Preface:

AMD announced the acquisition of World Labs for approximately $8.2 billion in an all-stock deal, marking the second-largest acquisition in AMD's history, second only to its approximately $50 billion takeover of Xilinx in 2022.

Fei-Fei Li has been appointed as AMD's Executive Vice President and Chief Scientist, reporting directly to Lisa Su.

Author | Fang Wensan

Image Source | Internet

Acquiring Insight into Future Workloads

The missing piece in AMD's puzzle does not lie within the wafer fab. In the second quarter of 2026, AMD reported revenue of $11.5 billion, a 50% year-over-year increase, with its data center business doubling.

The Helios rack, equipped with 72 MI455X GPUs, began shipping, packing 31TB of HBM4 memory per rack. OpenAI, Meta, and Anthropic are among the customers, with OpenAI alone signing a 6-gigawatt procurement agreement.

In terms of hardware, Lisa Su holds a strong hand.

However, NVIDIA possesses the CUDA ecosystem, which has been cultivating developers for 18 years, and the Omniverse, where simulated worlds operate day and night. AMD lacks this edge. ROCm has been chasing NVIDIA's lead for years, but it's always playing catch-up.

Reviewing Lisa Su's acquisition list over the past four years, the strategy is clear. $1.9 billion for Pensando to bolster networking, $665 million for Silo AI to enhance model engineering, $4.9 billion for ZT Systems, followed by a $3 billion divestiture of manufacturing operations to Sanmina, retaining the rack-level system design team. Then came MK1's inference engine, MEXT's memory optimization, and Taalas's dedicated inference chips.

These deals all address the same question: how to make models run better on AMD's chips.

In an interview after the official announcement, Lisa Su put it bluntly: 'The deeper your understanding of the end-to-end process, the better systems you can build. That's why we acquired World Labs.'

What Su is buying is this 'lead time,' bringing the most knowledgeable people about cutting-edge models to her decision-making table, directly incorporating the direction of model evolution into the design trade-offs of subsequent MI series chips.

An analysis has calculated: In an all-stock deal, AMD is only diluting its equity by about 0.8%, but in return, it gains firsthand insights into 'what computing power models need,' which could translate into design advantages for multiple generations of chips.

Citigroup analyst Atif Malik maintains AMD as his top growth-oriented computing semiconductor stock for 2026, with a target price of $575; Wells Fargo's Joe Quatrochi sets a target price of $700, stating that the core of this deal lies in 'providing better forward-looking visibility.'

The Narrative of Computing Power Shifts from Reading Text to Understanding the World

Why World Labs, and not another large language model company? The answer lies in the changing workload structure.

World Labs specializes in spatial intelligence, enabling AI to understand, generate, reconstruct, and simulate 3D physical worlds. Its product, Marble, can generate an explorable 3D world from a single image or sentence; Atlas, released in early September, goes further by natively processing text, images, videos, camera poses, and 3D depth information, reconstructing complete scenes with physical scale from just a few photos.

The significance of these models lies not in showcasing technology but in addressing a critical bottleneck. Large language models consume internet text, but robots lack an 'internet' to consume. Collecting action data in the real world is expensive, slow, and dangerous. The robotics industry's biggest bottleneck is not hardware but the scarcity of training data.

In July, World Labs acquired robotics simulation company SceniX and unveiled the R2S2R (Real-to-Sim-to-Real) system, which trains strategies in generated virtual worlds based on real tasks and then deploys them back to real robots. Here, world models act as 'data factories' for physical AI.

For AMD, the appeal of this track (which can be translated as 'field' or 'domain') lies in its representation of a second growth curve in computing power demand. Training demands for text models will eventually peak, while training demands for robotics, autonomous driving, and industrial simulation are just beginning and are inherently voracious consumers of computing power.

Generating a physically consistent 3D world is far more computationally intensive than continuing a sentence. When a chip factory acquires a world model team, it's planting a flag at the source of demand.

Fei-Fei Li's Calculation: Without Hardware, AI Is Lame

In this deal, Fei-Fei Li is not merely cashing out. Over two years, World Labs' valuation soared from $1.3 billion to $5.4 billion, finally reaching $8.2 billion. Financially, she could have exited at any time. Her choice to join AMD is driven by technical logic.

