Alphabet’s Google is expanding its artificial intelligence chip portfolio as it seeks to turn growing demand for alternatives to Nvidia’s processors into a larger semiconductor business, with new products designed for secure computing and complex AI applications.
According to Bloomberg, the company is introducing two additional variations of its tensor processing units (TPUs), custom-designed chips used to train and run artificial intelligence models. The new TPU 8tx and 8i Max configurations target two emerging requirements in the industry: protecting sensitive information during processing and enabling AI systems to work with much larger volumes of data in a single request.
The launch marks a faster product-development cycle for Google’s chip operation, which historically introduced a new AI processor approximately every two years. The company released two TPU generations together this year, and the latest configurations extend that portfolio to four offerings.
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Google’s chip chief, Amin Vahdat, said the strategy reflects an industry increasingly moving towards specialized computing systems rather than relying on a single processor design for every workload.
“This really follows the thesis that we have here at Google, which is that the future is going to be about specialization, and the upsides for specialization continue to be significant,” Vahdat said in an interview.
The expansion comes as the AI semiconductor market faces competing pressures. Concerns about a slowdown in infrastructure investment have unsettled technology stocks, while growing political opposition to data center construction is raising questions about the pace of new projects. At the same time, demand for Nvidia’s processors remains greater than the company can satisfy, creating an opening for Google and other suppliers.
New Chips Target Security and Long-Context AI Applications
The TPU 8tx is designed for confidential computing, a capability that allows customers, including government agencies and companies in regulated industries, to process sensitive information with additional protection.
The system works alongside processors from Advanced Micro Devices (AMD) as part of a collaboration between the two companies. Its architecture enables data to be encrypted at the chip level while processing takes place.
Under conventional computing arrangements, information may need to be decrypted in memory while a system works on it, creating potential exposure. The 8tx is designed to address that vulnerability by maintaining encryption at the hardware level, according to Vahdat.
The second configuration, the 8i Max, targets applications that require long-context windows. These determine how much information an AI model can consider when responding to a request or carrying out a task.
Longer context windows are essential for applications involving complex coding, mathematical problems, and drug discovery, where models need to process extensive information before producing an output. The 8i Max combines two TPU chips in a single package to provide additional computing power and memory.
Neither product represents an entirely new silicon architecture. Instead, Google has reconfigured its existing chip designs to address specific customer requirements, allowing the company to broaden its offering without developing completely separate processors for every application.
Both variations are expected to become available next year.
“We could build a single chip that tries to address the whole market, or by building specialized versions of the chips, we actually can deliver much more value,” Vahdat said.
The move also follows a broader shift in the semiconductor industry as AI systems evolve beyond conventional model training towards inference and autonomous agents. These applications require processors that can handle different stages of computing, including retrieving information, coordinating tasks, and executing instructions.
Nvidia, which has historically marketed its graphics processing units (GPUs) as versatile processors for a wide range of AI workloads, is also expanding into specialized systems. The company has introduced its own central processing units (CPUs), a market traditionally dominated by Intel and other established manufacturers, to support the management of AI agents.
Nvidia is also developing more specialized inference systems using technology acquired from startup Groq. Inference is the stage at which trained models respond to requests and perform tasks, rather than learning from training data.
Google’s latest products therefore place it in competition not only with Nvidia’s established GPU business but also with the expanding range of specialized processors being developed for the next generation of AI applications.
Growing Customer Base Strengthens Google’s Chip Business
Google’s effort to become a major chip supplier has evolved from an internal technology project into a commercial business with substantial growth potential.
The company began developing TPUs more than a decade ago to improve the efficiency of its own computing operations. It worked alongside Broadcom, which assisted with chip design and manufacturing coordination. Google subsequently recognized the processors’ potential for AI workloads and began making them available to customers outside its own cloud-computing operations.
That commercial expansion is attracting major technology companies, including Apple, Meta Platforms and Oracle.
Bloomberg Intelligence estimates that Google’s TPU sales, including associated software and hardware, could exceed $135 billion by 2028, compared with an estimated $12 billion this year. The projection excludes revenue from customers that access the processors through Google Cloud, suggesting that the broader economic value of the business could be larger.
