How far ahead is Nvidia?
In our AI chip market deck, you will find everything you need to understand the market
SUMMARY
Nvidia is several years ahead as an AI platform, roughly one deployed rack generation ahead of AMD, and only narrowly ahead on some inference workloads.
The headline gap depends on what we measure. Nvidia’s commercial lead is enormous, but the gap between individual accelerators can shrink to a few percentage points on selected tests.
Nvidia’s Data Center business is about thirteen times larger than AMD’s. Its networking revenue alone is more than twice the size of AMD’s entire Data Center division.
AMD has moved beyond small trials. Agreements involving OpenAI, Meta, Anthropic and Microsoft make it a serious second platform, although most of that future capacity has not yet become installed systems or reported revenue.
The hardest part of Nvidia’s lead to copy is not one GPU. It is the combination of racks, interconnects, networking, software libraries, operational tools and engineers who already know how to run the platform.
Inference is where Nvidia looks most vulnerable. AMD can match or nearly match Nvidia on several standardized workloads, while custom chips can be tuned for predictable services running at enormous scale.
Training remains a tougher market for challengers because failures, communication delays and software problems become far more expensive when thousands of accelerators work on one job for weeks.
CUDA still creates meaningful lock-in, but the lock-in increasingly comes from operations rather than source code. Moving a model is easier than reproducing its previous speed, reliability and monitoring environment.
Nvidia’s financial strength reinforces the technical moat. It can fund several product lines, reserve manufacturing capacity, support customers and absorb product transitions at a scale few rivals can match.
The most likely outcome is fragmentation rather than a clean defeat. AMD should win major deployments, custom chips should absorb selected hyperscale workloads, and Nvidia can still grow rapidly while losing market share.
Our judgment is that Nvidia should remain first through the next two platform cycles. Its exclusivity will fade faster than its leadership.
This market map, featured in our AI chip market deck, highlights top companies and startups in the AI chip market
What does “ahead” actually mean for Nvidia?
Today, Nvidia is overwhelmingly ahead as a business and a platform, while the chip-level gap can be surprisingly small.
We can measure Nvidia’s lead in four useful ways.
Commercial scale tells us how much infrastructure customers are buying. Chip performance compares accelerators on defined workloads. System capability looks at how well thousands of chips, CPUs and networking components operate together. Ecosystem depth covers the software, tools and skills needed to run those systems.
Nvidia dominates the first measure. It usually leads the second, although AMD is now close on several inference workloads. Its advantage grows again when we examine complete systems and the software surrounding them.
A company can build a chip that performs well in a benchmark without being able to deliver thousands of reliable racks, support every popular model or help customers solve failures inside a giant cluster. That is the real divide.
So we cannot describe Nvidia as simply “three years ahead” across everything. It is more than ten times ahead of AMD in current Data Center sales, around one deployed rack-scale generation ahead in complete systems and much closer in certain inference tasks.
Why are people questioning Nvidia’s lead now?
Nvidia’s lead is being questioned because alternatives have moved from experiments to signed, gigawatt-scale deployments.
AMD has recently added Anthropic and Microsoft to a customer list that already included OpenAI and Meta. The quantified OpenAI, Meta and Anthropic agreements cover up to 14 gigawatts of future AMD infrastructure. Microsoft also plans to deploy AMD’s Helios rack-scale system on Azure for frontier-model inference.
Those announcements are much more serious than a startup testing a few AMD servers. A single gigawatt can support hundreds of thousands of accelerators and require several billion dollars of surrounding infrastructure.
Custom chips are advancing at the same time. Google, Amazon and several large AI companies are building processors around their own workloads, often with help from Broadcom. These buyers now have three ways to reduce their dependence on Nvidia: purchase AMD accelerators, design their own chips or use both.
Yet Nvidia’s latest Data Center revenue still grew 92% year over year. Challengers are securing future deployments while Nvidia’s existing business continues to accelerate.
The market is broadening rather than turning against Nvidia. Nvidia can lose future share and still sell far more systems.
As this chart shows, and as featured in our AI chip market deck, search interest in AI chips has grown significantly
How much bigger is Nvidia’s AI business than AMD’s?
Nvidia’s Data Center business is currently about thirteen times larger than AMD’s, and that comparison still flatters AMD.
