Who is the next Nvidia?
In our AI chip market deck, you will find everything you need to understand the market
SUMMARY
Broadcom is the closest company to becoming the next Nvidia today because it is gaining control over the custom chips and networks that several of the largest AI buyers are building around.
A fast accelerator is no longer enough. Nvidia became dominant by making its software, networking, rack architecture and developer tools part of the same purchasing decision, so a successor has to control a computing layer rather than win a benchmark.
AMD is now a real production alternative, not merely a backup plan. Its announced pipeline of up to 14 gigawatts shows that OpenAI, Meta and Anthropic want a second large GPU supplier, although most of that capacity has yet to produce revenue.
The commercial structure of those AMD agreements matters as much as their size. Warrants, investment commitments and future performance milestones suggest that AMD is spending heavily to secure the deployments that could establish its platform.
Broadcom has a different advantage: it can benefit from several competing accelerator strategies at once. It earns money when large customers design their own processors and when those processors need Ethernet, switching, optical connectivity and rack-scale networking.
The cloud companies are already taking predictable workloads away from Nvidia, especially inference jobs they run repeatedly at enormous scale. Their weakness is reach: Trainium, TPU, Maia and MTIA remain closely tied to the companies and clouds that created them.
Inference is where the market is most likely to fragment. Repeated workloads make specialization economically attractive, which opens room for custom chips, Cerebras, AMD and other architectures even if Nvidia remains the default for flexible training and broad software support.
Cerebras has the strongest startup case because it now has public revenue, a large OpenAI agreement, an AWS partnership and concrete deployment plans. The hard part begins now: turning those commitments into hundreds of megawatts of reliable, recurring infrastructure.
CUDA is becoming easier to work around, but it is still difficult to replace in production. Running a model on another chip is one task; keeping thousands of processors efficient, stable and debuggable is the part that preserves Nvidia's advantage.
The market probably will not produce one clean replacement. AMD can become the large second GPU platform, cloud companies can absorb internal workloads, specialists can win parts of inference, and Broadcom can become the essential supplier underneath much of that fragmentation.
That is why Broadcom is the best overall answer. It may never build a developer ecosystem as visible as CUDA, but it has the clearest route to becoming unavoidable inside the infrastructure choices that shape the next generation of AI systems.
This market map, featured in our AI chip market deck, highlights top companies and startups in the AI chip market
What would a company have to do to become the next Nvidia?
For this article, the next Nvidia must control a part of computing that customers cannot easily avoid; producing a fast AI chip alone would fall short.
AMD could sell tens of billions of dollars of GPUs and remain mainly a second supplier. Cerebras could dominate one type of inference without creating a broad computing platform. Broadcom could become essential inside AI data centers while remaining almost invisible to ordinary developers.
Nvidia combines several advantages. Developers build with CUDA, cloud providers sell Nvidia systems, manufacturers design around its rack architecture, and customers buy its networking alongside its GPUs. Each part makes the others harder to replace.
A credible successor would need recurring demand across several product generations, technology customers struggle to swap out and enough influence to shape how future AI systems are built. We care more about that position than about one year of spectacular revenue growth.
| Meaning of “next Nvidia” | What success would look like | Strongest candidate now | How useful is this definition? |
|---|---|---|---|
| The main alternative GPU company | Large customers routinely choose its GPUs instead of Nvidia’s | AMD | Useful, but narrow |
| The next company with explosive AI revenue | AI adds tens of billions of dollars in new sales | Broadcom | Important |
| The company controlling a new computing layer | Customers design infrastructure around its technology | Broadcom | Strongest definition |
| The breakout AI-chip startup | A new architecture reaches large commercial deployments | Cerebras | Promising, but early |
If you want more recent data on this point, please see our latest AI chip market report.
How did Nvidia become much more than a GPU company?
Nvidia built its current lead by tying GPUs to CUDA, networking and complete systems that customers can deploy as one package.
