Signals Inbox·July 25, 2026·AI Chips

Can Google turn TPUs into an Nvidia rival?

Yes—Google can turn TPUs into a serious Nvidia rival for frontier training and hyperscale inference, but only if it can turn a few enormous deployments into a broader software and distribution platform.

Q1Why are Google TPUs suddenly being treated as an Nvidia threat?

Google TPUs look far more threatening today because Google has started selling them as outside infrastructure instead of keeping them mostly inside its own cloud and products.

The change is easy to see in three recent moves. Anthropic plans to use up to one million Google TPUs, with the expansion worth tens of billions of dollars and expected to add well over one gigawatt of capacity. Alphabet has also told investors that selected customers will receive TPU hardware in their own data centers. Soon after, Blackstone committed an initial $5 billion to a separate TPU cloud that aims to bring 500 megawatts online in 2027.

Taken together, those moves go well beyond another chip launch. Google has offered Cloud TPUs for years, but customers generally had to consume them through Google Cloud. Now Google is testing three routes at once: its own cloud, installations inside large customers’ facilities and a separately financed compute provider.

The hardware roadmap has become more serious as well. Google is now on its eighth TPU generation and has split it into one system for large training jobs and another for inference and reinforcement learning. That is the kind of product planning we expect from a company building a long-term computing platform.

So the question has moved on. Google already knows how to build an excellent AI accelerator. What remains uncertain is whether it can turn that accelerator into a widely used business before Nvidia’s software, distribution and system-level lead becomes even harder to challenge.

Q2What would it actually mean for Google TPUs to rival Nvidia?

Google TPUs would count as an Nvidia rival once they take a meaningful and lasting share of high-value AI training and inference, even if Nvidia remains much larger overall.

Several different claims are often mixed together. Google has already reduced its own need to buy Nvidia GPUs for every major AI workload. It is also winning some external work through Google Cloud. The much harder step is becoming a supplier that customers can use across clouds, partner facilities and their own data centers.

Google does not have to replace Nvidia in gaming, scientific computing, robotics, rendering and every other accelerated workload. For this article, “rival” means becoming one of the small number of platforms that a frontier AI lab or very large enterprise can seriously choose for its next training cluster or inference fleet.

That still sets a demanding bar. The choice must work at scale, remain attractive across several chip generations and avoid trapping the customer in an awkward software environment. A single giant contract proves that TPUs are commercially real. A broad platform requires repeated wins from unrelated customers.

What it would mean for TPUs to rival Nvidia

Q3Are TPUs powerful enough for frontier AI models?

TPUs are already powerful enough for frontier AI models, so raw capability is no longer the main doubt.

Google has trained and served generations of Gemini models on its own infrastructure, alongside demanding systems such as Veo, Imagen and AlphaFold. Those workloads require more than fast matrix multiplication. The chips, memory, network and software have to keep thousands of accelerators working together for long periods without wasting too much capacity.

Anthropic provides the stronger outside test. It trains and runs Claude across Google TPUs, AWS Trainium and Nvidia GPUs, then assigns work to each platform according to cost, availability and technical fit. Its decision to expand Google capacity to as many as one million TPUs would make little sense if the platform could not handle production models at frontier scale.

Google’s published system work shows that this performance has improved across several generations, so it is no one-off success. A recent paper from Google engineers traces roughly a 100-fold rise in peak performance per TPU node and a 3,600-fold increase in supercomputer performance across five generations. Earlier TPU v4 research described clusters of 4,096 chips and compared them favorably with contemporary A100 systems on both speed and power.

None of this proves that TPU 8 beats Nvidia Rubin on every model. Public comparisons rarely use identical software, scale, utilization and pricing. It does settle the basic question: a leading AI lab can build and operate frontier models on TPUs today.

Q4Has Google convinced outsiders to use TPUs at meaningful scale?

Google has won meaningful outside TPU demand, although the public customer base is still too concentrated to resemble Nvidia’s.

Anthropic’s planned Google expansion is measured in gigawatts and tens of billions of dollars, far beyond a small trial cluster. The company says its teams have used TPUs for several years and cites their price-performance and efficiency.

