Alphabet, Google's parent company, is quietly advancing the development of a new server chip designed specifically to boost the efficiency of its Gemini artificial intelligence models. The processor, internally code-named 'Frozen v2,' is expected to ship around 2028 and could provide a six- to tenfold improvement in performance per watt compared with Google's current-generation AI accelerators, according to a report from The Information citing anonymous sources.

The reported efficiency gains are measured by the number of tokens—the fundamental units of text or image data that AI models process—generated per unit of power. If realized, such a leap would significantly reduce the cost and energy consumption of running large-scale AI workloads, a critical concern as Google and other tech giants pour tens of billions of dollars into AI infrastructure.

Google responded to inquiries about the chip with a carefully worded statement that neither confirmed nor denied the project. 'Our teams are constantly researching and experimenting with new innovations to deliver maximum performance and efficiency for our users and customers,' the company said. '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.'

Background: Google's Long History with Custom AI Chips

Google has been designing custom silicon for machine learning for more than a decade. Its Tensor Processing Units (TPUs), first introduced in 2015, were originally built to accelerate internal services like search, translation, and image recognition. Over time, Google made TPUs available through its cloud platform, competing directly with offerings from Nvidia, Intel, and AMD. The current iteration, the TPU v5p, launched in 2023 and was designed for training and inference at scale.

The Frozen v2 chip appears to be part of a broader strategy to create even more specialized hardware that can handle the unique demands of generative AI models like Gemini. While Google has not publicly detailed the architecture, industry analysts speculate that Frozen v2 may rely on advanced process nodes, novel memory hierarchies, or custom interconnects to achieve its dramatic efficiency claims.

The Race for AI Chip Independence

The push for custom silicon is not unique to Google. Across the AI industry, companies are racing to reduce their reliance on Nvidia, which has dominated the market for AI training and inference chips. Nvidia's graphics processing units (GPUs) are widely considered the gold standard, but they come with high costs and limited availability. Supply shortages have forced many AI firms to wait months for hardware, and the expense of running thousands of GPUs has become a major financial burden.

In June, OpenAI unveiled its first custom inference chip, code-named 'Jalapeño,' designed to lower the cost of running its GPT models. Earlier this month, reports emerged that Anthropic was in talks with Samsung to co-develop a new AI chip. These moves highlight a growing trend: major AI labs are seeking to bring hardware design in-house to gain more control over performance, cost, and supply chains.

For Google, the stakes are particularly high. The company operates one of the largest AI fleets in the world, powering everything from search to Cloud AI services. According to a report from The Financial Times, Alphabet plans to invest between $180 billion and $190 billion in capital expenditures this year—much of it aimed at building data centers and acquiring computing hardware. Any improvement in chip efficiency could translate into hundreds of millions of dollars in annual savings.

Wall Street Takes Notice

News of the Frozen v2 chip appeared to reassure investors who have grown anxious about the enormous spending required to stay competitive in AI. Following publication of The Information's report, Alphabet's shares rose approximately 3% on Monday morning, reflecting optimism that the company can improve its AI infrastructure economics. The stock move came ahead of Alphabet's quarterly earnings report later in the week, which analysts expected to scrutinize for signs of cost overruns or delayed returns on AI investments.

The efficiency metric highlighted in the report—tokens generated per unit of power—is becoming a key differentiator for AI hardware companies. As models grow larger and more complex, the energy required to run them has skyrocketed. Data centers already account for about 1% of global electricity consumption, and AI workloads represent one of the fastest-growing segments of that demand. By designing chips that deliver more tokens per watt, Google could not only lower its own electricity bills but also reduce its carbon footprint.

Technical Challenges and Unknowns

Despite the promising targets, many technical hurdles remain. Scaling a custom chip from design to mass production is notoriously difficult and expensive. The reported 2028 timeline gives Google several years to refine the architecture, validate it in test environments, and secure manufacturing capacity from foundries like TSMC or Samsung. The chip may also face competition from Nvidia's next-generation architecture, which is expected to arrive in 2026 under the code name 'Rubin.'

Google's statement that 'not every project moves into production' serves as a cautionary note. The company has a history of developing impressive prototypes that never see commercial launch. For example, its first self-driving car project eventually became Waymo, but other initiatives like Google Glass and several internal processor designs have been shelved or repurposed.

Moreover, the reported 6x-10x efficiency gain is based on internal projections, which can change as engineering teams encounter unforeseen obstacles. Achieving such large improvements typically requires breakthroughs in materials, cooling, or chip architecture rather than incremental optimizations.

Implications for the AI Ecosystem

If Google succeeds in bringing Frozen v2 to market, the impact could extend far beyond its own operations. The company could potentially offer the chip through Google Cloud, enabling enterprises and startups to run their own AI models more cheaply. That would intensify competition with cloud rivals Amazon Web Services and Microsoft Azure, both of which have their own custom chips: AWS Trainium/Inferentia and Microsoft Maia.

The move also underscores a broader shift in the semiconductor industry. Traditional chip designers like Intel and AMD are facing pressure from hyperscale cloud providers that are increasingly designing their own hardware. This trend, sometimes called 'vertical integration,' allows companies to optimize every layer of the technology stack from the chip up to the software framework. Google's full-stack approach—where hardware and software are co-designed—has already yielded benefits in its earlier TPU generations, which offered competitive performance against Nvidia's GPUs for specific workloads.

Meanwhile, the AI chip market as a whole is expected to grow from roughly $80 billion today to more than $300 billion by the end of the decade, according to McKinsey. That growth is attracting a wave of startups and incumbents alike, each vying for a slice of the pie. Google's long-standing expertise in machine learning and its willingness to invest heavily in custom silicon position it as a formidable player—provided Frozen v2 lives up to its promise.

For now, the details remain sparse. Google has not disclosed the chip's architecture, foundry partner, or exact target specifications. The 2028 release date leaves room for many changes in the competitive landscape. But the mere prospect of a highly efficient Google-designed AI chip has already stirred excitement among analysts and investors, suggesting that the battle for AI hardware supremacy is far from over.

Source: TechCrunch News