Signals Inbox·July 20, 2026·AI Chips

Google TPU vs Nvidia GPU: who will win?

Nvidia is winning the broad AI computing market today, but Google TPUs are becoming the most credible custom-chip threat in the hyperscale workloads where cost, power and tight system design matter most.

Q1Why are Google TPUs and Nvidia GPUs being compared so much now?

Google TPUs and Nvidia GPUs are being compared more seriously now because Google has started turning its private chip advantage into a business that outside companies can actually buy.

For years, Google mainly designed TPUs to run Search, recommendations, advertising systems and its own AI models. Nvidia played a much broader role, selling GPUs to nearly every cloud provider, research laboratory and company building AI products. The overlap was real but limited.

That boundary has lately become much less clear. Anthropic first committed to using up to one million Google TPUs, representing well over one gigawatt of capacity. It then signed another agreement for multiple gigawatts of next-generation TPU capacity starting in 2027. Alphabet has also said it will deliver TPU hardware to selected customers for installation inside their own data centers.

Google has strengthened the product side at the same time. Ironwood, its seventh-generation TPU, is now available through Google Cloud. The company has also presented TPU 8t for large training jobs and TPU 8i for inference and reasoning workloads, although those eighth-generation systems are not yet generally available.

Nvidia has responded with Vera Rubin, a platform combining GPUs, CPUs, networking, storage processors and a dedicated low-latency inference accelerator. The competition now covers chips, rack-scale systems, networking, software and complete data-center architecture.

Q2Why is it still hard to say whether Google TPU or Nvidia GPU is winning?

Nvidia GPUs are clearly ahead in sales today, while Google TPUs create a lot of value that Alphabet never reports as chip revenue.

Every GPU that Nvidia sells to Microsoft, Amazon, Oracle, CoreWeave or an AI laboratory appears somewhere in Nvidia's financial results. A TPU used to train Gemini, answer a Search query or avoid an Nvidia purchase may generate value through advertising, cloud services or lower internal costs instead.

That accounting difference can make Google look smaller than it really is. Alphabet does not disclose TPU revenue, the number of TPUs deployed or the savings created by using its own chips. It combines TPU services with GPUs, Workspace, cybersecurity, databases and many other products inside Google Cloud.

We therefore need three scoreboards. Commercial sales show which company has captured customer spending. Actual workloads show which hardware is trusted with important AI systems. Software and distribution show which platform can spread beyond a few giant buyers.

Google TPU vs Nvidia GPU: current scoreboards

Q3Who is making more money from AI chips today: Google TPU or Nvidia GPU?

Nvidia GPUs currently generate vastly more commercial revenue than Google TPUs. The gap is large enough that accounting differences cannot explain it away.

In Nvidia's latest reported quarter, Data Center revenue reached $75.2 billion. Compute products contributed $60.4 billion, while networking added another $14.8 billion.

Alphabet reported $20.0 billion of Google Cloud revenue over its latest quarter. That number includes Google Cloud Platform, Workspace, databases, cybersecurity, Nvidia GPU services and TPU services. Nvidia's Data Center business was therefore 3.76 times larger than Google's entire cloud division.

Even Nvidia's compute revenue alone was three times Google Cloud revenue. Since only part of Google Cloud comes from TPUs, the gap between externally sold Nvidia computing and externally sold TPU computing must be considerably larger.

Google's internal TPU fleet still has substantial economic value. Each workload moved from a purchased GPU onto a Google-designed chip can reduce costs and supplier dependence. Those savings do not put Google close to Nvidia in the commercial market, though.

Latest reported quarterly revenue comparison

Q4Which is growing faster right now: Google TPU or Nvidia GPU?

Nvidia's AI infrastructure business is currently growing faster than Google Cloud, despite already operating from a much larger base.

Nvidia Data Center revenue rose 92% year over year and 21% from the previous quarter. Google Cloud grew 63% year over year and approximately 13% sequentially, from $17.7 billion to $20.0 billion.

