Why AMD Will Outperform NVIDIA Over the Next Decade (And It’s Not Just About Chips).
The demand for computation, is in principle, unlimited in nature.
Introduction:
AMD is set to report its Q2 2025 earnings on August 5th. But before we look ahead, I want to take a moment to reflect on the past quarter and highlight a few broader themes that shape my long-term thesis.
As I’ve mentioned before, my conviction in AMD is built on several core ideas - which can be summarized in just a few sentences:
1) The demand for computation is - in principle - unlimited.
2) Everything around us - the order - is a function of applied computation (the brain).
3) Now computation is replicated via chips.
The central takeaway here is that AMD is not merely in the business of chips: they are in the business of computation, at the core.
The demand for computation, is in principle, unlimited in nature. Fyrther, due to the evolving nature of the computation market, the company with the greatest agility is the company that is poised to capitalise on this evolving market best. AMD – uniquely due to their chiplet architecture – has this corporate agility, of which has manifested throughout the whole culture.
In sum, the future of the computational market is, by nature, uncertain. New architectures, paradigms, and use cases will emerge. Therefore, the company best positioned to lead is not necessarily the one with the biggest market share today, but the one with the greatest agility. Success will hinge on iteration speed, agility in architecture and product design, and the capacity to pivot toward new opportunities. In this regard, AMD’s chiplet architecture, heterogeneous computing model, and culture under Lisa Su provide it with a structural advantage.
We have some really interesting snippets from past earnings calls that I want to reiterate and go over.
AMD Q1 25: Earnings Call.
Management on FPGAs & Edge Computing:
CEO: Growth accelerated for the fourth consecutive quarter year over year, driven by strength in our core businesses and expanding data center and AI momentum.
As a result, first quarter revenue increased 36% year over year to 7.4 billion as our data center and client and gaming segments both grew by a large double digit percentage.
Despite the uncertain macroeconomic backdrop, our first quarter performance highlights the strength of our differentiated product portfolio and execution and positions us well for strong growth in 2025.
Turning to the segments, data center segment revenue increased 57% year over year to 3.7 billion.
More than 30 new instances launched from Alibaba, AWS, Google, Oracle, Tencent, and others in the quarter, including the initial wave of 5th Gen EPYC Turin instances. In addition, AWS launched new FPGA accelerated instances in the quarter, powered by EPYC processors with Xilinx Virtex FPGAs that are optimized for data and compute intensive workloads like genomics, multimedia processing, network security, and cloud-based video broadcasting.
The really vital takeaway here is that: AWS launched a new FPGA accelerated instance in the quarter, powered by processors with Xilnix. So what does this mean?
The Xilinx Acquisition: AMD’s Strategic Expansion Into Adaptive Compute:
In early 2022, AMD completed its $49 billion all-stock acquisition of Xilinx, the leading provider of FPGAs (Field-Programmable Gate Arrays) — highly flexible, reprogrammable chips used across aerospace, data centers, automotive, and telecommunications.
This wasn’t just about acquiring more silicon. It was about extending AMD’s architectural reach beyond CPUs and GPUs into adaptive computing — a category of compute where hardware can be optimized post-deployment, depending on the use case.
Here’s why this matters:
Diversification of Product Portfolio: Xilinx brings AMD exposure to high-margin, mission-critical markets like aerospace, 5G infrastructure, and industrial automation - beyond AMD’s traditional gaming and PC markets.
Heterogeneous Computing Vision: Combining AMD’s CPUs/GPUs with Xilinx’s FPGAs allows AMD to build heterogeneous systems — systems where different types of processors (general-purpose and programmable) work in tandem. This is crucial for AI, edge computing, and real-time processing.
Software and Ecosystem Leverage: Xilinx has deep software tools (like Vivado and Vitis), long-term enterprise relationships, and robust IP portfolios — all of which expand AMD’s ecosystem defensibility and customer stickiness.
Think of it this way: NVIDIA owns the AI narrative with GPUs. But AMD is building the full stack — CPUs, GPUs, FPGAs, and the software glue to bind them. The Xilinx acquisition is a major step toward making AMD the platform company for next-gen compute — from data centers to defense systems.
Semi-conductors are no longer merely about C/GPUs. Rather, they are about moving electrons around and generating insights. Artificial intelligence shall become pervasive across our world.
As Su has said in the past, the future of the chip market is not merely a singular chip that shall cover all work-loads and use-cases. Rather, different types of computing shall be required for the novel job and workloads that are demanded. Thus, she explained, you need CPUs, GPUs, FPGAs, and other forms of chips, in which shall aid this diverse range of workloads for the future.
