AMD buys Talus: burning AI models into silicon, 17,000 tokens per second
AMD has acquired Talus, a startup that builds chips customised for specific AI models and hits 17,000 tokens per second. It marks AI inference moving toward non-programmable, model-specific silicon.
The argument · tap a timestamp to hear it
Talus's founding idea: burn the model into the chip
Talus was founded by Ljubisa Bajic, the founder of Tenstorrent, and its core concept is burning an AI model directly into the chip, producing non-programmable custom silicon. The idea drew scepticism at launch, because models iterate fast and custom chips can go obsolete quickly. But Talus argues that even if a chip only works for one specific model, its 100x speedup is enough to offset the cost of replacing it.
— HostHardcore One: 53 billion transistors, 17,000 tokens/s claimed
Talus's first chip, Hardcore One, is built on TSMC's 6nm process, is close to reticle size, packs 53 billion transistors, and draws at least 300 watts. On Llama 3.1 8B it claims 17,000 tokens per second, far above Nvidia H200's 230, Groq's 600 and Cerebras's 2,000. That figure was verified in a demo, where it actually reached 14,200 tokens/s.
— HostAMD keeps Talus as an independent team under Boppana
AMD has signed an agreement to acquire Talus. The team will report to AMD AI head Vamsi Boppana and will keep operating as an independent team. Ljubisa previously worked at AMD and knows Jim Keller, so the acquisition can be seen as a homecoming. The deal value was not disclosed and is expected to close by the end of the year.
— HostTalus slots into AMD's mix-of-chips AI strategy
AMD CEO Lisa Su stresses that AI deployment is a mix of many kinds of chips. Talus's chip suits the small-model calls inside agentic AI, complementing AMD's Instinct GPUs, EPYC CPUs and the rest. AMD already partners with Cerebras, but Talus's model-specific nature offers another option.
— HostSally Ward Foxton: all-SRAM, best paired with a GPU
Sally Ward Foxton argues Talus resembles Groq in using an all-SRAM architecture, needs multiple chips to support large models, but is well suited to handling the decode portion alongside a GPU. She says the acquisition is not surprising, because Talus's software stack requirements are low and the technology is distinctive.
— Sally Ward FoxtonThe price guess: billions, because Nvidia reset the bar
Sally guesses the acquisition price could be in the billions of dollars, because Nvidia's $2 billion deal for Groq pushed up valuations across chip startups. Talus had previously raised only tens of millions of dollars, so a multi-billion-dollar valuation is reasonable.
— Sally Ward FoxtonIn their own words · checked verbatim
The point about Talus, when they came out of stealth, was to say, "Hey, we've got an idea which is going to break the internet, break how fast you can run AI tokens. And it's going to be this concept of baking the AI model into the chip."
Host2:05
It says, "Chat Jimmy, how can I help you today?" And let's ask it a question. Say, "What is in the bubbles in pop?" Generated in 33 milliseconds, the answer came at a speed of 14,200 tokens per second.
Host6:11
But, is the speed up from 200 tokens per second to 14,200 in this demo, or based on their numbers, 17,000 tokens per second, worth the effort for a company that has an established workflow, who is going to deploy to millions of users? Yeah. You would buy a new infrastructure set every 6 months if your speed up goes from 200 to 17,000.
Host7:11
Thing is, if a tool chain is fixed and it knows what models are being called, this is where Talus can really help. This is where on their first generation chip, if they're only doing 17,000 tokens per second, who knows what they're going to do with second generation chips, third generation chips.
Host10:16
I mean, there's no doubt in my mind that it works. They also, because they only run one model at a time per chip, they don't really need much of a software stack, which is the other sticking point for startups usually in this kind of scenario.
Sally Ward Foxton13:18
Figures
| Transistor count of the Talus Hardcore One chip | 53 billion | 5:09 |
| Tokens per second claimed by Talus | 17,000 | 5:09 |
| Tokens per second actually reached in the demo | 14,200 | 6:11 |
| Tokens per second of Nvidia H200 | about 230 | 5:09 |
| Tokens per second of Groq | 600 | 5:09 |
| Tokens per second of Cerebras | 2,000 | 5:09 |
| Talus chip power draw | at least 300 watts | 5:09 |
| Talus funding raised | tens of millions of dollars | 14:18 |
Glossary
- SRAM
- A high-speed memory used for on-chip cache; fast but small in capacity.
- reticle size
- The maximum size of the mask in lithography; a chip near this size has a very large die area.
- agentic AI
- AI systems that can decide and act autonomously, usually involving multiple model calls.
How to listen
AI chip founders, investors, data centre architects, and technical decision-makers focused on inference performance and cost optimisation.
The first 3 minutes of AI accelerator background can be skipped; go straight to the Talus introduction.