Is Nvidia becoming too powerful?
In our AI chip market deck, you will find everything you need to understand the market
SUMMARY
Yes. Nvidia is becoming too powerful because its lead now extends beyond AI chips into software, networking, rack design, supply allocation and the technical standards customers build around.
The company’s financial scale is only the surface of the story. Nvidia can earn more in one quarter than AMD earns in a year while keeping gross margins near 75%, giving it unusually large resources to fund the next product cycle before rivals catch the current one.
CUDA remains the hardest part of Nvidia to dislodge. A rival can offer a cheaper or faster accelerator, but customers still have to account for rewritten kernels, retested models, new operating tools and a thinner support ecosystem.
Nvidia’s expansion into networking and full-rack systems makes that software advantage stronger. The customer is no longer choosing one chip; it is increasingly choosing an architecture in which GPUs, links, switches, adapters, CPUs and software arrive on the same roadmap.
That integration lowers deployment risk, which is exactly what buyers of billion-dollar clusters are willing to pay for. Nvidia’s premium therefore reflects not only raw performance, but the fear that a cheaper alternative will take longer to make useful.
The strongest counterweight comes from Nvidia’s own largest customers. Google, Amazon, Microsoft and Meta remain dependent on Nvidia today, but their huge recurring workloads give them both the incentive and the money to build custom chips for stable training and inference jobs.
Custom silicon is already taking workload share without visibly slowing Nvidia’s growth. AI infrastructure spending is rising fast enough for cloud companies to deploy TPUs, Trainium, Maia or MTIA while still buying more Nvidia systems than before.
Inference is the most exposed part of Nvidia’s position. Repeated, predictable workloads reward specialized hardware and make small savings per token worth the migration effort, especially at hyperscaler scale.
Regulation and geopolitics can constrain Nvidia faster than an ordinary chip rival. Export controls have already removed the company from much of China’s accelerator market, while competition authorities are beginning to examine software, bundling and acquisitions rather than looking only at GPU share.
Nvidia can keep growing while becoming less unavoidable. The healthiest outcome is not to punish the company for building the best platform, but to make switching practical enough that customers can leave without months of engineering work or a serious rise in deployment risk.
This market map, featured in our AI chip market deck, highlights top companies and startups in the AI chip market
Why does Nvidia look so powerful right now?
Nvidia currently looks more powerful than any previous chip company because it is capturing the growth of an entire computing shift, rather than enjoying a normal semiconductor boom.
Its latest quarterly revenue reached $81.6 billion, up 85% in one year. Data-center revenue alone reached $75.2 billion and now provides roughly 92% of the company’s sales. Nvidia earned more in that one quarter than AMD earned across the whole of 2025, even though AMD also sells server CPUs, gaming chips and embedded processors. Put the two businesses side by side and the gap is already measured in multiples rather than percentages.
Nvidia’s annual revenue rose from $60.9 billion in fiscal 2024 to $215.9 billion two years later. A company already worth tens of billions in annual sales became more than three times larger while keeping gross margins above 70%.
Chip booms usually bring falling prices or a strong rival product cycle. Nvidia has lately combined rapid growth, scarcity pricing and wider control of the surrounding system. That is why its rise feels different from a normal semiconductor upswing.
| Period | Nvidia revenue | Data-center revenue | What changed |
|---|---|---|---|
| Fiscal 2024 | $60.9B | $47.5B | Generative AI started changing the size of the business |
| Fiscal 2025 | $130.5B | $115.2B | Revenue more than doubled |
| Fiscal 2026 | $215.9B | $193.7B | AI infrastructure became the clear core of Nvidia |
| Latest reported quarter | $81.6B | $75.2B | One quarter exceeded AMD’s full-year company revenue |
What would “too powerful” mean for Nvidia?
Nvidia becomes too powerful when companies can choose another supplier on paper, yet doing so would put them at a real disadvantage.
A high market share alone does not settle the question. A company can lead because its products are better while customers remain free to move when a stronger option appears. The problem grows when the leader controls several bottlenecks together and each one reinforces the others.
Nvidia now has influence over the accelerator, the programming environment, the links between chips, the networking between servers and the design of the full rack. A customer may begin by choosing a GPU and discover that the easiest route also includes CUDA, NVLink, InfiniBand or Spectrum-X, BlueField and an Nvidia reference system.
