Google’s secret weapon in the artificial intelligence race is a chip that has helped the search group’s models leap ahead of OpenAI, prompting tech investors to reassess a new threat to Nvidia’s dominance.
The Big Tech group’s “tensor processing unit” has been central to efforts to boost the performance of Google’s new Gemini 3 AI models, which have outperformed OpenAI’s GPT-5 in independent benchmarking tests.
That development is one factor behind the ChatGPT maker’s “code red” last week, as chief executive Sam Altman told staff to refocus resources into improving its chatbot and models.
Google plans to more than double production of its TPUs by 2028, analysts predict, as it steps up its ambitions for a processor that chip consultancy SemiAnalysis says is now “neck and neck with king of the jungle Nvidia” for building and running cutting-edge AI systems.
Nvidia investors have also been rattled by the prospect of Google offering TPUs to customers beyond its own cloud computing platform. This includes a recent deal to provide AI start-up Anthropic with 1mn TPUs, worth tens of billions of dollars.
Google argues vertical integration — developing AI hardware, software and chips largely in-house — will deliver both technical advantages and huge profits.
Gemini 3, like Google’s previous models, was trained largely on TPUs. OpenAI relies mainly on Nvidia graphics processing units to build the large language models that power ChatGPT.
“The most important thing is . . . that full stack approach,” said Koray Kavukcuoglu, Google’s AI architect and DeepMind’s chief technology officer, “I think we have a unique approach there.”
Combining that with understanding how billions of consumers use Gemini, AI overviews in search and other products gives Google a huge advantage, he added.
Morgan Stanley estimates that every 500,000 TPUs sold to external customers could generate as much as $13bn in revenue for Google. The Big Tech company works primarily with chip design partner Broadcom, as well as with MediaTek, to develop the processors.
Morgan Stanley analysts also predict that Taiwan Semiconductor Manufacturing Company will produce 3.2mn TPUs next year, growing to 5mn in 2027 and 7mn in 2028.
“Growth in 2027 is significantly stronger than previously anticipated,” the bank’s analysts said in a recent research note.
Nvidia’s stock fell sharply last month following a report in The Information that Meta was in talks with Google to buy TPUs. Meta has declined to comment on the report.
Some analysts believe Google could also strike such deals with OpenAI, Elon Musk’s xAI or start-ups such as Safe Superintelligence, potentially driving upwards of $100bn in new Google revenues over the coming years.
Experts add that AI-enabled coding tools could make it easier for potential TPU customers to rewrite their software, which until now has largely been built on top of Nvidia’s proprietary Cuda platform.
Nvidia has sought to assuage market concerns, saying it was still “a generation ahead of the industry” and was “the only platform that runs every AI model”, adding: “We continue to supply to Google.”
The chipmaker added it offers “greater performance, versatility, and fungibility” than processors such as TPUs, “which are designed for specific AI frameworks or functions”.
The origins of Google’s TPU date back to an internal presentation in 2013 by Jeff Dean, Google’s long-serving chief scientist, following a breakthrough in using deep neural networks to improve its speech recognition systems.
“The first slide was: Good news! Machine learning finally works,” said Jonathan Ross, a Google hardware engineer at the time. “Slide number two said: “Bad news, we can’t afford it.”
Dean calculated that if Google’s hundreds of millions of consumers used voice search for just three minutes a day, the company would have to double its data-centre footprint just to serve that function — at a cost of tens of billions of dollars.
Ross, who now leads AI chip start-up Groq, said he started working on TPUs in 2013 as a side project as he happened to be sitting next to a team working on Google’s speech recognition technology.
“We built that first chip I think with 15 people,” Ross told a podcast interviewer in December 2023.
The project has scaled rapidly. One early application was 2016’s famous victory by Google DeepMind’s AlphaGo programme against the world champion of the board game Go, Lee Sedol. The match is considered a significant AI milestone.
The chips have for several years powered many of Google’s core services, including search, advertising and YouTube.
Google typically releases a new generation of TPU every two years, though that cadence has shifted to annual updates since 2023.
“Google Cloud is experiencing accelerating demand for both our custom TPUs and Nvidia GPUs,” said a Google spokesperson. “We are committed to supporting both, as we have for years.”
Additional reporting by Melissa Heikkilä and Hannah Murphy
Read the full article here
