AMD 2026 年第三季法說會：資料中心營收大增 97%，上調 2027 年成長展望

重點摘要

• AMD 報告 2026 年第三季營收創歷史新高，達 134 億美元，年增 45%、季增 16%，主要得益於資料中心營收幾乎翻倍。
• 資料中心營收達 85 億美元，年增 97%，約佔總營收的 64%。部門營業利益率擴增至 34%（相較於先前的 25%）。
• Non-GAAP 營業利益成長 75%，達 39 億美元，而 Non-GAAP 營業利益率上升超過 5 個百分點至 29%。
• AMD 預估 2026 年第四季營收約為 170 億美元（上下浮動 3%），代表以中位數計算，年增約 66%、季增約 27%。
• 管理層目前預期 AMD 營收將在 2027 年成長超過 80%。伺服器營收預計將近乎翻倍，而資料中心 AI 營收預計成長將大幅超過兩倍。
• 需求持續高於可用供應。AMD 正努力爭取額外的晶圓、基板、先進封裝、記憶體及其他零組件，同時應對記憶體與相關投入成本的上漲。

關鍵財務數據

業務與營運表現

資料中心

資料中心營收成長 年增 97% 且 季增 27%，創下 85 億美元 的歷史新高。部門營業利益成長 168% 至 29 億美元，營業利益率達到 34%。

雲端與企業伺服器銷售額年增超過一倍。 AMD 表示，超大規模雲端業者擴大了 EPYC 在內部服務、公有雲產品與 Agentic 代理工作負載上的部署。Turin 已在全球超過 540 種公有雲實例類型 中上線。

AMD 在本季開始出貨產生營收的 EPYC Venice 處理器。OEM 與 ODM 夥伴已承諾採用超過 240 個 Venice 平台，比 Turin 多出 20%，首批 OEM 系統預計於 2026 年第 4 季登場。

該公司亦開始量產出貨 Instinct MI-455X。Anthropic、Meta、微軟 (Microsoft)、OpenAI 與甲骨文 (Oracle) 已上線首批 Helios 機架，管理層預期隨著更多 OEM 與 ODM 系統開始出貨，第 4 季部署量將大幅增加。

管理層表示，第 4 季將是資料中心 AI 營收首次超越伺服器 CPU 營收的季度，且以金額計算，資料中心 AI 預計將「大幅提高」。

客戶端與遊戲

客戶端與遊戲營收下滑 年增 8% 至 37 億美元。部門營業利益為 5.73 億美元，或 15% 的營收，相較於去年同期的 8.67 億美元與 21%。

客戶部門營收成長 年增 7% 至 29 億美元，受惠於 Ryzen 行動處理器支持，但由於記憶體成本上升打壓 PC 需求，季減 4%。

遊戲部門營收下滑 年減 40% 至 7.73 億美元，主要因為處於遊戲主機生命週期此階段的半客製化營收下降，以及顯示卡銷售較弱。Radeon GPU 通路銷量季增改善，而 Radeon AI Pro 營收則較上一季顯著成長。

嵌入式部門

嵌入式部門營收成長 年增 29% 且 季增 14% 至 11 億美元。 營業利益為 4.53 億美元，營益率為 41%。

成長主要由航太與國防、汽車、通訊及工業客戶帶動。AMD 仍按計畫進行，預計將達成超過 2026 年 180 億美元的設計獲案，這將創下連續第三個年度新高紀錄。

管理層財務預測

針對 2026 年第四季，AMD 提供以下展望：

展望 2027 年，管理層預計公司總營收將成長 80% 以上。資料中心業務預計將佔 總營收的 80% 以上，並推動公司大部分的營收與獲利成長。

AMD 預期 2027 年伺服器營收將比前一年翻倍成長，資料中心 AI 營收更將成長顯著超過兩倍。營業費用預估成長率不到營收成長率的一半，進一步支撐營業利潤率顯著擴大。

公司預計 2027 年 PC 市場出貨量將出現雙位數百分比下滑，不過管理層預估 AMD 將透過商用、頂級消費級及桌上型電腦通路產品，表現優於整體市場。

風險與關注焦點

• 供應瓶頸： 管理層表示需求大於供應，目前正積極尋求更多晶圓、基板、先進封裝、記憶體及相關零組件。
• 零組件價格上漲： 記憶體與其他投入成本上漲正影響 PC 和資料中心系統。AMD 正與客戶合作共同分擔部分成本。
• PC 市場疲軟： 系統價格調漲放緩了終端銷售 (sell-through)，尤其是在消費級市場。AMD 正調整出貨量 (sell-in) 以配合市場需求，確保客戶與通路庫存維持健全。
• 營運資金需求： 庫存大幅增加，而自由現金流則反映了為實現 2027 年成長所進行的庫存、設備及產能投資。
• 毛利率產品組合： 資料中心 AI 的利潤率低於 AMD 的整體平均水準。單季毛利率將取決於 AI 加速器、伺服器 CPU 及嵌入式 (Embedded) 產品之間的組合情況。
• 認股權證會計波動： 單季 GAAP 財報可能包含與 AMD 股價及其他假設相關的可變非現金公允價值調整。這些調整並不影響 Non-GAAP 結果或自由現金流。

分析師問答集錦

供應與 2027 年需求： 管理層表示，公司已與客戶及供應商規劃產能提升多年。AMD 已爭取到額外供應，但需求依然高於可用產能。公司持續在整個製造生態系統中爭取額外供給。

AI 基礎設施就緒度： AMD 表示正與核心客戶直接合作，著手規劃資料中心選址、電力、設施與部署時程。公司也與合作夥伴共同進行選擇性投資，以提前數年確保基礎設施與產能。

