AMD(AMD.US)2025Q4 Conference Call: MI450 Series Development Progress is Very Smooth, Launch and Production to Begin in the Second Half of the Year

Recently, AMD (AMD.US) held its Q4 2025 earnings call. The development of the MI450 series is progressing very smoothly, and the plan to launch and begin production in the second half of the year is proceeding as scheduled. Customer collaborations are continuing smoothly, and the partnership with OpenAI remains solid. The related capacity expansion plan will extend from the second half of this year through 2027, and everything is on track.

At the same time, AMD is working closely with many other customers. Given the product advantages, they are also very interested in rapidly increasing the usage of MI450. The company sees opportunities in both inference and training domains. Therefore, the company is very satisfied with the overall data center growth in 2026 and is confident in achieving hundreds of billions of dollars in data center AI revenue by 2027.

Q&A Session

Q: Can you share updates on customer collaborations regarding the 2027 AI revenue outlook and the demand for MI455 and Helios platforms in the second half of the year?

A: The development of the MI450 series is progressing very smoothly, and the plan to launch and start production in the second half of the year is proceeding as scheduled. Customer collaborations are ongoing smoothly, and the partnership with OpenAI remains strong. The capacity expansion plan will extend from the second half of this year through 2027, and everything is on schedule.

Meanwhile, we are working closely with many other customers. Given the product advantages, they are also very interested in quickly increasing MI450 usage. We see opportunities in both inference and training. As a result, we are very pleased with the overall data center growth in 2026 and are confident in reaching hundreds of billions of dollars in data center AI revenue by 2027.

Q: Could you provide details on the March quarter guidance and the overall growth outlook for data center GPUs for the full year?

A: We typically provide guidance for one quarter at a time, but I can share the first quarter outlook. Although overall revenue is expected to decline by about 5% quarter-over-quarter, the data center business is actually expected to grow. CPU revenue, which normally declines in line with seasonal patterns, is projected to see good quarter-over-quarter growth in this guidance. GPU revenue, including in the Chinese market, is also expected to increase. Therefore, the overall data center outlook is quite positive. Client, embedded, and gaming businesses are experiencing seasonal declines.

Regarding the full-year outlook, we are very optimistic about this year. The key theme is that data center growth remains very strong, driven by two main vectors: server CPU growth is very robust, and as AI continues to advance, CPUs are critical. Over the past few quarters, especially in the last 60 days, CPU orders have been consistently strengthening, which will be a strong growth driver for us. Server CPU revenue is expected to grow from Q4 into Q1 (which is typically a seasonal dip), and this trend is expected to continue throughout the year.

Data center AI is a crucial focus for us this year and represents a real inflection point. MI355 performed well, and we are satisfied with Q4 results, with continued improvements in the first half. But in the second half, MI450 will be the real inflection point. Its revenue will start from Q3 and will see a significant increase entering 2027. This roughly outlines the growth trajectory for the entire data center business for the year.

Q: What are the expectations for Chinese MI308 sales after Q1, and can the 60%+ growth target for data center revenue in 2026 be achieved?

A: Regarding China: We are satisfied with some MI308 sales in Q4, which were based on orders from early 2025 and obtained relevant licenses. We expect about $100 million in revenue in Q1. Due to dynamic changes, we are not forecasting additional Chinese revenue beyond that.

We have submitted licensing applications for MI325 and continue to communicate with customers to understand demand. We believe that, aside from the $100 million forecast in Q1, it is prudent not to make further predictions.

Regarding overall data center: As previously mentioned, we are very optimistic. Our growth drivers include the continued strong growth of the EPYC product line (Turin and Genoa), the upcoming launch of Venice in the second half, which we believe will further strengthen our leadership, and the significant capacity expansion for MI450 in late 2026. While we haven’t provided specific segment guidance, exceeding 60% long-term growth in 2026 is indeed possible.

Q: How is the capacity for server CPUs, especially regarding additional capacity from TSMC and other suppliers? How long does it take to produce wafers? What does this mean for the growth trajectory in 2026? Also, how should we think about the price inflection point?

A: Regarding the server CPU market: We believe the overall server CPU market will see strong double-digit growth in 2026. Concerning our supply capacity: over the past few quarters, we have observed this trend, which is why we have increased our server CPU supply capabilities. This is also one of the reasons for the positive guidance for Q1 server revenue.