CEO: Then, we also believe that you need different types of computing for all the jobs or workloads that there are in the world, and so you need CPUs, GPUs, FPGAs. That was our acquisition of Xilinx. I’ve always believed that you need all of these components, and our investments in this area have really accelerated as AI has become such a bigger platform. My conversation about AI is: We’re still in the early innings. There is no one size fits all when it comes to computing. Depending on what you’re trying to do, you’re going to need different technology.
Vitally, there is no one-size-fits-all model for chips. Depending upon what you are trying to do, you are going to need different technologies.
The issue with the market today is that everyone believes that the market is merely concerned with CPUs and GPUs. This is because, computation is not thought of at a fundamental level. Computation at the most fundamental level is merely the manipulation of information based on physical processes.
We will enter into a world whereby more and more technologies are connected to the internet. This is commonly touted as “Edge Ai”. In order for C/GPUs to drive incremental productivity, they must sit in a highly connected and self-optimising environment.
In the following decades, the economy will shift towards collecting data from endpoints and processing it (training AI) to then yield insights (AI inference). This will mean that our economy shall move from being terrible at making predictions to being great at them.
The Xilinix & Pensado acquisitions set AMD up very well for this future.
FPGA stands for Field-Programmable Gate Array. Imagine it as the LEGO of computing — a chip that can be reprogrammed even after it's manufactured, allowing it to be tailored for different tasks as needs evolve. Unlike traditional CPUs or GPUs, which have fixed functions baked in from the start, FPGAs offer a high degree of post-production flexibility.
Why does this matter? First, FPGAs are incredibly versatile. You can reconfigure them to perform specific functions long after deployment — whether in a data center, a satellite, or an autonomous vehicle. Second, they offer speed advantages for specialized tasks like signal processing, networking, or AI inference. Because they’re customized for the job, they avoid wasting compute cycles on irrelevant instructions. Third, FPGAs are highly efficient. For repetitive, low-latency workloads — like those found in edge devices, telecom infrastructure, or real-time systems — they consume less power than general-purpose processors.
This is exactly why AMD’s acquisition of Xilinx, the global leader in FPGA technology, was so strategically important. It gives AMD a differentiated foothold in high-performance, real-time computing — in markets that CPUs and GPUs alone can't fully address.
The FPGA market is projected to grow from ~$7B in 2022 to ~$15B+ by 2028, driven by demand for real-time, programmable, and low-latency processing. But the more important point is strategic relevance — FPGAs are becoming mission-critical across industries where standard chips fall short.
FPGAs (Field-Programmable Gate Arrays) are becoming increasingly essential in the modern computing landscape — not as competitors to CPUs and GPUs, but as powerful complements that unlock new capabilities across specific domains. Their true value lies in their flexibility, reprogrammability, and efficiency — especially at the edge, in AI inference, and in mission-critical systems.
One of the most important applications for FPGAs is edge computing. Edge devices — such as smart cameras, drones, industrial sensors, and medical scanners — can’t afford the latency of cloud-based processing. They need local, real-time compute. FPGAs are uniquely suited for these environments because they deliver high performance at low power, and can be reprogrammed in the field to adapt to new models, tasks, or regulatory requirements. This makes them ideal for fast-moving environments like autonomous vehicles, industrial IoT, smart cities, and 5G base stations.
FPGAs are also gaining traction in AI inference — the stage where trained AI models are deployed in the real world to make decisions. While GPUs dominate AI training, FPGAs can outperform them in certain inference workloads, especially those that are repetitive and latency-sensitive. Compared to GPUs, FPGAs can be more power-efficient and customized for very specific use cases — such as voice recognition on edge devices, barcode scanning in airports, or smart surveillance systems. And unlike ASICs, which are hardwired for specific tasks, FPGAs can be adapted as models evolve.
In mission-critical and high-security systems, FPGAs are already well-established. Their deterministic performance, low latency, and reprogrammable nature make them the go-to hardware for sectors like aerospace, defense, telecommunications, and advanced medical imaging. These are domains where failure isn’t an option, and the ability to tweak hardware-level functionality without replacing physical components is invaluable.
From an investment standpoint, AMD’s acquisition of Xilinx — the market leader in FPGAs — is not about competing head-to-head with NVIDIA on GPUs. It’s about building out the next layer of the compute stack. FPGAs give AMD a strategic foothold in areas where flexibility and adaptability are more important than sheer power. This positions AMD to lead in the era of heterogeneous computing — systems that blend CPUs, GPUs, and FPGAs to deliver optimal performance across an increasingly diverse range of workloads.
In short, FPGAs are a key pillar of AMD’s strategy to move beyond traditional processor markets and into the future of customizable, scalable, and adaptive computing.
What Is Edge Computing — And Why It Matters
Edge computing refers to the process of handling data closer to where it is generated — at the “edge” of the network — rather than relying on centralized cloud data centers. Whether it’s a car, a camera, a factory robot, or a smartphone, the idea is to enable real-time decision-making and computation locally, where it's needed most.
