Is a Two-Year-Old Lab Really Worth $8.2 Billion?
On September 28, AMD agreed to buy World Labs, the AI lab co-founded by Fei-Fei Li, for about $8.2 billion in stock. Li will join AMD as chief scientist.
We are excited to announce that World Labs is joining @AMD. The research and technical breakthroughs we have achieved since our founding in 2024 have given us a clear vision for AI’s potential to solve problems in the spatial and physical world. Accelerating the future of… pic.twitter.com/hIDM20vQsy
The deal is one of the most closely watched in tech this year. It showed how determined AMD is to catch Nvidia, the runaway leader in AI chips, and how much weight the industry now puts on world models, a field still in its early days.
Li led the creation of ImageNet, the image dataset whose 2012 competition, won by AlexNet, set off the deep learning boom, and she is often called the godmother of AI. World Labs has drawn intense attention, and its latest model, Atlas, has earned serious reviews. Even so, is a two-year-old startup that has not disclosed its revenue really worth $8.2 billion?
Patrick Moorhead, an analyst and former AMD executive, summed up the deal on X.
"This is a talent and model-insight buy, not a revenue play. AMD gets the people building the models its future chips will have to run."Patrick MoorheadFounder of Moor Insights & Strategy
The companies' own announcements point the same way. AMD said World Labs would give it deeper insight into how workloads are evolving and help shape its future technology roadmaps. In the same announcement, Li said that advancing the next generation of AI "requires close collaboration across model research, systems and compute."
But what do "roadmap," "talent," and "insight" actually mean here? To find out, it helps to start with how far ahead chips are designed.
What a Model Team Inside a Chip Company Actually Does
Nvidia and AMD now unveil new chips roughly once a year. But according to Damnang, a semiconductor engineer and independent analyst, design work on each chip's test vehicle usually starts two to three years earlier. Choosing a foundry, choosing a process node such as 3nm or 2nm, and confirming that simulation values hold up on that node take at least that long.
So the chip that ships next year was largely shaped by bets made two or three years ago, about models that may not have existed yet.
AMD’s Helios rack-scale system, built with Instinct MI455X GPUs.Source: AMD.
Eugene Cheah, CEO of Featherless AI, an AI inference platform, made the same point: GPUs, like any other chip, are designed years in advance. What matters for a chipmaker, in his view, is a strong team that can give internal feedback: people with experience training models and running inference on those GPUs. (Featherless counts AMD among its investors and gives AMD feedback as an outside partner.)
The catch is who sees what, and when. Cheah described the order in which a chipmaker shares its plans: first with its internal team, then with partners like Featherless under nondisclosure agreements, and only then with the public. For an outside partner, that means "by the time it reaches me, it could already be one plus year in R&D."
A chipmaker with researchers of its own, by contrast, can get that feedback much earlier.
"So at least they have someone in the company who says, 'Hey, I have this idea for the chip for 2027. Will this make sense?'"Eugene CheahCo-founder and CEO of Featherless AI
Damnang described what that feedback looks like in practice. Researchers inside a chip company train models to test questions such as how compute precision trades off against a chip’s throughput, or how to balance model quality against the amount of compute. The answers shape the chip’s architecture, and architecture decisions feed the design roadmap.
Cheah gave an example of the first kind of question. Chips have a property called precision: how many bits they use to represent each number. Before the current AI boom, he said everything ran in BF16. Then the industry moved to FP8, and now there is a lot of discussion about FP4 and FP6. The number in each name is the bit count.
Fewer bits buy speed and give up accuracy, and some proposed formats go so low that, as far as Cheah knows, they don't work for training at all. For a chipmaker that sells into both training and inference, that is a trap.
"If they were to invest too much time into that particular quantization and support for it, while it gives faster performance, potentially for inference, it might actually be useless for their biggest customer, which is training."Eugene CheahCo-founder and CEO of Featherless AI
How the same number is stored in FP16, BF16 and two FP8 formats.Source: NVIDIA.
Decisions like this have a direct impact on a chipmaker's future revenue. Unfortunately for AMD, its rival has made long-term technical bets at a scale no one else has matched: Nvidia.
Nvidia's 20-Year Bet, and How Far AMD Has Caught Up
Nvidia's lead in the AI era is widely traced to a bet it placed two decades ago. In 2006, when GPUs were still mostly for games, Nvidia introduced CUDA, software that lets developers use its chips for general-purpose computing, and set up NVIDIA Research the same year.
For years, the bet looked like a mistake. Putting CUDA into every chip raised costs, but gamers, its main customers, had no use for it. CEO Jensen Huang later recalled that shareholders doubted CUDA and wanted the company to focus on profitability.
Jensen Huang, founder and CEO of Nvidia.Source: Nvidia.
In the end, the bet let Nvidia build relationships with AI researchers before anyone else. Around 2008 at Stanford, Andrew Ng's students introduced him to CUDA and built a GPU server in a dorm room for their first deep learning experiments. Geoffrey Hinton's lab in Toronto and Yann LeCun's lab at New York University were doing the same, and they reached out to Nvidia at roughly the same time. When AlexNet set off the deep learning boom in 2012, the people driving it were already working on Nvidia's chips.