In plain terms, Nvidia can keep its position without blocking rivals by making the alternative feel risky. A cheaper chip loses its appeal when engineers must rewrite code, retest models, rebuild operating tools and accept a smaller support ecosystem.
The useful test is whether Nvidia keeps winning because every new generation is clearly better, or because too much of the industry has already been built around it. Right now, both forces are working together.
If you want more recent data on this point, please see our latest AI chip market report.
As this chart shows, and as featured in our AI chip market deck, search interest in AI chips has grown significantly
Is Nvidia already a monopoly in AI chips?
Nvidia currently has monopoly-like power in high-end, general-purpose AI accelerators, although the wider AI computing market still has several real alternatives.
The European Commission has already described Nvidia as likely dominant in the worldwide market for discrete data-center GPUs. That wording is important. It covers the merchant accelerators that companies can buy for many kinds of training and inference work, where Nvidia remains the default choice.
Custom chips keep the wider market open. Google runs TPUs, Amazon runs Trainium, Microsoft has Maia and Meta has MTIA. These processors handle large AI workloads and place a real ceiling on Nvidia’s future sales.
Most remain tied to their owner’s cloud or internal systems. A startup cannot buy Meta’s MTIA for its own data center, and using Trainium means adopting AWS infrastructure and its Neuron software. They restrain Nvidia without creating an open merchant ecosystem of similar reach.
Nvidia is still dominant in merchant data-center GPUs. Hyperscalers with chip teams have several paths around it, while smaller AI companies seeking flexible hardware and broad cloud access still see Nvidia as the safest answer.
Is Nvidia’s real advantage the chip or the ecosystem?
Nvidia’s strongest advantage today is the ecosystem around the chip, because several rivals can build fast silicon while very few can match the complete working environment.
CUDA is the deepest moat. Nvidia says more than 7.5 million developers use CUDA and its other software tools, while its computing platform supports around 6,000 applications. That base includes libraries, compilers, optimized kernels, debugging tools, training frameworks and years of code written inside companies and universities.
For a large model, the network can matter nearly as much as the chip. Slow communication leaves expensive accelerators waiting, so Nvidia tunes the GPU, NVLink connections, network adapters, switches and software together.
Blackwell and Rubin extend the same logic to the full rack. CPUs, GPUs, memory, networking and software follow one roadmap, giving customers a system already tested at scale instead of a box of parts they must integrate themselves.
A rival chip can beat Nvidia on a benchmark and still lose the purchase. Buyers care about how quickly their engineers can make the whole cluster work, how many models run without surprises and who will fix the problem at 3 a.m. That is where Nvidia’s lead looks strongest.
This chart, featured in our AI chip market deck, shows annual VC investment in AI chip startups
How hard is it to leave Nvidia CUDA today?
Leaving Nvidia CUDA is possible today, but the move still costs enough time and engineering effort to protect Nvidia from many cheaper alternatives.
Frameworks such as PyTorch and JAX hide part of the hardware underneath, so a model may run on AMD, Google or Amazon hardware without a complete rewrite. Open tools are also improving, and large buyers increasingly build their own abstraction layers.
The difficult work appears once a team has optimized performance. Companies often rely on CUDA-specific kernels, libraries and monitoring tools. Moving them can mean replacing unsupported operations, tuning memory use, checking whether numerical results still match and rebuilding the way jobs are scheduled across a cluster.
Those costs vary sharply by customer. A hyperscaler can spread the work across billions of repeated inferences. A startup may burn months of runway before the saving on chips becomes meaningful. “Compatible” and “easy to switch” describe very different realities.
The lock-in also grows with time. Every new CUDA-trained engineer, optimized model and internal tool makes the next Nvidia purchase easier than the first. Customers will move selected workloads, especially stable inference, but a clean platform-wide exit remains uncommon.
Is Nvidia quietly taking over AI networking too?
Nvidia is rapidly turning AI networking into a second source of market power, and that expansion deserves almost as much attention as its GPU lead.
The Mellanox acquisition gave Nvidia InfiniBand, high-speed adapters and deep expertise in connecting large computing clusters. Nvidia has since added Spectrum-X Ethernet, NVLink switches and BlueField data-processing units, all designed to work closely with its accelerators.
Its latest results show how quickly the strategy is paying off. Data-center networking revenue reached $14.8 billion in one quarter, up 199% from a year earlier, with InfiniBand, Spectrum-X and NVLink all contributing. The company has stopped presenting networking as an accessory; it now treats the network as part of the computer.
