Alphabet Inc.’s Google is expanding its artificial intelligence chip lineup as it challenges Nvidia Corp. in a market valued at hundreds of billions of dollars. The company is introducing two specialized versions of its tensor processing units, custom semiconductors designed to develop and operate AI software.
The TPU 8tx and 8i Max are reconfigured versions of existing designs rather than entirely new silicon. Google plans to announce the variations Monday, with availability expected next year, according to Google chip chief Amin Vahdat.
The rollout accelerates a development schedule that once produced a new Google AI processor every two years. Google released two processors at the same time this year, and the new variations will add more options to that pair.
Each of the four TPU offerings is intended to optimize different customer needs. “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.
The 8tx is designed for secure and confidential computing by government agencies and companies operating in regulated industries. It works with chips from Advanced Micro Devices Inc. and allows information to remain encrypted at the chip level while it is processed.
The 8i Max targets applications that require long context windows, meaning the amount of information an AI system can consider within one request. Vahdat said that capacity is needed for drug discovery, mathematics and coding applications.
Google developed TPU technology more than a decade ago with Broadcom Inc., initially seeking a more efficient way to manage its expanding computing requirements. The company later recognized its value for AI and began selling the processors beyond its cloud computing operation.
Bloomberg Intelligence projects that TPU sales, including related hardware and software, will exceed $135 billion by 2028. That compares with an estimated $12 billion this year and excludes revenue from customers accessing the processors through Google Cloud.
The chips have attracted customers including Apple Inc., Meta Platforms Inc. and Oracle Corp. Anthropic PBC also uses TPUs and has an option to buy as much as $150 billion of the chips between 2028 and 2031, according to a recent regulatory filing.
Oracle recently agreed to let its cloud customers use TPUs beginning next year, adding Google’s processors to existing choices from Nvidia and AMD. Oracle senior vice president Karan Batta said the partnership gives customers more options as AI workloads become increasingly demanding.
Jump Trading is another TPU customer and has thousands of the processors for training AI models used in areas including trading research. Chief Technology Officer Alex Davies said the firm still primarily uses Nvidia for training, although a “vocal minority” of employees strongly prefers TPUs for certain tasks.
Davies said TPUs perform well for a subset of workloads and can offer greater power efficiency where their performance matches or exceeds Nvidia chips. That efficiency can reduce costs, while using different processor types can help the firm approach problems in different ways.
Mandeep Singh, an analyst at Bloomberg Intelligence, said TPUs cost about 30% less than Nvidia GPUs while delivering competitive throughput. That price to performance proposition has appealed to customers in finance and other industries, even as Nvidia retains the dominant position in the market.
Nvidia is also moving toward more specialized systems, including its own central processing units and chips focused on inference, where trained AI models perform tasks. Google’s approach pairs the 8tx with an AMD central processing unit, while the 8i Max places two TPU chips into one package for additional memory and computing power.
Google still faces component shortages, manufacturing constraints and possible weakness in AI infrastructure spending. Singh estimated that opposition to data centers could reduce the projected $1.4 trillion in technology company spending by roughly 5% to 7% in 2027, while Google’s TPU profit margins remain below Nvidia’s.
Every available Google TPU is already contracted to a customer, according to the company’s earnings report. Google has signed an agreement with Marvell Technology Inc. that includes some TPU work, while Vahdat’s team is seeking additional supply sources to prevent manufacturing bottlenecks.