CUDA has now been developed for two decades. Nvidia says more than six million developers use its computing platform, creating software, libraries and tools that work best on Nvidia hardware. A developer choosing Nvidia gains access to work accumulated by millions of other developers.
Nvidia then moved beyond individual accelerator cards. Its current Vera Rubin platform brings together GPUs, CPUs, low-latency inference processors, Ethernet, InfiniBand, storage processors and rack-scale systems. Seven new chips are already in production as part of the platform.
Networking has become another large business. During Nvidia’s latest full fiscal year, Data Center compute revenue increased 59%, while Data Center networking revenue jumped 142%. Nvidia is capturing the money spent moving information between processors as well as the money spent on the processors themselves.
A rival may build a competitive accelerator and still depend on Nvidia-compatible software, outside networking or a cloud partner to deliver a complete system. Nvidia controls far more of the final product.
As this chart shows, and as featured in our AI chip market deck, search interest in AI chips has grown significantly
Is AMD now a real Nvidia rival?
AMD has finally become a credible second choice for large AI buyers, although Nvidia still operates on a completely different scale.
AMD’s Data Center revenue reached $5.8 billion in its latest quarter, up 57% from the previous year. The division has become AMD’s largest business and now drives much of the company’s growth.
Nvidia reported $75.2 billion in Data Center revenue during a comparable quarter. Nvidia’s division was therefore around 13 times larger than AMD’s. The accelerator gap is wider because AMD’s Data Center figure also includes EPYC server processors and other products.
AMD has nevertheless crossed an important line. Microsoft is preparing to deploy AMD Helios systems at scale on Azure. Meta, OpenAI and Anthropic have also placed AMD inside their long-term infrastructure plans. These are large production commitments rather than isolated tests.
ROCm, AMD’s software environment, has improved enough for leading AI companies to treat AMD as deployable. It still lacks CUDA’s depth, but the question around AMD has changed. Large customers are now deciding how much AMD capacity to install, rather than whether AMD belongs in the data center at all.
Do AMD’s 14 gigawatts of AI deals prove it can catch Nvidia?
AMD’s pipeline of up to 14 gigawatts proves that the biggest AI buyers want a serious second supplier, but most of that capacity still has to be built and switched on.
OpenAI has agreed to deploy six gigawatts of AMD GPUs, with the first gigawatt expected to begin in the second half of 2026. Meta has announced another six gigawatts on a similar schedule. Anthropic has now added up to two gigawatts, with its first gigawatt expected during the first half of 2027.
Fourteen gigawatts represents an extraordinary volume of computing equipment. A single gigawatt-scale AI site can require several billion dollars of chips, networking, power systems and construction. The announcements place AMD in projects that could generate tens of billions of dollars over several years.
The word “could” carries real weight here. AMD must manufacture the MI450 systems, finish the Helios architecture, improve ROCm, help customers move their models and maintain performance once thousands of racks are operating together.
The commercial terms also show how hard AMD is pushing. Meta received a performance-based warrant covering up to 160 million AMD shares at an exercise price of one cent. AMD separately committed to invest up to $5 billion in Anthropic.
The deals have made AMD a serious challenger. Whether they create Nvidia-like profits will depend on actual shipments, pricing and utilization.
| Customer | Announced AMD capacity | First expected deployment | What remains uncertain |
|---|---|---|---|
| OpenAI | 6 GW | Second half of 2026 | Speed of the full rollout |
| Meta | Up to 6 GW | Second half of 2026 | Performance milestones and warrant conditions |
| Anthropic | Up to 2 GW | First half of 2027 | Deployment schedule and AMD’s equity investment |
| Combined | Up to 14 GW | Mostly future capacity | Revenue, margins and real utilization |
This chart, featured in our AI chip market deck, shows annual VC investment in AI chip startups
Is Broadcom already the strongest challenger to Nvidia in AI infrastructure?
Broadcom currently has the strongest overall claim because its quarterly AI semiconductor revenue has already crossed $10 billion and continues to accelerate.
