Google is turning the custom AI chips it once reserved for its own use into a product it sells to outside companies - and bankrolling the customers who buy them, just as Nvidia did on its way to a $5 trillion valuation - in its most direct challenge yet to the chip leader's dominance.
Alphabet, Google's parent and now the second-most-valuable company in the world by market value, spent roughly a decade designing its Tensor Processing Units, or TPUs, as a private edge - chips used to train and run its own models and rented out by the slice through Google Cloud. It is now selling them directly into customers' data centers and bankrolling the AI labs that buy them, a shift that puts it on a collision course with Nvidia and borrows heavily from the strategy that carried its rival to the top of the global market.
A shift in strategy, lifted from the leader's manual
The clearest signal came on Alphabet's first-quarter earnings call in late April, when chief executive Sundar Pichai said the company would begin shipping TPUs "to a select group of customers in their own data centers." For a business that had always kept its silicon firmly inside its own facilities, the move marked a real pivot - and a deliberate expansion of the market Google is willing to chase.
Pichai framed the decision as opportunistic rather than a wholesale reinvention, noting that leasing capacity through the cloud often delivers a better long-term return because Google keeps earning from a chip across its entire lifespan. But demand, he said, was coming from AI labs, financial-trading firms and high-performance computing users who want the hardware on their own premises. Chief financial officer Anat Ashkenazi cautioned that revenue from hardware sales would be lumpy from quarter to quarter and would only become material to Alphabet's books in 2027.
The deeper borrowing from Nvidia is financial. Over the past two years Nvidia did not just sell chips; it bankrolled the companies that buy them - taking stakes in OpenAI, Elon Musk's xAI and Mistral, backing "neocloud" providers such as CoreWeave, Nscale and Nebius, and pairing long-term purchase commitments with investment. Google is running the same play. It has ploughed money into Anthropic - reported to run to tens of billions of dollars across successive deals - and Anthropic, in turn, has committed to consuming enormous quantities of TPU capacity.
Critics call this the "circular financing" that increasingly defines the AI economy, in which a small group of giants act at once as suppliers, customers, investors and cheerleaders. Supporters counter that tying capital to supply contracts is simply how you lock in scarce computing power and guarantee demand. Either way, it is unmistakably the Nvidia model - and Google has the cash to run it. Alphabet has guided to capital spending of roughly $180 billion to $190 billion this year, most of it on AI infrastructure, with another increase flagged for 2027.
New silicon, split in two
Google's hardware has grown more ambitious as well. At its Cloud Next conference in Las Vegas in April, it unveiled its eighth generation of TPUs and, for the first time, split the lineup in two: the TPU 8t, built for training models, and the TPU 8i, tuned for inference - the everyday work of answering user prompts once a model is live.
Google says the training chip offers about 2.8 times the performance per dollar of its seventh-generation "Ironwood" predecessor, which became broadly available late last year, and can shrink the development cycle for cutting-edge models "from months to weeks." The inference chip, which leans on fast on-chip SRAM in a manner similar to rivals such as Cerebras and Groq, is pitched as delivering roughly 80% better performance per dollar. The company says more than a million TPUs can be linked together in a single cluster. Tellingly, Google declined to benchmark the new chips directly against Nvidia's top hardware.
Behind the scenes, Broadcom turns Google's designs into manufacturable products and supplies key components, while TSMC handles fabrication - which is why Broadcom's shares tend to move on TPU news.
The customers Nvidia cannot afford to lose
The most striking part of Google's push is who is signing up. Anthropic, the maker of the Claude models, has agreed to take multiple gigawatts of next-generation TPU capacity, though it has not abandoned anyone: it runs a deliberate three-way strategy spanning Google's TPUs, Amazon's Trainium chips and Nvidia's GPUs to avoid being tied to a single supplier.
More worrying for Nvidia, Meta - one of its single largest buyers - has reportedly struck a multiyear, multibillion-dollar deal for TPUs, with capacity rented from Google Cloud first and chips bound for Meta's own data centers from 2027. Google is also said to be lining up TPU capacity for OpenAI, the standard-bearer of the rival Nvidia-and-Microsoft camp. Factor in trading firms and other frontier labs, and the addressable market starts to look enormous: the investment bank D.A. Davidson has estimated that Google's chip operation, together with its DeepMind research arm, could ultimately be worth around $900 billion.
Nvidia still towers - but the ground is shifting
None of this means Nvidia is in trouble today. It still controls an estimated 80% to 90% of the AI data-center chip market, depending on whose figures you use, and its data-center division alone generated roughly $194 billion in its most recent fiscal year. The company has guided to a combined $1 trillion in orders for its Blackwell and forthcoming Vera Rubin platforms across 2026 and 2027, and its CUDA software ecosystem remains a formidable source of lock-in that Google's tools do not yet match.
The relationship is also more tangled than a straight duel. Google has agreed to work with Nvidia on networking so that Nvidia-based systems run more efficiently inside Google Cloud, and Nvidia is itself among Anthropic's backers. Many analysts therefore see TPUs as diversification and negotiating leverage for big buyers rather than a wholesale replacement for Nvidia's chips, at least for now.
The longer-term threat is subtler. As more of Nvidia's biggest customers design or buy their own silicon, the danger is less a collapse in volumes than an erosion of the fat margins and pricing power the company enjoys today. And hanging over the whole contest is Jensen Huang, Nvidia's famously combative co-founder, whose readiness to defend his turf has long made rivals and partners alike think twice before crossing him.
What 2026 looks set to mark, then, is not the end of Nvidia's reign but the start of a genuinely contested market - one in which the world's second-biggest company is wagering tens of billions of dollars that it can claim a slice of the most important hardware franchise of the era, using the very flywheel the leader built first.
Comments
Join the discussion and share your perspective.