Nvidia generated $75.2 billion of Data Center revenue in its latest quarter. AMD generated $5.8 billion. AMD’s figure includes both Instinct accelerators and EPYC server processors, while Nvidia’s figure is overwhelmingly tied to AI compute and networking.
Nvidia’s commercial lead is therefore wider than the headline ratio suggests.
Its Data Center networking revenue reached $14.8 billion during the quarter. Networking alone was 2.6 times larger than AMD’s whole Data Center division. It was also larger than AMD’s total company revenue.
The comparison gives us a clearer picture than market-share estimates. AMD may be a credible technical competitor, but it operates at a completely different scale today.
At AMD’s current quarterly pace, it would need almost three years to generate as much Data Center revenue as Nvidia generated in one quarter. AMD will probably grow much faster than that comparison assumes, but the starting distance is huge.
| Latest quarterly measure | Nvidia | AMD | Nvidia’s scale |
|---|---|---|---|
| Data Center revenue | $75.2B | $5.8B | 13.0× |
| Data Center networking revenue | $14.8B | $5.8B total Data Center | 2.6× |
| Total company revenue | $81.6B | $10.3B | 8.0× |
Is Nvidia still ahead in raw chip performance?
On raw inference performance, Nvidia is ahead by far less than its revenue lead suggests.
The latest standardized MLPerf inference results show AMD’s MI355X reaching 92% to 104% of Nvidia B300 performance across three Llama 2 70B scenarios. Against Nvidia’s older B200, AMD tied the offline result and won the interactive test.
AMD was slightly further behind on the new GPT-OSS-120B test, reaching 82% to 91% of B300 performance. That is close enough for price, availability and energy use to influence the customer’s choice.
A buyer does not always need the fastest chip. A processor delivering 90% of the performance may be more attractive when it is cheaper, easier to obtain or reduces dependence on one supplier.
MLPerf compares carefully defined systems and models. Results can change with different sequence lengths, model architectures, networking setups or software optimizations. Nvidia also submitted results across a broader range of workloads.
Still, the result is hard to dodge: AMD has closed much of the raw inference gap. Nvidia’s wider lead now comes from covering more workloads and connecting the chips into mature production systems.
| MLPerf inference scenario | AMD MI355X versus Nvidia B300 |
|---|---|
| Llama 2 70B offline | 92% |
| Llama 2 70B server | 93% |
| Llama 2 70B interactive | 104% |
| GPT-OSS-120B offline | 91% |
| GPT-OSS-120B server | 82% |
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 VC investment in AI chip startups
Is Nvidia further ahead in training than inference?
Nvidia is clearly further ahead in large-scale AI training than it is in inference.
Training a frontier model requires thousands of accelerators to work on one job for weeks or months. A slow connection, unreliable node or software failure can leave hundreds of expensive chips waiting and waste a large amount of compute.
Nvidia’s advantage grows under those conditions. In MLPerf Training 6.0, Nvidia submitted results across all available benchmarks and demonstrated systems running at far larger scales. AMD made meaningful progress and showed multi-node training, but its coverage and maximum demonstrated scale remained narrower.
The training gap comes from several layers working together. NVLink connects GPUs within a system. InfiniBand and Spectrum-X connect systems across a data center. CUDA libraries optimize the model, while Nvidia’s management tools detect and isolate hardware failures.
AMD can catch individual layers faster than it can reproduce the whole operating environment. Its new hyperscale customers should speed up that work because they will expose ROCm and Helios to much larger production workloads.
For now, Nvidia remains the safer choice for companies building the largest training clusters. AMD is becoming credible, but it has less proof at the scale where small reliability differences get very expensive.
Is Nvidia one chip generation ahead or one system generation?
Nvidia is currently one deployed rack-scale generation ahead of AMD and is already ramping the next one.
Blackwell systems are operating at large scale, and Blackwell now represents most of Nvidia’s Data Center compute shipments. Nvidia has also moved its Vera Rubin platform into production.
AMD’s comparable rack-scale platform, Helios, combines MI455X accelerators, EPYC Venice processors, Pensando networking and ROCm. Shipments to Microsoft and other initial customers are expected to start during the second half of the year.
The timing is awkward for AMD. Helios is designed to compete with Nvidia’s latest integrated systems, but Nvidia is moving customers from Blackwell toward Rubin as Helios begins shipping.
