Signals Inbox·July 25, 2026·AI Chips

Who is Nvidia's most serious competitor today?

AMD is Nvidia’s most serious competitor today because it is the only challenger turning alternative AI accelerators into a broadly available second platform across clouds and private data centers. Google is the deeper long-term threat.

Q1What does “Nvidia’s most serious competitor” mean?

Here, Nvidia’s most serious competitor means the company with the best realistic chance of taking large, repeatable AI infrastructure orders from Nvidia across several customers.

Chip speed alone does not settle the question. Large buyers purchase complete systems containing accelerators, processors, memory, networking, cooling and software. They also care about how quickly engineers can move models onto the platform, how reliably thousands of chips run together and whether the same system is available from more than one provider.

That definition separates four kinds of challenger. AMD sells alternative accelerators and rack systems to outside customers. Google, Amazon and Microsoft use custom chips mainly inside their own clouds, although Google has started opening TPUs to selected on-premises buyers. Broadcom designs custom silicon and networking for other companies. Huawei competes in a Chinese market shaped heavily by export controls and domestic procurement policy.

By that standard, AMD fits better than anyone else.

Q2Why is the Nvidia competition question harder now?

The Nvidia competition question is harder now because its biggest customers are funding several alternatives while continuing to buy Nvidia systems in huge volumes.

Nvidia’s latest annual filing shows how much leverage a few buyers could eventually have. One direct customer accounted for 22% of annual revenue and another for 14%. Nvidia also said hyperscalers produced slightly more than half of Data Center revenue in its final quarter of the year. These are precisely the companies building TPUs, Trainium, Maia and other custom accelerators.

Meta captures the new pattern. It plans to deploy millions of Nvidia Blackwell and Rubin GPUs, yet it has also agreed to deploy up to six gigawatts of AMD infrastructure and continues developing its own MTIA chips. Microsoft is preparing Nvidia Rubin systems, AMD rack-scale systems and Maia accelerators. Google is expanding both TPUs and its Nvidia cloud offering.

These companies are deliberately avoiding a single winner. They want enough Nvidia capacity to keep moving quickly, plus enough alternative capacity to lower costs, secure supply and improve their negotiating position.

That makes the scoreboard slippery. A competitor can gain billions of dollars in business without causing Nvidia’s revenue to fall, because total AI infrastructure spending is still growing at an extraordinary pace.

Q3How far ahead is Nvidia today?

Nvidia currently operates in a different revenue league from every direct AI chip rival.

In its latest reported quarter, Nvidia generated $75.2 billion from Data Center products, up 92% from the previous year. AMD’s Data Center segment generated $5.8 billion during its latest quarter, and that figure included EPYC server processors as well as Instinct accelerators. Nvidia’s comparable lead over AMD in AI accelerators was therefore wider than the roughly thirteen-to-one headline ratio.

Broadcom’s figure is also imperfect because it combines custom AI accelerators with networking products supplied across several customers. Google Cloud generated much more revenue than AMD’s Data Center segment, although that business contains cloud software, storage, databases, security and Nvidia services alongside TPUs.

The scale gap is enormous. Nvidia added more Data Center revenue in one year than AMD has generated from its Data Center segment across several years. AMD can become strategically important without getting close to Nvidia’s size. Even a durable minority share of accelerator spending at several hyperscalers would change buying negotiations and weaken Nvidia’s control over the market.

Latest relevant quarterly revenue

Q4Is AMD now Nvidia’s strongest direct competitor?

AMD is Nvidia’s strongest direct competitor right now, largely because customers can buy its AI platform without committing themselves to one rival cloud.

Google TPUs mainly strengthen Google Cloud. Trainium mainly strengthens AWS. Maia mainly strengthens Microsoft Azure. AMD can sell Instinct accelerators and complete rack systems through cloud providers, equipment makers, specialist infrastructure companies and customer-owned data centers.

That neutrality gives AMD a valuable position. OpenAI may use several infrastructure partners. An enterprise may want capacity from Azure and Oracle while retaining the option to run hardware on its own premises. Supporting AMD can reduce reliance on Nvidia without handing more control to Google, Amazon or Microsoft.

AMD has also assembled more of the surrounding system. Its rack-scale offer combines MI455X accelerators, EPYC processors, Pensando networking and ROCm software. The company still depends on partners for parts of the broader ecosystem, but it is no longer presenting customers with a powerful chip and leaving them to build everything around it.

The financial gap remains brutal. AMD produced $16.6 billion in Data Center revenue for all of 2025, including CPUs. Nvidia produced $193.7 billion in Data Center revenue in its latest full year. Still, several of the world’s largest AI buyers now treat AMD as a serious second platform. That is new.

