Signals Inbox·July 25, 2026·AI Chips

Is Google replacing Nvidia?

Google is replacing Nvidia in parts of its own AI infrastructure and winning a handful of enormous outside workloads, but Nvidia remains the default platform across the wider market.

Q1What would “replacing Nvidia” actually mean?

Google is currently replacing Nvidia inside parts of its own AI stack. Across the wider chip market, Nvidia remains firmly ahead.

There are three different outcomes to separate. The first is internal: Google shifts Gemini, Search and other large workloads from Nvidia GPUs to its own TPUs. The second is commercial: outside companies choose TPUs for serious production work. The third is much harder: TPUs become an industry standard that developers can use across clouds, data centers and software stacks.

Google has already passed the first test and is making visible progress on the second. The third still belongs comfortably to Nvidia. Google can therefore remove billions of dollars of potential Nvidia demand without becoming the new default supplier.

What replacing Nvidia would mean

Q2Why does Google suddenly look like a real Nvidia threat?

Google looks more dangerous to Nvidia because TPUs have moved from an internal advantage to hardware that Google is willing to host, sell and place inside customer data centers.

Google introduced separate eighth-generation chips for training and inference, giving the two designs clearer jobs than earlier generations. Anthropic committed to as many as one million TPUs in an agreement worth tens of billions of dollars. Google and Blackstone also announced a new TPU cloud backed by an initial $5 billion equity commitment and designed to bring 500 megawatts online.

The biggest break came in Alphabet’s latest earnings call. Google said it would begin shipping TPU hardware to a select group of customers for use in their own data centers, with most of the related revenue expected later.

That is a big change. Cloud TPUs have existed for years, but customers generally had to consume them on Google’s terms and inside Google’s infrastructure. Direct installations make the Nvidia comparison much more serious than it was a year ago.

Q3Has Google already replaced Nvidia inside Google?

Inside Google, TPUs already handle massive AI workloads that would otherwise require far more Nvidia hardware.

Google says TPUs train and serve Gemini at scale and power large products including Search. Its first-party models currently process more than 16 billion tokens per minute through direct customer API use, up from 10 billion in the previous quarter. We cannot see what percentage runs on each chip type, but the workload is large enough to shape Google’s entire hardware roadmap.

Nvidia GPUs remain a core part of Alphabet’s accelerator portfolio. Google uses TPUs where it controls the model, compiler, network and data center, while keeping Nvidia available for teams and cloud customers that already rely on CUDA.

So yes, Google has cut its internal dependence on Nvidia. The size of that reduction remains undisclosed, and Google is in no hurry to tell us.

Q4Are Google TPUs now as capable as Nvidia GPUs?

For frontier AI, Google TPUs are capable enough to win major workloads from Nvidia. No public benchmark proves that either platform is the universal winner.

TPU 8t can connect 9,600 chips and about two petabytes of shared high-bandwidth memory in one superpod. Google says it delivers three times Ironwood’s processing power and up to twice the performance per watt. TPU 8i targets post-training and high-volume inference, with Google claiming 80% better performance per dollar than its previous generation.

Nvidia gives customers a broader package. Its Vera Rubin platform combines 72 GPUs, 36 CPUs, NVLink, networking, storage acceleration and software in a rack-scale design. MLCommons’ latest training round included Google alongside 23 other submitting organizations, while Nvidia systems appeared through many cloud and server partners.

That breadth shows that Nvidia is easier to deploy across vendors. It does not tell us which chip wins every workload, and there probably is no such answer.

Google TPU and Nvidia platform comparison

Q5Are major AI companies actually choosing Google TPUs?

Major AI companies are choosing Google TPUs, usually as one part of a mixed Google, Amazon and Nvidia fleet.

Anthropic provides the strongest proof. Its commitment covers up to one million TPUs, more than one gigawatt of capacity and tens of billions of dollars. The company later expanded its Google and Broadcom partnership to cover multiple gigawatts of future compute. Google also names Midjourney and Salesforce among the teams using TPUs intensively.

Yet Anthropic describes a three-platform strategy: Google TPUs, Amazon Trainium and Nvidia GPUs. Amazon remains its primary cloud and training partner.