AMD’s chip architecture is not two generations behind. Its MI355X already competes closely with Blackwell accelerators on several inference workloads. The larger delay appears when we look at production racks, networking, software and customer deployments together.
A fresh example comes from Japan’s new national AI infrastructure project. The announced system includes 27,500 Rubin GPUs and 13,750 Vera CPUs across 140 megawatts of capacity. Rubin has moved beyond a product roadmap and into concrete infrastructure plans.
Our estimate is that Nvidia holds a 12-to-18-month deployment lead in complete rack-scale systems, even though the gap between individual chips can be much smaller.
This chart, featured in our AI chip market deck, shows how Nvidia is leading in AI chips
Is CUDA still Nvidia’s biggest advantage?
CUDA still gives Nvidia its deepest advantage, but the lock-in now comes from operations more than source code.
Nvidia says more than six million developers use CUDA. They work with libraries, profilers, communication tools, optimized kernels and applications built over two decades.
A company already using Nvidia has engineers who know the tools, scripts written around the platform and internal systems designed to monitor Nvidia clusters. New employees are also more likely to arrive with CUDA experience than ROCm experience.
Moving the basic model code has become easier. PyTorch deliberately uses the same torch.cuda interface on AMD’s HIP platform, so many applications can start running on AMD hardware with few code changes. ROCm now officially supports widely used tools including PyTorch, JAX, vLLM and SGLang.
The harder work begins after the model runs. Engineers still need to match the previous speed, memory use and reliability. They must test every dependency, tune the kernels and make sure monitoring systems behave correctly.
Large companies such as Microsoft, Meta, OpenAI and Anthropic can assign hundreds of engineers to that job. Most enterprises cannot. Not even close.
CUDA is a broad convenience and confidence advantage. It saves customers time across thousands of small technical decisions, and those savings often outweigh a lower accelerator price.
If you want more recent data on this point, please see our latest AI chip market report.
Is Nvidia’s networking lead even bigger than its GPU lead?
Nvidia may be further ahead in networking than it is in GPUs.
Its latest quarterly Data Center networking revenue reached $14.8 billion, up 199% from the previous year. That operation already generates more than twice AMD’s entire Data Center division.
Networking becomes more valuable as clusters grow. Thousands of accelerators constantly exchange model updates, retrieve data and coordinate work. Faster GPUs provide limited benefit when they spend too much time waiting for information from another rack.
Nvidia controls much of that route. NVLink handles communication inside its systems. InfiniBand and Spectrum-X Ethernet connect racks. BlueField processors manage infrastructure and data movement.
This lets Nvidia optimize compute and communication together. AMD currently depends on a more open mix involving Pensando networking and technology from several outside suppliers.
Open systems could eventually become an advantage for AMD. Large customers dislike being locked into one networking architecture, and Broadcom remains a powerful supplier of Ethernet components. Microsoft is already integrating Pensando hardware more deeply into Azure.
Today, though, Nvidia can offer customers a complete path from one accelerator to an entire cluster. Its networking lead may be the hardest part of the platform to copy quickly.
This chart, featured in our AI chip market deck, shows annual funding in AI chip startups
Has Nvidia’s financial strength become unfair?
Nvidia’s financial strength now shapes the competition almost as much as its technology does.
The company generated $53.5 billion of operating income in its latest quarter. That was more than five times AMD’s total revenue. Nvidia also produced $50.3 billion of operating cash flow in just three months.
Nvidia spent $6.3 billion on quarterly research and development, slightly more than AMD’s whole Data Center division generated in revenue. It can fund several chips, networking products and software platforms at once without putting the business under financial pressure.
That money also helps outside the laboratory. Nvidia can reserve manufacturing capacity, build huge internal computing clusters, invest in customers and support cloud providers that need help financing new infrastructure.
Its latest filing showed $27 billion of investment commitments and more than $42 billion invested in private companies. These relationships can strengthen demand for Nvidia infrastructure while giving the company visibility into emerging AI businesses.
The 74.9% gross margin adds another advantage. Nvidia can absorb expensive product transitions, export restrictions and temporary inventory problems while continuing to spend aggressively. AMD’s latest gross margin was 53%.
A competitor may produce an excellent chip and still lack the money needed to secure packaging, build software, support customers and manufacture enough systems. Nvidia can fight across all those areas at the same time.
If you want more recent data on this point, please see our latest AI chip market report.
