Google Frozen v2 Targets 6–10x Gemini Tokens per Watt by 2028

Google is reportedly designing a new server processor, internally known as “Frozen v2,” that could generate six to ten times more Gemini tokens per unit of power than the company’s latest AI chips. Targeted for deployment around 2028, the project appears to go beyond another routine Tensor Processing Unit upgrade: engineers are said to be exploring how parts of Gemini’s architecture or model information could be embedded more directly into silicon. If Google can deliver anything close to the claimed efficiency, Frozen v2 could lower the cost of Gemini, relieve severe data-center capacity constraints, strengthen Google Cloud, and push the AI industry toward far tighter integration between models and the hardware running them.

---

Overview​

Google has not formally announced Frozen v2 or guaranteed that it will become a production product. The company’s response to reports about the project emphasized that its teams continually research new technologies, that not every experiment reaches production, and that hardware-software co-design remains central to its infrastructure strategy.
That careful wording matters. Designing an advanced processor is a multiyear undertaking involving architecture, simulation, physical design, manufacturing, packaging, networking, software enablement, and testing. A project intended for 2028 may still change substantially, miss its schedule, be folded into another chip family, or fail to meet its original performance target.
Nevertheless, the reported goal is credible in the context of Google’s history. The company has spent more than a decade building custom silicon for workloads that are too large, repetitive, or economically important to leave entirely to general-purpose processors.

From the first TPU to Ironwood​

Google introduced its first Tensor Processing Unit internally to accelerate machine-learning inference, later expanding the TPU family into large systems capable of both training and serving sophisticated models. Successive TPU generations have improved memory capacity, interconnect bandwidth, numerical formats, reliability, and performance per watt.
The current Ironwood generation represents Google’s seventh major TPU generation and is designed for large-scale training, reasoning, and inference. Google says Ironwood offers substantially higher per-chip performance than Trillium, its previous generation, while improving energy efficiency and supporting enormous interconnected TPU deployments.
Frozen v2 would reportedly sit alongside this TPU roadmap rather than simply replace it. That suggests Google may be moving toward a more diversified AI hardware portfolio in which different processors handle different stages, models, latency targets, and customer requirements.

What “Frozen” may imply​

The project’s name has not been explained publicly, but the underlying idea appears to involve freezing some model-related information into the processor design. Conventional accelerators remain highly programmable, allowing operators to load and run many models. That flexibility is commercially valuable, but it consumes silicon area, memory bandwidth, power, and software overhead.
A processor optimized around a narrower and more stable set of Gemini operations could eliminate some of that generality. Google might hardwire common mathematical paths, data movement patterns, sparsity structures, quantization schemes, attention operations, or other characteristics that remain consistent across a Gemini model family.
That is not necessarily the same as placing an entire Gemini model permanently inside a chip. The phrase “Gemini baked into silicon” is an accessible shorthand, but the engineering reality is likely to be more nuanced.

---

How Frozen v2 Could Work​

Modern generative AI is constrained by more than raw arithmetic. The processor must repeatedly move vast quantities of model data among high-bandwidth memory, on-chip caches, interconnects, and computing units. Data movement can consume enormous amounts of energy, sometimes making memory access and communication as important as the multiplication hardware itself.
Frozen v2’s opportunity is therefore not merely to calculate faster. It is to reduce how much information must move, how often it must move, and how much general-purpose hardware must remain active while Gemini generates a response.

Hardware-software co-design​

Google controls several layers of the Gemini delivery system:

• It develops the Gemini model family and understands its internal computational structure.
• It creates compilers, runtimes, orchestration systems, and machine-learning frameworks.
• It designs custom accelerators, networking components, servers, and data-center infrastructure.
• It operates consumer services and Google Cloud, giving it extensive production telemetry.
• It can modify future models to align with the strengths and constraints of future chips.

This control allows Google’s model researchers and chip designers to work toward the same target. Instead of building a processor and later adapting Gemini to it, the company can potentially modify both hardware and software simultaneously.
That could mean designing Gemini layers around supported low-precision number formats, predictable memory layouts, specialized matrix units, compressed weights, structured sparsity, or accelerated attention mechanisms. It could also involve changing how requests are batched and how different parts of a model are distributed across a large cluster.

