Google is suddenly at the center of the artificial-intelligence trade. It started with the celebrated release of the company’s new Gemini 3 AI model, which was trained on Google’s own AI chips. The rally picked up steam on a report from The Information that Meta Platforms was in talks with Google to buy those chips, known as Tensor Processing Units, to fill an artificial-intelligence data center—the domain of Nvidia’s red-hot graphics processing units, or GPUs.
Alphabet’s shares are up 12% since the debut of Gemini 3 on Nov. 18 while Nvidia’s are down 3.4%. Google-parent Alphabet is now worth $3.86 trillion, making it the third largest company in the world. It’s hot on the heels of No 1 Nvidia ($4.38 trillion) and No. 2 Apple ($4.1 trillion). Google’s success has also brushed off on Broadcom, which helps design the TPUs. Its stock is up 16% since the latest Gemini launch.
As Google was ramping its own AI efforts in the 2010s, it had the same problem that other Nvidia customers have since run into: Traditional servers weren’t up to the task, and Nvidia’s hardware was expensive, and hard to get in the large quantities Google needed. At the scale Google operates, it needed an in-house solution.
Google’s TPU made its debut in 2015. Before anyone outside the company even knew about the new hardware, it was starting to power the back end of many Google products like Maps, Photos, and Translate.
Fast forward to 2025, and Google is on its seventh generation of TPUs. The company continues to use them internally for its own products. More recently, Google has been able to find a few outside customers who otherwise would probably have been doing the same AI work on Nvidia hardware. Apple trained its Apple Intelligence models on TPUs, and AI start-up Anthropic, which has a $350 billion valuation, has a TPU deal as part of its multicloud strategy.
Gemini 3 has impressed users, and refocused attention on TPU as an alternative to Nvidia for running AI workloads. Just as Google discovered in the 2010s, Nvidia GPUs remain expensive and in short supply compared with ravenous demand, and no one wants to be reliant on one vendor for AI’s most important infrastructure.
Since 2021, Nvidia’s data-center revenue has risen by 2,400%. Nvidia hardware became the standard for AI research in 2012. When the current AI boom began in late 2022, all those years of hard work placed Nvidia right in the middle of the action.
Nvidia’s chips are still called graphics processing units though their use has gone well beyond driving personal computer monitors. GPUs are very good at splitting tasks into many pieces and running them side-by-side, which is how gamers can play at 120 frames per second. This is also what makes them so useful for AI calculation, as well as other high-performance domains.
TPUs, on the other hand, do one thing—matrix math for deep learning—but they do it very well. Under the right circumstances they can provide a much better cost structure than Nvidia GPUs. Deep learning has been the main thrust of AI research for over a decade now, leading us to the large language models that are powering new AI applications like chatbots and coding assistants. GPUs can do a lot more, but for many current AI loads, TPU is like a bullet train. It only goes from one place to another, but if that’s all you want to do, it’s very fast.
There is no question that customers would prefer to not be so dependent on Nvidia, but, for now, they are. That’s how Nvidia achieved an astounding gross profit margin of 73% in the third quarter, nearly a 300% markup. Customers are paying up for Nvidia products because alternatives like TPUs still don’t meet their full needs.
But nothing lasts forever. Intel long dominated the data-center chip market. In a three-way race with Nvidia and Advanced Micro Devices, Intel earned 65% of data-center chip revenue in the first quarter of 2021. By 2024, Nvidia had over 80% of the market, and Intel’s share had dwindled to single digits.
The other major cloud players have AI chips of their own, and there is a rush into this space from other parties. The strategy employed by Anthropic may be a glimpse into the future. It has large contracts with Amazon Web Services, Microsoft Azure, and Google that employ Nvidia GPUs, Google TPUs, and Amazon.com’s custom hardware called Trainium. Anthropic is spreading out its vendors and chip use, lowering its counterparty risk.
At some point, Nvidia’s market share and gross margin will come under attack. Google’s Gemini success is causing some investors to wonder if that time is now. But I’m betting it’s still a ways off, because Nvidia’s market share is protected by the company’s software, not its hardware.
Starting in 2004, Nvidia began building software known as CUDA. The idea was that GPUs could be powerful for tasks beyond graphics, but they were difficult to program. CUDA allowed developers to use common programming languages like C, which then got compiled into something a GPU can understand.
Today, just about all AI researchers know how to use CUDA; far fewer understand Google’s software, which is much less mature.
We will know when, and if, the competition becomes material to Nvidia when the company’s prized gross margins begin to slide, indicating that Nvidia is lowering prices to protect sales.