GPU 經濟效益與客製化 ASIC 競爭： 管理層將總擁有成本 (TCO)、每美元 Token 產出、能源效率及工作負載最佳化視為部署關鍵考量。AMD 正與前沿 AI 實驗室共同設計 MI-500 及 MI-600 產品，以針對客戶的工作負載調校未來系統。

伺服器成長結構： 預計 2027 年伺服器營收翻倍主要是由出貨量增加與核心數增多所帶動，而非價格提升，儘管 Venice 較高的核心數應能帶來部分平均售價 (ASP) 的提升。

Helios 與 Helion 戰略： AMD 旨在成為前沿實驗室、超大型雲端服務業者 (hyperscaler) 及企業更大的算力供應商。Helion 可配置為專用於代理 (agentic) 工作負載的高密度純 CPU 機架，或 CPU-GPU 混合系統，而 Helios 仍是 MI-450 量產引進的核心。

PC 戰略： AMD 計劃保持自律，而非將焦點放在波動較大的低階市場。管理層強調企業、頂級消費級和桌上型電腦通路市場，因為公司在這些領域看到了更高的價值與獲利能力。

電話會議全文紀錄

---

完整財報電話會議逐字稿

管理層陳述

Operator

Greetings. Welcome to the AMD Third Quarter 2026 Conference Call. [Operator Instructions] And please note that this conference is being recorded.

I will now turn the conference over to Matt Ramsay, VP, Financial Strategy and IR. Thank you, Matt. You may begin.

Matthew Ramsay

Thank you, and welcome to AMD's Third Quarter 2026 Financial Results Conference Call. By now, you should have had the opportunity to review a copy of our earnings press release and the accompanying slides. If you have not had the chance to review these materials, they can be found on the Investor Relations page of amd.com. We will refer primarily to non-GAAP financial measures during today's call. The full non-GAAP to GAAP reconciliations are available in today's press release and slides posted on our website.

Participants on today's conference call include Dr. Lisa Su, our Chair and CEO; and Jean Hu, Executive Vice President, CFO and Treasurer. This is a live call and will be replayed via webcast on our website.

Before we begin, I would like to note that AMD will participate in the following events with the financial community this quarter. The UBS Technology and AI Conference on Tuesday, December 1; and Barclays 24th Annual Global Technology Conference on Wednesday, December 9.

Today's discussion contains forward-looking statements based on our current beliefs, assumptions and expectations, including forward-looking statements regarding financial projections, business and industry trends that speak only as of today and as such, involve risks and uncertainties that could cause actual results to differ materially from our current expectations. Please refer to the cautionary statement in our press release for more information on the factors that could cause actual results to differ materially.

With that, I will hand the call over to Lisa.

Lisa Su

Thank you, Matt, and good afternoon to all those listening today. The third quarter was the largest and most profitable in our history as we began ramping our next-generation data center products. Revenue increased 45% year-over-year to $13.4 billion, driven by record EPYC and instinct sales and strong embedded growth. Earnings per share grew even faster, increasing 74% to $2.09. We executed very well in the quarter with Helios production ramping as planned and EPYC adoption expanding across cloud, enterprise and super intelligence workloads.

Data center revenue growth has accelerated sequentially every quarter this year, with the segment contributing 64% of total revenue in the third quarter, up from 47% just 1 year ago. We expect that acceleration to continue in the fourth quarter, driven by increased demand for inference and agentic workloads and the ramp of our newest products.

Turning to our segments. Data center revenue grew 97% year-over-year to a record $8.5 billion. In server, both cloud and enterprise sales more than doubled year-over-year. We significantly exceeded our previous outlook of more than 80% growth as we gained revenue share. Customers deploy EPYC across more applications with agentic workloads contributing significant growth. In cloud, hyperscalers deployed substantial volumes of EPYC servers in the quarter to run more internal services and expand their public cloud offerings. Turin is now the most widely deployed EPYC generation in the public cloud with more than 540 instance types available globally from the largest providers.

Cloud providers also added significant EPYC capacity in the quarter for personal and enterprise agents, which are quickly becoming the next killer app for CPUs in the cloud. Meadows Muse, an Open AI's dots illustrate how quickly personal agents are growing with more than 10 million people activating Muse agents in the first 6 weeks.

Enterprises are moving fast as well, putting agents to work across software development, supply chain and IT operations. These agents work continuously and in parallel across applications and enterprise systems, adding to accelerate our demand while creating a substantial increase in CPU demand.

Adoption is just starting, and we expect demand to grow sharply as agents are deployed more broadly and take on more complex work. In head nodes, we won major deployments with several of the largest model developers and cloud providers in the quarter, which will ramp over the coming quarters. EPYC's leadership performance and mature ecosystem proven at scale have made it the host CPU of choice for custom accelerators. As 1 example, OpenAI selected turn EPYC processors as the host CPUs for its first custom accelerator platforms.

In enterprise, EPYC server sell-through was a record with strong demand from new and existing customers. We added dozens of first-time EPYC customers in the quarter, including large financial services, communications, retail and health care companies. Existing customers are also making EPYC their primary compute platform, extending adoption from specific workloads to a much larger share of their overall infrastructure.

We began revenue shipments of EPYC Venice processors in the third quarter, and the first OEM systems are on track for availability this quarter. Early demand is the strongest we have seen for any new EPYC generation. Our OEM and ODM partners have committed more than 240 Venice platforms, 20% more than turn and cloud providers are deploying Venice this quarter ahead of instance rollouts planned for early 2027.