The appeal for customers is obvious. A cluster with thousands of premium GPUs can waste enormous amounts of money when data moves too slowly, and one supplier can tune the full system while taking responsibility for performance.
For Arista, Broadcom and other networking companies, the same integration creates a disadvantage. Nvidia’s private GPU roadmap is the key edge here: it can tune the network before rivals even know every requirement of the next system.
This chart, featured in our AI chip market deck, shows how Nvidia is leading in AI chips
Is Nvidia turning the AI chip into an entire data center?
Nvidia is now selling the architecture of the AI data center itself, which gives it influence far beyond the processor socket.
Rubin shows how far the strategy has moved. Nvidia’s latest platform brings together Vera CPUs, Rubin GPUs, NVLink switches, ConnectX adapters, BlueField storage and infrastructure processors, Spectrum Ethernet and the Groq 3 low-latency inference accelerator. The company describes seven chips and several rack types as one production platform.
Customers get a faster route to deployment because power, cooling, networking and software have already been planned together. Cloud providers can also repeat the same design in several locations without solving every integration problem again.
The trade-off is that Nvidia sets the shape of the purchase. Once its reference design becomes the standard rack, suppliers of memory, servers, cooling equipment and networking must fit Nvidia’s schedule and technical choices.
At that point, Nvidia looks more like a platform owner than a component vendor. Intel once shaped the PC through processor standards and Microsoft through Windows; Nvidia is trying to shape AI infrastructure through hardware and software together.
Do Nvidia’s margins prove it has too much pricing power?
Nvidia’s current margins show exceptional pricing power, and the burden is now on competition to bring them down.
The company reported a 74.9% gross margin in its latest quarter. AMD’s full-year gross margin was 50%, while manufacturers such as TSMC generally carry the cost and risk of running the factories that physically produce advanced chips.
Nvidia is fabless, AMD sells a broader mix of lower-priced products, and a complete AI system includes software and support that customers value. Even with those caveats, a margin near 75% at Nvidia’s scale is extraordinary.
Customers are paying for several things at once: leading performance, scarce supply, working software and a lower chance that a billion-dollar cluster fails to deliver. Nvidia can charge for that reduction in risk.
The real test comes when supply gets easier and alternative platforms mature. If margins stay near today’s level, the premium will increasingly reflect the cost of leaving Nvidia rather than the quality of the newest product.
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
Are Nvidia’s biggest customers trapped?
Nvidia’s biggest customers are dependent today, but they are rich and concentrated enough to push back harder than ordinary buyers could.
Three direct customers generated 21%, 17% and 16% of Nvidia’s latest quarterly revenue. Together, they accounted for 54%. The same three represented 64% of accounts receivable, which shows how much of Nvidia’s growth flows through a small group of giant buyers.
Those customers need Nvidia because their own clients ask for CUDA-compatible capacity and their frontier AI teams want the fastest proven systems. A cloud provider that lacks new Nvidia clusters risks losing valuable customers to another cloud.
Nvidia also needs them. Losing even part of one large account would matter more than losing thousands of small customers. The buyers have enough volume to demand special designs, better supply terms and long-term roadmaps.
Most importantly, they can fund alternatives. Google, Amazon, Microsoft and Meta each have recurring workloads large enough to justify custom chips and software teams. Their dependence is real, but it has already produced a coordinated response: buy Nvidia now while building a route around it.
Why is every major cloud company building its own AI chip?
Cloud companies are building custom AI chips because Nvidia currently takes too much of the value, and relying on one supplier has become a strategic risk.
Google’s Ironwood is its seventh TPU generation and is now available for large training and inference clusters. Amazon says almost one million Trainium2 chips are already training and serving Anthropic’s Claude, while Trainium3 expands the design to much larger systems. Microsoft’s Maia 200 is aimed at inference and, according to Microsoft, delivers 30% better performance per dollar than the previous leading hardware in its fleet. Meta uses MTIA for high-volume recommendation and advertising workloads.
These projects have moved well beyond defensive experiments. They now run important production services, support external cloud customers and have their own compilers, networking systems and development roadmaps. They are operating businesses now, not insurance policies sitting in a laboratory.
