Alphabet stock rises as report says it's building a more efficient AI chip
Reports say Alphabet is building a server chip, nicknamed “Frozen v2,” to run Gemini models more efficiently—possibly cutting power per token by six to 10 times. Deployment is not expected until 2028 and the design is still in flux.
Alphabet shares jumped on Monday after reports that the company is developing a new server chip intended to run its Gemini models more efficiently, potentially easing pressure on cloud capacity and lowering AI serving costs.ReutersCNBC
The move matters because custom silicon has become a strategic lever for big cloud providers: companies that cut power per token can serve generative models more cheaply and free up capacity as demand rises. Alphabet’s reportedly internal project—referred to in coverage as “Frozen v2”—aims to hardwire parts of Gemini into hardware to boost inference efficiency.Reuters
Reported efficiency gains: “six to 10 times” versus current chips
Sources briefed to Reuters say the chip could be “six to 10 times more efficient” than Google’s latest custom AI chips when measured by AI tokens served per unit of power, though engineers are still finalising how much of the model would be embedded in silicon.Reuters This figure comes from reporting, not a tested product benchmark, and should be treated as an internal estimate rather than independently verified performance data.
CNBC reports the new processors aren’t expected to be deployed until 2028 and that they would not replace Google’s existing tensor processing units (TPUs), which remain central to Alphabet’s training and inference strategy.CNBC Alphabet has previously pointed to material efficiency gains from TPUs—recent generations were said to be up to three times faster for training and to deliver roughly 80% better performance per dollar, according to company statements cited by CNBC.CNBC
Why this would matter for Google Cloud and Nvidia rivalry
The reported project responds to a broader industry squeeze on inference capacity and the economics of model serving. Reuters frames the effort as partly aimed at alleviating an “AI computing capacity crunch” that has forced Google Cloud to turn down some customer deals.Reuters That pressure has strategic implications: the fewer tokens a provider must power per dollar, the more aggressively it can price services and the less reliant it becomes on general-purpose GPUs from Nvidia.
Alphabet already co-designs TPUs with Broadcom for its current ASICs; reporting says the new design would be additive rather than a wholesale replacement of that stack.CNBC Competitors such as Amazon and Meta are also investing in custom silicon, but Nvidia remains the dominant supplier for many customers—so a materially better inference chip would shift economics, not instantly dethrone incumbents.
Reporters and industry analysts caution that significant caveats remain. Engineers are still finalising the design and how much model knowledge will be embedded in silicon, which affects versatility and upgrade cycles.Reuters Embedding model elements in hardware can improve efficiency but risks locking in specific architectures; one former chip designer warned that such trade-offs can lead to faster obsolescence if model architectures shift (commentary to Reuters).
Alphabet’s stock reaction on Monday reflected investor appetite for actions that reduce AI operating costs; whether the market’s optimism proves warranted will depend on engineering progress and deployment timing. The next concrete markers to watch are any engineering disclosures at Google Cloud events, product roadmaps from Broadcom partnerships, and whether Reuters or CNBC produce follow-up reporting with technical benchmarks or a clearer 2028 timeline.ReutersCNBC