There are other customers, including Midjourney and Salesforce, but Google does not publish TPU revenue, rented capacity or the number of large outside users. We cannot yet tell whether Anthropic sits at the front of a growing group or accounts for a very large share of the visible business.

Anthropic’s own strategy keeps the result in perspective. Amazon remains its primary cloud and training partner, and Anthropic now says it uses close to one million Trainium2 chips. It has committed more than $100 billion over ten years for up to five gigawatts of additional AWS capacity while continuing to use Nvidia GPUs.

Frontier labs are becoming multi-chip buyers. They want capacity, bargaining power and resilience more than loyalty to one architecture. Google has earned a seat at that table, which is already a major achievement. It has yet to prove that many less specialized customers will follow.

Q5Are Google TPUs actually cheaper than Nvidia GPUs?

Google TPUs can be cheaper for stable, high-volume AI workloads, but there is no honest public evidence that they cost less for everyone.

Google’s strongest published numbers compare new TPUs with older TPUs. TPU 8i is advertised at up to 80% better inference performance per dollar than the previous generation, while TPU 8t is designed to improve training economics. Those figures show rapid internal progress, but they do not give buyers a clean TPU-versus-Rubin comparison.

Production results are more revealing. Alphabet said the cost of core AI responses in Search fell by more than 30% after an upgrade to newer Gemini models, helped by hardware and engineering changes. In an earlier 18-month period, Google reduced the machine cost of AI Overviews by more than 90% even as the underlying model became larger.

Those savings come from the full serving system: model design, quantization, batching, caching, compilers, networking and chips. Google’s control of that entire chain is where its cost advantage can come from. A buyer pays for completed training runs or generated tokens, not a chip specification in isolation.

Nvidia can recover some apparent hardware disadvantage through mature libraries, easier debugging and better utilization. A cheaper accelerator becomes expensive when engineers spend months adapting code or leave part of the cluster idle. TPUs can win on cost in workloads Google understands deeply; Nvidia remains the safer economic choice across a wider range of jobs.

Q6Why did Google split TPU 8 into separate training and inference systems?

Google split TPU 8 because training and inference now create different bottlenecks, and one compromise chip would leave too much performance unused.

Large pre-training jobs need huge amounts of compute and fast communication across thousands of chips. TPU 8t therefore targets throughput, embedding-heavy models and clusters of 9,600 chips in one superpod. Google also designed its Virgo network to connect many superpods for much larger jobs.

Inference and reinforcement learning put more pressure on memory, latency and repeated communication. TPU 8i carries more high-bandwidth memory and much more on-chip SRAM, while a dedicated collectives engine is meant to reduce the time chips spend waiting for one another.

The split gives Google a chance to undercut a more general platform on the workloads that dominate its own spending. Search, Gemini APIs and large cloud customers produce enormous volumes of fairly predictable requests. Small improvements in token cost can save serious money.

There is a trade-off. Nvidia supports a much wider variety of models and non-AI workloads, which makes its platform easier to reuse when plans change. Google gains efficiency by specializing more aggressively, but customers need confidence that their future models will still map well onto the TPU design.

How Google split its eighth-generation TPU systems

Q7Can Google build and deploy enough TPUs?

Google can fund TPU deployment on an Nvidia-rival scale, but power, construction and equipment availability could still slow it down.

Alphabet now expects capital spending of roughly $180 billion to $190 billion this year and says spending should rise significantly again next year. About 60% of its latest quarterly technical-infrastructure investment went into servers, with the remaining 40% going into data centers and networking.

Demand is already pressing against supply. Alphabet has acknowledged that limited compute capacity held back Google Cloud revenue. At the same time, Google must divide new infrastructure among Gemini, Search, DeepMind research and Cloud customers.

A million accelerators cannot simply be shipped into an empty building. Large AI clusters need high-bandwidth memory, optical networking, substations, cooling systems, backup equipment and trained operators. Delays in any one of those areas can leave expensive chips waiting for the rest of the site.

