Huang Renjun's cousin once removed is surging ahead on the road of overcoming difficulties and realizing her ambitions.

Author: Liu Yangnan

Editor: Wang Bo

On November 4th, local time in the United States, AMD released its third-quarter financial report for the fiscal year 2025.

In AMD's official financial report press release, the expression 'record' appeared a total of seven times.

According to this financial report, AMD has set new records in total revenue, profit, free cash flow, and consumer revenue: with total revenue reaching $9.246 billion, a year-on-year increase of 36%; gross profit of approximately $4.78 billion, a yea

AMD GAAP Quarterly Financial Performance (Jiazi Guangnian)

AMD's other business areas also performed better than expected: the datacenter business generated $4.341 billion in revenue, a year-on-year increase of 22%, while the gaming business generated $1.298 billion in revenue, a surge of 181% year-on-year.

"We have presented an outstanding quarterly report with record-breaking revenue and earnings, reflecting the widespread demand for our high-performance EPYC, Ryzen processors and Instinct AI accelerators," said Dr. Lisa Su, AMD's chairman and CEO.

Su Zifeng (Lisa Su)

It is worth mentioning that Suzhongfeng and NVIDIA founder and CEO Huang Renxiong are distant relatives. By birth, Suzhongfeng should call Huang Renxiong as her "second cousin", which means she is Huang Renxiong's second cousin once removed.

In her speech at National Taiwan University in April, Su Zifeng once spoke highly of Huang Renxun, saying, 'I highly respect Huang Renxun for building an amazing company.'

AMD, which set multiple records in the third quarter of fiscal year 2025, has quietly taken market share from Nvidia in the AI sector. AMD's recent important strategic partnerships with Oracle and OpenAI have set high expectations for the market. The

What the market is most concerned about now is when these orders will be truly placed from paper to account.

How many AI chips does AMD hold?

According to the Jiazi Century, AMD's rapid rise in the AI market is mainly based on its strategic deployment of GPU, CPU, and software. The ambition of Dr. Su is to build a complete computing power ecosystem that can compete with the current market

AMD Instinct GPU

First, in the main battlefield of AI computing, GPU, AMD is chasing closely behind Nvidia.n

AMD's data center business has achieved record growth, benefiting from the popularity of its Instinct MI350 series GPUs and increased server market share. The upcoming MI450 GPU, scheduled for release in 2026, has also attracted NVIDIA's attention.

According to Wccftech, the initial TGP (Total Graphics Power) of Instinct MI450X was 2300W, but AMD later increased it to 2500W to further enhance performance; NVIDIA has also raised the TGP of its VR200 based on the Rubin architecture from 1800W to

TGP refers to the total power limit allowed for all components of a graphics card, including the graphics core, frame buffer memory, power supply module, fan, etc., under typical load, representing the energy supply capability of the GPU.

AMD not only rolls out parameters but also competes with Nvidia for customers.

Oracle has recently pledged to deploy 50,000 MI450 GPUs, with deliveries expected to commence from the third quarter of 2026, and plans to further expand its collaboration with AMD in 2027 and beyond. This order ensures that AMD has locked in a key c

Major clients like Oracle choose AMD not only to balance NVIDIA's ecological binding but also because AMD is a more cost-effective choice. The core advantage lies not in the cost of a single product, but in the total cost of ownership (TCO).

In AI large model training and inference scenarios, TCO calculation includes not only the chip purchase price but also factors such as energy consumption, rack space, interconnection bandwidth, and HBM memory capacity.

AMD leverages the market with higher HBM memory capacity and efficiency. For clients running large models with billions or trillions of parameters, higher HBM capacity means that the entire model can be loaded with fewer chips, reducing the number of

At the recent financial report conference, Su Zifeng revealed that the deployment of AI developers' MI300 series GPUs has also expanded this quarter. IBM and Zifra will train multiple generations of future multi-modal models on large MI300X clusters,

The full performance of the chip's paper specifications can be achieved with the support of the software ecosystem.