We recently launched Helion, a Venice reference design that enables turnkey EPYC rock scale systems for agentic infrastructure. Customers can configure Helion for their workloads from all CPU rocks for hyperscalers to enterprise systems combining Venice and instinct GPUs that can run agents and models together. In the all CPU configuration, a single Helion rack has up to 144 Venice processors with nearly 74,000 threads, delivering the highest agent density in the industry.

Every major server OEM and ODM plans to offer Helion with the first system shipping this quarter. Looking more broadly, demand for server compute is accelerating across cloud, enterprise, agentic and super intelligence infrastructure. Efficacy industry's broadest portfolio spanning all of these markets with per-socket leadership for cloud and agents, per core leadership for enterprise and head notes and memory and I/O options tuned to each workload.

Venice is build on that leadership with significantly better performance per watt, performance per core and performance per dollar than any other x86 or ARM server CPU, positioning us to gain share and significantly grow our server business as the TAM grows to approximately $220 billion by 2030.

Turning to our data center AI business. Frontier labs, hyperscalers and cloud providers expanded instinct deployments for inferencing and training driving record revenue. We started shipping MI-455x in volume in the quarter, and lead customers, including Anthropic, Meta, Microsoft, OpenAI and Oracle brought their first Helio racks online quickly.

Customers are now running their latest models and workloads on Helios, and we are seeing very strong performance with the initial deployments. We expect Helios deployments to increase substantially in the fourth quarter as production continues to ramp and HPE, Dell, Lenovo, Super Micro and other OEM and ODM partners begin shipping systems.

We are very pleased with the overall performance we are seeing at the rack level. MI-455x leads in compute flops, memory capacity and scale-out bandwidth. And as Helios comes to market, it delivers an even larger lead in memory bandwidth and higher scale up bandwidth than the competition. We are working closely with our manufacturing and OEM partners to scale production as customers expand their deployments.

In supercomputing, Euro HPC selected instinct MI-430x GPUs and 6-gen EPYC processors for LUMI AI, Europe's next flagship supercomputer planned to come online in the second half of 2027. MI-430X delivers nearly 9x the double precision performance of the competition, which is critical for scientific simulations. As a result, MI-430x has now been selected to power next-generation systems around the world, and we are well positioned to win additional designs as governments and research centers plan their next major investments.

In software, we launched ROCm 10 with major enhancements for production deployments at scale, including faster start-up across large GPU clusters and greater resilience for long-running workloads. ROCm 10 also adds agentic performance tuning with ROCm.AI, which profiles workloads, identifies bottlenecks and applies and validates improvements, turning weeks of engineering optimization into hours.

Customers are also using their own agents to optimize models on our open software stack. For example, Anthropic highlighted that 1 engineer used Cloud to bring up their latest model on instinct MI-355x racks over a single weekend with strong performance.

Our work with Frontier Labs is already enabling their models to handle complex software development and performance tuning for Instinct. And as those capabilities continue to improve, the benefits extend to everyone in the AMD ecosystem. We announced 2 acquisitions in the quarter to strengthen our Super Intelligence portfolio. Taalas brings a highly experienced engineering team and specialized silicon built for low latency efficient inference. We plan to integrate their technology across our solutions portfolio to address the growing demand for low latency super intelligence applications. World Mo Founder and CEO, Dr. [indiscernible], brings a world-class research team with expertise in spatial intelligence and models that understand the physical world, giving us deeper insight into how frontier models and workloads are evolving.

[indiscernible] will join AMD as Chief Scientist, leading our model research and helping shape our future hardware, software and systems road maps. We expect both acquisitions to close by the end of the year.

Looking ahead, our focus is on execution at scale. The racks of Venice, the MI-450 series and Helios are still early, and all 3 are on track to grow significantly. We now expect data center segment revenue growth to accelerate in the fourth quarter and full year segment revenue to more than double from 2025.

Our visibility into 2027 also continues to strengthen. Last quarter, we said we expected annual server revenue to grow more than 70% and data center segment revenue to more than double for the year. Since then, customers have expanded their EPYC and instinct deployment plans, and we have secured more supply to support that growth.

We now expect server revenue to double year-over-year in 2027 and data center AI revenue to grow even faster than we projected 90 days ago. As a result, we expect data center to be more than 80% of our total revenue in 2027 and to drive the majority of our revenue and earnings growth.

Looking further out, customer engagements for MI-500 continued to progress very well. Interest is very high as it is another significant leap forward in performance. We are working very closely with multiple frontier labs to optimize MI-500 for their workloads as they plan multi-gigawatt deployments.

Turning to client and gaming. Segment revenue decreased 8% year-over-year to $3.7 billion. In clients, revenue increased 7% year-over-year to $2.9 billion, driven by strong sales of our Ryzen mobile processors. Commercial is becoming a larger part of our client business as large enterprises adopt Ryzen Pro PCs across their employee base. Eight of the 10 largest banks in the world and 7 of the 10 largest health care and automotive companies now use Ryzen Pro PCs.

OEMs continue expanding their AMD-based commercial lineups from everyday business PCs to high-end workstations. HP, Lenovo and others launched the first commercial notebooks and small form factor desktops featuring our newest Ryzen AI Max Pro 400 series processors. These systems deliver workstation class performance and run super intelligence models with up to 300 million parameters locally.

Looking at the market, PC sell-through slowed late in the quarter as higher memory and component costs weighed on demand. We are managing the business accordingly, aligning sell-in with sell-through to keep customer and channel inventory healthy. In 2027, we are planning for PC market unit shipments to decline by a double-digit percentage. In this environment, we expect to outperform the market as we deliver our rising road map and grow our commercial footprint.