Mandeep Singh, an analyst at Bloomberg Intelligence, said the chips’ price-performance proposition is a major selling point. TPUs cost about 30% less than Nvidia’s GPUs while delivering competitive throughput, or the volume of data that can be processed over a given period.
That cost advantage is necessary for companies operating large AI workloads, where computing expenses can become a significant part of the cost of developing and deploying applications.
Jump Trading is among the customers using Google’s processors. The trading firm has spent several years testing TPUs before signing a larger agreement with Google about a year ago, according to its chief technology officer, Alex Davies.
Jump has thousands of the chips, which it uses to train AI models for applications such as trading research. The company continues to rely mainly on Nvidia for training, although Davies said a vocal minority of employees prefer TPUs for particular workloads.
He said the processors perform well in selected applications and that using different chip architectures can help companies approach technical problems in new ways.
“It’s actually attractive to have some problems being done differently,” Davies said. “You learn stuff.”
Davies also pointed to the energy-efficiency benefits of TPUs in workloads where they perform as well as or better than Nvidia’s processors. Lower power consumption can reduce the cost of running AI systems, particularly as data center operators face increasing electricity requirements.
Google’s customer base is expanding through partnerships with other cloud providers. Oracle announced last month that its cloud customers would be able to access TPUs for the first time starting next year. Previously, Oracle offered AI chips from Nvidia and AMD.
“We partnered with Google because customers need more choice as AI workloads become more demanding,” Karan Batta, an Oracle senior vice president overseeing cloud infrastructure, said in a statement. “This partnership expands our ability to help customers build and run AI on the infrastructure that best fits their needs.”
Anthropic, the company behind the Claude chatbot, is another major customer. A recent regulatory filing indicated that it has an option to purchase as much as $150 billion worth of Google’s chips between 2028 and 2031. OpenAI is also planning to purchase TPUs, Bloomberg News has reported.
These agreements give Google access to customers that might otherwise depend heavily on Nvidia, while increasing the commercial reach of its semiconductor business.
Supply Constraints and Competition Threaten Growth Projections
Google’s position as both an AI model developer and chip manufacturer provides an advantage in developing its processors. TPU engineers have early access to the company’s Gemini models, allowing them to test hardware performance, identify software bugs, and improve power efficiency.
That integration matters in semiconductor development, where design errors can lead to costly production problems and months of delays. Vahdat said advances in AI are helping reduce some of those risks.
However, the company faces growing competition from startups and major AI developers seeking to develop their own processors.
Tenspresso, a startup co-founded by two former Google TPU engineers, is developing a chip focused on long-context inference, putting it in direct competition with the market targeted by Google’s 8i Max. Bloomberg reported last week that the startup’s product was still at least a year away from completing its design.
OpenAI and Anthropic are also pursuing in-house chip development programmes staffed by former TPU engineers. Their efforts could eventually reduce their dependence on external suppliers, including Google, although the commercial viability and timing of those projects remain uncertain.
Another major constraint is the availability of components and manufacturing capacity. Google said in its earnings report that every TPU it currently has available is already contracted to a customer. Shortages of memory and limited semiconductor production capacity have made it difficult to meet demand.
The company is attempting to diversify its supply chain beyond Broadcom. It recently signed an agreement with Marvell Technology that includes some TPU-related work, although the announcement provided few details. MediaTek has also been reported to be involved in Google’s TPU business. Vahdat declined to comment on the roles of either company.
William Tong, an analyst at Bloomberg Intelligence, said Google was likely working towards longer-term supply agreements to secure manufacturing capacity and reduce bottlenecks.
The broader investment environment presents another risk. Bloomberg Intelligence’s Singh estimates that opposition to data centers could reduce by about 5% to 7% the $1.4 trillion technology companies were expected to spend in 2027. That projected expenditure covers chips, networking equipment, land and other infrastructure required for data center construction.
Google must also contend with the fact that its TPU profit margins are less robust than Nvidia’s, according to Singh. A larger market share will therefore not necessarily translate into equivalent profitability.
The company’s immediate advantage is that demand for specialized AI computing remains strong and its existing supply is already committed. Its challenge is to expand production, secure manufacturing capacity and maintain a competitive cost structure as customers seek alternatives to Nvidia.
The latest TPU configurations strengthen Google’s position in two areas likely to become more important as AI systems evolve: secure processing of sensitive information and the ability to handle increasingly large workloads.