Their goal is narrower than replacing Nvidia everywhere. Each cloud can start with stable workloads that repeat millions of times, where a modest saving per query becomes enormous at scale. It also gains bargaining power when Nvidia negotiates price and supply.
| Company | Custom chip | Where it is strongest now | First part of Nvidia’s business under pressure |
|---|---|---|---|
| Ironwood TPU | Large Google and Google Cloud training and inference | Internal and cloud workloads with a mature JAX stack | |
| Amazon | Trainium2 and Trainium3 | Claude training and serving inside AWS | Large AWS customers seeking lower costs |
| Microsoft | Maia 200 | Production inference for Microsoft services | Repeated inference inside Azure and Microsoft products |
| Meta | MTIA | Recommendation and advertising systems | High-volume internal inference |
This chart, featured in our AI chip market deck, compares the main business model options for AI accelerator chip companies
Are custom AI chips replacing Nvidia now?
Custom AI chips are already replacing Nvidia in selected workloads, although they have yet to slow the company’s overall growth.
Anthropic says almost one million Trainium2 chips now train and serve Claude. Google operates 9,216 Ironwood chips inside one superpod. Microsoft is deploying Maia for production inference, while Meta has moved parts of recommendation and advertising onto MTIA. These are large production workloads, not lab pilots.
At the same time, all four companies continue buying Nvidia systems. Microsoft’s current infrastructure strategy includes Maia, AMD and Nvidia. Google offers Nvidia GPUs beside its TPUs. AWS is expanding Trainium while continuing to market Blackwell capacity.
Rapid AI spending allows both trends to happen together. A cloud company can double its use of custom silicon and still buy more Nvidia hardware than it bought last year. Nvidia’s revenue remains a poor measure of whether alternatives are gaining ground.
The test will become clearer when infrastructure growth slows. Customers will then have to choose where the next dollar goes rather than buying every available accelerator. For now, Nvidia is losing workload share at the edges while keeping the most valuable general-purpose demand.
Can AMD become a real alternative to Nvidia?
AMD is now a credible second source for large AI customers, but its current scale remains far below Nvidia’s.
AMD generated $16.6 billion in data-center revenue during 2025, including both EPYC server CPUs and Instinct accelerators. Nvidia’s latest quarterly data-center revenue was more than four times that amount.
The gap in products is smaller than the gap in sales. AMD’s Instinct systems offer large amounts of high-bandwidth memory, ROCm has improved, and major clouds have been willing to deploy the hardware. Most recently, Microsoft announced that Azure will add AMD’s Helios rack-scale platform and MI455X systems for production AI inference.
Microsoft’s choice is revealing. It already works closely with Nvidia and builds Maia itself, yet it still wants another merchant supplier at scale. AMD gives Azure more choice and weakens the idea that every serious AI service must run on Nvidia.
AMD’s harder problem remains software. Customers compare the time needed to move and tune real workloads rather than relying on chip specifications. AMD can now win meaningful inference deployments, but it still needs several years of reliable production use before most buyers treat ROCm and CUDA as equally safe.
This chart, featured in our AI chip market deck, shows how revenue is split across customer segments in the AI chip market
Will inference crack Nvidia’s dominance first?
Inference is currently the most likely place for Nvidia’s dominance to weaken because repeated workloads reward specialization and ruthless cost control.
Training a new frontier model changes constantly. Researchers try unfamiliar methods, debug failures and need broad software support. Nvidia’s flexible platform remains especially valuable there.
Inference often runs the same model again and again. Once the workload stabilizes, engineers can tune a dedicated chip around latency, power use and cost per token. Google, Amazon, Microsoft and Meta have all focused much of their custom-silicon effort on that opportunity.
Nvidia has responded aggressively. Its Dynamo software manages inference across large clusters, Rubin targets sharply lower token costs, and the company has added the Groq 3 LPU architecture to the Rubin platform for low-latency serving.
Nvidia’s own roadmap confirms that inference is the contested ground. The company may keep the largest share, especially for new and fast-changing models, but the market should become far more mixed than training. Specialization can finally save more than migration costs.
If you want more recent data on this point, please see our latest AI chip market report.
Can Nvidia use its cash to shape the AI industry?
Nvidia can now use its balance sheet to shape the companies and technologies growing around its platform.
Its private-company investments reached $42.3 billion at the end of the latest quarter, almost double the $22.3 billion reported one quarter earlier. Nvidia also disclosed $27 billion in further investment commitments.
Nvidia can invest in model developers, cloud providers, software firms and infrastructure businesses that may later become major buyers or strengthen demand for its systems.