The dollar comparison is even more revealing. Nvidia added roughly $12.9 billion of Data Center revenue in one quarter. Google Cloud added approximately $2.3 billion across its entire product portfolio. Nvidia created about 5.6 times more incremental quarterly revenue.

Google's strongest growth metric sits further ahead in the sales cycle. Google Cloud backlog increased from $240 billion to $462 billion in one quarter. Alphabet said strong enterprise AI demand and the inclusion of TPU hardware agreements drove the increase.

Most of that backlog still comes from normal Google Cloud contracts, and Alphabet expects much of the TPU hardware revenue later rather than immediately. Nvidia is converting demand into reported revenue faster today. Google is building a pipeline that could narrow the gap later.

Q5Is real customer demand for Google TPUs finally taking off?

Google TPU demand has become commercially serious, although most of the disclosed volume still comes from a small group of enormous customers.

Anthropic provides the clearest proof. Its first expansion covered up to one million TPUs, tens of billions of dollars and well over one gigawatt of capacity. The subsequent agreement with Google and Broadcom added multiple gigawatts of next-generation TPU capacity expected from 2027.

That second commitment arrived only months after the first one. Anthropic described it as the company's largest computing commitment at the time, giving Google evidence of repeat demand rather than a single experiment.

Alphabet then added TPU hardware agreements to Google Cloud's backlog and announced that selected customers would receive TPU systems inside their own facilities. Those contracts will begin generating some revenue during 2026, with most expected during 2027.

Google has also named customers outside the usual frontier-model market, including financial-computing companies and robotics businesses. The disclosed scale around those relationships remains limited, so they do not carry the same weight as Anthropic.

The picture is clear enough: Google has found external buyers for TPUs, including at least one buyer willing to commit at multi-gigawatt scale. It still needs to show that ten or twenty independent companies will go nearly as deep.

Q6Are Google TPUs actually cheaper than Nvidia GPUs?

Google TPUs can cost much less than Nvidia GPUs on carefully optimized workloads, but customers have to include the engineering work required to reach those savings.

A recent technical study compared Google TPU systems with two Nvidia H100 GPUs while fine-tuning and serving the 31-billion-parameter Gemma 4 model. Under the researchers' setup, TPU training finished 1.61 times faster and cost 2.12 times less.

Inference throughput differed by less than 3%, but the TPU setup delivered the first token in 235 milliseconds compared with 475 milliseconds on the H100 setup. Across the combined training and serving workflow, the researchers calculated a 1.82-times cost advantage for the TPU configuration.

Those figures are useful, but they do not settle the wider debate. The experiment used a Google model and an implementation rebuilt around Google's hardware. The researchers changed the sharding setup, data pipeline, checkpointing process, module names and model-conversion workflow.

A company running one stable model millions of times can spread those migration costs across huge volumes. A smaller team changing models every few weeks may save more money by staying on familiar Nvidia infrastructure. It is a less glamorous advantage, but a very real one.

Google says TPU 8t will improve training performance per dollar by as much as 2.7 times over Ironwood, while TPU 8i should improve inference performance per dollar by up to 80%. Those figures compare Google's new chips with Google's previous generation. Independent comparisons with Nvidia Rubin will be needed once both systems are widely available.

Q7Can Google TPU beat Nvidia GPU at training frontier AI models?

Google TPUs can handle frontier-model training today, but Nvidia GPUs still have much stronger public evidence across different models, customers and computing environments.

Google has already trained major Gemini, Gemma and other DeepMind models on TPUs. Anthropic also uses TPUs to train and run Claude. These workloads prove that Google's architecture can operate at the highest level of AI development.

Ironwood can connect 9,216 chips inside one pod and supports both dense and mixture-of-experts models. Google's upcoming TPU 8t will increase that scale to 9,600 chips per superpod, with a network designed to connect more than 134,000 chips within one fabric and support training clusters beyond one million TPUs.

Nvidia currently has the stronger independent benchmark record. In the latest MLPerf Training round, Blackwell systems recorded the fastest submitted result across every benchmark and were the only platform represented on every test. Nvidia partners also demonstrated clusters using up to 8,192 Blackwell GPUs across production cloud environments.

