Signals Inbox·July 25, 2026·AI Chips

Can Google eventually kill Nvidia?

Google can break Nvidia's near-monopoly in hyperscale AI, especially in inference, but killing Nvidia itself would require TPUs to become a broad, portable platform far beyond Google's own cloud.

Q1What would Google killing Nvidia actually mean?

Google would have to make Nvidia replaceable across the wider AI market, and Google is currently nowhere near that threshold.

Google removing Nvidia from most internal workloads would hurt Nvidia, although it would settle only one part of the contest. Alphabet is one of the world's largest infrastructure buyers. Its expected capital spending now sits between $180 billion and $190 billion for the year, so every major workload kept on TPUs represents a large pool of GPU sales that Nvidia never receives.

A broader victory would require outside companies to treat TPUs as a normal choice. Developers would need to move models between TPUs and GPUs without a painful rewrite. TPU capacity would need to appear across more data centres and purchasing channels. Customers would also need confidence that Google would support their preferred frameworks, networking choices and deployment models for years.

The highest bar is making Nvidia irrelevant. Nvidia's latest quarterly Data Center revenue reached $75.2 billion, up 92% from a year earlier. That business spans cloud providers, AI laboratories, industrial companies, governments, robotics firms and scientific computing. Google would need to replace far more than one accelerator family to dismantle that position.

For this article, "kill Nvidia" means stripping Nvidia of its role as the default AI-computing platform and pushing the company into a narrower supplier position. Google can damage Nvidia badly without reaching that point.

What would count as Google killing Nvidia?

Q2Why is Google a much more serious threat to Nvidia now?

Google has become more dangerous to Nvidia because TPUs are moving beyond Google's walls and into large external infrastructure deals.

For years, TPUs mainly protected Google's own economics. Google could train models, rank content and serve recommendations without sending every dollar of accelerator spending to Nvidia. Cloud customers had access, although the hardware remained closely tied to Google's data centres and software stack.

Alphabet's latest quarterly filing shows a real change. Google Cloud has signed a limited number of agreements to supply multiple gigawatts of TPU hardware for specialised, high-scale workloads in on-premises environments. Alphabet says the first revenue should arrive during the year, with most coming later. Google has moved beyond renting TPU time inside its cloud and started supplying hardware for infrastructure that customers require or control.

The product roadmap is moving faster too. Ironwood, Google's seventh-generation TPU, is generally available and scales to 9,216 chips in one pod. Google has announced separate eighth-generation systems for large-scale training and for inference, although both remain listed as coming soon. The near-term threat is smaller than the marketing headlines suggest.

Google is designing separate systems for workloads whose economics have started to diverge. Training rewards throughput across very large clusters. Reasoning and serving care more about memory, communication latency and the cost of generating each token. A supplier that tunes hardware to each workload can take margin from a general-purpose platform.

Google now has mature internal demand, a credible cloud channel and the beginnings of external hardware distribution. A few years ago, that last part barely existed. Now it does.

Q3Has Google already reduced its dependence on Nvidia?

Google can already build and run frontier AI without relying on Nvidia for everything, although Google Cloud still needs Nvidia GPUs to win outside customers.

Google says TPUs power Gemini and flagship products including Search, Photos and Maps. That covers model training, inference and recurring consumer workloads measured at enormous scale. Google can develop a major model, serve it to users and improve the underlying hardware without waiting for Nvidia's roadmap.

The financial benefit reaches beyond avoided chip purchases. During Alphabet's latest earnings call, management said it had cut the cost of core AI responses in Search by more than 30% after hardware and engineering improvements. Alphabet did not break out how much came from TPUs, models, serving software or other infrastructure, so we should not assign the entire gain to custom silicon. Still, the result shows why Google wants tight control over the complete system.

Google Cloud follows a different logic. Outside customers arrive with existing software, trained engineers and procurement policies. Many have already standardised on Nvidia software or want the option to move between AWS, Azure and Google Cloud. A cloud provider that refused to offer Nvidia would lose business before it changed those habits.

So Google Cloud keeps both options available. Customers can rent TPUs for Google-optimised economics and Nvidia GPUs for compatibility, portability or specific model requirements. The same company can compete with Nvidia internally while acting as one of Nvidia's distribution partners externally.

Google now has choice. That is the valuable part. Nvidia remains useful to Google, but Nvidia no longer controls whether Google can build and run frontier AI at scale.

Q4Are Google's latest TPUs genuinely competitive with Nvidia GPUs?

Google's latest TPUs are good enough to beat Nvidia on some huge AI workloads, but nobody has proved that Google wins across the board.

