AMD, the new data center king with a $700 billion market cap

Author: Su Yang, Tencent Technology

“Su Ma” is shaping an AMD for the AI era.

Currently, it is inaccurate to define AMD solely by hardware. When AMD releases its Q1 2026 financial report, it will essentially mark its complete transformation from a chip manufacturer to a “computing power arms dealer” in the AI era.

In the first quarter of fiscal year 2026, AMD achieved $10.25 billion in revenue, with its data center business increasing from $5.38 billion in Q4 2025 to $8B in that quarter. It must be clear that this is a very critical milestone.

In Q1 2024, Nvidia was also at this juncture, after which its data center revenue began to surge, initiating a path of quarterly growth exceeding $10 billion.

Wall Street has long sensed AMD’s transformation.

Over the past 52 weeks, AMD’s stock price soared from a low of about $106.98 to a high of $469.22, with over an 80% increase in the past 30 days alone, and a market value exceeding $700 billion.

Bernstein analyst Stacy Rasgon upgraded AMD to “Outperform,” doubling the target price from $265 to $525; JPMorgan analyst Harlan Sur also pointed out that AMD’s forecast indicates a “multi-year CPU revenue trajectory far above previous expectations.”

Data is cold, investor enthusiasm is hot, but three structural signals are clear: CPU king, GPU:CPU ratio shifting from 8:1 to 1:1, and the dual-engine of GPU+CPU — this is the “homework” that Su Zifeng, CEO for nearly 12 years, has delivered to the board and Wall Street.

The new data center king

For decades, Intel has held an absolute dominance in the data center CPU market. Until 2017, AMD appeared as a challenger, launching its first-generation data center brand EPYC, called “Xiaolong” in Chinese, with the flagship product being AMD EPYC 7601.

In the nearly 10 years since, AMD has continued to catch up. Recently, in this fiscal quarter, AMD’s data center revenue reached $5.78B, continuing to surpass and widen the gap with Intel, whose data center revenue for the same quarter was $5.1 billion. In fact, this surpassing had signs in the past two quarters: in Q3 2025, Intel’s data center revenue was $4.1 billion, while AMD’s was $4.3 billion.

Digitime emphasizes that this is AMD’s first time surpassing Intel significantly in data center revenue in the first quarter, and also suggests that this trend may mark the beginning of AMD’s long-term dominance, a historic turning point in the x86 server market landscape, and that Intel may find it difficult to shake this position for some time.

Many call AMD the “CPU king,” but it should be more accurately described as the “data center king.”

In Mercury Research’s Q1 2026 report, AMD achieved a record 46.2% revenue share in the server market, up from 39.5% in the same period in 2025.

Rather than just looking at market revenue share, I prefer to combine it with shipment share — Mercury Research also reported that in Q1 2026, AMD’s shipment share in the data center CPU market was 33.2%.

33.2% shipment share corresponds to 46.2% market revenue share — a topic related to market share efficiency.

Bank of America Securities’ report in May 2026 pointed out that AMD is expected to hold about 50% of the server CPU market. Purely by proportional calculation, AMD could potentially capture over 70% of the revenue in the server CPU market, aided by a more flexible Chiplet architecture and a more stable roadmap — advantages in single-core performance, energy efficiency, and total cost of ownership.

AMD once set a goal to improve AI training and high-performance computing (HPC) server processors/accelerators’ energy efficiency by 30 times between 2020 and 2025. In June 2025, AMD disclosed that its combined configuration of fifth-generation AMD EPYC CPUs and Instinct GPUs has successfully achieved a 38-fold increase in energy efficiency compared to the 2020 baseline system. This means that, for the same computational workload, AMD’s product energy consumption has decreased by up to 97%.

Meanwhile, AMD also aims to have, by 2030, a single rack capable of training typical AI models that currently require over 275 racks, reducing power consumption by 95%, and cutting carbon emissions from about 3,000 tons to 100 tons.

Lower total cost of ownership leads to higher adoption.

AWS, Google Cloud, and Microsoft Azure have announced new and expanded fifth-generation EPYC cloud instances, including Google Cloud H4D virtual machines and Azure cross-workload optimized instances. Oracle announced it will deploy its first 50k MI450 GPUs in Q3 2026, becoming the world’s first publicly available large-scale AMD computing cluster.

With widespread customer adoption, only capacity might hold AMD back.

At the Morgan Stanley conference in March, when asked whether CoWoS packaging capacity was sufficient, Su Zifeng replied: “We definitely have enough CoWoS capacity. The best answer I can give you is that we have capacity, technology, strong customer relationships, and data center providers have already allocated space for this.”

CPU takes center stage

At the “Advancing AI” event in June 2025, “Su Ma” foresaw a surge in inference demand, although at that time, CPU demand was still overshadowed by the halo of GPU and ASIC inference performance.

As agent products like Openclaw exploded earlier this year, Su Zifeng also updated her view on the CPU narrative. She said at the Morgan Stanley conference in March: “Even large cloud providers are surprised. The CPU computing demand driven by AI might be a severely underestimated variable.”

AMD’s performance this quarter is not only about growth in numbers but also about management’s assessment of the market’s evolving growth potential.

In the Q1 earnings call, Su Zifeng said: “In the past, the ratio of CPU to GPU was basically 1:4 or 1:8 in head node configurations. Now, this ratio is approaching 1:1.”

This change in quantity essentially reflects that AI compute infrastructure is moving away from a single “accelerator mode” toward a “balanced computing era.” Goldman Sachs’ latest research report states that agents are fundamentally about “action,” not “prediction,” and this shift from vector computing to logical reasoning causes a structural regression of computational load toward CPUs.

The explosive growth of agents is precisely what is changing this narrative.

