AMD Completes Acquisition of World Labs for Approximately $8.2 Billion; Fei-Fei Li Appointed Chief Scientist
AMD announced the formal completion of its acquisition of World Labs, a transaction valued at approximately $8.2 billion, structured as an all-stock deal, which actually closed on November 13, with AMD issuing approximately 14.98 million shares of common stock to the original shareholders. World Labs was founded in 2024 by Fei-Fei Li and others and researches spatial intelligence and world models. Fei-Fei Li has joined AMD as Executive Vice President and Chief Scientist, reporting directly to CEO Lisa Su. The article also traces AMD's acquisition path from Xilinx and ZT Systems to Silo AI and Nod.ai, as well as its GPU deployment partnerships with OpenAI, Meta, and Anthropic.
Why read this
SelectedAI score 78
The article links AMD's several recent acquisitions into a single trajectory from chips toward systems and model capabilities, which can be used to understand how it is challenging Nvidia.
Topics
Body · AI translation
Today, AMD announced that it has officially completed its acquisition of World Labs. AI scientist Fei-Fei Li has formally joined AMD as Executive Vice President and Chief Scientist, reporting directly to CEO Lisa Su.
AMD Buys World Labs: The "Chip Queen" Is Also an Expansion Queen
Under the transaction agreement previously disclosed, this is an all-stock acquisition valued at approximately $8.2 billion. According to a filing AMD submitted to the U.S. Securities and Exchange Commission (SEC), the transaction was actually completed on November 13, with AMD issuing approximately 14.98 million shares of common stock to former World Labs shareholders.
This also makes it another major acquisition for AMD by announced transaction value, following its roughly $48.8 billion acquisition of Xilinx in 2022.
But unlike Xilinx, World Labs is not a chip company in the traditional sense. It was founded in 2024 by Fei-Fei Li and others, and focuses on spatial intelligence and world models, hoping to enable AI to not only understand text and two-dimensional images, but also generate, understand and simulate three-dimensional physical environments. Its research areas include interactive 3D world generation, robot learning and simulation.
From a business perspective, this is quite different from AMD's traditional CPU and GPU products.
But Lisa Su explained why she was drawn to World Labs. In the announcement of the completed acquisition, she said that to provide infrastructure for the next stage of superintelligence, it is necessary to optimize both the computing platform and the models. World Labs' research will help AMD shape future hardware, software and systems.
It is worth noting that this is not AMD's first acquisition of an AI model team. In 2024, it had already acquired European AI lab Silo AI for approximately $665 million.
Looking at these acquisitions and business changes over the past few years together, during Lisa Su's tenure at the helm of AMD, the company has undergone an expansion larger than outsiders might imagine, mainly through M&A, though the capabilities AMD bought differed at different stages.
In February 2022, AMD completed its acquisition of Xilinx. According to its annual report, the final purchase consideration was approximately $48.8 billion.
This transaction gave AMD technology assets such as FPGAs, adaptive computing and embedded systems, and also expanded its business coverage in areas including data centers, industrial, communications and automotive. In the same year, AMD also completed its acquisition of Pensando, bringing data center networking, DPUs and infrastructure processing capabilities into its product portfolio.
These two transactions share one thing in common: they both took place in the semiconductor and data center hardware fields, and their main purpose was to expand the variety of computing products AMD could offer.
In October 2023, AMD completed its acquisition of Nod.ai. This company does not make GPUs; it focuses on open-source AI compilers and model optimization software. The company was co-founded in 2013 by Anush Elangovan and Harsh Menon. Elangovan was formerly a member of Google's first Chromebook team and also served as a chief engineer at Cisco. Menon previously worked at Kitty Hawk, a company backed by Google co-founder Larry Page.
AMD noted in its announcement that the Nod.ai team had contributed to open-source compilation technologies such as Torch-MLIR and OpenXLA/IREE.
For ordinary users, these technologies may be quite unfamiliar. But for AI chipmakers, they directly affect whether models can run efficiently on different hardware. When the same model runs on different GPUs, it often requires solving issues such as operator adaptation, memory layout, computation graph optimization and kernel scheduling. The role of the compiler is to convert model computations into programs suitable for execution on the target hardware as much as possible.
