Nvidia dominates the AI chips market, but Google has been developing its own specialised hardware since 2013.
Nvidia’s graphics processing units (GPUs) are general-purpose processors with thousands of small cores that can handle many tasks in parallel, making them useful for graphics, high-performance computing, and most AI workloads.
In contrast, Google tensor processing units (TPUs) are chips designed exclusively for the mathematical operations used in machine learning. They are engineered to perform highly complex matrix and tensor computations efficiently, meaning TPUs can train or run very large AI models at potentially faster speeds and lower costs than GPUs.
GlobalData estimates that the total AI market will be valued at $642bn in 2029, up from $131bn in 2024 at a compound annual growth rate (CAGR) of 37%. The generative AI segment is forecast to record a CAGR of 93% over the same period. As spending on generative AI increases, so will the demand for efficient training and inference chips. Nvidia has cornered the AI chip market, but its lead could be threatened by competition.
Google has been using TPUs internally since 2015. Since then, the company has improved memory, compute power, and TPUs’ ability to interconnect with other chips. Google’s TPUs are used today to power its Gemini AI assistant and all of Google’s AI-enabled products, including Search, Photos, and Maps.
Thus far, TPUs have been reserved for Google’s internal use or leased as a service to customers like Anthropic, Cohere, OpenAI, and Salesforce. In October 2025, Anthropic confirmed it would access up to one million TPUs.
In 2024, Apple opted to train its Apple Intelligence models on Google TPUs.
For the first time, Google may now be selling the TPUs to customers, not as a service but as a product.
On November 24, 2025, it was reported that Google was in talks with Meta to supply it with a large-volume order of TPUs. Meta, which is largely dependent on Nvidia GPUs, is ostensibly diversifying its compute resources after acquiring AI chip start-up Rivos in September 2025. OpenAI is following a similar strategy, making multibillion-dollar deals with Broadcom and AMD in October 2025. Diversifying their chip supply is an astute move by large language model (LLM) developers, driving market competition that will ultimately lower chip prices and incentivise new, more competitive chip designs.
In the final days of October 2025, another competitor entered the fray. Qualcomm launched its own AI chips; the AI200 series will be released in 2026, and the AI250 in 2027. These chips will be designed specifically for inferencing tasks rather than training.
Until now, Qualcomm has concentrated its chip efforts on the smartphone, laptop, and tablet markets. This move towards the AI chip market marks a strategic shift from its core business and adds another pressure point for Nvidia. These successive reports of AI chip diversification in the final quarter of 2025 have investors concerned about Nvidia’s future.
How do TPUs compare to GPUs?
While both GPUs and TPUs accelerate AI workloads, TPUs can reportedly handle more computations per second. They are also more cost-efficient and consume less power. However, LLM developers will need to factor in flexibility and deployment needs in addition to raw performance benefits when choosing between the two. For example, GPUs are highly versatile, widely supported across AI frameworks like CUDA, PyTorch, and TensorFlow, and capable of handling multiple workloads beyond AI, including graphics, simulations, and data analytics. GPUs are available for on-premise, cloud, and edge deployments, making them ideal for companies experimenting with different models or requiring multi-purpose compute.
TPUs, in contrast, are purpose-built by Google for large-scale deep learning. Despite their performance benefits, they are primarily limited to Google Cloud and optimised for TensorFlow workloads. As a result, businesses typically favor GPUs when they need flexibility, framework compatibility, or hybrid deployment options, and reserve TPUs for large AI projects where maximum efficiency and scalability are critical. Therefore, TPUs will not replace GPUs en masse but will be preferred for specific AI workloads.
The beneficiaries of Google’s TPU development
Diversification in the AI chip space provides greater security to LLM developers. It has also increased the competitive pressure on Nvidia, which may explain its numerous investments in AI and quantum companies as it seeks to remain a technology leader.
Despite its staggering revenue growth, the future of Nvidia’s monopoly is now in question. Of its total revenue, nearly 50% comes from just four customers and 85% from just six companies.
Nvidia does not disclose the names of its largest customers, but they are likely to include Amazon, Meta, and Microsoft. Google itself is also expected to be a major customer of Nvidia. With large sums of revenue dependent on so few buyers, shifts in market dynamics can have rapid consequences.
Should leading customers diversify their chip supply to the likes of AMD, Broadcom, Google, and Qualcomm and reduce reliance on Nvidia, the chip designer will find itself in a vulnerable position.
In contrast, other AI chip companies are enjoying tailwinds on the back of Google’s potential TPU commercialisation. Broadcom, a design partner of Google, saw its share price jump 10% immediately following the news. This, along with the news that Broadcom will be co-designing AI chips with OpenAI, puts the company at the center of the AI hardware conversation.
Google has worked with Broadcom on TPU development for nearly ten years; however, reports in March 2025 suggested it will also work with MediaTek to develop future generations of TPUs. A partnership with the Taiwanese chip designer may help further strengthen the existing liaison with TSMC, which has served as a foundry for Google since the TPU’s inception.
Similar to Nvidia, TSMC’s revenue is primarily derived from a few select major customers, with estimates that Nvidia and Apple account for a combined 40% of annual revenue. The wider commercialisation of TPUs will help to diversify TSMC’s revenue streams. Also based in Taiwan, Foxconn has become one of the world’s largest server vendors. On December 1, 2025, it was announced that it would supply large volumes of server equipment for Google TPUs.
If TPUs become more accessible, rival chip companies will need to demonstrate superiority across performance, cost-efficiency, and power consumption.
Nevertheless, the development will likely provide momentum to other chip designers who, until now, have been obscured by Nvidia’s dominance. These will include incumbents like AMD and Qualcomm as well as start-ups like Cerebras, Groq, Mythic, and Tenstorrent. There will also be the opportunity for these companies to develop processors that target niche markets, where both GPUs and TPUs struggle to deliver.