The hardware is serious. Ironwood is available in pods reaching 9,216 chips, and Google designed it for dense models, mixture-of-experts systems, training and decode-heavy inference. The upcoming TPU 8t scales to 9,600 chips per superpod and adds features aimed at large pre-training jobs. TPU 8i takes a different route, with 288 GB of high-bandwidth memory, much more on-chip memory and a lower-latency topology for serving and reasoning.

Google claims TPU 8t can deliver up to 2.7 times better training performance per dollar than Ironwood. It claims TPU 8i can improve inference performance per dollar by up to 80%. Those comparisons show a rapid internal roadmap, but they compare new Google hardware with older Google hardware. They do not settle the contest against Nvidia's latest Blackwell and Rubin systems.

The newest MLPerf Training round also needs careful reading. Google and Nvidia both participated among 24 submitting organisations, while Nvidia submitted across the complete suite. Public benchmark results can compare time to a defined training target, yet they rarely reveal the full cost of power, networking, engineering time, capacity reservations and utilisation inside a real production service.

Google has the strongest case when it knows the workload deeply and can tune the model, compiler, interconnect and data centre together. Nvidia remains the easier choice for teams that run varied models, experiment constantly or need the same environment across several providers.

TPUs clearly belong in the top tier of AI accelerators. The public evidence still falls well short of a universal Nvidia replacement.

Where Google's latest TPUs stand against Nvidia

Q5Do major AI companies really want Google TPUs?

Major AI companies now want Google TPUs in enormous quantities, but the customer base is still narrow and highly concentrated.

Anthropic supplies the clearest evidence. The Claude developer first announced plans to use up to one million Google TPUs, representing tens of billions of dollars and more than one gigawatt of capacity. Anthropic later expanded the relationship with Google and Broadcom to multiple gigawatts of next-generation TPU capacity, calling it the company's largest compute commitment so far.

Anthropic's order is strong evidence that TPUs can handle serious frontier workloads. The company trains large models, serves demanding enterprise customers and understands infrastructure costs at scale. Reserving gigawatts of capacity would make little sense as a ceremonial partnership gesture.

The way Anthropic buys compute is equally revealing. Anthropic says Claude runs across Google TPUs, AWS Trainium and Nvidia GPUs. Amazon remains its primary cloud and training partner, and a separate agreement gives Anthropic access to as much as five gigawatts of Amazon capacity. The company is matching workloads to different systems and protecting itself from dependence on one supplier.

This buying pattern is spreading among large AI customers. They want custom silicon for price-sensitive or predictable workloads, GPUs for flexibility, and several infrastructure partners for resilience. Google wins a large order without forcing Nvidia out of the customer's architecture.

Breadth is still missing. We have one exceptionally large named customer and a limited number of undisclosed on-premises contracts. Hundreds of ordinary enterprises, universities and AI startups have yet to reorganise their stacks around TPUs.

TPUs now look credible to elite AI buyers. Ordinary customers are another story.

Q6Can Google sell TPUs beyond Google Cloud?

Google can now place TPU hardware outside Google Cloud, although today's model consists of custom deals for giant buyers rather than a normal chip business.

Alphabet's disclosure of multiple-gigawatt on-premises agreements changes the old TPU model. Large customers can secure Google accelerators for infrastructure they require or operate, rather than consuming every TPU through a standard Google Cloud service.

The scale tells us who these customers are likely to be. A gigawatt-class deployment involves vast data-centre capacity, energy contracts, networking and long planning cycles. Only a small group of model companies, cloud operators, governments and major technology businesses can make commitments of that size.

Nvidia reaches a far wider market. Its systems are sold through cloud providers, server manufacturers, original equipment manufacturers, distributors, specialised AI clouds and direct infrastructure projects. A mid-sized company can access Nvidia hardware through several routes and hire engineers who already know the environment.

Google's TPU sales and support network remains much smaller. Google may prefer to keep it that way, since broad merchant distribution would weaken some Google Cloud differentiation and force support for many more hardware configurations outside Google's carefully controlled data centres.

A selective hardware strategy may suit Google better. Google can keep its deepest optimisation inside its own products, offer TPU services through Google Cloud, and negotiate direct systems deals with customers large enough to justify custom support.

This selective approach can remove substantial Nvidia demand. It will struggle to replace the everyday hardware platform used across the rest of the market.

Q7Is Google Cloud big enough to spread TPUs widely?

Google Cloud is now big enough to turn TPUs into a major business, but Nvidia still reaches far more customers across far more environments.

Google Cloud produced just over $20 billion of revenue in Alphabet's latest quarter, up 63% from a year earlier. Operating income reached $6.6 billion, compared with $2.2 billion a year before. Alphabet also reported $462.3 billion of Google Cloud backlog, although that figure includes long contracts, short contracts and much more than TPU demand.