By 2024, AMD's acquisitions began to extend further into model development.
In August of that year, AMD completed its acquisition of Finnish AI company Silo AI for approximately $665 million.
Founded in 2017, Silo AI has experience developing AI models and applications for enterprises, has worked with companies such as Allianz, Philips, Rolls-Royce and Unilever, and has also developed open-source multilingual large models such as Poro and Viking. The significance of the acquisition is not just adding a model R&D team.
AMD can have more models trained and deployed on its own hardware, accumulate real-world model operation experience, and further improve its software tools and optimization solutions.
In June 2025, AMD announced the acquisition of Brium.
By 2026, AMD's acquisitions went further at both ends.
In August, AMD announced that it had signed a definitive agreement to acquire Taalas. This Canadian company researches specialized AI inference chips and hopes to reduce the compute and memory bottlenecks that exist in general-purpose computing architectures by optimizing data flow. According to AMD's latest quarterly report, the transaction is valued at approximately $2.9 billion, including up to approximately $1.4 billion in contingent consideration. As of the disclosure in that report, the transaction was still pending completion.
Then in September, AMD announced the acquisition of World Labs; in November, the transaction was officially completed.
Unlike Silo AI, which mainly targets language models and enterprise applications, World Labs' research focuses on spatial intelligence, interactive three-dimensional environments and robot simulation. These new types of models may impose different requirements on how GPUs compute, on memory, on data transfer and on system organization.
From these moves, a gradually clearer path can be seen: AMD is no longer just buying companies that can expand its existing hardware product lines, but is extending into different links of the AI computing chain. Compilers, models, inference chips, server system design, and even modeling of the physical world for the future are all beginning to enter AMD's technology map. Compared with its past chip-centric expansion, AMD is trying to build a more complete AI technology system.
Why does AMD increasingly look like an AI systems company?
If acquiring model companies represents AMD extending upstream in AI technology, then acquiring ZT Systems is an important step in its expansion toward complete computing systems.
In March 2025, AMD completed its acquisition of ZT Systems. According to financial filings, the total purchase consideration for the transaction was approximately $4.4 billion, including cash, AMD stock and contingent consideration.
ZT Systems has long provided data center computing infrastructure to hyperscale cloud customers and has accumulated experience in server and rack-level system design, validation and deployment. For AMD, this capability has practical value.
In the past, when people measured two GPUs, they usually compared peak compute, memory capacity, bandwidth or power consumption, but once AI clusters reach the scale of thousands or tens of thousands of cards, the performance of a single GPU is only part of the picture. How GPUs interconnect, how CPUs coordinate, how data flows between memory and the network, and how the system is powered, cooled and maintained can all affect final computing efficiency. Especially for large-model inference, what users ultimately care about is often not the theoretical performance of a single GPU, but how much time it takes to complete a request, how many users can be served simultaneously, and how much each token costs.
This means that chip vendors are increasingly participating in system-level optimization. The acquisition of ZT Systems precisely fills in AMD's capabilities in this regard, and what better illustrates AMD's strategic intent is that after completing the acquisition, it did not retain all of ZT Systems' manufacturing business for the long term.
In October 2025, AMD sold its data center infrastructure manufacturing business to Sanmina, while retaining some intellectual property and the system design team. Sanmina became AMD's preferred partner for manufacturing complex AI systems.
In other words, AMD prefers to retain system architecture design, customer adaptation and R&D capabilities, while continuing to leave large-scale manufacturing to specialized partners. This is somewhat similar to its long-standing fabless operating model: mastering core design capabilities while relying on the external manufacturing ecosystem to achieve scale delivery.
This transformation has already begun to show in its products.
AMD's Helios rack-level AI system integrates Instinct GPUs, EPYC CPUs, Pensando networking technology and ROCm software, with the goal of providing an AI computing platform that can be deployed at scale.
In October 2026, AMD released the Helion rack solution for Agentic AI. Unlike systems built mainly around GPU inference computing, Helion particularly emphasizes the role of the CPU in Agent workloads.