This architectural shift is driven by fundamental constraints in latency, bandwidth, and privacy. In many scenarios, sending data back and forth to the cloud is not only inefficient but practically impossible — especially in time-sensitive or highly regulated environments.
In applications like autonomous vehicles, robotics, and remote surgery, every millisecond counts. Waiting for a cloud server to process information and send back a response can introduce fatal delays. Edge computing eliminates this round-trip latency, delivering ultra-fast performance exactly where it’s required. The cloud may be powerful, but it’s often too slow. Edge computing, by contrast, is fast — and increasingly intelligent.
By 2025, Gartner predicts that 75% of enterprise data will be created outside traditional data centers. This explosion of data at the edge — from sensors, devices, and machines — requires real-time processing. Edge computing enables organizations to analyze and act on data at the point of creation, rather than overwhelming centralized systems.
As AI continues to proliferate across industries, inference — the process of applying trained AI models — is increasingly happening on the edge. Devices like drones, industrial machines, wearables, and AR/VR headsets require local AI processing for speed, efficiency, and responsiveness. This is where specialized hardware like FPGAs, ASICs, and compact accelerators (many now part of AMD’s portfolio through its Xilinx acquisition) become essential. They provide the computing power required for intelligent decision-making directly on the device.
In sectors like healthcare, finance, and defense, strict data privacy laws often prohibit sending sensitive information to the cloud. Edge computing enables secure, local data processing — ensuring compliance with regulations while maintaining performance and confidentiality.
What This Means for the Market
Edge computing is rapidly emerging as one of the most important megatrends in technology. The global edge computing market is projected to grow from $53 billion in 2023 to over $155 billion by 2030 and beyond. This surge is being driven by sectors such as automotive, manufacturing, smart cities, defense, healthcare, and telecommunications — all of which require high-performance, low-latency computing close to where data is generated.
In this evolving landscape, the winners will be those companies that can deliver powerful, energy-efficient, and reprogrammable compute capabilities at the edge. These solutions must handle real-time processing, adapt to changing software needs, and operate reliably in decentralized environments.
AMD is uniquely positioned to capitalize on this shift, thanks to its strategic acquisitions of Xilinx and Pensando. Xilinx brings industry-leading FPGA technology to the table — enabling real-time, programmable compute at the edge. Pensando adds a layer of software-defined networking, tailored for edge environments that demand flexibility and security.
At the same time, AMD’s Zen architecture provides the efficiency and scalability needed to power a wide range of edge devices and workloads. Together, these technologies enable AMD to deliver a comprehensive, heterogeneous compute stack — blending CPUs, GPUs, and FPGAs in a way that no other company currently can.
This positions AMD not just to participate in the edge computing market, but to lead it — by offering the most versatile and adaptive solutions for a rapidly evolving, decentralized world.
Management comments on inference:
CEO: Turning to our data center AI business, revenue increased by a significant double digit percentage year over year, as MI325X shipments ramped to support new enterprise and cloud deployments.
Several hyperscalers expanded their use of instinct accelerators to cover an increasing range of generative AI search, ranking, and recommendation use cases.
We also added multiple tier one cloud and enterprise customers in the quarter, including one of the largest frontier model developers that is now using instinct GPUs to serve a significant portion of their daily inference traffic. The depth and breadth of our customer engagements continues to expand as breakthroughs in large scale AI models like OpenAI's O3 and DeepSeek's R1 drive increased demand for traditional inferencing and increasingly as a critical part of pre-training.
Training engagements also ramped in the quarter, as multiple tier one hyperscale AI and enterprise customers scaled instinct GPU clusters to train internal and next gen frontier models.
In parallel, we're making meaningful progress with sovereign AI deployments as countries expand investments to establish domestic nation scale AI infrastructure. In February, we announced a strategic partnership with G42 to build one of France's most powerful AI compute facilities powered by instinct accelerators.
On the AI software front, we significantly accelerated our release cadence in the first quarter, shifting from quarterly ROCm updates to delivering ready to deploy training and inferencing containers on a bi-weekly basis that include performance optimizations and support for the latest libraries, kernels, and algorithms.
Why AMD is set to dominate the inference market:
Firstly, what actually is inference?
The AI market can be divided into two separate phases: a) inference, b) training. The training phase - “learning” - is when AI models are being built; the model learns from massive datasets. This includes learning from millions of images, texts, or videos. The goal of this market is the training of neural networks to recognise patterns, generate predictions or understand context.
The training market for AI is currently dominated by NVIDIA.
However, the AI market is solely concerned with "training”. In fact, the market also has another side, referred to as the “inference” market. The inference market is the “using” phase of artificial intelligence, namely when a trained AI model is actually used in the real world. This is inference - namely, giving predictions based on new inputs.