There is no public proof of circular demand, but the filings reveal an awkward overlap. Nvidia can be a startup’s supplier, technical partner, investor and gateway to scarce computing capacity at the same time.
Groq 3 shows the same instinct on the technology side. Nvidia brought a different inference architecture into Rubin, giving customers more choice inside its platform while reducing the chance that the technology grows into a separate ecosystem.
This chart, featured in our AI chip market deck, shows how AI accelerator chip technology has evolved over time
Can Nvidia influence who gets enough AI compute?
Nvidia has real influence over access to advanced AI compute, especially during shortages, even though allocation is shared with manufacturers and cloud providers.
A delayed shipment can decide whether an AI lab trains its next model this quarter or watches a rival move first. Large customers naturally receive more attention because they place bigger orders, can deploy whole clusters and commit to future generations.
Nvidia has backed its forecasts with $119 billion in manufacturing, supply and capacity commitments. That amount lets it reserve wafers, advanced packaging, memory and assembly capacity long before finished systems reach customers.
TSMC fabricates the leading GPUs, SK Hynix, Micron and Samsung provide memory, system builders assemble racks, and cloud providers decide which users receive rented capacity. One missing component can hold up the whole cluster.
Nvidia still sits in the middle of those relationships. Its roadmap influences what suppliers expand, what clouds build and what AI companies expect to access. During a shortage, ordinary allocation decisions can decide which AI labs get ahead.
How much power does Nvidia have over its suppliers?
Nvidia has enormous buying power, yet its dependence on a small Asian manufacturing chain remains one of the clearest limits on its dominance.
The company designs its chips and systems without owning leading-edge semiconductor factories. TSMC and Samsung manufacture wafers, memory comes from SK Hynix, Micron and Samsung, and advanced packaging such as CoWoS is essential to connect processors with high-bandwidth memory.
Nvidia can reserve capacity years ahead and pay premiums for priority. Its huge orders also give suppliers a strong reason to build around Nvidia’s product roadmap.
Money cannot create an advanced factory or packaging line overnight. Nvidia warns that some lead times have exceeded twelve months, and moving a complex chip to another foundry requires redesign, qualification and production risk.
Much of the critical supply chain remains in Taiwan and South Korea. A serious disruption there would hit Nvidia quickly, regardless of demand or software strength.
Nvidia captures an unusually large share of the profit, but it still relies on partners that control the physical production of every major system it sells. That dependency prevents the company from acting like a fully self-contained monopoly.
In our AI chip market deck, we identify pain points entrepreneurs should prioritize
Can governments rein in Nvidia faster than competitors can?
Governments can limit Nvidia faster than chip rivals can, and export controls have already proved it.
United States restrictions removed Nvidia from most of China’s data-center accelerator market. Nvidia recorded a $4.5 billion charge tied to H20 inventory and purchase commitments, and its latest filing says the company is currently unable to offer a competitive data-center product that satisfies both US and Chinese authorities.
Nvidia itself warns that its absence has helped Chinese competitors build larger developer and customer ecosystems, turning a lost market into a future competitive threat elsewhere.
Competition regulators are also watching the software and system layers. The European Commission considered Nvidia likely dominant in discrete data-center GPUs when it reviewed the Run:ai acquisition. It approved the deal, but the analysis showed that authorities are looking beyond chip sales and asking whether Nvidia could use software to favour its own hardware.
A forced breakup would be a blunt answer at this stage. Much of Nvidia’s value comes from making GPUs, networking and software work together, and separating them could damage customers before it creates a stronger rival.
Targeted rules would work better: protect interoperability, examine acquisitions of emerging alternatives, stop restrictive bundling and require fair access where Nvidia controls an essential interface. Governments have the leverage to do this now; waiting for a perfect chip competitor could take far longer.
| Force limiting Nvidia | How strong is it today? | Main weakness |
|---|---|---|
| Export controls | Very strong | Can fragment markets and strengthen foreign rivals |
| Competition enforcement | Moderate | Cases move slowly compared with product cycles |
| Hyperscaler custom chips | Growing quickly | Mostly tied to each company’s own cloud |
| AMD and merchant rivals | Credible but smaller | Software and ecosystem gap remains large |
| Suppliers | Structurally important | They depend heavily on Nvidia’s order volume too |
If you want more recent data on this point, please see our latest AI chip market report.
Will Nvidia keep becoming more powerful?