Google participated in the benchmark round, but the public results did not provide a comparable TPU sweep across the new DeepSeek, GPT-OSS and other workloads. Buyers can examine a wider set of standardized results for Nvidia than for Google TPUs.

Nvidia also lets an AI laboratory move the same broad software stack between several clouds, specialist GPU providers and privately owned clusters. Google can deliver excellent training performance for customers willing to build around its architecture. Nvidia is still easier to trust before anyone knows what the next model will look like.

Q8Is Google TPU now better than Nvidia GPU for AI inference?

Google TPU has a credible architectural edge for some large, repetitive inference workloads, while Nvidia GPU remains the safer choice across a mixed set of models and applications.

Google's current Ironwood systems support both training and inference, but the design of TPU 8i shows where Google thinks the market is going. TPU 8i is built specifically for serving, sampling, reinforcement learning and long reasoning chains.

Each TPU 8i chip will contain 288 GB of high-bandwidth memory and three times more on-chip SRAM than Ironwood. More SRAM allows frequently reused information, including parts of a model's key-value cache, to remain closer to the processor during long-context generation.

Google has also changed the network. Its Boardfly design reduces the maximum network path in a comparable 1,024-chip setup from 16 hops to seven. A separate collectives engine is designed to accelerate the communication steps that often slow autoregressive generation.

Nvidia's Rubin GPU offers broader raw capability and sits inside a more flexible system. Vera Rubin NVL72 combines 72 GPUs with 36 CPUs, high-speed networking and dedicated data-processing hardware. Nvidia claims up to ten times the inference throughput per watt and one-tenth the token cost of its previous Blackwell platform on selected reasoning workloads.

Both companies are comparing their newest systems with their own older products, so the percentages should be treated as vendor claims. TPU 8i also remains a future product, whereas Google's currently available offering is Ironwood.

Google TPU vs Nvidia GPU for AI inference

Q9Is CUDA still Nvidia's biggest advantage over Google TPU?

CUDA remains Nvidia's biggest advantage because it lets customers use GPUs without reorganizing their entire engineering operation around the hardware.

Nvidia said at GTC that more than six million developers had contributed to the CUDA ecosystem. CUDA also connects to thousands of optimized applications and a large collection of libraries for model training, inference, data processing, simulation, robotics and scientific computing.

The value comes from accumulated habits as much as code. Engineers already know the debugging tools. Cloud platforms already support the software. Companies can recruit people with CUDA experience, reuse existing kernels and move workloads between providers.

Google has recently made TPUs much easier to use. Ironwood supports JAX and PyTorch, while Google has placed native PyTorch support for its next TPU generation into preview. Google is also building around vLLM, Kubernetes and familiar model-serving interfaces.

The gap has narrowed, but the migration documentation still warns developers that differences between GPU execution and TPU execution can require changes to code and workflow. The Gemma comparison that showed strong TPU economics also documented considerable adaptation work.

Google's deeper advantage sits inside its own organization, where chip designers, model researchers and data-center teams can optimize together. Nvidia's advantage follows the customer wherever the customer chooses to run. That portability is still the stronger moat.

Q10Which has better distribution today: Google TPU or Nvidia GPU?

Nvidia GPUs currently have far better distribution than Google TPUs, even after Google's move toward selling hardware for external data centers.

Customers can access Nvidia systems through the largest cloud providers, specialist GPU clouds, server manufacturers and direct enterprise deployments. Nvidia says more than 80 MGX ecosystem partners are working on Vera Rubin infrastructure.

This breadth creates competition between Nvidia resellers. A company can choose a major cloud for integration, a specialist provider for price and availability or its own hardware for control. The underlying programming environment remains largely familiar when the provider changes.

Google TPUs are available through Google Cloud services such as Compute Engine and Google Kubernetes Engine. Alphabet's plan to deliver hardware to selected customers expands that route, but the word “selected” limits its significance for now.