One example of this is Google Translate generating a translation, or Tesla detecting a pedestrian in real time. However, we are still in the infancy stages for the inference market. This was most recently emphasised by Jensen Huang’s view, of which he articulated that AI will be everywhere - not just in training labs, but instead at “the edge”. This means that AI will be; in your fridge, car, factory, phone, lights, and frankly ubiquitous across society.
This even led Jensen Huang to state this:
“Inference will be thousands of times larger than training in terms of deployments.”
Now, this quote - in and of itself - warrants serious consideration. The rise in stock price for NVIDIA - predicated upon their dominance in the GPU training market - frankly has been outstanding. Yet, Jensen is asserting that the inference market - a market many investors are currently overlooking - shall be even bigger than the current training market - citing the potential for at least” thousands” of times larger than training! This is surley shocking.
Thus, we have learned that the AI market consists of two central principles: the a) training, and the b) inference. Both refer to different segments of the market, and have separate use cases.
ARK Invest have estimated that inference costs are already declining by 90% yearly, while training costs are falling at 75% per year. This means that inference is becoming cheaper at a much faster rate than training.
This means that: suppose it costs $1.00 today to run an inference task. Next year, it would cost ~$0.10 to do the same task. A year after that, it would be ~$0.01.
This is an exponential cost decline, and it has massive implications.
This matters because training occurs occasionally - it often happens once per model, or per version. It is centralised and more irregular. However, inference happens millions of billions of times per day. Every time a chatbot answers, an AI camera sees something, or a device makes a decision.
So, even if training is expensive, inference is where most of the AI activity happens - and therefore, where most of the cost lives at scale.
Lower inference costs = mass deployment becomes viable.
This therefore indicates an inflection point: ARK projects that by 2026, the cost of running large models (like GPT-4) will be so low that inference can be done locally on consumer devices - no cloud server needed. This means that AI will move to the edge. Your smartphone or laptop will be able to run powerful AI models locally - fast, cheap, and privately.
As inference costs continue to plummet, AI is poised to become truly ubiquitous in daily life. Instead of being limited to cloud-based applications, AI will be embedded across a vast array of consumer technologies - apps, homes, vehicles, augmented reality glasses, personal assistants, and more. This shift to local, on-device AI (also known as edge inference) removes traditional barriers such as latency, data privacy concerns, and the high cost of cloud-based compute. Users will benefit from faster, more responsive experiences with zero reliance on internet connectivity or centralized servers. This is the foundation for seamless, personalized, and real-time AI across everyday tools.
Inference To Dominate:
CEO: We expanded our open source community enablement in the quarter, making significantly more instant compute infrastructure available to enable developers to automatically build, test, and deploy updates to ROCm code nightly. As a result, more than two million models on Hugging Face now run out of the box on AMD.
We're also enabling an increasing number of models to launch with day zero support for instant accelerators, including Meta's Llama 4, Google's Gemma 3, and DeepSeek's R1 models that were released in the first quarter. Beyond launch, we are delivering regular software updates that increase performance for new models. For example, in the weeks following the launch of DeepSeek's R1 model, we introduced ROCm optimizations that enabled MI300 to deliver leadership inferencing throughput.
Turning to our AI solutions capabilities. Earlier this quarter, we completed our acquisition of ZT systems, adding world class systems design expertise to complement our silicon and software leadership. With ZT, we can provide ready to deploy rack level AI solutions based on industry standards built with AMD CPUs, GPUs, and networking, reducing deployment time for hyperscalers and accelerating time to market for OEM and ODM partners.
Once again:
CEO: The depth and breadth of our customer engagements continues to expand as breakthroughs in large scale AI models like OpenAI's O3 and DeepSeek's R1 drive increased demand for traditional inferencing and increasingly as a critical part of pre-training.
Inference is set to dominate from here out – and AMD is uniquely positioned to capitalise on this trend.
Full Stack Solutions:
CEO: We have received significant interest in ZT's manufacturing business and expect to announce a strategic partner shortly.
Customer interest in the MI350 series is very strong, setting the stage for broad deployment in the second half of this year. As one example, we are partnering with Oracle to deploy a large scale cluster powered by MI355X accelerators, 5th gen EPYC Turin processors and Polara 400AINCs.
This multi-billion dollar initiative highlights the expanding AMD and OCI partnership and the growing demand for AMD instinct to power the next wave of large scale AI infrastructure. Looking ahead, our MI400 series development remains on track to launch next year. The MI400 series is designed to deliver leadership performance for both inferencing and training, scaling seamlessly from single servers to full data center deployments.
AMD’s acquisition of ZT Systems represents a strategic leap forward. With this move, AMD is no longer just a semiconductor company selling CPUs and GPUs to third parties. It now has the internal capability to design and deliver full AI infrastructure systems — complete data center racks, fully integrated, optimized, and ready to deploy.