While AMD's AI chips are developing rapidly, they are also accelerating the improvement of the software ecosystem. At present, the ROCm ecosystem launched by AMD has been iterated to the seventh generation. Compared with the sixth generation, its inf

Comparison between ROCm 7 and ROCm 6

Guannian Jizi has found that AMD has strengthened its AI software capabilities through a series of acquisitions, including the acquisition of AI software companies Mipsology, Nod.ai, and Silo AI last year, all of which are aimed at accelerating the d

In the future, AMD will continue to increase its investment in AI business hardware and software to strengthen its TCO advantage. AMD will launch the Helios rack-level solution, which will be equipped with AMD Instinct MI450 series GPUs, the next-gen

It can be said that GPU is a compulsory course that AMD must make up. But GPU is not AMD's only way out. In the long run, AMD's real "base" may still be CPU.

In the third quarter financial report call, there was another key piece of information revealed. AMD's traditional server business growth slightly exceeded its renowned AI chip department in the previous quarter. Su stated that server CPU revenue hit

This also indicates that AMD will take a different path from NVIDIA that better aligns with current market demands to catch up with NVIDIA.

As AI computing centers have grown to clusters of tens of thousands of cards, the core market demand is no longer simply single-card performance, but how to provide the most cost-effective computing power scheduling and business adaptation. The CPU i

In recent years, while AMD has been accelerating its pursuit on GPUs, it has not slowed down its innovation on CPUs.

For example, after taking office, Su Zifeng began to develop the Zen architecture, whose core innovations lie in its Chiplet modular design and Infinity Fabric high-speed interconnect bus. In traditional design, increasing the number of cores would l

Meanwhile, AMD's 3D V-Cache technology (through vertical stacking of additional L3 caches) also provides unique value in certain AI workloads. While its help for large-scale deep learning training tasks is limited, it can significantly enhance CPU-si

Overall, AMD is well-versed in playing to its strengths and avoiding weaknesses. By utilizing its dominance in the CPU field, it attracts customer attention to the overall rack TCO rather than mere GPU performance. While selling easier-to-market, hig

Overall, AMD is transforming from a PC chip supplier to an AI system-level solution provider, which is a critical part of its AI strategy and aligns with the current demand for AI computing power in the domestic and international markets.

Between China and the United States

Insufficient computing power - this could be the core motivation for North American giants to pool resources in AI computing power investment before.

The bottleneck of computing power deployment is no longer merely technology or chip manufacturing, but has shifted to infrastructure and energy supply. Faced with the crisis of computing power, the Chinese and American technology industries have take

The core strategy of the United States is to increase supply by investing more money to create more computing power. Taking OpenAI as an example, its plan to invest billions of dollars in computing power is actually a microcosm of the thinking of US

China places more emphasis on 'organizational optimization'. Due to the export control of high-end chips by the United States, China needs to effectively utilize every bit of computing power. By optimizing algorithms, scenarios, and precise matching

AMD has established different ecological niche markets across both China and the United States.

In the North American market, AMD has joined forces with giants such as OpenAI, Oracle, Meta, and Microsoft in the computational power reunion to provide the most advanced MI450 and MI350 series GPUs and directly participate in the cutting-edge compe

In the Chinese market, AMD must design compliant customized chips to maintain its market share in China. For example, AMD has designed data center GPUs such as the Instinct MI308, whose performance and interconnect speed have been intentionally limit

However, AMD explicitly stated that its third-quarter results did not include any revenue from the sales of AMD Instinct MI308 GPUs to China.

In general accounting practices, when a company is uncertain about whether its final shipments will be realized or if its revenue recognition meets accounting and regulatory requirements, it may choose not to record uncertain sales revenue. From AMD'

However, in this complex environment, AMD is also trying to bypass non-commercial factors through commercial means. It partners with leading cloud providers such as Alibaba Cloud in China to focus on scenario-based computing power and chip co-creatio

At this year's Cloud Town Conference, AMD set up seven exhibition units in the exhibition hall, covering CPU, embedded computing solutions, ROCm, Mini AI workstations, as well as Alibaba Cloud's ApsaraDB RDS, ECS, and ACS ecosystem collaboration cont

Liu Libin, a senior technical expert in server research and development at Alibaba Cloud Intelligence Group, said, 'Our collaboration with AMD is deepening. Through the ninth-generation AMD cloud instances, we provide customers with a choice that bal

AMD's booth at the Cloud Native Computing Conference

According to Jiazi Guangnian, AMD's strategy in China's AI market is to participate in a compliant manner, provide an alternative to CUDA, cooperate deeply with cloud providers, and seek growth through an open ecosystem.