Google does not yet have Nvidia's network of equipment manufacturers, resellers, financing partners and independent cloud operators. Companies also cannot move a TPU workload onto AWS or Azure in the same way they can relocate an Nvidia workload.

Distribution channels

Q11Is Google or Nvidia shipping new AI chip technology faster?

Nvidia is currently better at turning a new architecture into a broadly available product, while Google is making the more aggressive bet on specialized chips for separate AI workloads.

Nvidia moved from Hopper to Blackwell, Blackwell Ultra and Vera Rubin on an annual rhythm. Vera Rubin's seven-chip platform entered full production before the systems began reaching the wider market, and Nvidia can reuse much of its existing cloud and manufacturing network during each transition.

Google has also increased its pace. Trillium was followed by Ironwood, then by the announcement of TPU 8t and TPU 8i. Separating training and inference into different chips is a major design choice and suggests Google no longer believes one accelerator should be optimized for every phase of AI.

Availability separates the two roadmaps today. Google's own Cloud TPU documentation still describes Ironwood as its latest available TPU. Access to eighth-generation systems is being handled through interest forms and future deployment plans.

Nvidia's latest products generally move into established customer channels more quickly. Google's newest architecture may eventually deliver larger gains for its chosen workloads, but the market cannot fully test those gains yet.

Q12Are major AI companies leaving Nvidia GPUs for Google TPUs?

Major AI companies are spreading workloads across Nvidia GPUs and custom chips rather than abandoning Nvidia outright.

Anthropic illustrates the pattern better than any other customer. The company has committed to Google TPUs, uses Nvidia GPUs and runs more than one million Amazon Trainium2 chips. Anthropic says it assigns workloads to different processors according to performance, cost and resilience.

Its Amazon commitment is even larger than the currently disclosed Google agreement. Anthropic has agreed to spend more than $100 billion across ten years on AWS technology and secure up to five gigawatts of capacity. Amazon remains its primary cloud and training provider.

That does not reduce the importance of Google's win. A multi-gigawatt TPU agreement takes workloads that could otherwise have run on Nvidia hardware and proves that a frontier laboratory can operate across several architectures.

The wider pattern points toward partial substitution. AI companies need so much computing capacity that relying on one supplier creates operational and negotiating risk. Nvidia remains present in the largest accounts, but those accounts increasingly refuse to give Nvidia every workload.

Q13Can Google TPUs force Nvidia to lower AI chip prices?

Google TPUs can put real pressure on Nvidia's pricing among large buyers, even without coming close to Nvidia's overall sales.

Nvidia reported a gross margin of roughly 75% in its latest quarter. Part of that margin reflects superior technology, software and support. It also shows how valuable limited GPU capacity has become to customers.

Hyperscalers can attack that margin by moving predictable workloads onto their own chips. Google can use TPUs internally, rent them through Google Cloud and begin placing them inside customer data centers. Amazon, Microsoft and Meta are following the same general logic with their own processors.

A large customer does not need to replace Nvidia completely to gain leverage. Moving even a minority of training or inference volume elsewhere can strengthen its position during price, capacity and financing negotiations.

Nvidia is answering with rapid reductions in cost per token. The company says Rubin will produce reasoning outputs at one-tenth the cost of Blackwell under its selected benchmark conditions. Nvidia's ability to lower its own prices through better performance may become more important than defending a premium on each individual GPU.

Google's most likely near-term impact is margin pressure and workload fragmentation. Taking the market lead would require a much broader customer base.

Q14Which company can afford the Google TPU vs Nvidia GPU chip race for longer?

Google and Nvidia can both finance this competition comfortably, although they earn their returns in very different ways.

Alphabet expects capital spending of $180 billion to $190 billion during 2026, with most of the investment going into technical infrastructure. The company ended its latest quarter with $126.8 billion in cash and marketable securities and generated $45.8 billion of operating cash flow during the quarter.

Google can justify TPU spending without selling a single physical chip. Cheaper computing can improve Google Cloud margins, lower the cost of Gemini, support more AI features in Search and reduce the amount paid to Nvidia.