This matters because in today’s AI landscape, customers don’t just want chips — they want end-to-end solutions. That’s what companies like NVIDIA have been doing so well with their DGX and HGX systems. These aren't just graphics cards; they’re full-stack offerings: integrated packages that include hardware, system design, drivers, software frameworks, and everything needed to run complex AI workloads at scale.
In technology, this is what’s referred to as the “stack.” A stack is the collection of layers that make up a complete computing system. At the base, you have hardware — CPUs, GPUs, and accelerators like AMD’s. Above that is system design — how those components are packaged, cooled, and interconnected. Then comes firmware, drivers, operating systems, and finally software frameworks like PyTorch or TensorFlow, all the way up to the AI applications themselves. Controlling more of the stack means controlling more of the value.
ZT Systems gave AMD deep capabilities in system-level architecture and integration, allowing it to package its chips into full-blown AI servers and racks. This is a huge unlock. But AMD didn’t stop there — it immediately sold off the manufacturing operations to a trusted partner, Sanmina. This allows AMD to focus on designing and optimizing the stack, while Sanmina handles production. It’s a capital-efficient way to scale without becoming the hardware manufacturer side of this segment.
In essence, AMD is now positioned to compete directly with NVIDIA, not just at the chip level, but at the system level. It can deliver AI infrastructure as a service — not just a part. With its leadership in chiplet design, its acquisition of Xilinx for programmable compute, and now the ZT Systems integration, AMD has quietly built the foundation to become a full-stack AI systems provider.
This is what makes the ZT acquisition so important: it marks AMD’s evolution into a company that can design the chips, build the systems, deliver the software, and partner for manufacturing — all while staying lean, agile, and modular. It’s not just a semiconductor company anymore. It’s a compute company — end to end.
CEO: Embedded demand continues to recover gradually. We expect improving demand in the test and measurement, communications, and aerospace markets will drive a return to growth in the second half of 2025. We completed initial shipments of our cost optimized parts and Ultrascale plus FPGAs and second generation versal AI edge SOCs to meet growing demand for AI at the edge.
FPGAs again & edge computing in inception.
CEO: We're also doubling down on our execution to deliver, and where possible, accelerate our industry leading roadmaps. We view the current environment as a strategic opportunity to further differentiate AMD as we deliver an expanding product portfolio that combines leadership, compute, and AI capabilities for data centers, edge PCs, and embedded end devices.
The future of the chip market – will not be confined by one singular chip, or even player. But rather, the market will have a variety of different manifestations and chips. AMD has a highly differentiated chip portfolio, which means that the company is poised to capitalise on this evolving trend nicely.
Analyst Key Questions:
CEO: Sure. So on the instinct ramp, I would say the a two, one performance of data center GPU was in line with maybe a little bit better than expected. I think the key point that we've always said about the instinct ramp is very excited about the MI350 launch. We're right on track for that launching mid-year.
I would say, customer interest has been very high. So from a competitiveness standpoint, we feel really good about where it's positioned. Overall, I think one of the advantages that we have with the MI350 launch is that from a systems overall environment, it's actually very similar to the MI300. So we believe it's going to ramp fast.
And we already have a couple of deals that have been announced, including a very important relationship with Oracle in terms of the MI350 series for a number of joint customers. So we're excited about the overall AI business. I think we continue to see strength there. I know there are some uncertainties as it relates to tariffs and other things.
AMD’s roadmap looks compelling. The MI350 (CDNA 4) launch in 2025 promises to provide as much as 35x greater AI inference performance. If this target is reached, AMD can increase its market share to 15–20% in two years. It won’t topple Nvidia, perhaps, but it’s enough to establish a multi-billion-dollar business segment with higher margin and diversification opportunities for AMD.
Once again, inference is the market to dominate here – and this is the real market in AI. Albeit, training is vital, inference is more vital – and is where the real money will be made.
Q: Hey, good afternoon. Thanks for taking my question. I know there's been a lot of focus on your upcoming MI350 series, Lisa, but MI400 next year is where you potentially close the competitive gap in a big way, right? You're bringing frontier class model training performance GPU, interact scale, solution.
CEO: I think look, we're excited about the MI350 series launch that's coming up, but we are extremely excited as well about the MI400 series and the roadmap there. I think we've been very active with customers on our roadmap. As you know, this is one of those areas where you absolutely have to be planning many quarters in advance for that.
I would say, the MI400 series enthusiasm from customers is high, and there's a lot of activities that are going on right now to ensure that we do in fact learn from some of the-- let's call it some of the challenges that have occurred with some of the recent deployments.
Q: Great. Thank you. One of the things your cloud customers have been talking about is this kind of growth in inference costs, the sort of reasoning models using a lot of inference compute and some tightness. Can you talk about that from AMD's perspective? Are you seeing that in your business? Does that change the focus that you have going forward?