In gaming, revenue declined 40% year-over-year to $773 million, primarily due to lower semi-custom sales as expected at this stage of the console cycle. Gaming graphics revenue declined year-over-year, while Radeon GPU channel sell-through increased sequentially. Beyond gaming, Radeon AI Pro revenue grew significantly from the prior quarter as more developers choose AMD to run and fine-tune models locally. We expect strong growth for Radeon AI Pro to continue, driven by our leadership performance per dollar and an open software stack optimized for the latest open source models.

Turning to our Embedded segment. Revenue increased 29% year-over-year to $1.1 billion, the fourth consecutive quarter of accelerating growth. Demand was very strong across all of our end markets, led by strong growth with aerospace and defense, automotive, communications and industrial customers. Data center build-outs are also driving higher demand for our embedded CPUs and adaptive SoCs in control plane and optical networking applications. We remain on track to close more than $18 billion of design wins in 2026, our third consecutive record year.

Adaptive design wins have increased every year since we acquired Xilinx with embedded x86 and semi-custom wins growing even faster. As these wins ramp over the next few years, we expect our embedded business to deliver strong double-digit annual revenue growth and continue outgrowing the market.

In summary, we delivered an excellent third quarter and expect an even stronger fourth quarter as data center and embedded segment growth more than offsets lower client CPU sales. We entered 2027 with the strongest product portfolio in our history and significant customer momentum. We have worked across our supply chain to secure the wafers, substrates, advanced packaging, memory and other key components needed to support much higher growth next year.

Based on our current visibility, we now expect revenue to grow more than 80% in 2027, which is more than twice the long-term financial model we outlined at our Financial Analyst Day last November. Data center will be the largest contributor, driven by expanding Helios deployments and continued EPYC share gains. We also expect embedded growth to accelerate next year driven by adoption of our adoptive and x86 solutions. Beyond 2027, we are increasing R&D investment across our silicon software and systems road maps and working closely with our largest customers to support their ambitious infrastructure plans for the next 3 to 5 years. We are also deepening our collaboration with our supply chain partners and investing alongside them to secure the long-term capacity our growth will require.

Taken together, these give us strong confidence in our ability to sustain exceptional revenue growth for years to come, with earnings growing even faster.

Now I will turn the call over to Jean to provide additional color on our third quarter results. Jean?

Jean Hu

Thank you, Lisa, and good afternoon, everyone. I'll start with a review of our financial results and then provide our outlook for the fourth quarter of fiscal 2026.

We delivered an outstanding third quarter. Revenue increased 45% year-over-year to a record $13.4 billion, exceeding the high end of our guidance, driven by our data center segment, which nearly doubled year-over-year. Revenue increased 16% sequentially, driven by strong growth in the data center and embedded segments, partially offset by a decline in the client and gaming segment. Our data center segment accounted for approximately 64% of total revenue, up from 58% last quarter, underscoring the continued shift in our business mix and its increasing contribution to AMD's group.

Gross margin was 56%, up 190 basis points year-over-year, reflecting a richer product mix. Operating expenses were $3.6 billion, an increase of 29% year-over-year as we continue to invest in R&D to expand our data center silicon systems and software capabilities. Operating income grew 75% year-over-year to a record $3.9 billion, significantly outpacing our revenue growth. Operating margin was 29%, expanding more than 5 percentage points year-over-year, underscoring the operating leverage in the business as we scale.

In the third quarter, we began recognizing the accounting impact of warrants issued under our previously announced customer agreement as we shipped MI-450 series GPUs. Our GAAP diluted earnings per share were $1.46, which included $1.1 billion noncash charge in other income and expense, primarily reflecting the increase in our share price and the changes in the estimated probability of meeting the warrant conditions.

Going forward, this quarterly noncash fair value adjustment are expect to increase as we and our customers progress under this agreement. The amount will vary each quarter based on AMD's stock price and other valuation assumptions. These noncash adjustments would not affect our non-GAAP results or free cash flow. Our non-GAAP diluted earnings per share were $2.9.

Now turning to our reportable segment, starting with the data center. Data center segment revenue was a record of $8.5 billion, up 97% year-over-year, driven by record server revenue and strong demand for our 5th generation EPYC processors and by record data center AI revenue, highlighted by the launch of our AMD Instinct MI-450 Series GPUs.

On sequential basis, data center revenue increased 27%, primarily driven by initial MI-450 Series sales and strong server revenue growth. The data center segment operating income was $2.9 billion, up 168% year-over-year. And the operating margin increased to 34% compared to 25% a year ago, demonstrating strong operating leverage of data center business.

Client on the Gaming segment revenue was $3.7 billion, down 8% year-over-year due to lower semi-customer and graphic sales, partially offset by higher demand for the latest generation Ryzen mobile processors. In the client business, revenue was $2.9 billion, up 7% year-over-year, driven by higher sales of Ryzen mobile processors. Sequentially, client revenue was down 4% as higher memory cost weighted on PC unit demand, partially offset by a richer mix of high-end processors.

Gaming revenue was $773 million, down 40% year-over-year due to lower semi customer revenue and the lower graphics demand. Sequentially, gaming revenue was essentially flat. Client and the Gaming segment operating income was $573 million or 15% of revenue compared to $867 million or 21% a year ago.

Embedded segment revenue was $1.1 billion, up 29% year-over-year and 14% sequentially as demand strengthened across several end markets, led by growth in aerospace and defense, automotive, communication and industrial end markets. Embedded segment operating income was $453 million or 41% of revenue compared to $283 million or 33% a year ago.