The flexibility trade-off​

The fundamental compromise in any application-specific integrated circuit is efficiency versus adaptability. A GPU can run an enormous range of workloads because it retains substantial programmable flexibility. A more specialized processor can outperform it on a defined task, but may struggle when that task changes.
Google consequently must decide how much of Gemini it can safely encode into hardware several years before mass deployment. Models evolve rapidly, and a chip designed around today’s architecture could become inefficient if Gemini adopts a radically different structure before 2028.
A sensible design would freeze only durable elements: operations, dataflows, and memory behavior that Google expects to survive multiple model generations. More volatile details could remain in software, firmware, external memory, or configurable processing blocks.

---

Tokens per Watt Becomes the Critical Metric​

AI companies once emphasized training performance, parameter counts, benchmark scores, and cluster size. Those measures still matter, but the economics of widely deployed generative AI increasingly revolve around inference, the process of using a trained model to answer a request.
Every Gemini prompt triggers computation. A longer conversation, research task, coding job, multimodal analysis, or agentic workflow may consume thousands or millions of tokens across visible output and hidden reasoning. Multiply that by large numbers of users, applications, and automated agents, and inference becomes a recurring industrial expense.

Why six to ten times would be extraordinary​

The reported Frozen v2 target measures efficiency by the number of AI tokens produced for each unit of electrical power. If the processor genuinely delivers a sixfold to tenfold improvement over Google’s current hardware under comparable workloads, the effects could extend across the company.
Google could use the gain in several ways:

1. It could serve the same workload with less electricity and fewer processors.
2. It could support more users within an existing data-center power envelope.
3. It could run larger or more computationally intensive Gemini models without multiplying operating costs.
4. It could lower prices to make Gemini APIs more attractive to developers.
5. It could reserve premium hardware for training and use Frozen systems for high-volume inference.
6. It could increase hidden reasoning budgets while maintaining acceptable response costs.

Those benefits would not necessarily arrive in full. A sixfold chip-level improvement does not automatically create a sixfold reduction in total service cost because data centers also contain CPUs, memory, networking, storage, cooling, power conversion, and redundant systems.
Even so, processor efficiency is a major lever. At Google’s scale, a modest percentage gain can produce material savings; a multiple-generation leap could alter the economics of entire product categories.

Efficiency is also capacity​

Power efficiency is often presented as an environmental concern, but for hyperscalers it has become a capacity constraint. Data centers cannot draw unlimited electricity, and new grid connections, substations, generators, cooling systems, and transmission infrastructure can take years to build.
A chip that produces more useful work from each watt effectively creates additional capacity without requiring an equivalent expansion in physical infrastructure. Frozen v2 could therefore help Google serve more Gemini requests in locations where power, space, or cooling limits prevent the installation of additional conventional accelerators.

---

Google’s AI Capacity Crunch​

The reported chip project arrives amid claims that Google has struggled to secure enough computing capacity for its internal AI products and external cloud customers. Demand for accelerators has grown faster than the industry’s ability to manufacture chips, package them with high-bandwidth memory, connect them into clusters, and provide the electricity needed to operate them.
Google has reportedly faced difficult allocation decisions between its own services and paying Google Cloud customers. If cloud deals have been delayed or rejected because the necessary hardware was unavailable, the constraint is no longer an abstract infrastructure problem. It directly affects revenue and competitive positioning.

Why buying more GPUs is not a complete answer​

Google can and does use third-party processors, including Nvidia hardware, where they make technical and economic sense. Purchasing more accelerators, however, cannot solve every constraint.
The leading AI chips depend on a concentrated supply chain involving advanced fabrication, high-bandwidth memory, sophisticated packaging, networking components, substrates, and server manufacturing. Hyperscalers compete for many of the same limited production slots.
Even if processors are available, deploying them requires suitable buildings and power. Construction schedules, permits, grid interconnections, cooling capacity, and local opposition can slow projects for years.
Custom silicon gives Google another path. It can optimize the processors for its own needs, negotiate long-term manufacturing capacity, and avoid competing exclusively for the same finished systems purchased by every other AI company.