Is AMD finally a real alternative to Nvidia?
AMD is now a real alternative for hyperscalers, although most of its biggest wins have yet to become installed systems or reported revenue.
OpenAI and Meta have each agreed to deploy up to 6 gigawatts of AMD GPUs. Anthropic has added an agreement covering up to 2 gigawatts, with the first gigawatt scheduled for the first half of 2027. Microsoft will start deploying Helios on Azure as AMD begins shipping the platform.
Together, the three quantified agreements cover as much as 14 gigawatts. That would represent hundreds of thousands of accelerators and tens of billions of dollars in infrastructure if fully completed.
The word “if” deserves attention. These are multi-year agreements with phased rollouts. Some contain warrants, strategic investments or performance conditions. The full amounts should not be treated like guaranteed purchase orders arriving next quarter.
The first Meta and OpenAI gigawatts are scheduled to begin with MI450-based systems. Anthropic starts later. AMD still needs to manufacture the accelerators, deliver Helios racks and prove that ROCm works reliably at that scale.
Even so, customer behavior has changed. The largest AI companies are helping AMD design systems, optimizing models for ROCm and reserving large amounts of future capacity.
AMD has crossed the line between backup supplier and serious second platform. It remains far behind Nvidia commercially, but dismissing it as a benchmark-only competitor no longer fits the evidence.
This chart, featured in our AI chip market deck, compares the main business model options for AI accelerator chip companies
Are custom chips a bigger threat than AMD?
Custom silicon is the bigger long-term threat because it can remove entire workloads from Nvidia’s addressable market.
AMD competes with Nvidia by selling another general-purpose accelerator. A customer switching from Nvidia to AMD still buys a merchant GPU platform.
A custom chip changes the calculation. Google, Amazon or OpenAI can design hardware around workloads that run repeatedly at enormous scale. Once those workloads move onto internal silicon, the customer may stop shopping for general-purpose accelerators altogether.
Broadcom’s latest quarterly AI semiconductor revenue reached $10.8 billion, up 143% year over year. That figure includes custom accelerators and networking. It is already almost twice the size of AMD’s Data Center revenue, and Broadcom expects it to reach roughly $16 billion in the following quarter.
Google has now developed several TPU generations. Amazon continues expanding Trainium. These programs have survived long enough to show that custom AI chips are part of long-term infrastructure plans rather than short-lived bargaining tactics.
Nvidia keeps one major advantage. General-purpose infrastructure adapts more easily when models, data types or training methods change. A custom chip works best when the owner has a stable workload large enough to repay the design cost.
Custom silicon will take selected workloads rather than replace Nvidia everywhere. Those selected workloads could still be the largest and most profitable ones.
If you want more recent data on this point, please see our latest AI chip market report.
Will inference weaken Nvidia’s lead?
Inference will probably reduce Nvidia’s market share while making Nvidia’s business much larger.
Frontier-model training is concentrated among a small number of companies operating huge clusters. They value flexibility, reliability and rapid deployment, which plays directly to Nvidia’s strengths.
Inference is more fragmented. A customer-service chatbot, a coding agent and a video generator use compute differently. Some workloads prioritize latency, others throughput, memory capacity or electricity costs.
This variety gives AMD and custom chips more openings. AMD already comes close to Nvidia on several inference tests. Google and Amazon can also tune their own chips for predictable services running billions of times.
Customers become more sensitive to cost once a model reaches production. A 15% saving per query can be worth far more than a small performance advantage when the service handles hundreds of millions of requests.
Nvidia is adapting by optimizing complete inference systems and introducing processors aimed at specific workloads. Its latest MLPerf results also show how much software improvements can raise performance without changing the hardware. Nvidia reported that software tuning increased GB300 throughput on one reasoning benchmark by as much as 2.7 times within six months.
Inference will make the market less uniform. Nvidia should remain the largest supplier, but customers will increasingly mix Nvidia GPUs, AMD accelerators and internal chips according to workload.
This chart, featured in our AI chip market deck, shows how revenue is split across customer segments in the AI chip market
Can Nvidia’s biggest customers turn against it?
Nvidia’s largest customers are also the only companies rich enough to weaken its moat.
Three direct customers represented 21%, 17% and 16% of Nvidia’s latest quarterly revenue. Together, they accounted for 54%.