The Blackstone venture offers a practical way around part of the problem. Outside capital will pay for facilities and operations while Google supplies the TPUs, software and services. Nvidia expanded rapidly because clouds, server makers and specialist providers invested their own money around its platform. Google is starting to create a similar multiplier, although the first 500 megawatts will not come online until 2027.

Q8Is CUDA still Nvidia’s strongest defense?

CUDA remains Nvidia’s strongest defense because moving a model involves far more than translating a few lines of code.

Nvidia says CUDA now has more than six million developers and supports nearly 6,000 accelerated applications. The important part is the accumulated work behind those figures: libraries, optimized kernels, debugging tools, documentation, trained engineers and production habits built over two decades.

Google has strong tools of its own. JAX and XLA are widely respected among advanced AI teams, and PyTorch/XLA lets PyTorch models run on TPUs. Google has also put a native, eager-first PyTorch backend into preview. That is useful progress, although the official migration guide still warns that TPUs use a different execution and compilation model, so developers may need to change code and working habits to get good results.

That friction affects customers unevenly. Anthropic or Google DeepMind can assign specialist teams to optimize a cluster worth billions of dollars. A smaller company usually prefers the system its engineers already know, especially while model code is changing every week.

Google can partly avoid a direct CUDA fight by hiding the hardware. Companies using Gemini APIs, Vertex AI or other managed services buy tokens and outcomes without choosing the accelerator. Alphabet says customer API use of its own models rose from 10 billion to more than 16 billion tokens per minute in one quarter. At that level, Google can route stable workloads to TPUs behind the scenes.

CUDA remains a major barrier for raw infrastructure sales. Managed AI services give Google another path: the more customers buy an API rather than a cluster, the less often Nvidia’s developer advantage decides the sale.

Q9Is Google Cloud large enough to make TPUs a global platform?

Google Cloud is large enough to build a substantial TPU business, although its current reach cannot match Nvidia’s presence across almost every major provider.

Synergy Research Group estimates that Google held 14% of worldwide cloud infrastructure spending in the latest quarter, behind AWS at 28% and Microsoft at 21%. Google can reach a huge number of companies from third place, but an Nvidia buyer can choose AWS, Azure, Google Cloud, Oracle, CoreWeave and many other providers.

Google Cloud is growing unusually fast. Its latest quarterly revenue reached $20 billion, up 63% from a year earlier, while operating income tripled to $6.6 billion. Backlog almost doubled in one quarter to $462 billion. Alphabet said AI solutions had become Cloud’s largest growth contributor and that the backlog now includes TPU hardware agreements.

We should not treat the whole backlog as TPU demand. Most of it still comes from ordinary Google Cloud contracts covering infrastructure, software and services. Even so, the inclusion of hardware sales shows that TPUs are becoming commercially large enough to influence reported figures.

Cloud growth gives Google a strong starting point. Distribution beyond Google Cloud will determine whether TPUs remain a differentiated feature of the third-largest cloud or become an independent platform that customers can obtain wherever they prefer to operate.

Q10Can TPUs really escape Google Cloud?

TPUs are finally moving beyond standard Google Cloud rentals, but Google is opening the doors selectively rather than creating a free market for its chips.

Alphabet says a small group of customers will install TPU hardware in their own data centers, with most of the related revenue expected next year. The company has described financial institutions and frontier AI laboratories as likely buyers, which points toward large, tightly managed deployments.

The Blackstone partnership creates another route. Its new company will sell TPU-based compute as a service from facilities financed and operated outside Alphabet, while Google provides the chips and technical stack. That should give customers a way to buy TPU capacity without using an ordinary Google Cloud region.

These arrangements remove two real objections. Some customers need dedicated infrastructure for regulation, security or predictable capacity. Others may want TPU economics without handing every part of their operation to Google Cloud.

Google still controls the circle closely. TPUs are not widely sold through Dell, HPE, Lenovo, Supermicro and independent resellers. Customers cannot yet mix the chips with any server, network and operating model they choose. Nvidia’s hardware is available through a broad supply chain; Google is working with a small number of chosen partners.