Capital spending raises the stakes​

Alphabet expects its 2026 capital expenditures to reach between $180 billion and $190 billion, with another significant increase anticipated in 2027. That staggering commitment covers more than AI processors, but AI infrastructure is a central driver.
Investors will expect this spending to produce revenue, durable competitive advantages, or lower operating costs. It is not enough for Google to build impressive data centers; those facilities must support profitable Cloud workloads, strengthen Search and advertising, improve Gemini, and protect the company’s position against Microsoft, Amazon, OpenAI, Anthropic, and Meta.
Frozen v2 offers a compelling narrative for that investment. If better silicon allows each dollar of infrastructure to produce substantially more AI output, Google can argue that its spending is building an increasingly efficient platform rather than merely feeding an endless demand for expensive chips.

---

Frozen v2 Versus Google’s TPUs​

It would be easy to interpret Frozen v2 as a successor to Ironwood, but reports indicate that Google envisions a separate processor family. That distinction may reveal how the company expects AI infrastructure to evolve.
A general TPU must support numerous Google workloads and, when offered through Google Cloud, models and frameworks created by outside customers. Frozen hardware could be much more narrowly aligned with Google’s own Gemini services.

Training and inference require different priorities​

Training constructs or updates a model by processing huge datasets and coordinating computation across large clusters. It benefits from programmability, high throughput, massive memory bandwidth, fast interconnects, and the ability to accommodate changing research techniques.
Inference serves a completed model. Its priorities often include:

• Low cost per token.
• Predictable latency.
• High utilization under fluctuating demand.
• Efficient model-weight storage.
• Fast communication between model partitions.
• Support for quantization and lower-precision computation.
• Reliable operation across millions of requests.

Ironwood is designed to cover both training and inference at considerable scale. Frozen v2 could be tuned more aggressively for the latter, sacrificing flexibility to improve the economics of production Gemini traffic.

A portfolio rather than a universal chip​

Google’s future AI stack may resemble a heterogeneous factory. TPUs could train frontier models and handle adaptable workloads, Frozen processors could serve stable Gemini variants, CPUs could manage orchestration and data preparation, and specialized networking hardware could move information among them.
This approach acknowledges that no single processor is ideal for every phase of AI. A premium reasoning request, a short Search summary, an image-generation job, and an enterprise document-analysis workflow place different demands on hardware.
The orchestration layer then becomes crucial. Google must route each request to the most suitable processor while meeting latency, privacy, reliability, residency, and cost requirements.

---

The Race to Reduce Dependence on Nvidia​

Nvidia remains the central supplier of AI accelerators because it combines powerful hardware with networking, mature software, development tools, optimized libraries, and a vast engineering ecosystem. Its advantage is not simply that its GPUs perform calculations quickly; customers can deploy diverse models with a comparatively familiar toolchain.
The hyperscalers nevertheless have strong incentives to develop alternatives. They spend tens of billions of dollars on infrastructure, operate predictable workloads at enormous scale, and possess enough internal demand to justify custom designs.

Every major platform wants custom silicon​

The competitive field now includes several distinct programs:

• Google has developed multiple TPU generations and is reportedly exploring Frozen processors specifically for Gemini.
• Amazon Web Services offers Trainium for model training and Inferentia for inference, alongside its Graviton server CPUs.
• Microsoft has developed Maia accelerators for Azure AI workloads and has continued expanding its own silicon portfolio.
• Meta is deploying successive generations of its Meta Training and Inference Accelerator across recommendation and generative AI systems.
• OpenAI has moved toward custom inference silicon to gain greater control over the cost of running its models.
• Other AI developers are pursuing partnerships with semiconductor designers, foundries, memory suppliers, and cloud platforms.

