Google is developing a new server chip codenamed Frozen v2, intended to run its Gemini AI models more efficiently. According to a TechNews report citing tech outlet The Information, the chip is meant to "substantially raise the efficiency of running Gemini models" (E1). The same account was corroborated by Central News Agency (CNA), whose reporter Chang Hsin-yu filed from San Francisco on July 20 with an identical description of the project (E7).
Per the TechNews report, Frozen v2's approach is architectural: the chip embeds parts of the Gemini model architecture directly and permanently into the silicon itself — effectively hardwiring segments of the model into hardware. The stated goal is to reduce both the number of computations needed to answer a user query and the volume of data that must move within the chip while doing so (E3).
According to both TechNews (E4) and CNA (E9), Google engineers estimate that, at the same power consumption, Frozen v2 could process 6 to 10 times as many tokens as Google's current AI chip, the TPU (Tensor Processing Unit). The figure is reported identically across both outlets, which independently cite the same underlying claim traced to The Information.
TechNews reports that Frozen is positioned as a more specialized chip within Google's custom silicon lineup, rather than a replacement for the broader-purpose TPU. Google is not currently planning to mass-produce Frozen at the same scale as TPU (E5).
CNBC, citing The Information, reported that Google is targeting 2028 for deployment of the new chip — a timeline relayed by both TechNews (E2) and CNA (E8).
In a statement to CNBC, Google's parent company Alphabet said its teams "continue to research and test various new technologies, hoping to provide the best performance and efficiency for our users and customers." Alphabet added that "while not every project will ultimately reach mass production, rigorous exploration is core to our full-stack technology strategy" (E6).
Both TechNews and CNA report that news of the 2028 deployment target — first surfaced via The Information and CNBC — coincided with a rise in Alphabet's stock price on the day the report emerged (E2, E8). Neither report specifies a percentage move.
The reporting places two Alphabet-linked signals side by side: a hedged official framing in which Alphabet itself stresses that "not every project" it explores reaches mass production (E6), and a market reaction — Alphabet shares rising the same day (E2, E8) — that followed a specific, unhedged efficiency claim of 6 to 10 times TPU's token throughput (E4, E9) and a firm 2028 deployment target (E2, E8). At the same time, TechNews reports Frozen is designed as a specialized addition to Google's chip lineup rather than a TPU replacement, with no plan for TPU-scale production (E5) — a narrower scope than the scale of the performance multiplier being reported alongside it.
No. According to the TechNews report, Frozen is positioned as a more specialized chip within Google's custom chip lineup rather than a replacement for the broader-purpose TPU, and Google does not currently plan to mass-produce it at TPU's scale (E5).
Per CNBC's report, cited by both TechNews and CNA, Google is targeting 2028 for deployment of the new chip (E2, E8).
Google engineers estimate that, at the same power consumption, Frozen v2 could process 6 to 10 times as many tokens as Google's current TPU chip, according to both TechNews and CNA (E4, E9).
According to the Liberty Times and UDN, the White House pushed Apple, NVIDIA and SpaceX toward Intel after Washington took an roughly $8.9 billion, near-10% equity stake in the chipmaker. Deals followed, but Intel's foundry unit lost $10.4 billion over four quarters, and Apple still relies on TSMC for most custom chips.
According to United Daily News (UDN), Nikkei Asia reported that TSMC (台積電) has discussed raising 2027 wafer prices by a base 5%–10%, plus a 10%–15% surcharge on rush high-performance-computing (HPC) orders. Central News Agency (CNA) reports talks ran from June to July, effective early next year. TSMC told Nikkei it does not comment on pricing and calls its approach 'strategic, not opportunistic.'
According to a UDN Money report, TSMC (台積電) holds its earnings call on July 16, 2026, with analysts expecting about $56 billion in 2026 capital expenditure and over $150 billion across the next three years. Liberty Times Net reports TSMC guided Q2 revenue to $39.5 billion, up roughly 10% quarter over quarter.