "Alphabet stock pops on report it's developing a more efficient AI chip." Bloomberg likewise reported that "Google Shares Gain on Report of Chip to Boost AI Efficiency." The consistent market reaction underscores how sensitive investors have become to any sign that a major AI lab can lower the cost of serving models.
Why Efficiency Now Matters More Than Raw Power
The chip push lands at a moment when the economics of AI are under intense scrutiny. As Tom's Hardware noted in a separate analysis the same day, the industry faces "the enduring paradox of the AI economy" — models keep getting better and more efficient, yet the total cost of running them can still spiral out of control as usage scales. The problem is not training alone; it is inference, the ongoing expense of answering every query, which dominates the bill as products reach millions of users.
A chip purpose-built to make Gemini inference cheaper would address exactly that pressure. Google serves Gemini across consumer products like Search and its chatbot as well as through its cloud division, meaning even modest gains in efficiency translate into large savings at scale. Investors clearly read the report as a sign that Google intends to defend that economic position with custom silicon rather than relying solely on commercially available accelerators.
Google's Long Bet on Custom Silicon
The Frozen project fits neatly into Google's well-established strategy of designing its own AI chips. The company has for years built its Tensor Processing Units, or TPUs, which power both internal workloads and are offered to cloud customers. Google has also designed custom silicon for its Pixel phones and has worked to shrink AI models so they can run on consumer devices rather than only in vast data centers.
What reportedly distinguishes the Frozen effort is its explicit focus on efficiency for serving Gemini. While TPUs have historically been optimized for both training and inference, a chip tuned specifically to the cost of answering queries signals where Google believes the next competitive battle will be fought: not just who can build the smartest model, but who can deliver it most cheaply. That logic aligns with a broader industry shift in which inference cost — not peak capability — increasingly decides which products survive.
The move also tightens Google's vertical integration. By owning the model, the software stack, and the silicon beneath it, Google can co-design each layer for the others in ways that buyers of off-the-shelf chips cannot easily match.
The Wider Chip Rivalry
Google is hardly alone in the race. Nvidia continues to dominate the market for the most powerful AI accelerators, and rivals including Amazon, which has pushed its own custom AI chips, and a growing field of startups have all concluded that controlling silicon is essential to competing on cost. Google's report surfaced on a day when the broader AI hardware story was in motion: ASML's plans to raise prices on its advanced lithography machines were said to frustrate TSMC, a reminder that the entire AI supply chain — from the machines that print chips to the chips that run models — is being repriced in real time.
For Alphabet, the strategic appeal is twofold. Cheaper inference protects the profitability of AI features woven through its consumer products, and it strengthens Google Cloud's pitch to enterprise customers who are intensely focused on the cost of deploying models. Both matter at a time when investors are demanding proof that the enormous capital pouring into AI infrastructure will eventually produce durable profits rather than a commoditized race to the bottom.
What to Watch
The Information's report means that, for now, the details remain limited and Google has not publicly confirmed the Frozen project, its specifications, or a timeline. What is clear is the intent: Google views efficient inference as a decisive competitive frontier and is willing to design dedicated silicon to win it. As the cost of running AI models increasingly determines which products thrive, the companies that control their own chips — and can tune them to their own models — will hold a structural advantage.
The signal for the rest of the industry is that the AI chip race is shifting from a contest over raw training power to a contest over cost-per-query, and AI hardware may determine the winners.
Stay Ahead of AI
Custom silicon is quietly becoming the deciding factor in who profits from AI. For ongoing coverage of the chips, models, and economics shaping the industry, visit our homepage at aibuzzwire.news.
Read more AI news →