These projects do not mean Nvidia is about to disappear from AI data centers. General-purpose accelerators remain essential for research, rapidly changing models, new workloads, and customers that require portability.
Instead, custom chips can absorb high-volume, predictable workloads. That may reduce the number of Nvidia GPUs needed for certain services while reserving those GPUs for jobs where their flexibility delivers the greatest value.

Nvidia’s ecosystem remains a formidable barrier​

A chip’s theoretical efficiency does not guarantee useful deployment. Developers need compilers, debuggers, profiling tools, optimized kernels, libraries, frameworks, schedulers, and monitoring systems. Data-center operators also need reliable servers, networking, spare parts, and support.
Google is better positioned than most companies because it controls the applications running on its hardware. It does not need to persuade the entire developer community to abandon established tools before using Frozen v2 internally.
The challenge becomes greater if Google wants Frozen systems to host third-party models through Cloud. External customers will expect compatibility, predictable performance, transparent pricing, and an easy migration path from Nvidia, AMD, TPU, or other platforms.

---

What Frozen v2 Means for Gemini​

Gemini’s competitiveness depends on more than benchmark results. Google must deliver the model at acceptable latency and cost across Search, Workspace, Android, Chrome, Cloud, developer APIs, and emerging agentic services.
More efficient hardware could let Google improve the experience without asking users to think about the infrastructure beneath it. Faster responses and higher usage limits would appear to be software improvements even when custom silicon enabled them.

More reasoning within the same budget​

Advanced models increasingly use additional inference-time computation to reason through difficult questions, evaluate possible answers, call tools, verify outputs, and revise mistakes. This can improve quality, but it also increases token consumption and cost.
Frozen v2 could allow Gemini to spend more computation on each request while holding the power budget relatively stable. Google could apply that capacity to coding, research, mathematics, planning, security analysis, or complex multimodal tasks.
This possibility is particularly relevant following reports that the next flagship Gemini release has been delayed while Google improves its performance, especially in coding. Better hardware cannot compensate for weaknesses in model design or training data, but it can enable more ambitious inference strategies once the underlying model is ready.

Stable model families may be necessary​

Highly specialized silicon works best when the target workload changes predictably. Google may therefore need to create longer-lived Gemini branches designed for specific Frozen generations.
A stable enterprise model could remain supported for years, while experimental frontier models continue to run on flexible TPUs. Google could periodically translate successful architectural changes into the next Frozen design rather than attempting to hardwire every rapidly evolving research idea.
This would create a cadence in which software innovation leads and specialized hardware follows. The approach resembles long-term platform engineering more than the current race to release a new model every few months.

---

Consumer Impact​

Consumers are unlikely to purchase a Frozen v2 chip or install one in a Windows PC. The processor is reportedly intended for Google’s servers, so its effects would arrive through cloud-delivered products rather than local hardware.
For Windows users, the most visible changes could appear in Gemini’s web application, Chrome, Google Workspace, developer tools, Search, and third-party services that call Google’s APIs.

Faster and more widely available Gemini services​

If Google lowers its inference costs, it can potentially offer:

• Higher usage limits for free and paid Gemini users.
• Faster responses during peak periods.
• Longer context windows without proportionally higher prices.
• More capable reasoning modes.
• Improved image, audio, video, and document processing.
• Better availability in regions with constrained infrastructure.
• More agentic features that perform multiple steps on a user’s behalf.

The key word is potentially. Companies do not always pass infrastructure savings directly to consumers. Google could instead use the efficiency to improve margins, increase model quality, or serve rapidly growing demand without lowering subscription prices.
Competition makes user benefits more likely. If Microsoft, OpenAI, Anthropic, Amazon, and others also reduce inference costs, providers may compete through lower pricing, generous quotas, faster models, and richer features.

Cloud AI versus local Windows AI​

Frozen v2 would also highlight the continuing divide between cloud and on-device AI. Microsoft and PC manufacturers are promoting neural processing units in Windows PCs for private, low-latency local workloads. Google, meanwhile, is pursuing extreme specialization inside hyperscale data centers.
Neither approach eliminates the other. A future Windows application might use the PC’s NPU for transcription, indexing, privacy filtering, or lightweight generation, then send complex reasoning tasks to Gemini running on Frozen hardware.
The best user experience may come from hybrid systems that decide automatically where each operation should run. Cost, battery life, connectivity, privacy, model capability, and data sensitivity will all influence that decision.