Q5Did AMD’s OpenAI, Meta and Microsoft deals change the race?

Yes. AMD’s agreements with OpenAI, Meta and Microsoft have moved it from occasional alternative to planned infrastructure supplier at gigawatt scale.

OpenAI agreed to deploy six gigawatts of AMD GPUs across several generations, beginning with a one-gigawatt MI450 deployment in the second half of 2026. Meta later agreed to deploy up to six gigawatts, with its first gigawatt built around a customized MI450-based accelerator and AMD’s rack-scale design.

The latest development is Microsoft’s decision to deploy Helios in Azure for frontier-model inference, Azure AI services and customer applications. AMD says Helios shipments to Microsoft and other customers will begin in the second half of 2026. This gives AMD a major distribution channel as well as another buyer.

The combined OpenAI and Meta commitments reach twelve gigawatts. The caveat is large: most of that infrastructure has not been installed, and the agreements contain milestones linked to future purchases, technical performance and deployment schedules. They show serious intent, not completed market share.

Even so, the pattern has changed. AMD now has three routes into large AI workloads: direct alignment with a frontier model company, a customized platform for a hyperscaler and broad access through Azure. One announcement could be bargaining pressure. Three linked commitments from very different buyers are harder to wave away.

Q6Can AMD offer a complete alternative to Nvidia?

AMD has the pieces of a complete Nvidia alternative, although its ability to deliver them smoothly at massive scale remains unproven.

The modern contest happens at rack level. Customers need accelerators, CPUs, high-bandwidth memory, network interfaces, switches, cooling, power management and software that can keep thousands of devices busy. Poor communication between chips can erase an impressive benchmark result.

Helios directly addresses that problem by combining MI455X GPUs, EPYC “Venice” CPUs, Pensando networking and ROCm in an integrated rack. Meta is using the architecture for its customized deployment. Microsoft plans to use it in Azure. AMD and TCS have also announced a 200-megawatt Helios project in India, giving the platform another route into large deployments.

Nvidia still has the stronger record. Its Blackwell systems are already shipping widely, and Rubin extends the same approach across GPUs, CPUs, networking, storage and system software. Nvidia can coordinate these components through one mature product roadmap and a much larger base of deployment partners.

AMD now has to ship. Customers must be able to install Helios on schedule, run it reliably and expand after the first clusters. A rack diagram and a large commitment get AMD into the contest. Repeat orders after real production use would prove it belongs there.

Q7Has AMD really closed the CUDA software gap?

ROCm has become good enough for major AI companies to deploy AMD, but Nvidia still offers the easier and safer software choice for most buyers.

AMD has made visible progress. ROCm support is built into the official PyTorch repository, and the platform supports large-scale training, mixed precision and popular inference engines. AMD’s latest annual report said ROCm offered out-of-the-box support for more than two million Hugging Face models and recorded a tenfold increase in downloads during 2025.

The customer list gives those claims more weight. Those buyers would not plan large MI450 deployments if basic software compatibility remained a deal-breaker. Engineers at those companies can run serious workloads on AMD today.

The real test is how much work the customer must do. A hyperscaler can assign large teams to rewrite kernels, tune communication libraries and fix platform-specific failures. A smaller AI company usually wants the platform that works with the fewest changes and has the largest pool of experienced engineers.

Nvidia’s CUDA stack covers far more than a programming toolkit. It includes specialized libraries, model frameworks, debugging tools, deployment software and years of community knowledge. Nvidia also updates that stack alongside its own chips, networking and rack systems.

ROCm is usable for sophisticated buyers. CUDA remains the default because it is still more convenient and predictable across the wider market.

Q8Is Google a bigger long-term threat than AMD?

Google is the bigger long-term threat to Nvidia because it can design the model, compiler, chip, network, cloud and data center as one system.

Google has developed TPUs for more than a decade and already uses them for Gemini and other internal workloads. Its eighth generation splits the architecture into TPU 8t for training and TPU 8i for inference and reinforcement learning. Google says TPU 8i delivers 80% better inference performance per dollar than the previous generation.

Google has also started widening distribution. Alphabet said it will begin delivering TPU hardware to selected customers for use in their own data centers. Most related revenue is expected in 2027, so this remains an early move, but it expands Google’s opportunity beyond workloads hosted inside Google Cloud.

Google also has guaranteed internal demand. Search, Gemini, YouTube and Cloud can support large chip investments before outside adoption becomes broad. That gives Google time to refine hardware and software together, while AMD must convince customers to fund each generation through purchases.

AMD remains the stronger direct rival today because its platform is neutral and available through more channels. Google could eventually build a second full computing ecosystem around TPUs, especially if direct hardware sales become routine. That would challenge Nvidia more deeply than AMD winning a minority share of GPU orders.