The reason is that when an Agent completes a task, it often needs to continuously carry out task planning, tool invocation, state management and interaction with external systems. In addition to model inference, there is also a large amount of general-purpose computing and system coordination work.
AMD therefore proposed different configurations: both racks that mainly rely on EPYC CPUs to handle Agent execution tasks, and enterprise-level solutions that integrate both CPUs and GPUs. This shows that AMD has begun to reorganize its product portfolio according to the computing characteristics of different AI applications.
But whether Helios or Helion, neither is simply putting existing chips into a larger cabinet. Behind them is another kind of value that AMD hopes to provide: customers do not need to independently solve all hardware integration problems, but can deploy and optimize AI workloads based on AMD's system solutions.
From selling a single chip to participating in designing an entire computing system, what AMD is competing for is no longer just GPU market share, but also the right to design AI infrastructure systems.
AMD certainly wants to have its own CUDA ecosystem.
There is one question that always remains unavoidable: if AMD's GPUs are competitive enough in performance and price, why does it still need to keep acquiring software companies? The answer is related to the software ecosystem Nvidia has built over the past decade or more.
CUDA is not just a set of GPU programming interfaces. Around CUDA, Nvidia gradually built up compilers, math libraries, deep learning acceleration libraries, development tools, model optimization and deployment frameworks, and a vast developer community.
In the era of large AI models, the value of a GPU can no longer be measured by hardware performance alone. For developers, whether a GPU can directly run mainstream frameworks, whether it supports common models, and whether there are mature optimization tools when performance problems arise are often just as important as the hardware price. This is also the problem AMD has been trying to solve.
But in fact, it is not that AMD has no software platform of its own. Its long-developed ROCm is precisely an open software stack for GPU computing and AI, providing compilers, runtimes, compute libraries, and support for mainstream AI frameworks.
But having a software platform and having a mature developer ecosystem are two different things.
Especially in large-model training and inference, even if both platforms support PyTorch, practical deployment may still encounter problems such as operator compatibility, GPU kernel optimization, framework version adaptation, and performance debugging. If the engineering investment required for migration exceeds the hardware cost saved by using another GPU, customers may not be willing to switch platforms.
This is also the real-world background for AMD's acquisitions of software teams such as Nod.ai and Brium.
Nod.ai's open-source compiler technology can help AI models be compiled and optimized more effectively on AMD hardware; Brium focuses on reducing model deployment's dependence on specific hardware configurations. The two acquisitions are not simply adding software engineers, but filling in the technical links developers are most likely to encounter when using AMD GPUs.
In July 2026, AMD and Anthropic announced a strategic partnership, taking this approach a step further. Under the agreement between the two sides, Anthropic plans to deploy up to 2 gigawatts of AMD Instinct GPUs in the future, with the first 1 gigawatt planned to start in the first half of 2027.
But the partnership is not only about purchasing GPUs. The two sides also plan to use Claude to optimize AI workloads on AMD GPUs and accelerate ROCm software development. AMD also plans to widely adopt Claude in its own engineering and product R&D teams. The signal is fairly clear: AMD is having its leading AI customers directly participate in building its software ecosystem.
On the one hand, large-model companies provide real workloads and development needs; on the other hand, AMD improves compilation, runtime, and inference efficiency based on these needs.
These moves more or less bear the shadow of CUDA.
In addition to acquisitions, another obvious change for AMD over the past year is that its relationships with large AI model companies are deepening.
In October 2025, AMD and OpenAI announced a multi-year, multi-generational GPU deployment agreement involving 6 gigawatts of computing infrastructure, with the first 1 gigawatt planned to start in the second half of 2026.
The cooperation between the two sides is not limited to procurement. According to information published by AMD, OpenAI will work with AMD on next-generation GPUs and rack-scale AI systems to help optimize future computing platforms.
In February 2026, AMD reached a cooperation agreement of similar scale with Meta, planning to deploy up to 6 gigawatts of AMD GPUs. This cooperation has a noteworthy technical detail: Meta's first deployment will use custom Instinct GPUs based on the MI450 architecture and optimized for its own workloads.