Nvidia’s research organization went on to produce Nvidia's own language models, Megatron in 2019 and Nemotron in 2023, and even a world model, Cosmos, in 2025. More recently, Nvidia agreed to buy Hugging Face, the model platform, extending its reach into model distribution and cementing its place at the top. Research firm IDC estimates that Nvidia will take about 81% of AI data center chip revenue this year. AMD's share, by the same estimate, is around 10%.
Lisa Su, chair and CEO of AMD.Source: AMD.
That gap did not open overnight. While Nvidia was placing that bet, AMD was fighting on a different front. In 2006, the year CUDA launched, it took on debt to buy the graphics chipmaker ATI. When Lisa Su became CEO in 2014, the company was losing money, and her first big bet went to processors to win back ground from Intel. AMD did release ROCm in 2016 as its answer to CUDA, but many still see it as falling short of CUDA’s grip on developers.
AMD is now trying to make up for lost time through acquisitions. In 2022, it bought Xilinx for $48.8 billion, a deal that also brought in AMD’s current AI chief, Vamsi Boppana. In 2024, it paid $665 million for Silo AI, a Finnish lab whose team had already trained large language models at scale on AMD chips.
On models, AMD is catching up in a hurry. In 2025 it released Instella, its own open models, though they remain small: the largest uses fewer than 3 billion active parameters. With World Labs in 2026, AMD gets its first research team with real name recognition in physical AI and world models.
The same pattern holds in the layers this article does not cover in detail: the networking that links thousands of GPUs, and the rack-scale systems that package them as one machine.
| Layer | Nvidia | AMD |
|---|---|---|
| GPU software platform | CUDA (built, 2006) | ROCm (built, 2016) |
| Corporate research lab | NVIDIA Research (built, 2006) | AMD Research (built, around 2008) |
| Open language models | Nemotron (built, 2023) | Instella (built, 2025), Silo AI team (acquired, 2024) |
| World models | Cosmos (built, 2025) | Atlas (World Labs, acquisition pending, 2026) |
| Networking and interconnect | Mellanox (acquired, 2020) | Pensando (acquired, 2022), Enosemi (acquired, 2025) |
| Rack-scale systems | DGX (built, 2016), NVL72 (built, 2024) | Helios (built, 2026), ZT Systems team (acquired, 2025) |
| Model distribution | Hugging Face (acquisition pending, 2026) | None |
* What Nvidia and AMD have built or bought at each layer of the AI stack. Compiled by EO from company announcements.
Line the two companies up layer by layer, and the full picture comes into view. Nvidia started with CUDA, built its own models, and is now buying distribution. AMD is assembling its model stack later and faster. World Labs, AMD's second-largest acquisition after Xilinx, is its latest attempt to close the gap, and the deal raises the question of whether money can buy back time.
Part of the answer is in the table. While Nvidia is years ahead on most layers, in world models the gap is only about 20 months. Perhaps AMD has concluded that world models and physical AI, a field just now opening up, are where it has to make its stand.
"Going forward, AMD will build up its model research and software capabilities and try to move half a step early into new markets where Nvidia's lead has not yet hardened. There, it will develop models, software, and hardware together to give customers a reason to choose the AMD platform, and to set up a new competitive dynamic with Nvidia."DamnangSemiconductor engineer and independent analyst at Damnang Research
The Fight to Seat Researchers Across the Hall
Jensen Huang presents Nvidia's open models at GTC Berlin, October 2026.Source: NVIDIA.
Of course, Nvidia is a formidable opponent in physical AI, too. It released its Cosmos world models in January 2025, and at GTC Berlin this month, it announced Cosmos 3.5, simulation tools for testing robot and self-driving policies before real-world trials, and a reference humanoid for its Isaac robotics platform.
AMD enters that fight in earnest with World Labs. Where Cosmos grew out of video, World Labs’ Atlas, released in September, starts from consistent 3D worlds, generating images and video while reconstructing a space in 3D from just a few photos. In evaluations World Labs published, it beat rival models on both counts, though it is still in early access and the results have not been independently verified.
AMD also gets a foothold in robotics. In July, World Labs acquired SceniX, a startup that turns real environments into digital copies for training and testing robots.
Beyond the rivalry, both bets point to the same conclusion: for chipmakers, world models are becoming a source of demand, not just a research topic. Keith Adams, a seed investor who was the founding engineering lead of Meta's AI research lab, was asked what it signals when a chipmaker buys a world model lab.
"The semiconductor industry now believes world models have graduated from a research area to a driver of real silicon demand."Keith AdamsCo-founder and Managing Partner at Pebblebed
Fei-Fei Li, co-founder and CEO of World Labs.Source: Stanford HAI.
The most valuable asset in this fight may be people. Li's arrival as chief scientist also gives AMD a chance to draw other top researchers in her wake. Boppana, as Moorhead relayed, called the World Labs team a nucleus to attract more talent. If Nvidia's real moat was the relationships with researchers it built early through CUDA, AMD is gaining someone who can stand at the center of such relationships.
Once the deal closes, Li and the World Labs researchers will be working close to the people drawing AMD's chips for two or three years out. For a two-year-old lab, $8.2 billion is still a lot of money. Whether it was worth it may not be clear until the chips they helped shape reach the market.