The agreement is a huge vote of confidence in TPUs. It is not a mass exit from Nvidia, and Anthropic has been unusually clear about that.

Q6Can Google TPUs finally escape Google Cloud?

TPUs are beginning to leave Google Cloud through a small number of channels that Google still tightly controls.

Alphabet has confirmed plans to ship hardware to selected customers for installation in their own data centers. The Blackstone joint venture adds a separate TPU cloud, with Google supplying the chips, software and services. Traditional Cloud TPU access remains available through Google Compute Engine, Google Kubernetes Engine and Vertex AI.

Nvidia’s reach is much wider. Its newest systems are due through AWS, Google Cloud, Microsoft Azure, Oracle Cloud and specialist providers including CoreWeave, Lambda, Nebius and Nscale. Cisco, Dell, HPE, Lenovo and Supermicro are also preparing systems.

Google now has a few serious routes to market. Nvidia has dozens.

Routes to the customer

Q7Can Google break Nvidia’s CUDA advantage?

Google has finally attacked the software problem seriously. CUDA still gives Nvidia the easiest path for most AI teams.

Google now promotes native support for PyTorch, JAX, SGLang and vLLM on its newest TPUs. TorchTPU, developed with Meta, is meant to run standard PyTorch code with minimal changes. Bare-metal access and better Kubernetes support also make TPUs feel less like a special research environment.

The installed base remains Nvidia’s biggest defense. AI teams have years of CUDA code, optimized kernels, monitoring tools and employee experience. Google’s documentation still contains TPU-specific steps around compilation, distributed workers and orchestration.

TorchTPU can make migration realistic. Only repeated production migrations will show whether it makes migration routine. There is quite a gap between those two things.

Q8Is inference where Google can hurt Nvidia most?

Inference is Google’s best chance to take meaningful volume from Nvidia because Google already controls some of the world’s largest repeated AI workloads.

Training changes constantly, so developers value flexibility. Mature inference is more predictable: the same models answer requests millions or billions of times, making cost per token, memory use, latency and electricity decisive. Google can study that traffic across Gemini, Search and its APIs, then design the chip and software around it.

Alphabet says it cut Gemini serving costs by 78% during 2025. Google also claims TPU 8i improves inference performance per dollar by 80% over the previous TPU generation.

Those comparisons do not tell us whether TPU 8i beats Nvidia on a neutral workload. They do explain why Google keeps specializing its silicon. At enough volume, an internal cost advantage becomes a very large external business.

Q9Why does Google still buy Nvidia GPUs?

Google still needs Nvidia because many cloud customers arrive with CUDA software and expect their code to run immediately.

Alphabet calls Nvidia GPUs a core part of its AI accelerator portfolio. Google Cloud already offers Hopper and Blackwell instances and expects to be among the first providers with Vera Rubin systems. Refusing Nvidia would send valuable customers to AWS, Azure, Oracle or a specialist GPU cloud.

Selling both also protects Google. TPUs can serve workloads that fit Google’s stack especially well, while Nvidia covers customers that want portability, familiar software or a different model architecture.

Google has no good reason to turn this into an ideological fight. It makes money either way.

Q10Are custom AI chips getting big enough to hurt Nvidia?

Custom AI chips are now large enough to pressure Nvidia’s future market share while Nvidia’s own business keeps growing.

Broadcom gives us the best public numbers because it supplies custom accelerators and AI networking to several large customers. Its latest quarterly AI semiconductor revenue reached $10.8 billion, up 143% from a year earlier. It then guided to $16 billion for the following quarter, which would equal a $64 billion annualized pace if sustained.

That total includes networking and several customers, so it cannot be treated as Google TPU revenue. Even so, the scale is hard to dismiss. Google, Amazon, Microsoft and Meta are all building internal accelerators, and Broadcom’s reported AI business moved from $8.4 billion to $10.8 billion in consecutive quarters before the $16 billion guidance.

Custom silicon is no longer a side project used to annoy Nvidia during contract negotiations. It has become one of the largest pools of AI infrastructure spending outside Nvidia.

Q11Is Nvidia already losing business to Google?