Nvidia funds research from direct infrastructure profits. Its latest quarter produced $53.5 billion of operating income on $81.6 billion of revenue. That gives Nvidia enormous resources for chip development, networking, software, investments and acquisitions.

Google treats custom silicon partly as a strategic cost advantage. Nvidia treats AI computing as its main profit engine. Neither company is likely to run out of money, so execution, software and customer adoption will decide the contest.

Q15Is Google now the biggest threat to Nvidia GPUs?

Google is currently Nvidia's strongest individual custom-chip challenger, but Nvidia faces more pressure from the combined custom-silicon movement than from Google alone.

Google has the longest record of deploying proprietary AI accelerators at hyperscale. It also controls frontier models, a major cloud platform, global data centers and enormous internal workloads across Search, YouTube, advertising and Gemini.

No other Nvidia customer combines those pieces quite as fully. Google can learn from its own workloads, change the chip, adjust the compiler and deploy the result inside the same organization.

Amazon has nevertheless secured an even larger disclosed capacity commitment from Anthropic. Microsoft, Meta and other major technology companies are also building custom processors, while AMD continues to compete in the general-purpose accelerator market.

These alternatives can divide Nvidia's market without any one competitor replacing it. Google may capture specialized inference, Amazon may win workloads tied to AWS, AMD may serve buyers demanding a second GPU supplier, and newer architectures may target low-latency applications.

Google presents the most complete version of the threat. The broader danger for Nvidia is that its customers are gradually becoming chip suppliers themselves.

Q16Could Google TPU eventually beat Nvidia GPU?

Google TPUs could eventually lead large-scale inference and tightly integrated AI systems, but Google has not yet shown a credible path to replacing Nvidia GPUs as the general market standard.

Google's best route runs through workloads with enormous volume and limited variation. When one company controls the model, compiler, networking and data center, a specialized processor can remove unnecessary flexibility and produce better economics.

That route could create a very large business. A handful of multi-gigawatt customers, combined with Google's internal use and cloud services, could move tens of billions of dollars away from Nvidia.

Winning the broader platform race requires more. Google would need several independent customers deploying TPUs deeply in production, a wide external hardware channel, smoother PyTorch portability and repeated benchmark results across models Google did not create.

Nvidia's position looks stronger when customers want freedom. GPUs work across clouds, model families, industries and research methods. That flexibility becomes especially valuable when the next important AI architecture remains uncertain.

Google has a plausible path to owning important parts of the market. Nvidia currently owns the platform around which most of the market organizes itself.

Q17Who is winning right now: Google TPU or Nvidia GPU?

Nvidia GPUs are clearly winning the overall AI computing market right now, while Google TPUs are becoming a serious threat in hyperscale workloads where cost and power efficiency matter most.

Nvidia leads on the three dimensions carrying the most weight. Its commercial business is several times larger than Google's entire cloud division. CUDA remains the familiar environment for most developers. Nvidia hardware can also be bought through far more clouds, manufacturers and infrastructure providers.

Google's progress should still worry Nvidia. Anthropic has made repeated TPU commitments at gigawatt scale, Alphabet has started signing external hardware agreements, and Google is building separate processors for training and inference. Those steps move TPUs well beyond their old role as internal Google equipment.

The lead remains wide because Google's strongest announcements mostly describe capacity arriving later. Nvidia's latest products already benefit from a mature channel, standardized benchmarks and a huge installed software base.

Three developments would change our judgment. Google would need to disclose several new large TPU customers, show substantial revenue from hardware installed outside Google Cloud and demonstrate that ordinary PyTorch teams can migrate without major rewriting.

Nvidia, meanwhile, needs Rubin deployments to deliver the promised cost reductions. It must also keep Data Center revenue growing as Anthropic and other large customers spread their workloads across custom chips.

Our verdict is firm. Nvidia is winning the broad and commercially valuable race by a clear margin today. Google has the stronger chance of winning selected hyperscale inference workloads, reducing Nvidia's pricing power and creating a large second platform. That is still a narrower victory than becoming the default infrastructure for AI.

Google TPU vs Nvidia GPU: current position by criterion