CEO: Sure, Joe. So I think overall, what we're seeing is that with these new reasoning models, the inferencing is more important. And there's also a move to more distributed inferencing. So I think that plays into our strengths. I think we have demonstrated with MI300 that we are an excellent inference solution. And that holds true for 35 and 350 series as well. So we continue to see with our memory bandwidth and memory capacity advantages, that's a positive. I will say that as we're going into this, the number of workloads that we're seeing overall is expanding. So we're seeing both training and inferencing as important workloads that we're working on.
And our customers continue to demonstrate-- I think the desire that we're seeing probably from a trend standpoint is that there are many models that people are using today. So they're not necessarily using one model. They're actually using several different models. And so the optimizations around that are the things that we're doing with our ROCm software suite.
Once again here, highlighting importance of inference market.
Q: Great. And then just an update on your thoughts on competing with custom silicon with ASICs in the AI space. Most of your largest customers also have a custom silicon offering. So will they invest in both AMD and ASICs? And just how do they decide how to apportion that investment?
CEO: I mean, Joe, I view them as really two different things. I think one of the primary aspects, as we've talked about, the $500 billion TAM and the opportunities there, look, we think ASICs have a place. We happen to think GPUs have a larger piece of that because the models are changing so much. And from our standpoint, it's really important to have competitive TCO. And people want choice to get there, especially as inference costs become so important.
And we're working on trying to expand the overall inferencing capability out there. So I don't think it's an either/or. I think it's a let's get the best solutions out there, and we will certainly believe that we're very competitive in inferencing. And I think we are also becoming a much more solution for training as well.
MI350 Is The Most Powerful Inference Accelerator On The Market:
In her recent interview with CNBC, AMD CEO Dr. Lisa Su introduced the company’s latest AI chip, the MI350, calling it the most powerful inference accelerator on the market today. As artificial intelligence use cases expand rapidly across industries, customers are demanding greater efficiency and affordability. AMD’s latest generation of chips delivers 40% more tokens per dollar, making AI workloads more accessible and cost-effective at scale.
Looking ahead, AMD is committed to an annual cadence of AI accelerator launches, with its 2026 roadmap already in place. But the company’s strategy goes far beyond just chips. Through its acquisition of ZT Systems, AMD is expanding into rack-scale solutions, enabling it to offer full-stack AI infrastructure — not just components, but integrated systems optimized for performance, speed, and deployment at scale.
This strategy is already gaining traction. Today, 7 out of the top 10 AI companies — including OpenAI, Oracle, xAI, and Tesla — are using AMD’s AI products. What these organizations demand is clear: cutting-edge hardware that can immediately support production-scale AI loads. To meet this, AMD is investing heavily not only in hardware, but in the software stack and system integration, hiring across design engineering teams to reduce time-to-market and accelerate customer deployment.
A core part of AMD’s approach is its commitment to openness. Dr. Su emphasized that we are still in the early innings of the AI revolution. While there’s already been rapid adoption, the market is expected to grow at over 60% annually, reaching $500 billion by 2028. In this environment, customers want flexible, programmable, and open architectures — and AMD is delivering just that.
Dr. Su also stressed the importance of co-development with customers, where engineering teams from both sides collaborate deeply to shape the final solution. The MI350 itself offers 3x the performance of last year’s model, while AMD’s upcoming MI400 series is projected to deliver a 10x hardware improvement, setting a new bar for performance and scalability. At the heart of all this is AMD’s belief in the power of an open ecosystem — a foundation that drives faster innovation and more adaptable solutions for a rapidly changing AI landscape.
The MI350 and MI355 series — are not only competitive with NVIDIA’s industry-leading accelerators, but in many cases, they outperform them. According to Su, the MI355 chips, which began shipping earlier this month, deliver up to 35 times the performance of their predecessors, representing a major leap in computational capability.
While AMD still trails NVIDIA in overall AI market share, particularly in high-performance accelerators, the company is making a serious push to close the gap. Su now forecasts the AI chip market will exceed $500 billion by 2028, up from her earlier estimates, and positions AMD’s MI300 series as a key driver of that growth. The new chips are optimized to run AI software faster than NVIDIA’s B200 and GB200, and match or exceed their performance in training and code generation tasks — all while coming in at a significantly lower price point.
These advances are critical for AMD as it continues its strategic shift away from simply competing with Intel in the CPU market. With NVIDIA now dominating the AI landscape, AMD’s accelerator business has become the centrepiece of its future growth strategy. Although NVIDIA currently generates over $100 billion annually from AI hardware, AMD is starting to bring in billions of dollars as well, and expects its upgraded MI series to drive further adoption among major cloud providers and enterprise clients.
Despite previous concerns over slower-than-expected data center growth, AMD believes the new MI355 rollout will restore momentum, regain market confidence, and signal that it can go head-to-head with the world’s most dominant AI chipmaker.