Google is taking potential orders from Nvidia today. Nvidia’s reported demand is still accelerating.

Nvidia’s latest quarterly data-center revenue reached $75.2 billion, up 92% from a year earlier and almost double the $39.1 billion reported in the comparable quarter. The company also guided to roughly $91 billion of total revenue for the next quarter. Whatever Google is taking, Nvidia is still selling nearly twice as much data-center hardware as it was a year earlier.

The scale comparison is brutal for Google. Nvidia’s data-center revenue in one quarter was about 3.8 times Google Cloud’s entire $20 billion quarterly revenue, which includes Workspace, databases, cybersecurity, storage, Nvidia instances and many other services.

Google does not disclose TPU revenue separately. We can see displacement at the workload level, but there is no evidence yet that Google has materially slowed Nvidia’s overall business.

Q12Are Google TPUs actually cheaper than Nvidia GPUs?

Google has proved that TPUs cut its own AI costs. Outsiders still lack a clean modern comparison against Nvidia.

Alphabet’s 78% reduction in Gemini serving costs shows major progress, and its latest earnings call said the cost of core AI responses in Search fell by more than 30% after an upgrade to Gemini 3. Google Cloud customers have also reported strong TPU price-performance, although public examples usually involve older chip generations or specific workloads.

None of those figures isolates the chip. Better models, lower numerical precision, software changes, higher utilization and data-center engineering can all reduce costs. Google’s 80% performance-per-dollar claim for TPU 8i also uses the previous TPU generation as the baseline.

Google’s internal efficiency is credible. A neutral, apples-to-apples comparison with Nvidia is still missing.

Q13Why is Google spending up to $185 billion if TPUs are cheaper?

Cheaper TPUs are helping Google consume more AI while its total infrastructure spending keeps rising.

Alphabet expects full-year capital expenditure of between $175 billion and $185 billion. Google Cloud’s backlog has reached $240 billion, and management said Cloud revenue would have been higher if more computing capacity had been available.

Lower costs let Google serve more responses, use longer reasoning, train larger models and add AI to more products. It reinvests the savings into volume, so the total bill keeps rising.

This is the awkward economics of AI infrastructure: making each unit cheaper does not reduce spending when demand expands even faster.

Q14Could Google hurt Nvidia without overtaking it?

Google can weaken Nvidia’s pricing power and market share without ever becoming the largest AI-chip supplier.

Large buyers can move stable, predictable workloads onto their own chips and reserve Nvidia GPUs for projects that need flexibility. That gives them leverage when negotiating prices and capacity. It also reduces the number of Nvidia GPUs they would have bought in a world without custom silicon.

Anthropic already spreads Claude across TPUs, Trainium and Nvidia GPUs. Other hyperscalers are following the same broad logic with their internal chips.

Nvidia can remain the biggest supplier while losing part of each customer’s future expansion, especially in high-volume inference. That is probably the more realistic threat.

Q15What would prove that Google is truly replacing Nvidia?

Google would need broad, repeatable customer switching before we could say it is truly replacing Nvidia.

Several independent AI labs would need to make TPUs their main platform for training or inference. Direct TPU sales would have to grow beyond a few selected customers, and independent clouds or server companies would need to offer the hardware at scale.

Existing PyTorch and CUDA workloads would also have to move with little engineering effort and keep their performance in production. Finally, Nvidia’s hyperscale growth, pricing or margins would need to weaken in a way that could reasonably be tied to custom chips.

Current evidence has not reached that bar. Not close, really.

Q16Is Google replacing Nvidia?

Google is partly replacing Nvidia: it is taking selected hyperscale workloads while Nvidia remains the industry’s default AI platform.

The claim is true inside Google and increasingly true for a small group of enormous customers. Google now has competitive training and inference systems, serious outside commitments and direct hardware sales. Its software is becoming easier for ordinary PyTorch teams to use as well.

Across the full market, the claim is exaggerated. Nvidia still has the broader software base, many more buying channels, stronger public benchmark coverage and vastly larger disclosed AI-chip revenue. Its latest growth shows that Google’s progress has not materially slowed demand.

Google is reducing Nvidia’s grip on hyperscale AI. Nvidia still sets the standard everywhere else.