Several fresh usage figures show that AI is driving more than presentation slides. Alphabet said revenue from products built on its generative AI models grew nearly eightfold year over year. Over the previous 12 months, 330 Google Cloud customers each processed more than one trillion tokens, and 35 passed ten trillion tokens. Those customers create a natural audience for Google's own accelerators.

Google still has less distribution than the two larger public clouds. Synergy Research Group currently estimates global cloud-infrastructure shares at roughly 28% for AWS, 21% for Microsoft and 14% for Google. More importantly, Nvidia hardware appears across all three platforms, plus Oracle, specialised AI clouds and private data centres.

A developer can often keep an Nvidia-based stack while changing infrastructure provider. Moving deeply into TPUs usually creates a closer relationship with Google, even as framework support improves.

Google Cloud can grow a very large TPU business from its current position. Replacing Nvidia's reach would require either a much larger cloud share or a distribution model that lets TPUs spread well beyond Google's own platform.

Q8Can Google beat Nvidia by making AI cheaper?

Google's best route around Nvidia is cheaper production AI, because Google earns money from Search, Gemini and Cloud even when nobody buys a standalone TPU.

Nvidia's latest quarterly gross margin was about 75%. That number shows how much value Nvidia is capturing from scarce hardware, sophisticated systems and strong software. It also gives Google an enormous incentive to bring more of the stack in-house.

Google can judge a TPU by the cost of a useful outcome. The relevant unit might be a completed Search response, a Gemini answer, a recommendation, a translated sentence or a training step. A custom chip that looks less flexible on paper can still win if it handles a recurring Google workload at lower total cost.

The latest Search cost reduction gives this strategy a concrete shape. Alphabet says the cost of core AI responses fell by more than 30% following hardware and engineering improvements. Across billions of interactions, even a smaller sustained saving would be worth more than the revenue of many semiconductor companies.

Specialisation should become more valuable as AI workloads separate. Different memory systems, interconnects and on-chip components can remove costs that a broader architecture carries for flexibility.

Public claims leave a large blind spot. Google compares TPU generations with one another, while Nvidia compares new Nvidia systems with older Nvidia systems. Neither company publishes enough matched production data for outsiders to calculate the full cost per useful token under identical conditions.

Google does not need to win every benchmark. Lower costs on Google's largest recurring workloads would take a meaningful share of AI value away from Nvidia even if Nvidia remained the preferred platform elsewhere.

Q9Can Google finally loosen Nvidia's CUDA grip?

Google is making TPUs much easier for mainstream AI developers, but CUDA still gives Nvidia the stronger software ecosystem today.

CUDA has had roughly two decades to accumulate libraries, tools, training material, optimised kernels and experienced engineers. Nvidia publicly describes an ecosystem containing millions of developers and thousands of accelerated applications. The advantage reaches far beyond large language models into simulation, science, healthcare, industrial design, robotics and data processing.

Google has attacked the most obvious friction. TPUs support JAX, PyTorch, Keras, XLA and vLLM. Google's new TorchTPU stack runs ordinary PyTorch tensors directly on TPUs and aims to reduce the unusual code patterns developers previously associated with PyTorch/XLA. Google has also launched a dedicated TPU Developer Hub.

Google has just added first-class Ray support for Google Cloud TPUs. Ray is widely used to coordinate distributed Python and AI workloads, so official support removes another reason for teams to stay exclusively on GPUs.

Migration still requires work. PyTorch's own documentation explains that GPU execution and traditional PyTorch/XLA execution use different models. Compilation behaviour, sharding, synchronisation and changing input shapes can all affect performance. A model running successfully on a TPU does not guarantee that the complete production system will run efficiently without tuning.

Google is closing the accessibility gap faster than before. CUDA still has the deeper ecosystem, the larger labour pool and the reassurance that the same platform is available through competing infrastructure providers. Avoiding CUDA is getting easier for some workloads, but it remains the default for most serious AI teams.

Q10Is Nvidia still ahead of Google as a complete AI platform?

Nvidia remains ahead of Google as the complete platform for outside customers, largely because Nvidia can sell every layer into almost any environment.

Nvidia's position now includes accelerators, CPUs, NVLink, InfiniBand, Spectrum-X Ethernet, BlueField data-processing units, rack-scale systems and a large collection of software libraries. Calling Nvidia a GPU vendor misses how much of the data centre it already sells.