Multipronged Strategy & Future Of Molecular Biology:
AMD is executing a multi-pronged strategy to challenge NVIDIA’s dominance in the AI chip market. While AMD remains a distant second in market share, its moves are increasingly strategic and targeted. One key front is AI infrastructure. Last month, AMD led a $333 million investment in Vultr, a GPU-as-a-service provider, securing a strong foothold in the infrastructure layer where many AI applications are deployed. This partnership builds on Vultr’s earlier adoption of AMD’s Instinct MI300X GPU and ROCm software stack, integrating AMD’s silicon deeper into cloud-based AI services.
Beyond infrastructure, AMD is expanding into vertical-specific markets—particularly healthcare. In a move first reported by The Wall Street Journal, AMD invested $20 million in Absci, a public drug discovery firm. According to AMD CTO Mark Papermaster, this signals a broader pivot:
CTO: “We’re now expanding our focus into vertical markets and prioritizing healthcare, where we can immediately have an impact on society.” This approach mirrors NVIDIA’s sector-specific playbook and gives AMD a chance to build traction in areas where compute needs are specialized and underserved.
AMD is also leveraging M&A as a growth engine for its AI ambitions. In August 2024, the company announced its $4.9 billion acquisition of ZT Systems, a leading provider of compute and storage servers for AI-driven environments. As CEO Dr. Lisa Su noted, the acquisition brings “world-class systems design and rack-scale solutions expertise,” allowing AMD to offer more integrated, end-to-end AI infrastructure — from chips to full data center solutions.
Collectively, these initiatives underscore AMD’s broader shift: it’s not just selling GPUs — it’s building an ecosystem. From hyperscaler infrastructure to industry-specific applications and full-stack systems, AMD is positioning itself to capture deeper value across the AI supply chain, challenging NVIDIA’s lead with calculated, long-term plays.
AMD’s $20 million investment in Absci — a company at the forefront of AI-driven drug discovery — reflects a broader trend: the deepening convergence of computation and biology. This move underscores a key insight about AMD’s long-term positioning. The company’s growth is not tied to the fortunes of any single chip or sector. Instead, it is anchored to the underlying demand for computation itself — a demand that spans industries, disciplines, and frontiers. And in principle, that demand is unlimited. As every field — from healthcare to energy to finance — becomes increasingly computational, AMD’s relevance will continue to scale alongside the world’s most complex and urgent challenges.
By targeting specific areas where it can effectively compete, AMD is steadily gaining ground in the AI accelerator market, even as it remains a clear second to NVIDIA. And that’s not a bad place to be. According to a January 6 note from Jefferies analyst Blayne Curtis, AMD’s market share in AI accelerators has doubled since August 2024, even though sales in November remained flat. As of that month, AMD held 5.3% of the market, while NVIDIA captured 98% of new accelerator deployments, bringing its total share to 85%.
Still, AMD’s focused, high-leverage strategy is beginning to carve out a meaningful slice of a massive opportunity. In its announcement of the ZT Systems acquisition, AMD forecasted a total addressable market (TAM) of $400 billion for AI accelerators by 2027. While NVIDIA may dominate the headlines, AMD is positioning itself to claim a solid and profitable share of what’s shaping up to be one of the most important markets of the next decade.
Partnerships With OpenAI:
At AMD’s recent "Advancing AI" developer conference in San Jose, CEO Dr. Lisa Su announced a major strategic milestone: OpenAI will adopt AMD’s latest AI chips, marking a significant endorsement from one of the most influential players in the space. Onstage alongside Su was OpenAI CEO Sam Altman, who confirmed that his team is working closely with AMD to co-develop the MI450 series, optimizing the chips for large-scale AI workloads. "Our infrastructure ramp-up over the last year — and what's ahead — has just been a crazy, crazy thing to watch," Altman remarked, underscoring the scale and urgency of demand in the AI infrastructure market.
Su unveiled details about AMD’s MI350 and upcoming MI400 series, which are designed to go head-to-head with NVIDIA’s Blackwell line. The MI400 chips will power AMD’s new Helios server, set to launch next year. Each Helios unit will pack 72 MI400 GPUs, mirroring NVIDIA’s NVL72 systems — a move that signals AMD’s full entry into the race to provide rack-scale AI infrastructure, not just chips.
The shift in the AI hardware market is clear: it’s no longer just about individual GPUs — it’s about delivering complete systems that combine compute, networking, and integration in a seamless platform. AMD is answering that call. Executives from Meta, Oracle, xAI (Elon Musk), and Crusoe — a specialized cloud provider — also joined the event, sharing their plans to deploy AMD hardware. Notably, Crusoe disclosed that it plans to spend $400 million on AMD’s latest chips, further validating the company’s momentum in the hyperscale AI market.