The company sees the capacity for sustained growth throughout the year. Demand remains strong, and we are working with supply chain partners to increase supply. Based on current conditions, overall server demand is robust, and we are increasing supply to meet it.

Q: Can you discuss the gross margin framework for the full year and how to balance the ramp-up of the server CPU business with the potential acceleration of GPU business in the second half?

A: We are pleased with our gross margin performance in Q4, with guidance of 55% for Q1, representing a 130 basis point year-over-year increase, and our MI355 scale has expanded significantly. We are benefiting from a very favorable product mix across all businesses.

In data centers, we are advancing new generations, including MI355, which supports gross margin improvement. In client business, we are moving toward high-end products and gaining growth momentum in commercial markets, with gross margins steadily improving.

Additionally, we see a recovery in embedded business, which also contributes positively to gross margins. We expect these tailwinds to continue over the next few quarters. When MI450 begins ramping in Q4, our gross margin will be largely driven by product mix. We will provide more details then, but overall, we are very satisfied with our gross margin progress this year.

Q: Will the capacity expansion for MI455 be 100% rack-based? Will there be server products with eight GPUs based on this architecture? Is revenue recognized upon shipment to rack vendors?

A: Yes, we do have multiple variants of the MI450 series, including configurations with eight GPUs. But for 2026, the vast majority will be rack-scale solutions. Yes, revenue will be recognized upon shipment to rack assembly vendors.

Q: What are the risks in converting chips into rack systems, and have measures like pre-assembling racks been taken to ensure smooth capacity ramp-up?

A: Development is progressing very smoothly; both the MI450 series and Helios rack are on schedule. We have conducted extensive testing, including at the silicon wafer and rack levels. Everything is proceeding well. We have received substantial testing feedback from customers, enabling us to conduct parallel testing. We expect the product launches in the second half to go as planned.

Q: How should we think about the ongoing increase in operating expenses, especially as GPU revenue begins to grow? Can we expect operational leverage, or will expenses grow faster with AI revenue?

A: Regarding operating expenses, we are highly confident in our current roadmap. In 2025, as revenue grows, we did increase operating expenses, which was the right decision. Moving into 2026, with the expected significant growth, we believe operating expenses will grow at a slower rate than revenue in our long-term model.

We expect this to be the case in 2026 as well, especially as revenue growth accelerates in the second half. Based on current free cash flow generation and overall revenue growth, investing in operating expenses remains the right choice.

Q: Was the $100 million Chinese market revenue in Q1 also on a zero-cost basis like Q4? Could this pressure gross margins? Additionally, can you provide specific figures for the Instinct business in 2025?

A: Regarding the $100 million in Q1: The $360 million inventory reserve in Q4 not only related to Chinese revenue for that quarter but also covered the $100 million of expected shipments of MI308 to China in Q1. Therefore, the gross margin guidance for Q1 is very clear.

Regarding the scale of the Instinct business: We do not provide guidance at the business level. But to help your modeling, even excluding non-recurring Chinese revenue, you can see growth when looking at data center AI revenue from Q3 to Q4, which should be helpful.

Q: The client business performed strongly in Q4, but considering inflationary costs in memory markets, has the order pattern changed? What is the outlook for client market growth and health in 2026?

A: The client market in 2025 performed very well, with strong growth in ASPs toward high-end products and in unit sales. For 2026, we are closely monitoring the development. Based on current observations, considering inflationary pressures including memory, the overall PC market TAM may decline slightly. Our full-year model predicts that the second half will be slightly below the first half’s seasonal level.

Even in a declining PC market environment, we believe our PC business can still grow. Our focus remains on enterprise markets, where we made good progress in 2025 and expect to continue growth in high-end segments in 2026.

Q: How does AMD’s collaboration with SRAM architecture suppliers and competitors impact the competitiveness of HBM-based Instinct in inference markets? How does Instinct address the demand for low-latency inference?

A: This reflects the expected evolution as the AI market matures. As inference scale expands, the efficiency per dollar or per token becomes increasingly important. Our chiplet architecture offers strong optimization capabilities for different inference and training stages.

In the future, we will see more workload-optimized products, which can be implemented via GPUs or more ASIC-like architectures. We have a complete compute stack to meet all these needs. From this perspective, we will continue to focus on inference as a key opportunity beyond just training improvements.

