On Monday morning, Alphabet shares climbed some 3% after a report said a new Google Gemini chip, internally called “Frozen v2,” could make the company’s in-house AI models far more efficient by 2028.
That timing matters. Alphabet reports earnings later this week, and investors have already questioned the company’s planned AI spending after Google said earlier this year it plans to spend between $180 billion and $190 billion. A chip that can serve more Gemini output per unit of power gives Alphabet a cleaner answer to the question hanging over the stock: can AI scale without crushing the economics?
The new server chip is being designed to help Gemini models operate more efficiently, according to TechCrunch, citing The Information’s report. Google did not directly confirm the report to TechCrunch. It did not deny it either.
“Our teams are constantly researching and experimenting with new innovations to deliver maximum performance and efficiency for our users and customers,” Google told TechCrunch. “While not every project moves into production, this rigorous exploration is central to our full stack approach. By co-designing our hardware and software from the ground up, we ensure our systems are integrated and highly optimized for real-world workloads.”
Monday morning’s 3% move turned Frozen v2 into an earnings story
The reported Google Gemini chip is not scheduled for immediate release. The Information said Frozen v2 is slated for sometime in 2028, which makes the market reaction more interesting. Investors were not reacting to a product launch. They were reacting to a possible answer to Alphabet’s AI cost problem.
The reported metric is specific: Frozen v2 could be between six and 10 times more efficient than Google’s existing AI chips, measured by the number of tokens generated per unit of power. That is a cleaner metric than vague claims about AI performance because it points straight at serving economics.
XOOMAR analysis: this is why the report landed before earnings. For AI infrastructure, the market is no longer only asking whether models are smarter. It is asking whether the companies building them can afford to run them at scale. A model that wins benchmarks but burns too much compute becomes a margin problem. A model that produces more tokens per watt gives management a better cost narrative.
This also fits a wider Google moment. The company is facing pressure across multiple fronts, from app store economics, as covered in Google Opens Android App Stores but Keeps the Cash, to platform safety questions in Nudify Apps Drag Apple and Google Into Legal Crosshairs. Frozen v2 sits in a different category, but the theme is similar: Google is trying to preserve control where its distribution and infrastructure matter most.
The 2028 target puts the Google Gemini chip on a tokens-per-watt clock
The key distinction is inference. Training creates the model. Inference serves the answer. Once AI products reach heavy daily use, inference can become the recurring cost that keeps hitting the income statement.
The reported Frozen v2 benchmark focuses on output per power unit, not model quality. That matters because power is not a soft constraint. It shapes data center capacity, operating cost, and the number of model responses Google can serve from the same infrastructure footprint.
Reported figures and claims:
| Item | Reported detail |
|---|---|
| Chip name | Internally dubbed “Frozen v2” |
| Target release | Sometime in 2028 |
| Efficiency claim | Between six and 10 times more efficient than Google’s existing AI chips |
| Measurement | Tokens generated per unit of power |
| Company response | Google did not directly confirm or deny the report |
| Stock reaction | Alphabet stock climbed some 3% on Monday morning after The Information report |
The Economic Times, also citing The Information, reported that the chip is expected to help address an AI computing capacity crunch that has fueled internal tensions and prompted Google Cloud to decline deals with outside customers. That claim raises the stakes. If capacity is tight enough to affect customer deals, efficiency is not just an engineering target. It is commercial oxygen.
XOOMAR analysis: the most important number here is not the 2028 date. It is the six-to-10-times efficiency range. If that survives real deployment, Google gets more Gemini output from the same power envelope. If it does not, Frozen v2 becomes another long-horizon R&D project in a spending cycle already large enough to worry investors.
From TPUs to Frozen v2: Google is building a separate Gemini hardware lane
Google already has custom AI chips, including its Tensor Processing Units, or TPUs. The reported Frozen v2 project is different in one important way: The Economic Times report says the “Frozen” project is aimed at creating homegrown chips apart from Google’s TPUs, rather than replacing them.
That distinction matters. It suggests Google is not abandoning the TPU path. It is exploring a more specialized path for Gemini serving.