That cautious approach may be sensible at first. Large TPU systems are complex, and a badly supported rollout could damage confidence. For now, external hardware availability is evidence of a serious expansion, not proof of an open ecosystem.

Q11How far behind Nvidia is Google financially?

Google is nowhere near Nvidia’s AI infrastructure revenue today, even though TPUs may already create large savings and cloud income that Google does not report separately.

Nvidia’s latest quarterly data-center revenue reached $75.2 billion, up 92% in one year. Google Cloud produced less than one-third of that in the comparable quarter, and its figure includes databases, Workspace, cybersecurity, managed AI, Nvidia GPU rentals and many other services.

The comparison is imperfect but useful. Nvidia’s data-center business alone is roughly 3.8 times larger than Google’s entire Cloud segment. Nvidia also generated $14.8 billion from data-center networking in the quarter, close to three-quarters of Google Cloud’s total revenue.

Google captures TPU value in several places. It avoids some Nvidia purchases, lowers the cost of Search and Gemini, sells cloud compute and is beginning to recognize hardware revenue. None of those figures is broken out, so a direct TPU revenue comparison is impossible.

The commercial gap is nevertheless enormous. Nvidia has already turned accelerated computing into a $200-billion-plus annual business with gross margins around the mid-70s. Google is only now giving investors enough information to see that external TPU sales have become material.

Nvidia’s commercial scale compared with Google Cloud

Q12What does Nvidia sell that Google still lacks?

Nvidia sells a complete outside ecosystem. Google currently offers an excellent system that works best under Google’s supervision.

Nvidia’s data-center product stretches from GPUs and CPUs to NVLink, InfiniBand and Ethernet networking, BlueField processors, rack-scale systems, libraries and deployment blueprints. A cloud or server maker can buy a design that has already been tested across many customers.

Google has built comparable depth inside its own environment. TPU clusters use custom interconnects, optical switching, compilers, schedulers, storage and software developed together. The recent TPU 8 designs show that Google is engineering the whole machine rather than dropping a chip into a generic server.

The difference appears when another company has to operate the system. Nvidia equipment is supported by global manufacturers, cloud providers, integrators, consultants and a large labor market. A customer can hire people who already know the tools and find replacement suppliers if one partner disappoints.

Google is only starting to build that outside layer. The early partner-operated clusters will test whether the TPU stack can run reliably without the same level of direct Google control found in its own data centers. Technical quality got Google into the contest. Independent support will decide how broad the contest becomes.

Q13Is Google’s internal demand its biggest TPU advantage?

Google’s enormous internal demand is its clearest TPU advantage because it can test each generation on real workloads before outside customers take the risk.

Search, Gemini, YouTube, advertising, Photos, Maps and DeepMind research create a mix of training and inference jobs that very few companies can reproduce. Google can use a new TPU even when the external market is small, then improve its compilers, networking and operating procedures on live traffic.

That captive demand changes the economics of chip development. Designing a custom accelerator costs billions across engineering, software, manufacturing and data centers. Google can spread those fixed costs across consumer products used by billions of people, then sell the same platform through Cloud.

The feedback loop is especially valuable for inference. Search gives Google huge volumes of repeated requests, so inefficiencies show up quickly and small savings compound. The cost reductions reported for AI Search responses suggest that model and hardware teams are learning together rather than waiting for an outside supplier’s next roadmap.

Nvidia has broader feedback from many industries. Google has deeper control over a smaller number of exceptionally large workloads. That is enough to make TPUs hard to dismiss even while their external ecosystem is still developing.

Q14Could custom AI chips weaken Nvidia without replacing it?

Custom AI chips are already starting to weaken Nvidia’s monopoly on new AI capacity, and Google only needs to capture part of that shift to become strategically important.

Amazon now says close to one million Trainium2 chips are training and serving Claude. Anthropic has committed more than $100 billion over ten years for up to five gigawatts of additional AWS capacity, covering several future Trainium generations.

Broadcom offers another measure of the trend. Its latest quarterly AI semiconductor revenue reached $10.8 billion, up 143% from a year earlier, driven by custom accelerators and AI networking. The company expects that figure to rise to $16 billion in the following quarter.