Broadcom generated $8.4 billion from AI semiconductors in its first fiscal quarter of 2026, up 106% from the previous year. The figure rose to $10.8 billion in the following quarter, up 143%. Broadcom expects $16 billion in its third quarter, which would represent growth above 200%.
If Broadcom reaches that forecast, it will produce $35.2 billion in AI semiconductor revenue across three quarters. That is almost twice the $19.2 billion generated during the first two quarters, showing that the business is still accelerating from an already large base.
Nvidia’s latest Data Center quarter was still more than twice that entire projected three-quarter Broadcom total. Broadcom has a vast distance to cover before matching Nvidia’s scale.
Broadcom’s advantage comes from where it sits. It helps large customers create custom AI accelerators and supplies much of the networking needed to connect them. Broadcom benefits when OpenAI builds its own processor, when Meta expands MTIA and when data centers shift toward large Ethernet networks.
AMD needs customers to choose its GPUs. Broadcom can earn money from several accelerator architectures.
If you want more recent data on this point, please see our latest AI chip market report.
Can Broadcom become indispensable without its own version of CUDA?
Broadcom can become indispensable without owning the software developers use because it enters the hardware decisions customers make years before a data center opens.
OpenAI and Broadcom are working on ten gigawatts of custom accelerators and Ethernet systems. Their first processor, called Jalapeño, is designed for large language models and is expected to begin deployment by the end of 2026. OpenAI designs the accelerator, while Broadcom handles parts of the silicon implementation, connectivity and networking.
Broadcom has expanded its Meta partnership to support multiple gigawatts of MTIA chips and the Ethernet systems connecting them. Meta can own the visible processor and software while Broadcom earns revenue from the technology underneath.
Broadcom, Apollo and Blackstone have also created a financing platform intended to support more than 20 gigawatts of custom XPU infrastructure through 2028. The projects include capacity for leading AI laboratories such as OpenAI and Anthropic, so that total should not be added mechanically to the individual company announcements.
Once a company has designed a processor, network and rack architecture around Broadcom technology, changing suppliers becomes expensive and slow. The relationship can continue through several chip generations.
Broadcom will probably remain less visible than Nvidia. Developers may never describe themselves as part of a Broadcom ecosystem. Its position can still become extremely powerful because a growing share of AI infrastructure may depend on Broadcom before the first developer starts using it.
This chart, featured in our AI chip market deck, shows how Nvidia is leading in AI chips
Are Amazon, Google, Microsoft and Meta already replacing Nvidia?
Amazon, Google, Microsoft and Meta are already replacing Nvidia on selected internal workloads, especially large jobs they run repeatedly and understand in detail.
Amazon says almost one million Trainium2 processors are currently training and serving Anthropic’s Claude models. Trainium3 is already available and offers higher performance and better output per watt than the previous generation.
Google’s seventh-generation Ironwood TPU supports both training and inference. Google can connect as many as 9,216 Ironwood chips into a single pod and optimize the hardware around Gemini, its software and its own data centers.
Microsoft’s Maia 200 targets inference and includes 216 gigabytes of high-bandwidth memory. Meta uses a mixed fleet that includes Nvidia GPUs, AMD GPUs and its own MTIA processors. Meta’s engineers have built tools to optimize software across those different types of hardware.
These companies already control the models, cloud services and workloads. They can design a chip for a narrow purpose, use it millions of times and save enough money to justify years of development.
Nvidia’s concentration makes this push especially important. In one recent filing, four direct customers represented 22%, 15%, 13% and 11% of quarterly revenue, or 61% combined. Nvidia does not identify those customers, but the concentration shows how much influence a small group of buyers can develop.