It is not difficult to find that AMD has built a complete business architecture with clear strategic priorities and decisive actions. I believe it will be difficult for people to overlook AMD in the next AI computing ecosystem.

AMD's revelation: casting the net first, then gathering

AMD's successful transformation is almost a textbook enterprise revival in the history of global semiconductors. Its core driving force comes from two key business reorganizations led by Suzenn Feng: the first is 'diversification', and the second is

The ebb and flow of this process have outlined a clear development path for AMD. These experiences are of great significance to Chinese chip companies in the current complex environment.

At the beginning of Suzy Tong's term in 2014, AMD was on the brink of financial collapse, facing difficulties with revenue decline and cash flow depletion. The company's management was fully aware that it must break away from the excessive reliance o

To this end, AMD launched its first reorganization in 2014, targeting the high-performance computing market, and restructured its business into two divisions: “Computing and Graphics” and “Embedded and Semi-Custom”. The core goal was to diversify its

When revenue is in a downturn, this change is like "spreading the net" to quickly capture new market share and gain time for survival.

Despite facing financial pressure, Su Zifeng made a high-risk, heavy-investment decision to develop the Zen architecture.

AMD Zen

The success of the Zen architecture not only saved the company but also truly impacted Intel's long-term monopoly in the server CPU market through the EPYC processor series. It also provided AMD with the technical conditions to transition from PCs to

The second reorganization occurred in the second quarter of the 2022 fiscal year. Prior to the release of ChatGPT, AMD had presciently elevated its datacenter business to an independent unit that could respond quickly to market changes. This departme

To ensure strategic execution, the company has quickly promoted Forrest Norrod to Executive Vice President and General Manager of its Data Center Solutions Business Unit in terms of organization. With extensive experience in enterprise computing, Nor

From 'spreading out a wide net in high-performance computing' to 'focusing on' AI, we have learned some valuable lessons from AMD's transformation path.

First of all, being bold to merge and acquire does not mean “I want all of them”.

AMD's experience shows that mergers and acquisitions should essentially serve the company's long-term strategy and main business, clarify its own core boundaries, and be willing to make trade-offs.

AMD has acquired GPU, FPGA, networking, system design, and AI software capabilities through a series of mergers and acquisitions (Xilinx, ZT Systems, Silo AI). However, after acquiring ZT Systems, AMD made a critical move to quickly divest its manufa

ZT Systems' core values lie in its rack-level integration capability and system thermal design, which are crucial for deploying MI series GPUs on a large scale in cloud data centers. Physical manufacturing is a capital-intensive process with relative

Secondly, it is necessary to truly understand the market demand.

At present, the core demand of the AI computing market has shifted from focusing on the peak performance of a single chip to the TCO and business adaptability of large-scale clusters.

AMD's acquisition of ZT System indicates that it is not focusing on competing head-on with NVIDIA in terms of GPU performance. Instead, it emphasizes the advantages of its EPYC CPU in cluster computing, data coordination, and overall TCO, in response

Finally, openness is the key principle.

Although AMD is catching up quickly in GPU hardware, the relative lagging of the ROCm ecosystem is a core obstacle to its market share expansion. To address this pain point, AMD chose to embrace open source code and promote rapid iteration and practi

Overall, AMD's transformation is a classic case in the semiconductor industry, with a high-risk strategic bet, continuous R&D investment, and flexible business adjustments being indispensable.

Of course, success in one stage is not equal to permanent success. Nowadays, the global computing power market is bound to be turbulent, and AMD's radical transformation also buries some potential risks for its future long-term development.

Market sentiment can indicate something. Despite the comprehensive performance exceeding expectations, AMD's stock price fell more than 3.7% in after-hours trading. Wall Street analysts believe that this reflects the pressure of profit-taking after t

We believe that the maturity of the AMD ROCm ecosystem will take more time. Although the deep cooperation between OpenAI and Oracle has accelerated the iteration of ROCm, whether it can achieve stability, reliability, and development tool sets compar

In addition, AMD has acquired a large number of different businesses and technology stacks such as Xilinx, Pensando, ZT Systems, and Silo AI within a short period of time. The challenge lies in how to efficiently integrate these complex software and

However, instead of dealing with these long-term issues, AMD's top priority now might be to fulfill the large orders worth billions of dollars recently signed, which will directly determine market confidence and expectations for AMD.

（Source of cover image and accompanying images in the text: AMD）