In the past, a data center’s standard configuration was one CPU for 4 to 8 GPUs, with the CPU serving as a dispatcher and manager. From ChatGPT to large model training, this has been the case. After agents emerged, AI assistants shifted from “answering questions” to “autonomous task completion,” requiring continuous planning, tool invocation, result checking, and re-planning. These intensive logical judgments and task orchestration must rely on CPUs.

Agent inference demand is an objective factor. Under the GPU-centric narrative, Wall Street also needs a new story.

Bank of America’s report predicts that the global data center CPU market will soar from $27 billion in 2025 to $110 billion in 2030. Moreover, AMD management aggressively raised its 2030 server CPU total addressable market target from an 18% CAGR to over 35%, aiming for $120 billion.

Evercore ISI’s report introduces the concept of a CPU revival. Similar topics have been discussed recently, essentially about CPUs re-entering the spotlight.

The ratio is changing, the market size is changing, and AMD, now the data center king with lower total cost of ownership, is entering its best era.

Dual-engine acceleration

Rapid growth in CPU demand is pushing AMD toward a trillion-dollar market cap.

However, even as the demand ratio shifts, CPU growth does not mean replacing GPUs. In AMD’s ecosystem, they are more like two engines accelerating simultaneously.

In the earnings call, Su Zifeng clearly articulated the dual-drive perspective: “Strong demand for EPYC server CPUs and increasing shipments of Instinct GPUs will continue to rise.”

The CPU and GPU businesses are accelerating in tandem, a rare trend in the semiconductor cycle. It signifies that AMD is no longer solely competing based on single-chip performance but is transforming into a full-stack computing solution provider through combined CPU and GPU strategies, optimizing total cost of ownership for compute customers. Deep collaboration with Meta exemplifies this strategy.

In February 2026, Meta signed a deal with AMD not just for GPU procurement but for bundling 6 gigawatts of Instinct MI450 GPUs with sixth-generation EPYC server CPUs.

Meta’s choice of this deep bundling scheme is driven by the need for “end-to-end” optimization for trillion-parameter Llama 4 and subsequent agent clusters. The CPU+GPU bundle effectively reduces system-level instruction latency. For example, using Infinity Fabric interconnect technology, it overcomes the traditional “memory wall,” enabling seamless, ultra-low latency data exchange and unified memory sharing between EPYC processors and Instinct accelerators.

Su Zifeng explained in the conference: “We design chips based on actual workloads.” AMD’s customization from the bottom architecture to system level for Llama models exemplifies this.

Meanwhile, Goldman Sachs also pointed out in its report that “the ‘CPU+GPU’ package not only increases order value but also, through deep architectural integration, makes it difficult for cloud providers to switch to other solutions afterward.”

This trend is also reflected in the GPU supply agreement with OpenAI’s 6GW GPU capacity: CPU deployment plans are synchronized within the contract. “Customers are already planning CPU and accelerator deployments in tandem — this is a key market signal,” Su Zifeng said.

Financial data confirms the logic.

Server CPU revenue grew over 50% YoY in Q1, with guidance for over 70% growth in Q2. Meanwhile, the Instinct GPU series is ramping up shipments, with the MI450 entering large-scale volume in the second half. These two engines are jointly amplifying AMD’s data center growth potential.

In the past, NVLink and CUDA built a closed ecosystem, but AMD has always adhered to an open philosophy. In the new race defined by agents, AMD aims to leverage its existing advantages in the CPU market combined with GPU platform integration to deliver a more universal “dual-engine standard,” difficult for Nvidia to replicate in the short term. Take UALink as an example: as an open interconnect standard jointly promoted by AMD and global industry partners, it has attracted over 70 members worldwide, including several Chinese companies.

So, if you ask what differentiates AMD from Nvidia, this might be a tangible aspect.

Su Zifeng’s Second Half

With the stories of the data center king, CPU rise, and GPU+CPU dual engines, Wall Street and investors have given AMD a high valuation of $700 billion, also a positive reflection on Su Zifeng’s nearly 12-year CEO tenure.

In a recent interview, when asked about the “glass cliff,” she said: “When I took over, the outside world was generally pessimistic, thinking AMD was about to go bankrupt, but what I saw was a company with a great engineering heritage, just with unstable execution. I didn’t see it as a trap but as my dream job.”

If the goal of fully integrating AMD into the AI era was the first half, then the second half should be about explosive growth in performance after the transformation, with a recent milestone of reaching a trillion-dollar market cap.

At the Financial Analyst Day in November 2025, Su Zifeng said: “In the next three to five years, annual growth will exceed 60%, pushing data center revenue to $100 billion, and achieving an annual EPS of over $20 within the strategic timeframe.”

Based on this quarter’s $5.77 billion, AMD’s annualized data center revenue is about $23 billion. To reach $100 billion in five years, this figure needs to triple again. Supporting these targets is an error-free product chain: from EPYC Venice, MI450, to Helios rack systems, and the ROCm ecosystem.

Regarding ROCm, there is an interesting reversal. In December 2024, SemiAnalysis founder Dylan Patel published a report criticizing AMD’s ROCm software stack for numerous bugs. Less than 24 hours after the report, Su Zifeng personally contacted Patel. The next morning at 7 am, she sat there listening to SemiAnalysis’ engineering team report each bug and improvement suggestion — a full 90 minutes.

It is precisely because of this attitude that SemiAnalysis rewrote that article in April 2025, changing its conclusion from “unable to become an effective competitor” to “AMD has entered wartime mode, and the MI450X is expected to face Nvidia head-on in late 2026.”

Su Zifeng also invited OpenAI CEO Greg Brockman, Luma AI CEO, Liquid AI CEO, and AI pioneer Fei-Fei Li to CES.

Where compute power flows, Su Ma always invites the key players onto the stage.