---

Enterprise and Google Cloud Impact​

Enterprise customers care less about the internal codename than about price, availability, security, reliability, and contractual commitments. Frozen v2 will matter commercially only if its efficiency translates into measurable improvements in these areas.
Google Cloud could use the technology internally to improve managed Gemini services, or eventually expose Frozen-backed capacity through Vertex AI and related platforms.

Lower inference costs could unlock new workloads​

Many AI prototypes fail to reach production because their economics become unattractive at scale. An application that appears affordable during a small trial can become prohibitively expensive when thousands of employees or millions of customers use it.
A substantial reduction in cost per token could make previously marginal workloads viable, including:

• Continuous analysis of corporate communications and documents.
• Large-scale software modernization and code review.
• Customer-service agents handling long conversations.
• Security systems correlating events across extensive logs.
• Research agents that perform repeated searches and verification steps.
• Industry-specific assistants operating across large private datasets.

Enterprises would still need to evaluate accuracy, governance, integration, and compliance. Cheaper inference does not make unreliable output safe, but it can remove one barrier to deployment.

Capacity guarantees may be more important than price​

Organizations building critical systems need confidence that processing capacity will be available when demand rises. A low advertised price is of limited value if quotas tighten, regions run out of accelerators, or new deployments face long delays.
If Frozen v2 expands Google’s effective AI capacity, Cloud could offer stronger reservations, more predictable throughput, and better service-level commitments. That would help Google compete with Azure and AWS for customers that want large, dependable pools of AI inference.
However, Google must avoid creating a two-tier problem in which its own Gemini services receive priority while external customers face shortages. Cloud buyers will watch whether hardware investments support an open commercial platform or primarily protect Google’s consumer products.

---

The Economics of Vertical Integration​

Google’s full-stack strategy covers models, chips, compilers, networking, servers, data centers, cloud services, and consumer applications. Vertical integration can create efficiencies that a company purchasing each layer from separate suppliers may struggle to match.
It can also shift profit within the supply chain. Money that would have gone toward another chip vendor may instead fund Google’s own engineering and manufacturing contracts.

Why scale changes the calculation​

Custom chip development is expensive and risky. A smaller company might never deploy enough units to recover design costs, while Google can spread those costs across Search, YouTube, Workspace, Gemini, advertising, Android services, and Cloud.
The company also has detailed visibility into its workloads. It knows which operations consume the most power, which model configurations dominate traffic, where latency accumulates, and which software changes could simplify hardware.
That information can guide design choices more accurately than public benchmarks. Frozen v2 does not need to win every independent test; it needs to perform Google’s actual Gemini workloads efficiently enough to justify its development and deployment.

Lock-in moves in both directions​

Vertical integration can strengthen Google’s control, but it also creates dependence on its own decisions. If Frozen v2 is too specialized, Google may find it difficult to adopt a superior external model architecture quickly. A manufacturing delay could disrupt both hardware and software roadmaps.
Cloud customers may also become more deeply tied to Google if applications are optimized around Gemini and Google-specific processors. That can improve performance, but organizations should consider portability, exit costs, and whether equivalent workloads can move to another cloud.

---

Manufacturing and Supply-Chain Realities​

Designing the processor is only one part of the challenge. Google will need manufacturing capacity on an advanced process node, high-bandwidth memory, packaging, substrates, networking, servers, cooling equipment, and deployment space.
Most custom AI chip programs still rely on a relatively small number of suppliers. Vertical integration at the design level does not make Google independent of the semiconductor supply chain.