NDTV Profit, citing the same report, said Frozen v2 would directly integrate elements or blueprints of the Gemini model into hardware, with the goal of reducing data movement and cutting the number of decisions needed during inference. That is the technical heart of the story. Less data movement can mean better energy efficiency, which connects directly to the reported tokens-per-power-unit metric.
Google’s statement to TechCrunch also points in that direction. The company emphasized co-designing hardware and software “from the ground up” for “real-world workloads.” That is corporate language, but here it maps to a concrete strategy: build silicon around the work Gemini actually does, rather than relying only on general-purpose AI accelerators.
June’s Jalapeño and Samsung talks show why Nvidia pressure is part of the story
TechCrunch notes that AI companies have increasingly sought to produce their own chips to make in-house models run more efficiently and to address global shortages in AI computing capacity. It also says firms are trying to wean themselves off Nvidia, which has historically dominated the AI chip market and left major AI makers dependent on its hardware.
Google is not alone in this push.
In June, OpenAI announced its first custom chip, an inference processor called Jalapeño. Earlier this month, it was reported that Anthropic was discussing a new chipmaking partnership with Samsung.
That does not mean Nvidia gets displaced overnight. The supplied reporting does not support that conclusion. The cleaner read is narrower: every serious in-house chip effort gives major AI labs and platform companies another option for high-volume workloads. It reduces dependence at the margin first. If the chips work, that margin can grow.
XOOMAR analysis: Frozen v2 is best understood as a defensive wall. It aims to defend Gemini’s operating economics, defend Google Cloud capacity, and defend Alphabet from being fully exposed to external chip supply and pricing dynamics. The reported efficiency metric is the wall’s first brick.
Developers and cloud customers may benefit, but lock-in risk cuts both ways
If Frozen v2 reaches production and performs as reported, Google gains the clearest upside. It could serve more Gemini tokens per unit of power, reduce pressure from scarce AI computing capacity, and tighten the connection between its models and its infrastructure.
For Google Cloud customers and developers, the implications are more mixed.
Potential benefits, if the efficiency translates into commercial products:
- Capacity: More efficient inference could support more Gemini usage on constrained infrastructure.
- Reliability: Better alignment between hardware and model workloads could reduce pressure points in serving.
- Pricing optionality: Google could choose to pass some cost gains into commercial AI services, though no such pricing change has been reported.
The risk is dependence. A Gemini-optimized chip may make Gemini services work especially well inside Google’s infrastructure. That can be attractive for developers who want performance and scale. It can also deepen switching costs if workloads become tightly tied to Google’s model and serving stack.
Investors should separate the reported fact from the tempting leap. The reported fact is efficiency, measured in tokens per unit of power. The unproven leap is that this will automatically lead to cheaper APIs, faster user experiences, or better margins. Those outcomes require deployment, utilization, pricing decisions, and evidence from live workloads.
After the report, Alphabet’s next test is proving AI spending has a payback path
The next decision point is close: Alphabet’s earnings report later this week. Frozen v2 will not ship then. But management may face questions about AI capital spending, compute capacity, and whether custom silicon can change the cost curve.
The evidence that would strengthen the Frozen v2 thesis is specific: Google confirming production plans, giving more detail on deployment timing, showing efficiency gains in real Gemini workloads, or tying chip work to measurable capacity improvements. Evidence that would weaken it would be delays, silence around production, or efficiency claims that stay confined to internal research.
For now, the Google Gemini chip story signals a sharper phase in AI competition. Model releases still matter. But the harder contest is moving underneath them, into power budgets, server chips, and the cost of generating each token. If Frozen v2 becomes real in 2028, Google’s advantage will not be that it built another AI chip. It will be that Gemini’s economics start looking less like a spending problem and more like infrastructure control.
The Bottom Line
- Alphabet investors are looking for signs that Google can scale AI without overwhelming its cost structure.
- A more efficient Gemini chip could reduce power demands and improve the economics of AI model serving.
- The 2028 timeline means the chip is a long-term answer, not an immediate fix for Alphabet’s AI spending concerns.
Originally published on XOOMAR. For more news and analysis, visit XOOMAR.
Top comments (0)