Direct customers can include cloud providers, manufacturers and distributors, so those percentages do not map perfectly onto final users. Even with that caveat, Nvidia depends heavily on a small group of very large buyers.
Those buyers have strong reasons to find alternatives. They want better prices, more supply and greater control over future products. They can also afford specialized engineering teams that smaller customers lack.
Meta is a good example. It plans to deploy millions of Nvidia Blackwell and Rubin GPUs, while also working with AMD on a possible 6-gigawatt deployment and contributing designs to open rack standards. Nvidia remains central to Meta’s infrastructure, but it no longer has the account to itself.
The same pattern appears across the industry. Large customers may reduce Nvidia’s percentage of their spending while purchasing more Nvidia hardware in absolute terms.
A sudden revenue collapse looks unlikely. Gradual pressure on market share, pricing and contract terms is the more credible risk.
How badly has Nvidia been hurt by losing China?
So far, losing China has barely dented Nvidia’s growth, while giving Chinese chipmakers a protected market in which to improve.
Nvidia now assumes no Data Center compute revenue from China in its next-quarter outlook. Despite that absence, the company expects total quarterly revenue of roughly $91 billion.
That forecast shows how much demand Nvidia can currently find elsewhere. Its latest Data Center revenue still grew 92%, even as export restrictions largely closed one of the world’s biggest technology markets.
The immediate damage was real. Earlier restrictions on the H20 accelerator forced Nvidia to record a $4.5 billion charge related to inventory and purchase commitments.
The larger risk will take longer to appear. Chinese cloud providers and model developers now have stronger incentives to optimize their software for domestic accelerators. Local chip companies receive orders, feedback and funding that would otherwise have gone to Nvidia.
Even a weaker accelerator can improve quickly inside a large protected market. China could eventually develop a separate hardware and software ecosystem that competes across emerging markets and countries facing similar restrictions.
China will probably not cost Nvidia its global lead during the next product cycle. Over a decade, it could prevent Nvidia from remaining the universal standard.
This chart, featured in our AI chip market deck, shows how AI accelerator chip technology has evolved over time
Could supply constraints slow Nvidia before competitors do?
Nvidia is more likely to be slowed by power, packaging and deployment bottlenecks than by one rival chip.
AI accelerators depend on advanced semiconductor manufacturing, high-bandwidth memory, complex packaging, networking equipment and liquid cooling. All these components must arrive in the correct order before a customer can switch on a cluster.
Nvidia’s latest inventory reached $25.8 billion, up from $21.4 billion one quarter earlier. Work in progress alone approached $10 billion. Those figures show the scale of the pipeline Nvidia must manage as it shifts from Blackwell toward Rubin.
Annual product launches make that job harder. Nvidia must ramp a new architecture while customers continue requesting the previous one. Ordering too little loses sales, while ordering too much can leave the company with expensive components tied to an older product.
The constraint increasingly sits outside Nvidia as well. Customers need land, power connections, cooling infrastructure and construction crews. A delayed substation can hold back an AI cluster even when the GPUs are ready.
Supply problems create openings for competitors. A customer unable to obtain enough Nvidia systems may install AMD or custom chips rather than leave a data center empty.
Availability can turn a temporary second choice into a permanent part of the customer’s architecture. That may be AMD’s fastest route to market share.
Can Nvidia stay ahead through the next two generations?
Nvidia will probably remain first through the next two platform cycles, although it will no longer own the market by default.
Rubin is already entering production as AMD prepares the first Helios shipments. Nvidia starts the cycle with working Blackwell clusters, established networking and customers already planning large Rubin deployments.
AMD enters with its strongest customer pipeline so far. Meta, OpenAI, Microsoft and Anthropic will expose Helios and ROCm to workloads large enough to reveal problems quickly. Their engineers can then help fix those problems.
This should make AMD improve faster than it did during earlier accelerator generations. The company is also previewing the MI500 series, which gives customers some confidence that Helios will be followed by another serious platform rather than a one-off product.
Custom accelerators will progress alongside AMD. Google, Amazon and Broadcom-backed projects do not need to beat Nvidia everywhere. They only need to become the best option for several large, repetitive workloads.
No competitor currently combines Nvidia’s sales scale, training performance, networking, software and deployed experience. Catching one layer will happen regularly. Catching all of them during the same product cycle looks unlikely.
