AMD aims at NVIDIA with World Labs and Taalas for next-gen AI supremacy

Highlights

AMD closes World Labs and Taalas deals within 24 hours

Both acquisitions mirror Nvidia's Groq licence and Hugging Face buy

Hardwired inference chips give AMD an answer to custom ASICs

Just as the whole tech world couldn’t take their eyes off NVIDIA’s Jensanity, Anthropic’s record-breaking IPO, AI-fuelled cybersecurity lapses and of course GTA VI, AMD quietly made two announcements that show decidedly the scale of their aspirations and the future they’re willing to build – not just bet on. Within two days, AMD announced completing its acquisition of World Labs, the spatial AI startup founded by Dr Fei-Fei Li – dubbed as The Godmother of AI. The second one involved AMD’s purchase of Taalas, a startup that hardwires AI models into custom silicon. 

It may seem like a couple of routine corporate press releases on the surface, but AMD’s making moves to not just forecast what future AI will look like, it’s actually assembling the brain trust and expertise to build it both at the model layer and physical chips. World Labs to build the incoming models of the near future that will run more efficiently than ever on chips designed by Taalas. AMD wants to "optimize compute platforms and models together" as CEO Lisa Su suggested, and boy do they mean it.

AMD’s trying to match NVIDIA’s moves

The parallels are hard to miss, if you line up the dates. When NVIDIA reportedly struck a $20 billion deal with Groq – a US chipmaker – to licence their AI inference tech, AMD took notice. Just a few months later, AMD approached to acquire Taalas for $2.9 billion. Soon after that, NVIDIA agreed to buy Hugging Face for $12.9 billion, and barely days after AMD announced its intent to buy World Labs for $8.2 billion – the second most expensive acquisition in AMD’s history.

AMD has bought not only one of the foremost AI research startups in World Labs, it has also ensured it owns inference silicon that Taalas will build out. While both deals mirror NVIDIA’s moves in Hugging Face and Groq, there’s an important distinction to point out. NVIDIA has only licensed Groq’s inference tech, they don’t own it – whereas AMD has bought Taalas outright. This simply shows that AMD wants to be in full control of its tech stack, not rent it out of someone else’s yard.

World Labs to build tomorrow's workloads

The significance of the World Labs acquisition by AMD can be a potential gamechanger for both. And its significance can’t be fully appreciated until you realise the unmatched pedigree and calibre of Dr Fei-Fei Li – who now joins AMD’s executive leadership.

As the incoming chief scientist at AMD, Fei-Fei Li built ImageNet in the mid-2000s, which is the stuff of lore in the story of AI progress. It was the cornerstone on which the deep learning revolution of the early 2010s was built at large across research labs and the wider tech industry. It also established the GPU as a foundational piece of AI infrastructure going forward.

AMD knows this better than anyone in the business that semiconductor chips take years to go from drawing board to inside datacentres. Therefore, the surest way to know what future silicon must handle is to have frontier model researchers in-house. And that’s exactly what AMD has said it wants the World Labs team to do, which is cutting-edge foundational research that AMD engineers can better package inside future hardware and software systems.

World Labs’ foray into spatial AI is critical for AMD to start defining how physical AI takes shape and gets deployed inside future robots, autonomous systems and simulation environments. “The universe isn’t made up of words, it’s made of real things,” as Fei-Fei li said. That’s code for AMD will try very hard to ensure NVIDIA doesn’t build the CUDA of physical AI.

Fei-Fei Li will report directly to AMD CEO Lisa Su, ensuring her model research gets a seat at AMD’s top executive table, as they try to architect the next era of AI compute.

Taalas’ silicon bet is far-reaching

Taalas does something most chip designers would consider blasphemy, which is exactly why AMD has bought it. Taalas makes hardware chips with an AI model's weights and dataflow directly into the transistor layer, reducing the dependency on high-bandwidth memory for fast-paced operations. How fast, I hear you ask? At least 28 times faster than NVIDIA’s B200 chips, according to reports. 

While this will help provide unthinkable levels of power and productivity efficiency at a silicon level, there’s also a flipside to this equation. A finished chip will only run the model it was built for. Therefore, new models will need new silicon.

Why is AMD buying Taalas despite knowing this caveat? Simply because its biggest customers are designing their own AI accelerators. Google is already working on a hardwired chip for Gemini called Ironwood, on top of its Tensor TPU chips. Amazon has Trainium and Inferentia. Microsoft has its own Maia AI accelerator chip now, so does Meta in the form of MTIA. It isn’t AMD who’s benefiting from this growing trend, its brands like Broadcom and Marvell (apart from NVIDIA) who make them. And these custom chips, ASICs as they're called in the industry, had started eating into AMD’s pie.

This is where Taalas lets AMD get into that trend. Imagine a frontier lab’s production model getting etched into silicon and racked next to Instinct GPUs in a Helios system. That makes AMD suddenly very formidable. What’s unproven is when Taalas-fuelled custom-model silicon starts appearing into AMD’s product line – don’t think we will see anything until the second-half of 2027.

AMD’s going all in on AI

World Labs and Taalas acquisitions were completed by AMD within a fortnight of posting record-breaking quarterly revenue of $13.4 billions – up 45% than last year for the same period in Q3 – thanks in no small part to a booming datacentre business in its order book. In case you didn’t know, Anthropic, Meta, Microsoft, OpenAI and Oracle all had their first AMD-powered Helios racks come online in this period. Hence, the quick business with World Labs and Taalas, as both teams play a front and centre role into AMD's 2027 roadmap planning from day one.

And that roadmap matters more than ever, because it now carries the whole company to a large extent. AMD CEO Lisa Su expects datacentre to make up more than 80% of AMD's revenue in 2027, which will make it an AI infrastructure firm that also sells PC chips – pretty much like NVIDIA. One can only hope and wonder how World Labs will inform AMD’s AI strategy, just as Taalas embeds it into an unmatched silicon layer. Something tells me we just need to wait until the next Instinct roadmap to get a clue.

Jayesh Shinde

Jayesh Shinde Executive Editor at Digit. Technology journalist since Jan 2008, with stints at Indiatimes.com and PCWorld.in. Enthusiastic dad, reluctant traveler, weekend gamer, LOTR nerd, pseudo bon vivant.

View full profile