The market is likely to divide. Custom chips will take a growing share of predictable internal workloads, while Nvidia remains especially strong for changing models, outside customers and applications requiring broad software support.
| Company | Custom AI processor | Strongest natural use | Main limitation as an Nvidia successor |
|---|---|---|---|
| Amazon | Trainium | AWS training and inference | Closely tied to AWS |
| TPU | Gemini and Google Cloud workloads | Google controls the surrounding platform | |
| Microsoft | Maia | Azure and Microsoft inference | Still early in large-scale deployment |
| Meta | MTIA | Recommendation and Meta AI workloads | Primarily built for Meta’s own infrastructure |
Why is AI inference the easiest place to attack Nvidia?
AI inference gives Nvidia’s challengers their clearest opening because repeated workloads reward lower costs, lower power use and faster answers.
Training creates a model. Inference runs every time someone asks that model a question, generates an image, uses a coding assistant or sends an AI agent to complete a task. A popular product may run inference billions of times after completing one major training cycle.
That repetition makes specialization worthwhile. Microsoft built Maia 200 around token generation. Amazon markets Trainium3 partly through output per megawatt. Cerebras focuses heavily on fast responses, while Groq built its processor around predictable, low-latency inference.
Groq’s recent story shows how seriously Nvidia takes this threat. Groq licensed its inference technology to Nvidia, and its founder, president and several team members joined Nvidia. Nvidia has since incorporated Groq technology into its Groq 3 LPX inference system for the Vera Rubin platform.
The deal kept Groq operating independently, but it also allowed Nvidia to answer a specialist architecture by bringing its strongest ideas into Nvidia’s wider platform.
Inference should become the most fragmented area of AI computing. Different models and response-time requirements will favor different chips. Nvidia’s broad system lets it compete across those categories, yet challengers have a clearer economic reason to enter here than in general-purpose frontier training.
If you want more recent data on this point, please see our latest AI chip market report.
This chart, featured in our AI chip market deck, shows annual funding in AI chip startups
Is Cerebras the startup closest to becoming the next Nvidia?
Cerebras is currently the strongest startup candidate because it has moved beyond impressive demonstrations into public revenue, major contracts and concrete deployment plans.
Cerebras generated $193.4 million in its latest quarter, up 94% from the previous year. Its cloud and services revenue rose 178%, showing that growth is coming from usage as well as hardware sales. Cerebras expects roughly $855 million to $865 million in core revenue for the full year.
OpenAI has agreed to deploy 750 megawatts of Cerebras inference capacity over several years under an arrangement valued above $20 billion. Cerebras is also working with AWS on a system that splits inference between AWS Trainium processors and Cerebras CS-3 machines.
Lately, Cerebras has started adding evidence outside those two relationships. The company plans to build 200 megawatts of European capacity by the end of 2027, with its first European installations expected before then. CrowdStrike has also selected Cerebras to help run AI detection and response models where fast answers carry real operational value.
Those developments make Cerebras more credible than a startup supported mainly by benchmarks and funding rounds.
The scale gap remains severe. Nvidia generated $215.9 billion during its latest full fiscal year, while Cerebras currently expects around $860 million. Cerebras would therefore produce less than half of 1% of Nvidia’s annual revenue even after meeting its guidance.
Cerebras now has to deliver hundreds of megawatts of infrastructure, keep customers using it and turn the OpenAI agreement into recurring revenue. Its architecture has earned a place in the conversation. Execution will decide whether Cerebras becomes a lasting platform.
Is CUDA still the moat no one can copy?
CUDA still gives Nvidia the deepest moat in AI chips, even as PyTorch and open software make basic movement between processors easier.
Nvidia says CUDA now supports more than six million developers. That community has produced libraries, models, optimization tools and industry-specific software over 20 years. A challenger has to compete with that accumulated work as well as Nvidia’s newest processor.
PyTorch supports Nvidia CUDA, AMD ROCm and several custom backends. Many models can therefore start running on a different chip with far less rewriting than a few years ago.
Production remains harder. Meta operates infrastructure containing Nvidia, AMD and MTIA processors, yet it still needs hardware-specific kernels that translate a model into efficient instructions for each architecture. Meta has built an AI engineering system specifically to create and tune those kernels.