Advanced packaging may remain a bottleneck​

AI processors must be placed close to large pools of fast memory and connected efficiently to other chips. The resulting packages are technically difficult and expensive to produce.
A shortage of packaging or high-bandwidth memory could limit Frozen v2 deployments even if the processor design is successful. Google will need to secure capacity years in advance, balance Frozen production against TPU production, and ensure that its manufacturing partners can achieve acceptable yields.
The six-to-ten-times efficiency claim should also be evaluated at the system level. A processor may excel in a controlled test while encountering bottlenecks in memory, networking, software, or cluster utilization under production traffic.

The 2028 target is both distant and close​

Two years is a long period in AI software but a short one in advanced semiconductor development. Google must make architecture decisions while Gemini continues to evolve, then carry those decisions through physical design and fabrication.
The first silicon may expose defects or performance problems requiring another revision. Software teams must prepare compilers and runtimes before large-scale deployment, while data-center planners must design systems around power, cooling, and networking requirements that are not yet final.
A 2028 release is therefore plausible but far from assured. Even if chips begin deployment that year, meaningful fleet-wide impact could take longer.

---

Strengths and Opportunities​

Frozen v2 could give Google several important advantages if the reported project reaches production and meets its targets.

• It could dramatically improve Gemini’s performance per watt. Higher token output within the same power envelope would expand effective capacity and reduce pressure on data-center construction.
• It could lower the marginal cost of inference. Google could improve margins, reduce API prices, raise usage limits, or allocate more computation to reasoning and multimodal workloads.
• It could strengthen Google Cloud. More capacity and predictable economics would help Cloud compete for enterprise AI deployments that require reserved throughput.
• It could reduce dependence on a single accelerator supplier. A diversified mix of TPUs, Frozen chips, and third-party hardware would give Google more purchasing and deployment flexibility.
• It could deepen hardware-software integration. Google’s control over Gemini lets engineers optimize models around processor characteristics that competitors cannot easily access.
• It could turn infrastructure efficiency into a product advantage. Faster responses, richer features, and lower prices would make the benefits visible to consumers and developers.
• It could establish a repeatable specialized-chip family. If Frozen v2 succeeds, future versions could follow Gemini’s model roadmap and absorb stable innovations at regular intervals.

---

Risks and Concerns​

The project also carries technical, commercial, and strategic risks that should temper expectations.

• The efficiency target may apply only to selected workloads. Marketing-friendly comparisons can change when model size, context length, batching, latency, and output quality are held constant.
• Gemini may evolve beyond the chip’s assumptions. An architectural breakthrough could make portions of the Frozen design obsolete before widespread deployment.
• Specialization could reduce flexibility. Hardware tuned for Gemini may perform poorly with third-party models or new forms of AI computation.
• Manufacturing constraints could delay deployment. Advanced fabrication, memory, packaging, and networking remain concentrated among a small number of suppliers.
• System-level savings may be smaller than chip-level gains. CPUs, networking, cooling, storage, and idle capacity still consume resources.
• Greater vertical integration could increase customer lock-in. Enterprises may receive attractive performance while becoming more dependent on Gemini, Vertex AI, and Google Cloud.
• Cheaper inference can increase total energy consumption. Efficiency often stimulates additional use, potentially offsetting some environmental savings through much larger workloads.
• The project may never reach production. Google has explicitly noted that not every research effort becomes a shipping product.

---

What to Watch Next​

The first major checkpoint arrives on Wednesday, July 22, 2026, when Alphabet is scheduled to report its second-quarter financial results. Investors will look for updated comments on AI capacity, capital expenditures, Cloud demand, and the returns generated by Google’s infrastructure expansion.
Google is unlikely to disclose detailed Frozen v2 specifications so early, but its broader statements may reveal how urgently it views inference efficiency and supply constraints.

Five signals that will determine the project’s importance​

1. Google must clarify whether Frozen is an inference-only architecture. Its relationship with TPUs will remain uncertain until the company explains the division of labor.
2. The company must define the efficiency baseline. A six-to-ten-times gain is meaningful only when compared under equivalent model, latency, accuracy, and utilization conditions.
3. Gemini’s architecture must stabilize enough for hardware specialization. Continued major redesigns would make deep model-specific optimization harder.
4. Manufacturing partners must deliver at scale. A successful prototype will not resolve Google’s capacity problem unless thousands of systems can be produced and deployed economically.
5. Users must see practical improvements. Lower prices, higher limits, faster responses, stronger Cloud availability, or more capable reasoning would demonstrate that silicon efficiency is reaching the product layer.

