Alphabet Developing New Server AI Chip to Ease Computing Bottlenecks

Aimed at easing an internal shortage of computing resources

“Processes 6 to 10 times more tokens per unit of power than the latest TPU”

“Built to embed part of the Gemini architecture directly into silicon”

Alphabet Inc. is developing a new AI chip for servers to ease internal computing shortages and deliver artificial intelligence models more efficiently, according to a report. The news sent Alphabet shares up more than 3% in U.S. trading on July 20.

The Information reported on July 20, citing people familiar with the matter, that Google’s parent company is developing the chip under the name “Frozen v2” to address a shortage of AI computing capacity and improve the efficiency of model delivery.

Google engineers expect Frozen to process six to 10 times more tokens per unit of power than the company’s tensor processing units, or TPUs, the report said. Rather than replacing TPUs, the chip is expected to fill a more specialized role in Google’s custom chip portfolio.

The chip would permanently embed part of Gemini’s architecture directly into silicon, reducing the computation and data movement needed to answer queries.

In a statement sent to CNBC, Alphabet said co-designing hardware and software from the ground up helps ensure systems are integrated and optimized for real-world workloads. It added that not every project is ultimately deployed in production, but that such extensive exploration is central to its full-stack approach.

Google Cloud has recently been unable to expand new deals because its rental computing capacity for customers had reached its limit, the report said.

Google plans to deploy the chip by 2028, though engineers are still finalizing the design and how much model information will be embedded in it.

Bloomberg reported that Google last week delayed the release of Gemini’s latest model after it failed to meet internal targets, particularly in coding performance.

Google’s AI business is facing mounting challenges.

The next Gemini Pro release has been delayed, while lower-cost Chinese AI models are gaining share among U.S. companies. Chinese models now account for 45% of token usage by U.S. companies, the report said. Google has also lost several senior researchers to rivals including Anthropic and OpenAI.

Competition in AI models has intensified further following recent product launches from Moonshot AI and Alibaba.

Meanwhile, Demis Hassabis, the head of Google DeepMind, visited Congress this week and proposed that lawmakers create an AI oversight body similar to FINRA, the Financial Industry Regulatory Authority, to test the most advanced AI models before release. The body would operate under federal oversight, be funded largely by the industry, and aim to assess national security risks.

Kim Jung-a, guest reporter, Hankyung.com, kja@hankyung.com