Q9Is Broadcom the hidden threat to Nvidia?

Broadcom may be Nvidia’s most powerful indirect competitor because it helps several hyperscalers turn their own chip plans into working products.

Broadcom reported $10.8 billion in quarterly AI semiconductor revenue, up 143% from the previous year, and projected roughly $16 billion for the following quarter. Those figures combine custom accelerators and AI networking, yet they show that custom silicon has already become a large commercial business.

The attraction is straightforward. A hyperscaler running the same high-volume workload billions of times may gain more from a specialized chip than from a general-purpose GPU. The buyer can optimize memory, numerical formats and data movement around its own models while avoiding part of Nvidia’s margin.

Broadcom supplies design expertise, intellectual property and networking technology to companies capable of placing enormous orders. Several apparently separate challengers therefore rely on the same design and networking expertise, even when the finished chips carry different names.

Broadcom does not control the complete customer experience. Developers rarely decide to “move to Broadcom,” and each custom platform still needs its own software and deployment environment. Its impact appears in the spending that shifts toward Google TPUs, Meta accelerators and other private designs.

AMD remains the clearest company-to-company rival. Broadcom is helping many large buyers build an escape route from Nvidia.

Q10Do Amazon Trainium and Microsoft Maia seriously threaten Nvidia?

Amazon Trainium and Microsoft Maia already take real workloads from Nvidia, although both remain tied closely to their owners’ clouds.

AWS now says almost one million Trainium2 chips are training and serving Anthropic’s Claude through Project Rainier. That is one of the largest disclosed deployments of a non-Nvidia accelerator and shows that a custom chip can support a frontier model in production.

Microsoft’s Maia 200 focuses on inference. Microsoft says it delivers 30% better performance per dollar than the latest hardware previously used in its fleet. Maia is already serving workloads involving OpenAI models, Microsoft Foundry and Microsoft 365 Copilot, but its rollout began in a limited number of Azure regions and the software kit remains much younger than CUDA.

Both companies gain even when outside adoption stays narrow. Every workload moved to Trainium or Maia lowers their dependence on Nvidia, gives them more control over costs and strengthens their bargaining position during future purchases.

Their main weakness is portability. A customer choosing Trainium is also choosing AWS. Maia is part of Azure. Nvidia capacity is available across all three major clouds, Oracle, CoreWeave, equipment makers and private data centers. AMD is trying to offer similar flexibility from a much smaller base.

Where each alternative is strongest

Q11Is Huawei Nvidia’s biggest competitor in China?

Huawei is clearly Nvidia’s strongest competitor in China today, helped by domestic policy and Nvidia’s restricted access to the market.

Nvidia once held roughly 95% of China’s advanced AI accelerator market, according to Jensen Huang. An Associated Press report citing Bernstein estimated that Nvidia’s share could fall to around 8% in 2026 while Huawei reaches about 50%. The exact numbers are forecasts, but the direction is already visible: Huawei leads domestic sales while Nvidia has struggled to generate revenue from newer restricted products.

Huawei is building more than individual Ascend chips. It offers CANN software, networking and Atlas SuperPod systems that connect thousands of accelerators. Chinese model developers are also adapting software to Ascend, which gradually makes the platform more useful inside the domestic ecosystem.

Huawei still trails Nvidia’s leading technology in several areas. Chinese manufacturers face limits in advanced fabrication, high-bandwidth memory and production capacity. Researchers and companies continue seeking Nvidia hardware for demanding training workloads, sometimes through restricted channels.

In China, regulation shapes the commercial result. Buyers are being pushed toward domestic systems while Nvidia’s supply remains uncertain. Huawei has used that opening more effectively than any other Chinese company and may keep much of the market even if some Nvidia exports resume.

Q12Does the shift toward AI inference make Nvidia easier to challenge?

The shift toward inference gives Nvidia’s rivals a better opening because repeated production workloads are easier to optimize for specialized chips.

Frontier training changes constantly. Researchers alter model structures, numerical formats and communication patterns, which rewards flexible hardware and mature software. Nvidia performs well in that environment because CUDA and its networking stack can support a wide range of experiments.

Inference is often more stable. Once a model serves millions of similar requests, the operator can optimize hardware around token generation, memory access, latency and power use. Google designed TPU 8i specifically for inference and reinforcement learning. Microsoft built Maia 200 around token economics. Meta’s customized AMD deployment also targets its own workloads rather than a generic market.