The competitive response​

Nvidia will continue improving inference performance, memory systems, networking, and software. Microsoft, Amazon, Meta, OpenAI, and other major AI operators will advance their custom silicon at the same time.
Frozen v2 will therefore compete against the hardware available in 2028, not the processors installed in 2026. Google’s claimed multiple must survive a moving baseline in which every supplier is targeting better performance per watt and lower cost per token.
The likely outcome is not a single winner but a more fragmented market. Flexible GPUs will coexist with cloud-specific accelerators, model-specific inference chips, local NPUs, and specialized processors for video, recommendation, robotics, and scientific computing.

---

Google’s reported Frozen v2 project shows that the next stage of the AI race will be fought as much in data centers and semiconductor design as in chatbot interfaces. The ability to generate a better answer matters, but the ability to generate billions of useful answers within realistic power and cost limits may matter even more. If Google can safely encode durable elements of Gemini into specialized silicon and achieve the reported efficiency gains by 2028, it could turn a severe infrastructure constraint into one of its strongest competitive advantages. If it cannot, Frozen v2 will serve as another reminder that tightly coupling fast-moving AI software to slow, expensive hardware is among the industry’s most ambitious—and unforgiving—engineering challenges.

---

References​

1. Primary source: Bitcoin World
   Published: 2026-07-20T22:00:11+00:00
   
   

   Google Is Developing A New AI Chip To Make Gemini Far More Efficient

   Google is developing a new AI chip, Frozen v2, to boost Gemini efficiency by up to 10x by 2028, as reported by The Information.

   bitcoinworld.co.in
2. Independent coverage: WION
   Published: Mon, 20 Jul 2026 19:26:00 GMT
   
   https://www.wionews.com/technology/google-s-secret-frozen-v2-ai-chip-could-make-gemini-up-to-10-times-more-efficient-report-says-1784181857721
3. Independent coverage: the-decoder.com
   Published: 2026-07-20T18:08:33+00:00
   
   

   Google's "Frozen v2" chip reportedly bakes Gemini's architecture directly into silicon for efficiency gains

   Google is developing "Frozen v2," a server chip that bakes the Gemini architecture directly into hardware. According to internal sources, it could be 6 to 10 times more efficient than current TPUs. Scheduled for 2028, the chip would drastically cut Google's AI inference costs and...

   the-decoder.com
4. Independent coverage: Techgenyz
   Published: 2026-07-20T15:13:28+00:00
   
   

   Google Frozen v2 chip Delivers Vital AI Efficiency - Techgenyz

   Google Frozen v2 chip delivers vital efficiency boosts up to 10x higher to overcome cloud computing capacity bottlenecks for Gemini AI.

   techgenyz.com
5. Independent coverage: The Tech Buzz
   Published: Mon, 20 Jul 2026 15:17:00 GMT
   
   

   Google's Frozen v2 Chip Embeds Gemini AI Into Silicon | The Tech Buzz

   Alphabet stock jumps on report of custom AI chip embedding Gemini architecture directly into hardware

   www.techbuzz.ai
6. Independent coverage: Republic World
   Published: 2026-07-20T14:51:40+00:00
   
   

   Google May be Building a New AI Chip With Gemini Baked Directly Into the Hardware | Republic World

   Informally known as Frozen v2, Google's new server chip could arrive as early as 2028 and reportedly deliver six to 10 times greater efficiency than its latest custom AI chips.

   www.republicworld.com

 

[ChatGPT]

Additional coverage of this story: Google Frozen v2 Targets 6–10x Gemini Tokens per Watt by 2028
It frames Frozen v2 chiefly as an inference-economics play, emphasizing the continuous token load from Search, Workspace, Android, APIs, and agent systems rather than periodic model training.

https://en.bloomingbit.io/feed/news/116638/