Turning to balance sheet and cash flow. During the quarter, inventory increased approximately $3 billion sequentially to $11.5 billion to fulfill our strong data center demand. We generated $1.2 billion in cash from operating activities of continuing operations. Free cash flow was $554 million, reflecting increased working capital investment and ongoing capital expenditures to secure equipment and capacity for the rapid growth of our data center business, including capacity dedicated to support our data center demand. We expect to continue to invest in inventory and capacity to support our strong growth in 2027.

At the end of the quarter, cash, cash equivalents and short-term investment were $15.8 billion. Our total debt was $7.1 billion, including $4.75 billion senior notes offering we completed in August.

Now turning to our fourth quarter 2026 outlook. We expect revenue to be approximately $17 billion, plus or minus 3%. The midpoint of our guidance represents approximately 66% year-over-year revenue growth, primarily driven by the acceleration of our data center segment. Sequentially, we expect revenue to grow by approximately 27%, driven by a sharp increase in our data center segment of approximately 50% growth, a decline in our client and the gaming segment of greater than 20% and a double-digit growth in our Embedded segment.

In addition, we expect fourth quarter non-GAAP gross margin to be approximately 55%, and we expect non-GAAP operating expenses to be approximately $3.75 billion. We expect non-GAAP net interest and other expense to be again of approximately $4 million. We expect our non-GAAP effective tax rate to be 13%, and diluted share count is expected to be approximately 1.67 billion shares, including the estimated impact of closing the Taalas and the World Labs acquisition in the fourth quarter. We are not expecting our fourth quarter share count to be impacted by customer warrants.

In closing, the third quarter was an outstanding quarter for AMD with record revenue, significant operating margin expansion and earnings growth, reflecting disciplined execution across the business. As our data center business continue to scale, we are very well positioned to drive increased profitability and long-term shareholder value.

With that, I'll turn it back to Matt for the Q&A session.

Operator

[Operator Instructions] And the first question comes from the line of Tom O'Malley with Barclays.

分析師問答

Thomas O'Malley

Congrats on the great results. I guess, Lisa, let's start off with the fact that you're giving some more color into 2027. It looks like revenue is growing 80%, but the data center component of that is growing much faster where data center is going to be a greater than 80% of your business. So maybe can you talk about, first and foremost, like is this backed up by supply, whether that be wafers, packaging, substrate memory? And then if you look at the demand profile behind that, would demand outstrip that supply if you were able to get some more?

Lisa Su

Yes. Sure, Tom. Thanks for the question. So look, we have a lot of momentum going into the fourth quarter and into 2027. We do see a situation where there is very strong demand. So we see demand across the data center business, across our CPU business as well as our data center AI business. And when you look at the rate and pace of that growth, we've been planning for it for the last several years. So we've absolutely spent a significant amount of time with our customers planning what their needs are in terms of their deployments and their deployment schedules and also with our supply chain partners to ensure that we have the wafers, the substrates, the advanced packaging, the memory, all of the other components that go along with it.

It is true that demand is higher than supply. So we continue to work with our supply chain partners to increase supply. And I do believe that there is a good opportunity for us to do that, just given the strong relationships we have throughout the ecosystem. But sitting where we are today to be able to say that the overall company will grow greater than 80%, especially the server business will approximately double from a fairly high point in 2026. And then the data center AI business, we continue to have strong visibility, and we see strong demand from our customers, and we continue to see also good follow-through on all of the other aspects of the infrastructure that are needed. So overall, we're very positive about the data center growth into 2027.

Thomas O'Malley

Super helpful. And then just as a follow-up on the client side, you're saying down greater than 20% in the fourth quarter. That's kind of in line with what we've heard from others so far in the market. But if you look at '27, I think something that's difficult that normally you see seasonality down in Q1 and then you wouldn't get a recovery kind of until Q2, Q3. So maybe talk about like what the longevity of this correction kind of looks like and why you've been able to kind of outperform a bit up until this point and how you see that kind of coming out of this trough into next year?

Lisa Su

Yes. Sure, Tom. So talking a little bit about the PC market. The PC market continues to be an important market to us. I think we were signaling all along in the last couple of calls that we were expecting a subseasonal second half of the year, and that's how it's played out. Our view is this is one of those times when component prices go up, it affects overall system pricing. One would expect the consumer markets to be more sensitive to that. And so we have seen a bit of slowdown in the sell-through.

We're just managing the business very prudently, and we want to ensure that inventory stays healthy throughout the channel and throughout our customers. And so our planning assumption for 2027 is that the market will be double-digit percentage down. We will certainly focus on the segments where we can add the most value places like enterprise, where we continue to be underrepresented, premium consumer, desktop channel. These are really good markets for us. But we are planning with the current environment for a down year for 2027. And we will work through that.

The good part about it is, our supply chain is quite fungible across our PC business and our data center business. And so that actually does afford us the ability to move supply around as we see these signals in the market. And overall, as I said earlier, we do have more demand than we have supply. And so we're continuing to work all of the aspects to increase supply as we go through the next few quarters.

Operator

And the next question comes from the line of Vivek Arya with Bank of America Securities.

Vivek Arya

Lisa, there's a lot of market debate about the financing of data centers and availability of land, power and shell, et cetera. So kind of a 2-parter, what's the level of confidence, right, that AMD has that your anchor customers have the financing and the physical infrastructure to support, I think it's roughly 3 gigawatts or so of planned AMD deployments into '27?

And then, as we look beyond, many of your competitors are kind of investing in the ecosystem, providing residual value guarantees, and [indiscernible] saying there is a lot more fungibility of their products. So does that create kind of a competitive dynamic for AMD. So just curious to hear your perspective on '27 readiness and '28 competitive landscape?