In her official announcement, she wrote candidly: To take spatial intelligence to the next level, one must 'get closer to hardware.' AI lacking hardware collaboration is limited in efficiency.

This is not mere politeness. Training and inferencing world models have vastly different requirements for memory, communication, and mixed precision compared to text models. An independent lab attempting to influence underlying hardware can only shout across cloud providers and chip roadmaps.

Joining AMD means model requirements can directly influence chip design processes. In her Bloomberg interview, she called it 'accelerating the flywheel effect between software and hardware.'""For a scholar with two decades of research behind her, there's also a historical echo. ImageNet once proved that data could rewrite AI's trajectory, but data's value ultimately depends on computing power. Now, she firmly believes spatial intelligence is the necessary path for AI to understand the physical world, and this path is also paved with transistors at the most fundamental level.

From ImageNet to AMD's Chief Scientist, Fei-Fei Li has consistently pursued one thing: finding the infrastructure to support the paradigms she believes in.

$120 Million Per Person: The Exchange Rate of Talent Density

World Labs employs approximately 70 people. At $8.2 billion, that's $117 million per person.

Looking at the valuation curve, in September 2024, less than half a year after the company's founding, its valuation was $1 billion. In February 2026, it completed a $1 billion financing round, reaching a post-money valuation of approximately $5 billion, with both AMD and NVIDIA appearing on the investor list. Seven months later, AMD bought it outright for $8.2 billion, a premium of about 64%.

The company has no mature product lines, no visible revenue streams—only Marble and the multimodal world model Atlas, released on September 1st.

Patrick Moorhead, a former AMD executive and tech analyst, succinctly captured the pricing logic: this is a deal valuing talent and model insights over revenue.

The contrast is with Hugging Face 25 days earlier, which had an annualized revenue of about $150 million. NVIDIA paid $12.93 billion, over 80 times revenue.

Both deals share the same subtext: in the arms race for large models, the scarcest asset has shifted. Computing power has taken a backseat to those who know where computing power should be directed.

The Battlefield for Chipmakers Shifts from Wafers to Models

Zooming out, this deal is just a corner of the 2026 M&A frenzy. Meta invested $14.3 billion in Scale AI, bringing Alexandr Wang to lead its superintelligence lab; NVIDIA first acquired inference chip company Groq for $20 billion, then shelled out $12.93 billion for Hugging Face, the world's largest open-source model community.

Three giants, three approaches, worth examining together.

① NVIDIA buys access: Hugging Face controls the distribution of models—where they come from and where they go.

② Meta buys data and people: Scale AI's annotation capacity and Wang himself.

③ AMD buys judgment: World Labs doesn't sell shovels or charge admission; it only answers one question: What will the next generation of workloads look like?""These three approaches point to a consensus: AI competition has evolved from a mere computing power arms race into a full-stack war of 'chips + models + access.' Hardware vendors are collectively migrating toward the model side.

Jensen Huang has already deployed the Cosmos world model and robotics development stack, while AMD's previous model offerings were largely limited to text and video. This acquisition fills the gap in understanding the physical world.

Of course, the tension between academic ideals and a public company's quarterly earnings is the toughest nut to crack in every 'lab marries into wealth' case.

The commercialization path for world models remains stuck between demonstrations and imagination. No matter how impressive Marble and Atlas are, they haven't proven they can earn back $8.2 billion.

Conclusion:

Fourteen years ago, ImageNet fed algorithms, GPUs ignited computing power, and deep learning began. Fourteen years later, the recipe remains the same, but the ingredients have changed to spatial data, world models, and Instinct chips.

On the day of the signing, Fei-Fei Li quoted Tennyson: 'Come, my friends, 'tis not too late to seek a newer world.' The $8.2 billion buys not so much the present as a ticket to departure.

Partial References: The Paper: 'AMD's $8.2 Billion Bet on World Models: Competing for the Next Generation of AI Computing Power Access,' CCTV Finance: 'AMD Announces Approximately $8.2 Billion Acquisition of AI Startup World Labs,' Sina Finance: 'Lisa Su vs. Jensen Huang: Chip Sellers Go Rogue, All Competing in AI,' ET Net News Agency: 'AMD's $8.2 Billion Acquisition of World Labs: A Trillion-Dollar Market Cap Becomes a Double-Edged Sword for AI Development'