Nvidia’s influence will probably keep growing in the near term, even as its share of some AI workloads begins to fall.
The company still has a powerful loop working in its favour. More sales fund more research, software and supply reservations. More installed systems attract developers. More developers make the next customer more likely to choose Nvidia.
Its annual product cadence also forces rivals to chase a moving target. By the time a competitor catches one GPU generation, Nvidia is selling a new rack, network and software stack around the next one.
The counterforces are now concrete. Almost one million Trainium chips are running Claude, Ironwood is broadly available, Maia is in production, Azure is expanding AMD inference infrastructure and export controls have closed Nvidia out of China’s data-center market.
These developments attack different parts of Nvidia’s position rather than producing one direct replacement. That is probably how its power will erode: custom chips take stable workloads, AMD becomes a second source, regulators protect interfaces and suppliers retain control of manufacturing.
Nvidia can become larger while becoming slightly less unavoidable. Its absolute influence should rise for now, while the market slowly builds enough alternatives to prevent permanent control.
This chart, featured in our AI chip market deck, shows how revenue is split by region across Europe, Asia, North America, Africa, and South America in the AI chip market
So, is Nvidia becoming too powerful?
Yes, Nvidia is becoming too powerful because too many layers of the AI industry now depend on one company’s hardware, software and system roadmap.
Its lead still reflects real performance and years of better execution. Customers receive genuine value from a platform that works, scales and arrives with mature tools. Regulation should protect competition rather than punish success.
The worrying part is how the pieces reinforce one another. CUDA makes the GPU easier to choose. Nvidia networking makes CUDA clusters run better. Full-rack systems make Nvidia networking easier to adopt. Investments and supply commitments then help the whole ecosystem expand around the same roadmap.
The limits remain serious. Giant customers are building their own chips, AMD is winning fresh production deployments, governments can close markets and Nvidia cannot manufacture without a concentrated group of suppliers.
Those limits have yet to make Nvidia optional for most frontier AI work. Today, a company can choose another accelerator, but it often pays through engineering time, narrower software support or greater deployment risk. That is already more power than one private supplier should comfortably hold over a technology becoming basic infrastructure.
Nvidia has earned its lead, but the AI industry has allowed that lead to spread too far across the stack. The company is becoming too powerful now, and the best remedy is a market where leaving Nvidia becomes practical before regulators are forced to make the choice for everyone.
If you want more recent data on this point, please see our latest AI chip market report.
OUR METHODOLOGY
We assessed whether Nvidia is becoming too powerful by separating product leadership from structural market power. The analysis looks at the places where customers can still choose another supplier and the places where switching carries a real penalty in engineering time, software support, deployment risk or access to capacity.
We used Nvidia’s latest financial results and filings to track revenue, margins, customer concentration, networking growth, investment exposure, supply commitments and export-control effects. We treated these disclosures as the clearest evidence of how much economic power the company has today, rather than relying on market share alone.
CUDA, NVLink, InfiniBand, Spectrum-X, BlueField and Nvidia’s rack-scale platforms were reviewed together because the central question is whether control across several layers makes each individual product harder to replace. We used Nvidia’s developer and data-center documentation to understand how those layers connect in practice.
We compared that position with the strongest counterforces already operating at scale: Google TPUs, Amazon Trainium and Neuron, Microsoft Maia, Meta MTIA, AMD Instinct and ROCm, leading-edge manufacturing constraints, US export controls and European competition oversight. Custom chips were counted as competitive pressure where they run meaningful production workloads, even when they are not sold as open merchant products.
Margin comparisons were used as evidence of pricing power, not as proof by themselves. Nvidia is fabless, AMD sells a broader product mix, and TSMC carries manufacturing economics that are not directly comparable, so the stronger conclusion comes from margins combined with software dependence, system integration and switching costs.
Key sources include Nvidia’s quarterly results, Nvidia’s SEC filings, Nvidia’s CUDA documentation, Nvidia’s data-center platform materials, Nvidia’s networking materials, Google Cloud TPU documentation, AWS Trainium documentation, AWS Neuron documentation, Microsoft Azure product announcements, Meta engineering disclosures, AMD investor materials, AMD ROCm documentation, European Commission competition materials, US Bureau of Industry and Security rules, and investor disclosures from TSMC, SK Hynix and Micron.
This chart, featured in our AI chip market deck, shows annual VC investment in AI chip startups
Related blog posts