Lisa Su

Sure, Vivek. Thanks for the question. So as I said earlier, we're seeing very strong demand from our lead data center AI customers. I think you can see that there's a strong demand for more compute and MI-455 is positioned extremely well. When we look at Venice, MI-455, Helios, I think we feel extremely good about the competitiveness.

We have been planning for this ramp with our customers jointly. So we have good visibility into the 2027, as you said, land, power, shelves, all of those, the data center infrastructure, and we are very much working hand-in-hand with our customers. So we know where the GPUs are going in terms of which data centers and which locations and all of that stuff.

Going into 2028 and beyond, I guess, what I would say is the following. Look, I think we are seeing very strong growth overall. I think we have a very strong business. We have a very strong balance sheet. And from that standpoint, our goal is, first and foremost, we're investing in our growth to support that, and that means investing in R&D, investing in supply chain and capacity, investing alongside our partners to ensure that we have the capacity for all of those things that we talked about earlier that are necessary for this type of ramp.

And then we are selectively investing with our lead partners where we see it's helpful to have, for example, the data center infrastructure identified way in advance, and that's multiple years in advance. And I think that's the right thing to do to ensure that we are jointly ensuring that there's the right infrastructure as our new chips become available. But we are being selective and strategic in how we do that with the first principle being investing in our growth.

Vivek Arya

Got it. And for my follow-up, Lisa, one more on the competitive landscape. Many of your Frontier lab customers have been working on their custom ASICs. And I'm curious, how do they conceptually make the choice between using ASICs versus AMD products? Like what determines their allocation? Is it cost per token, power efficiency, software flexibility, time to deployment. Just -- what are the puts and takes in that decision making? And are there workloads where using an AMD GPU works better than the alternatives that they have?

Lisa Su

Sure, Vivek. So look, I think we feel extremely good about where we're positioned in terms of our performance and our total cost of ownership that we're delivering with Helios. That is fundamentally the most important factor when we think about sort of deployment scale and deployments going forward. I think there was a significant investment that we've made in the rackscale systems and ensuring that Helios really ramped nicely. Third quarter was the first quarter that we shipped MI-455 and Helios racks for revenue. I think this is a very major milestone. We have a big step up in Q4. We'll continue to ramp throughout 2027.

And the key is for us to continue to show that TCO advantage across workloads. It is true that from general purpose CPU standpoint, that we have more fungibility as you think about various workloads that occur.

There has been also very strong codesign. So beyond the MI-450 series, we're working on MI-500 and MI-600 as well to ensure that our technology gets even more, let's call it, tuned to what the Frontier Labs need. So I think that's -- those are the key factors in terms of deployment choices.

Operator

And the next question comes from the line of Joshua Buchalter with TD Cowen.

Joshua Buchalter

Congrats on the progress. I was wondering if you could give us insight into how we should be thinking about HBM costs impacting your model as the instinct business ramps. I think investors have been asking how your revenue per gigawatt contribution in the last year or so, your outlook has been so consistent in that $15 billion to $20 billion per gigawatt range given what's going on with HPM pricing. Is that because early builds are fairly programmed specific and pricing was agreed upon a long time ago. And again, as we think about that business longer term, do you have the ability to pass on these higher input costs even to your big 3 customers?

Lisa Su

Sure, Josh. So the way to think about it is as we ramp, we have had let's call it, long-term agreements with some of our suppliers as well as some of our customers. But given where the memory market is and not just the memory market, by the way, other component costs are also seeing some inflationary pressures in the sense of as there is more CapEx that's put into the overall supply chain environment. We are also working with our customers to include -- to ensure that those costs are shared as we go forward.

So from my perspective, back to the question that Vivek had, this is all about TCO and ensuring that from a tokens per dollar standpoint, that we continue to offer very strong advantages for choosing AMD Instinct in our overall Helios system. And even with some of the inflationary pressures that are out there, we certainly believe that we are offering a leading TCO for the AI workloads.

Joshua Buchalter

And maybe for Jean, was good to see OpEx come in light in the third and fourth quarter. It's been a sticking point for investors for a little bit. But I guess given how steep the revenue growth is expected to be in 2027, we're flying a little blind given you're kind of more than 2x your target model. How should we think about OpEx growth in '27? And I guess while I'm in the neighborhood, do you think you can hold your target gross margin range sort of that 55% floor as we move through the year and Helios continues to ramp?

Jean Hu

Yes. Josh, thank you for the question. On the OpEx, as you said, we have been investing aggressively. What you are seeing is the investments are translating to accelerating revenue growth we are seeing in 2026. We're actually quite pleased with the 2026 our operating leverage. We're expanding revenue much faster than OpEx growth, which is driving the operating margin increase.

And if you think about the Q4 guide, we are guiding very significant revenue growth more than OpEx. Now that dynamic, we believe, will carry into 2027. Lisa talked about is that we expect the whole company growing more than 80%. We should expect OpEx to grow less than the top line revenue growth rate less half less than half of the top line revenue growth rate. So that will drive a very significant operating margin expansion in 2027.

Going back to gross margin, we are actually very happy about where we are with our gross margin. In Q3, we are ramping MI-450 data center AI business is doing very well, actually better than our expectation when we got into the Q3 quarter. And in Q4, we are guiding sequentially revenue growing 50% for data center business. It's actually going to be the quarter data center AI business surplus server business. So we all know there's a margin little bit dilutive to corporate average, but the gross margin dollars are increasing very significantly.

I think looking ahead of 2027, we actually think that our data center business will be more than 80% of the total company revenue. So the mix of data center business is going to drive the company's gross margin significantly. The way to think about this data center AI will continue to ramp through 2027, and the server business is expected to double. So quarter-over-quarter, there may be variations because the mix of those businesses will be a little bit different.

