Google wants to chip away at that dominance with something very un-GPU-like: its own tensor processing units, or TPUs. The hardware has been around internally since 2015, but only in the past year has Alphabet begun pitching TPUs as a full-blown alternative to Nvidia’s flagship GPUs—backed by high-profile customers, aggressive cost claims and, most recently, talks with one of Nvidia’s biggest buyers, Meta Platforms. Investors are asking a simple question with a complicated answer: Do Google’s home-grown chips threaten Nvidia—or merely cap its upside?
A private rival goes public
TPUs started as a kind of house chip for Google’s own search and ad ranking. They are application-specific accelerators, tuned to chew through the linear algebra at the heart of modern AI. Unlike general-purpose GPUs, they strip away graphics baggage and much of the flexibility in favor of dense, low-precision matrix math and tight integration with Google’s TensorFlow and JAX software stacks.
Seven generations in, Google’s latest TPU family—codenamed Ironwood—boasts per-chip floating-point performance on par with Nvidia’s coming Blackwell-class parts, according to independent technical analyses. One such estimate pegs Ironwood at around 4.6 petaFLOPS of FP8 compute per chip, marginally above Nvidia’s B200 at about 4.5 petaFLOPS, both drawing roughly 600 watts.
Google isn’t just leaning on raw speed. It is selling the TPU story on economics. For certain inference workloads, third-party benchmarking suggests Google’s newer TPU v6e can deliver up to four times better performance per dollar than Nvidia’s H100, helped by tight coupling with Google’s software and data center fabric. Early adopters like Midjourney and Cohere have reported large cuts in inference costs and higher throughput after migrating portions of their workload to TPUs, while Google’s own Gemini models run at vast scale on internal TPU clusters.
The Meta moment
What was once an internal advantage is now becoming a market signal. Meta is in talks to rent Google Cloud’s TPUs as early as next year and potentially spend billions to deploy the chips directly inside its own data centers from 2027 onward.
For Nvidia, the timing is awkward. Meta is one of its largest customers, with an AI infrastructure budget estimated in the tens of billions of dollars this year. If a slice of that wallet migrates to Google silicon, Nvidia doesn’t just lose unit sales; it loses some of its leverage as the de facto standard for large-scale AI training. Analysts estimate Google is eyeing a prize equal to as much as 10% of Nvidia’s annual revenue if the Meta deal and other TPU wins scale.
Wall Street has noticed. Reports of a potential Meta-Google tie-up helped send Nvidia’s stock down and Alphabet’s up in late November, even as Berkshire Hathaway disclosed a new multibillion-dollar stake in Alphabet—widely read as a bet on Google’s vertically integrated AI stack, including TPUs.
What TPUs really threaten
Still, for all the noise, TPUs today represent a small fraction of deployed AI accelerators. One recent industry analysis estimates that Nvidia accounts for about 80% of the GPU market in 2024, while TPUs represent just 3–4% of deployment, potentially climbing to mid-single digits by 2025.
The gap isn’t just about silicon. Nvidia’s real moat is the CUDA software ecosystem, which has spent more than a decade cementing itself in university courses, open-source libraries and corporate AI stacks. Most cutting-edge models, frameworks and tools are built and optimized first for Nvidia hardware. Porting them to TPUs is possible—Google has invested heavily in tooling—but rarely free.
Then there is availability. Nvidia’s GPUs can be bought, leased or rented from a sprawling roster of cloud providers and system builders. TPUs, by contrast, are essentially a Google-only affair: They run in Google Cloud or, in the future, in the data centers of a few strategic partners such as Anthropic and, potentially, Meta. That limits their addressable market to customers willing to anchor their AI roadmap around Google’s cloud and software.
If Google wins big, it will be first in cloud, not in chips
In that sense, TPUs threaten Nvidia indirectly, by strengthening Google Cloud. Alphabet gains two forms of leverage: it keeps more of the AI-economics “stack” in-house—from models (Gemini) to chips (TPUs)—and it can use TPU pricing to undercut or pressure Nvidia-based offerings from rivals like Amazon Web Services and Microsoft. Analysts already see supply-chain beneficiaries—from TSMC and advanced packaging houses to memory suppliers—that would ride any TPU volume ramp.
But Nvidia has seen this movie before. Amazon’s own Trainium and Inferentia chips were also billed as cheaper, cloud-native alternatives for AI training and inference. They have found niches—Anthropic trains its Claude models on them, among others—but Nvidia’s share of the accelerator market remains dominant, and some startups report that AWS’s chips still lag in performance and ecosystem maturity.
The more likely outcome is not a sudden dethroning but a gradual pluralization of AI hardware. The very largest buyers—Google, Amazon, Meta, Microsoft, and a few others—have strong incentives to reduce dependence on Nvidia’s supply and pricing. They will push harder into in-house or semi-proprietary silicon for inference, where workloads are more predictable and price sensitivity is higher, while still leaning heavily on Nvidia for bleeding-edge training and for everything that needs CUDA’s flexibility.
What that means for Nvidia
For Nvidia shareholders, TPUs are best understood as a ceiling, not a cliff. They are a real check on how far Nvidia can push pricing and margins in the long run; they may divert some hyperscaler capex away from Nvidia’s highest-end parts; and they ensure that any stumble in Nvidia’s product roadmap will quickly be exploited by would-be rivals.
Yet the AI market itself is growing so fast—and is so supply-constrained—that alternative chips may simply expand the pie rather than slice it neatly. Recent estimates show total AI accelerator sales surging past $30 billion in a single quarter, with Nvidia still capturing the lion’s share.
At the same time, the stock of available Nvidia compute in the wild has been roughly doubling every 10 months, reinforcing its position as the default platform for AI research and deployment.
Google’s TPUs, then, look less like an existential threat and more like something subtler: a powerful bargaining chip for the cloud giants, a margin-squeezer at the high end of the market, and a reminder that in AI, the real contest is not just about the fastest chip—but who owns the stack around it.