A model running correctly on AMD or a custom chip only clears the first hurdle. Large deployments also need stable communication between thousands of processors, efficient memory use, debugging tools, monitoring and engineers who know how to fix failures.
Open software will continue to reduce Nvidia’s lock-in. Large customers with strong engineering teams will gain the most freedom. Smaller companies will often continue choosing Nvidia because the easiest development path still saves time and reduces risk.
This chart, featured in our AI chip market deck, compares the main business model options for AI accelerator chip companies
Is networking becoming more valuable than the AI chip itself?
Networking now determines enough of an AI cluster’s performance that Broadcom can win even when another company supplies the accelerator.
Modern AI systems can contain tens of thousands of processors. Those processors regularly exchange model parameters and intermediate results. A slow or unreliable connection leaves expensive chips waiting instead of computing.
Nvidia’s own numbers show how quickly that part of the system is growing. Its full-year Data Center networking revenue increased 142%, considerably faster than its compute revenue.
Broadcom combines custom accelerators with networking. Its Jericho4 router is designed to connect more than one million processors across multiple data centers. Broadcom is also shipping 102.4-terabit-per-second switch chips in production volume.
Broadcom has several routes into the same project. It can help create the accelerator, supply the switch silicon, provide optical components and connect separate buildings into one computing system.
Nvidia is defending that territory with NVLink, Spectrum-X Ethernet, InfiniBand and BlueField processors. Its goal is to keep customers buying a connected Nvidia system rather than assembling chips and networks from several vendors.
The company controlling the network can shape which processors fit easily into the data center. Broadcom’s position here is one of the strongest reasons it could become the next essential AI infrastructure company.
If you want more recent data on this point, please see our latest AI chip market report.
Can Nvidia keep outrunning the companies chasing it?
Nvidia is currently widening the contest faster than most challengers can close a single product gap.
A competitor may spend years building a faster inference chip, only to find that Nvidia now sells a specialized inference processor alongside its GPUs. Another company may focus on Ethernet while Nvidia expands Spectrum-X. A startup may simplify one type of deployment while Nvidia adds CPUs, storage processors and complete rack designs.
The Vera Rubin platform illustrates that pace. Nvidia has placed seven chips and five rack-scale systems inside one architecture covering training, inference, networking, storage and orchestration. Customers can adopt one part or purchase a much larger Nvidia system.
Nvidia also responds aggressively when a specialist develops something useful. The Groq licensing agreement brought a competing inference architecture and several of its creators into Nvidia’s roadmap. The finished Groq 3 LPX product now sits beside Rubin GPUs instead of competing with them from the outside.
This ability to absorb new approaches makes Nvidia unusually difficult to catch. Challengers need to improve their own technology while anticipating where Nvidia will move next.
Nvidia can still lose market share as custom chips, AMD and specialist systems grow. The AI infrastructure market is expanding quickly enough for Nvidia to keep growing while that fragmentation happens.
This chart, featured in our AI chip market deck, shows how revenue is split across customer segments in the AI chip market
Who is the next Nvidia?
Broadcom is the closest overall answer today, while AMD is the closest direct GPU rival and Cerebras is the strongest startup candidate.
Broadcom has the best chance of controlling a new layer of AI infrastructure. Its quarterly AI revenue already exceeds $10 billion, its growth is accelerating, and its technology sits behind custom processors and networks being built for OpenAI, Meta and other large AI buyers.
AMD offers the clearest direct challenge. Up to 14 gigawatts of announced deployments could turn AMD into a much larger AI company. The main test now is whether those announcements become operating systems, recurring orders and attractive margins.
Cerebras has built the strongest startup case through its OpenAI contract, AWS partnership, growing revenue and planned international capacity. Its unusual wafer-scale architecture can create a valuable inference platform, though the company remains tiny beside Nvidia and faces a demanding build-out.