In addition, embedded businesses, strengthening in 2027, I think that will also help our gross margin. Overall, when you put the whole business model together, what you should expect us to do with significant top line revenue growth, significant gross profit expansion and OpEx growing [indiscernible] revenue growth, which will result in very significant operating margin expansion and profitability expansion.

Operator

And the next question comes from the line of Timothy Arcuri with UBS.

Timothy Arcuri

Lisa, I'm wondering if you can break down a little bit how data center is going to look like next year. You did revise server CPU up from up 70% to more than 100%, but you're not being super specific about data center AI. So is the right way to think about it that the GPU business will still be bigger in dollars than the server CPU business, but not much. Is that the right way to think about the split? And then how to think about the shape of the year? I know that the GPU business is lumpy, but I imagine it's probably pretty ratable through the year. Is that kind of fair to think there?

Lisa Su

Yes, Tim, maybe let me break down a couple of things. So first of all, we think 2027 is going to be a very strong momentum year for both the CPU business as well as the data center AI business. And when we gave you some of the markers, you should just consider them markers, right? So we've said the overall company will grow greater than 80%. And for the data center CPU business, we will about double; and for the data center AI business, it will significantly more than double. And from that standpoint, you should expect that the data center AI business continues to be, as Jean just said, Q4 is the first quarter where the data center AI business will cross over in terms of size compared to the CPU business. And we would expect both businesses to grow nicely throughout the year. So hopefully, that gives you a bit of shape as we go through the year.

Timothy Arcuri

And then, Jean, I guess, can you help us on the warrants? How are they going to layer in through the year? Any sort of like guardrails you can give us on sort of how to roll that into our model through the year and into 2027?

Jean Hu

Yes. Thanks for the question. I think the simple way to think about it is there are 3 parts for the warrant. The first one is we report net revenue, the Q3 revenue, Q4 guide and where we talk about the revenue, it's always net revenue, which already incorporated the warrant revenue reduction piece, which is determined at the grant [indiscernible] fair value. So that one you should not worry about it. It's always incorporated in our revenue whenever we talk about the revenue.

The second piece is probably you're asking is the share count. The way to think about this the share count only come into diluted share once all the vesting conditions and the exercising conditions are met by our customers. That means the shipments need to be there and some commercial conditions that need to be met. We are going to provide the guidance in 2027 to give you a clear guidance on we believe the share of vesting and exercising conditions are met. That is something we'll provide transparency once we get into 2027.

The last piece is it's a noncash charge, quarterly noncash charge, which is a remeasurement of fair value, that actually does not impact our non-GAAP financial results and free cash flow. So on the share count, just wait for us to provide you a clear guidance when we get there.

Operator

And the next question comes from the line of C.J. Muse with Cantor Fitzgerald.

Christopher Muse

I guess first question, on the server CPU side, you've obviously taken your outlook higher. I'm curious how we should be thinking about units versus mix including higher core count as well as higher ASPs associated with [indiscernible]. Is there a way to break that down to help us understand that 100% growth?

Lisa Su

Sure, C.J. So on the server CPU side, when we talk about 100% growth, you should imagine that we are definitely shipping significantly more units in that. Of course, some of those units have higher core counts as we go into Venice. But I would say the growth in revenue is mostly in units and core count and less so in ASPs, although there are some -- there's an ASP uplift as we increase core count.

Christopher Muse

Great. And just as a follow-up, just to button down 1Q seasonality. Based on what you said tonight, I'm interpreting that we should see kind of normal client in gaming down seasonally in Q1 kind of ratable data center above December. And I guess as part of that, any sort of vision for how to think about embedded?

Jean Hu

For the embedded business, typically, there is seasonality. We do see overall, when we look at the 2027 for the whole year, embedded business is actually going to grow significantly and also strengthening through the year. But typically, Q1, actually, there is some seasonality for embedded business.

Operator

And the next question comes from the line of Stacy Rasgon with Bernstein Research.

Stacy Rasgon

I was wondering, I know you said that you had crossover for GPU versus CPU in Q4 within the context of that 50% guide. I was wondering if you could help us with the magnitude of that crossover? Is it just barely there? Or like is it like significantly higher? I'm just trying to get the sizing in my head in terms of the entrants trade in the next year?

Jean Hu

Yes. Stacy, maybe I'll give you some color, right? Is the way to think about it is if you think about the MI-450 ramp, it's a very significant step up sequentially from Q3 to Q4. And that is really compared to server side, service is growing really nicely, strongly. But the ramp of data center AI, which surpassed the [indiscernible] in Q4. That is a very significant step up. I think the way server business, of course, is very strong, but the data center AI will be much higher than server business in Q4.

Stacy Rasgon

In terms of dollars?

Jean Hu

Yes, in terms of dollars.

Stacy Rasgon

Right. And for my follow-up, there were some questions earlier about gross margins next year. I wasn't exactly quite sure how to interpret your answer in terms of the gross margins next year relative, I guess, to the Q4 rate. Do you think you could give us some color or some highlights on like a good entrance rate and a good exit rate for gross margins in 2027 -- given that the ramps of the different products and everything that are happening.

Jean Hu

Yes, Stacy, the reason we are not able to provide you quarter-over-quarter. It's really because -- when you think about the data center mix is the primary driver in 2027. When you have a data center business revenue is more than 80% of the -- total company business. And then within data center, the interesting parties you have a data center business continue to ramp significantly. On the server side, we did mention, Lisa mentioned that we're going to double in 2027.

Of course, we also know the supply constrained environment and Lisa also mentioned that we are working with our suppliers to ensure we get more supplies. So the mix between those 2 businesses actually determine what the company's gross margin will be. Right now, I actually don't have that visibility. I think we hope both businesses are going to do great. And we know they're going to both to [indiscernible].