Nvidia should remain the overall leader. The market around it will change: AMD as a genuine second platform, custom chips handling predictable hyperscale workloads and more open networking between them.
In our AI chip market deck, we identify pain points entrepreneurs should prioritize
So, how far ahead is Nvidia?
Nvidia is several years ahead as an AI platform, roughly one deployed rack generation ahead of AMD and only a few percentage points ahead on some inference workloads.
Commercially, the distance remains extreme. Nvidia’s Data Center business is about thirteen times larger than AMD’s, and its networking operation alone is much larger than AMD’s accelerator and server division.
Technically, the answer is less flattering to Nvidia. AMD’s MI355X can match or nearly match Nvidia’s latest accelerators on several inference tests. Any claim that Nvidia is universally “years ahead” in chip performance is no longer credible.
Nvidia’s durable advantage appears when we combine the layers. It can provide GPUs, CPUs, high-speed interconnects, Ethernet, complete racks, optimized libraries and experienced engineers at a scale no rival currently matches.
CUDA probably leaves Nvidia three to five years ahead across the broad market because millions of developers and thousands of applications have accumulated around it. Hyperscalers can shorten that gap dramatically by spending heavily on their own software.
Our final judgment is clear. Nvidia remains far ahead and should stay first through the next two generations. AMD can narrow the hardware gap and win major deployments, while custom chips gradually take high-volume workloads. Neither has yet built an equally complete platform.
Nvidia’s lead will probably shrink through fragmentation rather than disappear through one decisive defeat.
| Dimension | Our current judgment | Estimated Nvidia lead |
|---|---|---|
| Commercial scale | Overwhelming | More than 10× AMD |
| Individual inference workloads | Small and inconsistent | Near parity to around 20% |
| Frontier-model training | Clear | About one platform cycle |
| Rack-scale deployment | Strong | Roughly 12–18 months |
| Networking | Very strong | Several product and deployment cycles |
| Software and operations | Deepest advantage | Roughly 3–5 years for the wider market |
| Overall position | Durable leadership | Several years, with falling exclusivity |
If you want more recent data on this point, please see our latest AI chip market report.
OUR METHODOLOGY
This analysis tests how far ahead Nvidia is across the parts of AI infrastructure that customers actually buy and operate. We compare commercial scale, accelerator performance, large-scale training, rack deployment, networking, software, customer adoption and financial capacity.
We analyzed each dimension separately before bringing them together into one overall judgment. A single benchmark cannot explain platform leadership, just as revenue alone cannot show whether a rival chip has become technically credible.
We gave the greatest weight to recent, observable evidence. Reported revenue, standardized benchmark results, systems already in production and named customer deployments were treated as stronger evidence than product roadmaps, maximum contract values or uninstalled future capacity.
Future agreements were included when they materially changed the competitive picture, but we kept them separate from completed shipments and recognized revenue. Multi-year gigawatt commitments can establish strategic intent without proving that every planned system will be delivered.
For chip performance, we used MLCommons results to compare defined workloads while recognizing their limits. Benchmark performance can change with model architecture, sequence length, software tuning, networking configuration and system scale, so no individual result was treated as a universal ranking.
For commercial scale and financial strength, we relied on company filings and investor materials. AMD’s Data Center revenue includes both accelerators and EPYC server processors, which means the simple revenue ratio understates Nvidia’s lead in AI accelerators and networking.
For software and operations, we examined CUDA, ROCm and support across major frameworks and inference tools. We distinguished between getting a model to run and reproducing the performance, reliability, monitoring and operational confidence of an established production environment.
Our estimates for Nvidia’s 12-to-18-month rack deployment lead and three-to-five-year broader software advantage are editorial judgments rather than outputs from a scoring formula. They reflect the convergence of deployment timing, benchmark breadth, ecosystem maturity, customer evidence and operating scale.
We prioritized first-hand and authoritative sources, including Nvidia investor materials, AMD investor materials, MLCommons Inference results, MLCommons Training results, Nvidia’s CUDA documentation, AMD’s ROCm documentation, AMD’s newsroom, Microsoft’s official blog, Anthropic’s news releases, OpenAI’s announcements, Meta’s newsroom, Broadcom investor materials, and company filings available through the SEC’s EDGAR database.
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
Related blog posts
- Is Nvidia becoming too powerful?
- Why is Nvidia financing OpenAI?