Nvidia generated $60.4 billion from Data Center compute and $14.8 billion from Data Center networking in its latest quarter. Networking revenue grew 199% year over year and already represented nearly one-fifth of Data Center sales. Custom accelerators can displace some Nvidia compute while leaving room for Nvidia interconnects, switches or software.

Nvidia is actively designing for mixed infrastructure. NVLink Fusion allows customers building custom CPUs or accelerators to connect those components with parts of Nvidia's platform. Nvidia can earn money from a data centre even when every important processor does not carry an Nvidia logo.

Google also owns a full stack: TPUs, Axion CPUs, optical networking, compilers, data centres, models and applications. Inside Google's environment, that stack can be extremely efficient. Outside Google, customers encounter a narrower supplier network and fewer deployment options.

Nvidia can sell one broad platform into many companies' data centres. Google's strongest setup still works best inside Google's own data centres and services, which leaves Nvidia with the broader commercial position today.

Q11Are Nvidia's biggest customers building escape routes?

Nvidia's largest customers are actively reducing their dependence, and that trend now poses a bigger threat to Nvidia's margins than any single Google chip.

Nvidia's annual filing shows how concentrated the business has become. One direct customer represented 22% of fiscal-year revenue and another represented 14%. Those two direct customers accounted for 36% combined, up sharply from the concentration reported two years earlier. Nvidia does not identify them, and direct customers can include manufacturers buying for end users, so the percentages should not be assigned casually to familiar hyperscaler names.

The customers controlling the largest AI budgets have every incentive to create alternatives. They can save supplier margin, design around recurring workloads and gain leverage in negotiations. Google has TPUs. Amazon has Trainium and Inferentia. Microsoft has Maia. Meta says it is developing four MTIA generations within two years and already deploys hundreds of thousands of MTIA chips.

Yet large buyers are diversifying rather than making one clean switch. Meta's custom-silicon plan starts with more than one gigawatt and aims for several gigawatts over time. During the same period, Meta signed a long-term Nvidia infrastructure partnership and an agreement for up to six gigawatts of AMD GPUs.

A large technology company can use custom chips for stable internal inference, merchant GPUs for frontier training and several suppliers to reduce delivery risk. Nvidia loses exclusivity long before it loses the account.

The danger for Nvidia is gradual. More workloads become contestable, buyers gain bargaining power and Nvidia must provide more value around the chip to defend its margin. Google is one major force in that change, alongside several other customers that are becoming chip designers themselves.

Q12Will AI inference become Google's best chance to take Nvidia share?

AI inference gives Google its clearest chance to take large volumes from Nvidia because repeated production workloads reward narrow optimisation.

Frontier training changes quickly. Researchers alter model architecture, precision, routing and communication patterns, so flexibility remains valuable. A programmable platform with broad tools can justify a higher price when teams are still discovering what the next model needs.

Inference often becomes more predictable after deployment. The same model serves similar requests again and again, sometimes billions of times. Small improvements in memory use, latency, utilisation or energy consumption then compound across an enormous number of tokens.

Google's TPU roadmap follows that logic. TPU 8i is designed for post-training, serving and reasoning. Its architecture uses more on-chip memory and a network topology intended to reduce the delays created when mixture-of-experts models move information between chips. Google claims an 80% performance-per-dollar gain over Ironwood for selected low-latency inference workloads.

Google also owns the demand. Search, Gemini, Workspace, YouTube, advertising, Maps and Android can all generate recurring inference. The company sees the request patterns, controls the models and can change the service around the chip.

Other hyperscalers see the same opportunity. Microsoft built Maia 200 around inference economics, and Meta says its next MTIA designs will focus heavily on generative-AI inference. Nvidia will probably feel the pressure through thousands of production workloads moving onto several custom chips, not through one dramatic benchmark defeat.

Nvidia is responding with faster inference systems and lower claimed token costs. Google still has the best opportunity where Google controls both the workload and the bill.

Q13Does Google's control of chips, models and products give it an unfair advantage?

Google's control of chips, models and products gives Google a real cost advantage inside Google, but it does not automatically win outside customers.

Google DeepMind can work with TPU engineers before a model reaches production. Infrastructure teams can prepare networking, storage and serving systems for workloads that Google expects years ahead. Search and Gemini then provide immediate demand at a scale few companies can match.

Google can justify a TPU through lower infrastructure spending, faster Search responses, stronger Gemini products or more competitive Google Cloud pricing. The chip programme can create huge value without earning an Nvidia-style margin as a standalone business.

Alphabet can finance that strategy comfortably. The company generated $45.8 billion of operating cash flow in its latest quarter and spent $35.7 billion on capital expenditure. Its full-year capital-spending plan has climbed to $180 billion to $190 billion. Those figures cover data centres, servers, networking and other infrastructure rather than TPUs alone, but they show the scale supporting Google's hardware roadmap.