Q4 actually, is a good benchmark to think about it is because Q4 data center AI business is much higher than CPU business. So I think that gives you some idea at that benchmark. Then in 2027, different quarter could be different. There could be variation there.

Operator

And the next question comes from the line of Joe Moore with Morgan Stanley.

Joseph Moore

We've heard quite a bit of positive feedback on the Helion racks. And I wonder if you could talk a little bit about that. In the CPU only version of that, what's different for sort of these agentic workloads versus a traditional rack? And then do you anticipate an attach rate of the GPU blade in a lot of these?

Lisa Su

Sure, Joe. Yes, there is quite a bit of positive excitement around the Helion racks. And I think the main point there is these are genetic racks, they really want to be optimized for the environment that they're in. So if they are in, let's call it, hyperscaler or large cloud rocks, you're going to have them in all CPU configurations. It's going to be quite dense.

Venice is the -- from a performance per socket standpoint is the best in the industry. And so that will translate into a very strong agentic rack if you fully populate it with CPUs. But in enterprise, areas, you will also see the opportunity to mix CPUs and GPUs. And it just allows people to really optimize for the workloads that they're in. And it is -- we have contributed to OCP. So it's a fully open rack, and I think that is viewed as a positive. And from that standpoint, we've gotten very positive feedback from both OEMs and ODMs as well as customers.

Joseph Moore

Okay. And if I could follow up, the 100% growth in server, do you think you're meeting demand at that level? Or do you think you still have kind of a shortage condition through all of next year?

Lisa Su

Well, Joe, for us to grow that amount, we're meeting a lot of demand. And when you look at -- when we talked about the TAM growing, let's call it, about 50% CAGR over the next 3 or 4 years, we saw growth in sort of 3 areas, right? So traditional cloud and enterprise, agentic CPUs and agentic workloads and head nodes. We believe we're gaining share in each 1 of those segments as we give you that forecast of roughly 100%.

I do believe there's more demand, and we are actively working with our supply chain partners to satisfy additional demand. Obviously, we also have to make sure that the squared sets are there and DDR5 and all of the other components are there. But my view is we are taking a big step up in terms of the demand forecast from our customers. I'm excited about the fact that our demand is broad-based. So it is really virtually across every customer in terms of their expansion of their footprint with AMD. And so that's positive. And we're going to continue to work on increasing supply to meet additional demand as it comes to fruition.

Matthew Ramsay

John, I'm actually going to ask you to have 1 more caller join us, please. I think we have time for 1 more question here.

Operator

And the final question comes from the line of Jim Schneider with Goldman Sachs.

Our final question comes from the line of Harsh Kumar with BMO Capital Markets.

Harsh Kumar

Question is on Helios. So you're 1 of 2 companies shipping complete racks. Can I ask you about what your share aspirations is? And the reason why I'm asking is you've done a wonderful job in servers, taking share from basically a very difficult competitor? And also, the second part of that question is, do you have aspirations to address the enterprise market with this -- with the Helios rack in the setup? And then I have a follow-up.

Lisa Su

Sure, Harsh. So look, we're very happy with the progress that we've made with Helios. I think this is a significant -- the rackscale solutions are significant R&D efforts. I think Helios has ramped very nicely. We saw a very a good initial quarter in Q3. We're ramping nicely into Q4 and into 2027. Our aspirations are to be a very significant compute supplier to our largest customers, both frontier labs as well as hyperscalers. And from that standpoint, the focus on Helios M450 ramp is very much on our minds for Q4 and 2027. But we're in very active dialogues with our largest customers on MI-500 and MI-600 and beyond. And our aspirations are to continue to have a larger and larger piece of their compute needs. And from everything that we see, demand continues to go up. So I think that's the overall view on Helios.

Was there another part of that question [indiscernible]. Yes, certainly. I think we continue to see -- although enterprise is developing, let's call it, a little bit slower than the largest Frontier labs and hyperscalers, we do believe that the enterprise business is quite important. We have made a lot of progress in enterprise CPUs. Many of our enterprise customers are actually asking for our advice as to how to deploy AI and how to ensure that they're deploying it as efficiently as possible when you think about both on-prem as well as cloud deployments. So we are expanding our portfolio. We've introduced some enterprise-specific GPUs, and that will be a place where we would like to grow our footprint to, again, be a very strategic supplier to our largest enterprise customers.

Harsh Kumar

Understood. And my follow-up is on the client side. Historically, you've done a great job managing this business through situations like this. What would be your share keep strategy if your chief competitor was to cut ASPs, would you try to match them? Or would you be -- would you have to just gauge the situation as it goes? Just be curious about how you're going to manage this.

Lisa Su

Yes. I think, Harsh, our view is PCs are an important market for us in the standpoint of I'm a big believer in end-to-end AI deployment, so from cloud to edge to clients, and we have a unique portfolio in that fashion. That being the case, I think we're always quite disciplined in how we manage. Our focus has really been the most -- the highest value and most profitable segments of the PC market. Enterprise, premium consumer, desktop channel. These are places that we do very well. We do recognize that there's a good amount of volatility in the low end of the market, and that's a place where we have the least exposure, frankly, to that portion of the market.

Matthew Ramsay

All right. Thank you, John, for organizing the Q&A. Thank you, everybody, for joining our call. We really appreciate you spending some time with AMD and thank you very much for your questions and your attention. John, go ahead and close the call. Thank you very much.

Operator

Thank you, Matt. And ladies and gentlemen, thank you for your participation. This does conclude today's teleconference. You may disconnect your lines at this time, and have a wonderful day.