The financial incentive is enormous. A small reduction in the cost of each request becomes meaningful when a company serves billions of queries. Google said it cut Gemini serving costs by 78% during 2025 through a combination of model improvements, infrastructure work and better utilization. Custom hardware was only part of that reduction, but the result explains why hyperscalers keep pursuing deeper control.

Nvidia is adapting quickly. Blackwell and Rubin emphasize low-precision computing, memory and rack-scale networking for inference. Nvidia should remain a major inference supplier. The spending itself will probably spread across more architectures than frontier training spending.

Q13Are Nvidia’s biggest customers now playing both sides?

Yes. Nvidia’s largest customers are buying its newest systems while building alternatives that keep Nvidia from becoming their only option.

Meta is the cleanest example. It has committed to millions of Nvidia Blackwell and Rubin GPUs and up to six gigawatts of AMD infrastructure, alongside its internal MTIA program. Microsoft combines Nvidia, AMD and Maia. Google offers Nvidia GPUs while advancing TPUs. AWS sells Nvidia capacity while scaling Trainium.

This repeated behavior tells us more than any isolated product announcement. The hyperscalers still value Nvidia’s speed, software and broad compatibility. Their spending has also become large enough to justify several hardware platforms.

Nvidia depends heavily on a few buyers. With two direct customers representing 36% of annual revenue, a change in buying plans at only a few companies can redirect tens of billions of dollars. Those buyers know it and are investing accordingly.

For now, they are adding alternatives rather than removing Nvidia. Their first goals are supply security and bargaining power, followed by lower costs on workloads suited to custom chips. A full departure from Nvidia would create unnecessary technical risk while AI demand remains so strong.

Q14Can Nvidia keep growing while competitors take market share?

Nvidia can keep growing rapidly while competitors take market share because the total AI infrastructure market is expanding faster than any one supplier can serve it.

Nvidia’s Data Center revenue rose from $47.5 billion to $115.2 billion and then to $193.7 billion across its last three full fiscal years. Its latest quarter added another 92% year-over-year growth. AMD, Broadcom and the hyperscalers’ custom chips also expanded during that period.

Revenue growth alone therefore cannot tell us whether Nvidia’s competitive position is improving. A company can lose ten percentage points of a market that doubles and still increase sales sharply. The first alternatives are also absorbing unmet demand and specialized workloads rather than replacing installed Nvidia clusters.

Nvidia’s margins suggest it retains pricing power. Its latest quarterly gross margin stayed near 75%, and customers continue committing to Blackwell and Rubin systems even while developing substitutes.

The pressure becomes serious when alternatives grow faster than the market for several years, force visible price concessions or capture the majority of new capacity at major customers. The evidence today shows diversification while Nvidia continues advancing.

Q15What would prove that Nvidia is actually losing its lead?

We would only say Nvidia is losing its lead after customers repeatedly expand rival systems and the change begins to show up in Nvidia’s economics.

The first test is whether customers buy again. Large customers often announce alternatives to reserve capacity or strengthen negotiations. A second and third purchase after running the hardware in production carries much more weight than the initial commitment.

Rival chips also need to become easy for ordinary teams to use. AMD, TPU, Trainium and Maia would have to work smoothly with mainstream frameworks and deployment tools, without large amounts of custom engineering. Easier portability would let smaller AI companies follow the hyperscalers.

Then we would look at Nvidia’s numbers. Slower Data Center growth than overall AI infrastructure spending, sustained margin compression or a declining share of new hyperscaler capacity would show that alternatives are taking more than overflow demand.

None of those conditions is decisive yet. AMD’s largest deployments are only beginning, Google’s on-premises TPU business is selective, Maia remains limited and Nvidia’s latest Data Center growth accelerated.

What would show Nvidia is losing its lead

Q16Who is Nvidia’s most serious competitor today?

AMD is Nvidia’s most serious competitor today.

It has the best combination of outside customers, multi-cloud reach, hardware that can run in customer data centers, rack-scale systems and improving software. The commitments from OpenAI, Meta and Microsoft show that AMD has moved beyond small trials. Its next challenge is execution: shipping Helios on time, running it reliably and winning repeat expansions.

Google is the strongest long-term strategic threat. TPUs already operate at huge internal scale, Google can optimize the full technology stack and selected customers will soon receive TPU hardware in their own data centers. If that distribution widens, Google could build the second complete AI computing ecosystem.

Huawei dominates the China-specific answer because it leads the shift away from Nvidia there. Broadcom matters in a different way: it enables custom chips across several hyperscalers. Trainium and Maia mainly reduce Amazon’s and Microsoft’s dependence on Nvidia.

The direct answer remains AMD. It is far smaller than Nvidia, and CUDA still gives Nvidia a major advantage, but AMD is the only challenger positioned to become a broadly available second platform for the global AI industry.