The market is moving toward several large custom platforms rather than one clean Nvidia replacement. Hyperscalers and frontier laboratories have repetitive workloads, huge engineering teams and enough purchasing power to justify specialized chips. Smaller companies still benefit from Nvidia’s flexibility and easier access.

Nvidia can keep growing quickly while losing some control over the market’s direction. Each major workload placed on TPU or Trainium reduces the share of new AI spending that automatically flows through Nvidia. Google’s opportunity is wider than a head-to-head chip contest: it can help turn a near-default supplier into one choice among several for the world’s largest buyers.

Q15What could stop TPUs from becoming a real Nvidia rival?

Software friction and limited distribution could keep TPUs important but niche, even if the chips remain technically excellent.

Migration cost is the clearest danger. PyTorch/XLA works, yet its own documentation describes execution, compilation and synchronization differences that engineers must understand. Google needs routine models to move onto TPUs without a long optimization project.

Availability could be just as limiting. Customers will hesitate to depend on a platform if they cannot obtain the same systems across regions, clouds and replacement suppliers. Google still has a large gap to close between a few controlled deployments and Nvidia-style distribution.

Capacity could then turn a software problem into a supply problem. Alphabet is spending heavily, but it already reports more AI demand than current infrastructure can comfortably serve. Internal products may receive priority when supply is tight, creating uncertainty for outside buyers.

Google also has to convince partners that this is a lasting business. A hardware ecosystem needs stable roadmaps, clear pricing and room for other companies to make money. Google has often been strongest when it controls the full experience. An Nvidia rival must also work well when the customer, data-center owner and software team sit outside Google.

None of these problems is fatal. They are simply harder to solve than producing a fast benchmark, and the next two years will show whether Google is building a platform or a collection of very large special deals.

Q16What would prove that TPUs have become a true Nvidia rival?

TPUs will have become a true Nvidia rival when broad customer adoption catches up with the hardware, not when Google announces another faster generation.

The cleanest proof would be several unrelated customers operating very large TPU fleets. Anthropic shows that one frontier lab can commit at enormous scale; repeat wins would show that Google has a market rather than an anchor customer.

Software has to feel ordinary as well. Common PyTorch models should move across with predictable performance, fewer special fixes and familiar monitoring tools. Most customers will never accept a lower-level platform that only specialist teams can run efficiently.

Distribution is another decisive test. Google has opened routes into customer sites and a partner-operated cloud, but they still need to become repeatable products across several locations with clear support and expansion plans.

Google also needs to disclose enough commercial information for outsiders to measure progress. TPU revenue, external capacity, utilization or large-customer counts would separate genuine platform growth from a handful of eye-catching agreements.

The proof points that would establish TPUs as a true Nvidia rival

Q17Can Google turn TPUs into an Nvidia rival?

Yes. Google can turn TPUs into a serious Nvidia rival in frontier training and hyperscale inference, but Nvidia will remain the broader computing platform for the foreseeable future.

Google has already cleared the hardest technical hurdle. TPUs run Google’s largest AI products and support Claude at large scale. Google is also commercializing them far more aggressively than it did in the past.

Its best chance lies at the top of the market. Frontier labs and hyperscalers spend billions on predictable workloads, employ teams that can optimize custom hardware and care intensely about cost, power and guaranteed capacity. Google’s control of models, compilers, chips, networks and data centers can produce savings that a more horizontal supplier may struggle to match.

Nvidia still owns the easier buying decision. CUDA has a vastly larger developer base, its systems are available almost everywhere and its platform supports far more than generative AI. The revenue gap also leaves Nvidia with far more money to invest in software, networking and future architectures.

The likely outcome is a split market. Nvidia remains the default for broad access, flexibility and outside development. Google captures a meaningful share of the biggest training and inference fleets, while Amazon and other custom-chip programs take additional pieces.

Our final judgment is mostly true. Google can build an Nvidia rival where the most expensive AI workloads live. It is unlikely to recreate Nvidia’s universal ecosystem, and it does not need to. Several more large, multi-year customer wins would change Nvidia’s position even if Nvidia continued growing.