Amazon, Google, Microsoft and Meta will probably take more workloads onto their own chips. Together, they may weaken Nvidia more than any single outside competitor. Their processors are designed mainly around their own clouds and services, which limits their ability to become a neutral platform for the wider industry.
TSMC, CoreWeave and other AI infrastructure companies will also capture enormous value. TSMC manufactures the advanced chips, while CoreWeave provides access to large GPU clusters. Their current businesses make them vital partners within the AI boom, but neither controls the complete computing environment in Nvidia’s way.
The most likely future has several winners. Nvidia remains the main platform for flexible accelerated computing. AMD becomes the large second GPU supplier. Custom processors take predictable internal workloads. Cerebras and other specialists gain ground in inference. Broadcom earns money from the custom chips and networks connecting much of that infrastructure.
Broadcom therefore deserves the title of closest successor. Nvidia still has no single replacement, and Broadcom’s opportunity comes from helping create the more fragmented market that follows it.
| Candidate | Strongest claim | Biggest obstacle | Current verdict |
|---|---|---|---|
| Broadcom | Custom AI chips and networking across several major customers | No developer platform comparable to CUDA | Closest overall |
| AMD | The only large, widely available alternative GPU platform | Large revenue and software gap | Closest direct rival |
| Cerebras | Distinct architecture with major commercial agreements | Small installed base and major deployment risk | Strongest startup |
| Cloud companies | Workloads, capital and custom chips | Platforms remain tied to their own clouds | Strongest collective threat |
| Nvidia | Software, systems, scale and rapid product expansion | Customers increasingly want alternatives | Still the clear leader |
If you want more recent data on this point, please see our latest AI chip market report.
OUR METHODOLOGY
This analysis asks which company comes closest to Nvidia’s strategic position, rather than which company has the fastest chip, the highest recent growth rate or the largest market capitalization. We looked for control over a computing layer that customers build around and struggle to replace.
We separated the question into several practical definitions: the main alternative GPU supplier, the company adding the most AI revenue, the business controlling a new infrastructure layer and the strongest independent AI-chip startup. That distinction keeps a direct Nvidia rival such as AMD from being judged by the same standard as Broadcom or Cerebras.
We used recurring commercial adoption as the clearest test of platform strength. Revenue, production deployments, signed customer commitments, multi-generation roadmaps and infrastructure capacity received more weight than benchmark wins or one-off demonstrations.
Future capacity announcements were treated as evidence of strategic direction, not completed success. AMD’s announced gigawatts, Broadcom’s custom-silicon projects and Cerebras’s planned deployments matter because they show where customers are committing capital, but actual revenue, margins, utilization and operating performance will decide how valuable those projects become.
For Nvidia’s moat, we looked beyond the GPU itself. CUDA adoption, software libraries, networking growth, rack-scale architecture and the breadth of the Vera Rubin platform were used to show how hardware, software and systems reinforce one another.
We also separated internal custom chips from neutral industry platforms. Trainium, TPU, Maia and MTIA can take meaningful workloads from Nvidia, but their close ties to their owners’ clouds and services limit how directly they can become a broad replacement for outside customers.
Key sources included Nvidia investor materials, the CUDA platform, Nvidia’s technical blog, AMD investor materials, AMD’s newsroom, ROCm documentation, Broadcom investor materials, Broadcom’s newsroom, Cerebras announcements, AWS Trainium materials, Google Cloud TPU materials, Meta Engineering, PyTorch documentation, and company filings available through SEC EDGAR.
The final judgment is deliberately qualitative. Different companies lead in different parts of the stack, so we did not compress software depth, customer adoption, networking, custom silicon and deployment risk into a single numerical score. Broadcom ranks first overall because its position appears across several of those dimensions at once.
This chart, featured in our AI chip market deck, shows how AI accelerator chip technology has evolved over time
Related blog posts
- Is Nvidia becoming too powerful?
- Why is Nvidia financing OpenAI?