Q: Regarding the collaboration with OpenAI, is the 6 GW, three-and-a-half-year plan for the second half of the year on schedule? Can you share more details about the partnership?

A: We are working closely with OpenAI and our cloud service provider (CSP) partners to deliver the MI450 series and execute capacity expansion plans. These are on schedule to start in the second half of the year. The MI450 development is progressing well, and Helios is performing strongly. We are engaged in deep joint development with all these partners.

Looking ahead, we are optimistic about capacity increases for OpenAI’s MI450. But we also want to remind everyone that many other customers are very interested in the MI450 series. So, besides our partnership with OpenAI, we are also working with multiple other clients to expand capacity during the same period.

Q: How does AMD view the competition between x86 and ARM architectures in the server CPU market, especially regarding agents? Does x86 hold an advantage, and what are your thoughts on NVIDIA’s ARM CPUs?

A: The current market has a high demand for high-performance CPUs, especially for agent workloads, which generate many traditional CPU tasks. Most of these are currently running on x86. EPYC’s advantage lies in workload optimization; we have the best cloud and enterprise processors, as well as low-cost options for storage and other uses.

We believe that building a complete AI infrastructure will involve all these factors. CPUs will continue to be a critical part of AI infrastructure evolution. As we mentioned at our November analyst day, this is a multi-year CPU cycle, and we continue to see this trend. EPYC has been optimized for all these workloads, and we will keep working with customers to expand EPYC’s market share.

Q: Regarding the procurement timeline for HBM and other memory components, is it a year or half-year in advance?

A: Considering the lead times for supply chain components like HBM and wafers, we work closely with suppliers over multi-year horizons, covering demand forecasting, capacity expansion, and collaborative development. We are satisfied with our supply chain capabilities.

We have planned for several years of capacity expansion, and regardless of current market conditions, we have been planning significant increases in CPU and GPU capacity over the past few years. We believe we are well prepared for substantial growth in 2026. Due to supply chain tightness, we have also signed multi-year agreements beyond that scope.

Q: Will AMD follow these architectural shifts—such as system accelerators, KV cache offloading, and more discrete ASIC-style computing? How do you see your system architecture evolving?

A: With our highly flexible chipset and platform architecture, we are capable of providing different system solutions tailored to various needs. We recognize that future solutions will vary; there is no one-size-fits-all approach. Rack-scale architectures are excellent for high-end applications like distributed inference and training. At the same time, we see opportunities in enterprise AI using other forms of products, and we are investing across the entire relevant spectrum.

Q: How will gross margins change from MI300 to MI400 and MI500? Will they rise, fall, or fluctuate?

A: At a high level, each generation offers more capabilities and more memory, providing greater value to customers. Generally, with each new generation, gross margins should gradually improve.

Typically, gross margins are lower at the start of mass production for new products. As scale is achieved, yields improve, testing becomes more efficient, and overall performance increases, gross margins within each generation tend to improve. This is a dynamic process, but in the long term, each generation is expected to have higher gross margins.

Q: What is the expected decline in gaming revenue in 2026? What is the annual trajectory?

A: 2026 is the seventh year of the current product cycle. Usually, revenues tend to decline at this stage. We do expect a significant double-digit drop in semi-custom revenue in 2026. However, the launch of new generation products like Xbox is expected to reverse this downward trend.

Q: Will capacity constraints in rack systems in the second half affect or limit revenue growth, especially from Q3 to Q4?

A: We have planned at every component level. Regarding our data center AI capacity expansion, I believe we will not face supply constraints that hinder our plans. We have a proactive capacity expansion plan that is feasible. Given AMD’s scale and size, our top priority is to ensure smooth capacity growth for data centers, including AI GPUs and CPUs.

Q: What is the biggest investment area in 2025? And the largest incremental operating expense in 2026?

A: In 2025, the main investment focus is on data center AI, including the acceleration hardware roadmap, expanding software capabilities, and acquiring ZT Systems to enhance system-level solutions. We are also heavily investing in marketing to support revenue growth and expand our CPU business in commercial and enterprise markets. In 2026, we expect to continue investing actively, but revenue growth will outpace operating expense increases, driving EPS growth.