An outside customer pays for that efficiency with deeper dependence on one provider's capacity, tools and future priorities. A tightly integrated Google stack works best when the model, compiler, cloud and hardware all fit together.

Nvidia offers a different kind of safety. Its platform appears across rival clouds and private infrastructure, making supplier changes easier even when the underlying accelerator remains Nvidia.

Google's integration can make Google the most efficient operator of Google AI. Turning that strength into the universal preference of outside customers will be much harder.

Q14Could Amazon, Microsoft and Meta weaken Nvidia more than Google does?

The combined custom-chip push from Amazon, Microsoft, Meta and Google can shrink Nvidia's share faster than Google could on its own.

Amazon has already reached meaningful production scale. AWS says almost one million Trainium2 chips are training and serving Claude, and Anthropic has secured as much as five gigawatts of additional Amazon capacity. Microsoft's Maia 200 is running in Azure for inference workloads, including Microsoft products and OpenAI models. Meta deploys hundreds of thousands of MTIA chips and is accelerating its release cycle.

Custom AI silicon has become a normal hyperscaler strategy. Each company has enough recurring demand to absorb design costs and optimise around its own software.

The fragmentation also helps Nvidia. Google TPUs, AWS Trainium, Microsoft Maia and Meta MTIA do not form one shared alternative platform. Each comes with different availability, compilers, systems and commercial terms. A model company seeking portability may prefer Nvidia rather than maintaining four deep hardware optimisations.

The result could look odd: Nvidia's market share falls as custom chips grow, yet Nvidia's revenue keeps rising because total AI spending expands even faster. Nvidia's Data Center revenue nearly doubled in the latest quarter despite years of custom-chip investment by its largest customers.

Google benefits when workloads move specifically to TPUs. Google gains little when Amazon or Microsoft captures the same workload with its own silicon. The group can weaken Nvidia's dominance without creating a new Google monopoly.

How the major custom-chip platforms compare with Nvidia

Q15What would still have to happen for Google to truly kill Nvidia?

Google would need several difficult wins at the same time, while Nvidia would have to make a series of major mistakes.

TPUs would first need broad distribution. Gigawatt contracts prove Google can serve a few giant customers, but Nvidia reaches organisations of almost every size through clouds, manufacturers, distributors and private systems.

Google would then need near-frictionless software portability. The recent progress around native PyTorch, Ray, JAX and vLLM is real. Customers still face different compilation behaviour, debugging tools and optimisation work when they move a mature GPU stack.

The economics would also need public proof across diverse workloads. Google's internal performance-per-dollar claims are encouraging, though customers cannot infer a universal advantage from comparisons with earlier TPU generations.

Google Cloud would have to close much more of the distribution gap or let TPUs spread through competing channels. A 14% cloud share creates a major business; it does not match hardware offered across nearly the entire cloud market.

Finally, Nvidia would need to stumble. A serious loss of software momentum, delayed hardware and failure to support mixed systems could turn customer diversification into broad abandonment. Nvidia's current numbers still show rapid Data Center growth and gross margins around 75%.

Google can achieve some of these conditions. Achieving all of them before Nvidia adjusts looks remote.

Q16Can Google eventually kill Nvidia?

No. Google is unlikely to kill Nvidia, although it can break a large part of Nvidia's grip on hyperscale AI computing.

Google has already proved that a technology company with enough internal demand can build frontier-grade AI infrastructure outside Nvidia's platform. TPUs power Google's biggest products and attract gigawatt-scale external customers. Google has also started placing TPU hardware in selected on-premises environments, while the software barriers keep falling.

Those advances will cost Nvidia real business. Google can keep more Search and Gemini spending inside Alphabet, take selected training and inference workloads from outside customers, and force Nvidia to defend its pricing with better systems, software and networking.

Nvidia currently remains too broad, too large and too adaptable for Google to remove. Nvidia's platform is sold across competing clouds and private data centres. CUDA still carries the deepest software ecosystem. Networking has become a large business of its own, giving Nvidia ways to earn revenue even inside mixed clusters. Customers building custom chips continue to sign enormous GPU agreements because no single architecture covers every workload well.

What happens next will probably look messy. Google will run more of its own AI on TPUs and sell capacity to a small group of very large customers. Amazon, Microsoft and Meta will keep expanding their own chips. Nvidia loses exclusivity and some margin, but remains the broad platform connecting this fragmented industry.

Google can eventually kill Nvidia's near-monopoly, especially in hyperscale inference and tightly controlled cloud workloads. Google is very unlikely to kill Nvidia itself.