The two sides will also jointly coordinate GPU, CPU, system, and software roadmaps.
By July 2026, Anthropic joined this camp of cooperation, with its agreement with AMD involving up to 2 gigawatts of GPU deployment using the Helios rack-scale solution, while also launching long-term software engineering cooperation.
If the maximum planned scale of the three agreements is added together, the computing infrastructure involved reaches 14 gigawatts.
These specific partnerships still reflect a trend: AMD is entering the long-term technology planning of the world's largest AI model companies and cloud computing customers.
Previously, a chip company mainly designed products based on market forecasts and then sought customer purchases. Now, AMD hopes more to obtain customers' judgments about next-generation models in advance and have customers participate in product design. This model has two benefits: first, it can reduce the mismatch between chip R&D and real workloads. For example, changes in model architecture, context length, inference methods, and computing precision may all affect the chip's design trade-offs in memory capacity, bandwidth, compute units, and interconnect; second, it can carry out software optimization around the needs of important customers before the product goes on the market. When a customer plans to deploy tens of thousands or more GPUs, even a change of only a few percentage points in single-inference efficiency may have an important impact on overall computing costs.
Of course, customer participation in roadmap planning does not mean that a chip company must develop its products entirely around a single customer. AMD still needs to strike a balance between customized optimization and the demands of the general market.
But from OpenAI to Meta and then to Anthropic, it is clear that AMD is changing its position in the AI industry chain. It hopes to no longer merely compete when customers decide what GPUs to buy, but to go further and already be seated at the meeting table when customers discuss what kind of computing platform they will need in the future.
So the acquisition of World Labs can be said to be a natural choice. Previously, AMD relied more on external model companies to provide demand information. With the World Labs team joining, it will also have the capability to conduct internal research on frontier models.
As AMD continues to expand, what kind of company is it becoming?
Whether this round of expansion can succeed—at least on the financial level—AMD has already delivered a report card showing clear growth.
On November 3, AMD released its financial results for the third quarter of 2026. In the quarter, the company achieved revenue of approximately US$13.4 billion, up 45% year on year; of that, data center business revenue was approximately US$8.5 billion, up 97% year on year, setting a new quarterly high.
Based on these figures, the data center business already accounts for approximately 64% of AMD's total revenue. Unlike in the past, when the outside world mainly focused on how AMD was challenging Intel in the personal computer processor market, data centers have now become the core business determining AMD's growth performance.
AMD said in its financial report that the growth came from strong demand for Instinct GPUs and EPYC CPUs, as well as the launch of new-generation products. More notably, AMD's expansion has extended from procurement and R&D to capital investment in AI infrastructure. For example, when signing the GPU cooperation agreement with OpenAI, AMD also provided OpenAI with warrants to purchase up to 160 million shares of AMD stock. In other words, AMD's relationship with leading AI companies is no longer merely that of supplier and buyer; it also includes joint software development, product roadmap coordination, and capital arrangements.
On the other hand, continuous expansion and integration will also bring new challenges. From the system design engineers at ZT Systems to the compiler experts at Nod.ai, and then to the model researchers at Silo AI and World Labs, the R&D pace, technical goals, and organizational culture of these teams are not entirely the same. What AMD needs to solve is not only how to retain these teams, but also how to integrate technical capabilities scattered across different organizations into a unified product roadmap.
From the perspective of external competition, AMD is also far from shaking off the pressure brought by Nvidia.
Nvidia has already built long-term accumulated advantages around CUDA, GPUs, networking and interconnect, system platforms, and development tools. Even if AMD fills in more technical links, it will still need to compete in areas such as software maturity, developer habits, and the reliability of large-scale deployment.
Moreover, against the backdrop of large AI customers beginning to develop their own chips and inference workloads continuously diversifying, AMD not only needs to compete with Nvidia but also needs to prove that its products still have advantages over dedicated ASICs and other accelerated architectures.
The answers to these questions can only be left to time.
Reference links:
This article comes from the WeChat public account "AI Frontline" (ID: ai-front), author: Dongmei, published by 36Kr with authorization.
AI Trends