Dr. Su emphasized that AMD’s competitive edge lies not in a walled garden, but in an open, collaborative ecosystem. "The future of AI will not be built by any one company or in a closed ecosystem," she said. "It’s going to be shaped by open collaboration across the industry." AMD also reiterated its commitment to a rapid product release cycle, mirroring NVIDIA’s cadence — a signal that it is ready to compete not just on performance, but on timing, scale, and platform breadth.
AI Inference Market Is Key:
While much investor attention remains fixated on AI training, the true exponential growth lies in AI inference — the deployment of trained models at scale across cloud, enterprise, and edge environments. This is where AMD sees its opening. Although NVIDIA continues to dominate training workloads with its tightly integrated CUDA ecosystem, the explosion in inference demand is shifting priorities toward cost-effective, memory-rich, and power-efficient alternatives — precisely the segment AMD is targeting.
Rather than trying to compete everywhere at once, AMD is executing a targeted, multi-layered AI strategy. Its Instinct MI300X is designed for inference-heavy data centers, delivering high throughput with lower TCO. Ryzen AI brings advanced on-device capabilities to the next generation of AI-powered consumer PCs, while Xilinx-powered Versal chips bring programmable intelligence to the edge, where latency, power, and real-time processing are critical.
As the market moves from AI hype to AI monetization, AMD is strategically positioned to capture share from NVIDIA — not by brute force, but through a flexible, open ecosystem, a maturing software stack, and an architecture purpose-built for the next wave of scalable AI deployment. The market hasn’t fully priced in this transition — and that’s where the opportunity lies.
What AMD is building isn’t hype — it’s momentum grounded in execution. In 2024 alone, AMD’s data center business nearly doubled, driven primarily by its Instinct AI-optimized GPUs, which generated over $5 billion in revenue. What was once a rounding error on the company’s income statement has become a core revenue engine, firmly establishing AMD as NVIDIA’s only serious challenger in the high-stakes AI acceleration race.
Strategic partnerships are accelerating as well. Microsoft Azure has integrated AMD GPUs into its OpenAI services, and Meta is currently testing AMD silicon to power its own AI workloads. On the software side, AMD is actively working with Hugging Face and Microsoft to ensure broad model compatibility and performance optimization across AMD hardware. This software enablement is critical as AMD expands its reach into AI-powered PCs, where joint initiatives with Microsoft are ensuring that Copilot features run natively and efficiently on Ryzen AI.
Behind the scenes, AMD continues to invest aggressively in R&D, now surpassing $5 billion annually — a push made possible by healthy gross margins of around 50%. These investments are amplified by the company’s strategic acquisitions of Xilinx and Pensando, which give AMD a unique ability to deliver full-stack AI infrastructure — spanning compute, networking, and adaptive programmable hardware — all under one roof.
NVIDIA Still In Lead:
Let’s call it what it is: NVIDIA’s lead is massive. With over $114 billion in data center sales in FY2024, it continues to dwarf AMD’s roughly $12.6 billion in the same category. Its CUDA platform remains the default language of AI developers, and the upcoming Blackwell GPUs are likely to push performance even further. That said, the AI market is so hot, and so supply-constrained, that secondary suppliers like AMD aren’t just tolerated — they’re essential.
AMD’s roadmap offers a compelling case for share expansion. The launch of the MI350 (CDNA 4) in 2025 promises up to 35x improvement in AI inference performance. If AMD hits that mark, analysts believe it could grow its market share to 15–20% within two years. That wouldn’t dethrone NVIDIA, but it would be enough to create a multi-billion-dollar, high-margin business that adds resilience and diversification to AMD’s portfolio.
Meanwhile, AMD’s Xilinx-based edge solutions are delivering unmatched flexibility for power-sensitive applications across automotive, telecom, and industrial AI. To be clear, challenges remain: ROCm needs to become as developer-friendly as CUDA, the ecosystem must deepen its commitment, and MI300’s performance at scale still needs to prove itself in real-world deployment. But if AMD’s recent history is any indication, it knows how to rise to the occasion. It outmanoeuvred Intel in the x86 CPU market, built credibility with enterprise partners, and pioneered chiplet architecture that fundamentally redefined high-performance compute.
Now, AMD is betting big on AI inference as the next great inflection point. If AI becomes as ubiquitous as electricity or the internet — as many predict — then AMD’s move to power that future, across all layers of compute, may prove to be its most significant strategic decision yet.
While Wall Street still chases the flash of AI training, the real monetization opportunity lies in inference — and AMD is laying the infrastructure now. With the MI300X in-market and the MI350X on the horizon, AMD isn’t trying to replace NVIDIA. It’s quietly positioning itself as the indispensable second engine powering global AI infrastructure. The market hasn’t priced in this pivot yet — but it will.