AMD Buys Chip Startup Taalas to Build AI Models Directly Into Silicon

AMD announced Thursday it has agreed to acquire Taalas, a Toronto-based chip startup that takes a fundamentally different approach to running artificial intelligence models than the GPUs AMD is best known for selling.

Instead of building general-purpose processors that can run any AI model, Taalas hardwires a specific model directly into the silicon. This means less flexibility, but according to CNBC, the tradeoff is speed and cost: Taalas says its chips can produce output thousands of times faster than a traditional GPU for the specific model they're built to run.

AMD did not disclose a purchase price. Taalas, founded in 2023, has raised $219 million in venture funding to date, according to CNBC.

Why this deal, and why now

AMD's CEO Lisa Su has been telling investors and reporters for months that GPUs will keep driving the bulk of the company's data center growth as cloud giants buy up every advanced AI chip they can get their hands on. That hasn't changed. But Su made clear at a product launch in July that she doesn't see GPUs as the only answer going forward.

"I'm a big believer that there's no one-size-fits-all as it comes to chips," Su said, according to CNBC. She added that AMD still expects GPUs to make up the majority of the AI chip market because of their flexibility to handle new models as they're released.

Taalas' pitch is aimed at a narrower but increasingly important slice of the market: low-latency applications, where how fast an AI model produces its first response actually matters to the end user. Think real-time voice assistants, trading systems, or anything where a half-second delay is the difference between usable and useless.

Taalas CEO Ljubisa Bajic described the company's approach on its website as a "platform for transforming any AI model into custom silicon." Bajic wrote that once the company receives a previously unseen model, it can turn that into working hardware in about two months. The startup's current chip runs a smaller version of Meta's Llama 3.1 model and is manufactured on an older Taiwan Semiconductor Manufacturing Co. process node, using fast SRAM memory built directly onto the chip.

AMD isn't the only one making this bet

This strategy is not new in the chip industry. Nvidia, the current king of AI hardware with a market cap north of $5 trillion, spent $20 billion earlier this year buying assets from Groq, a rival designer of high-performance AI chips built around the same specialized, low-latency philosophy. CNBC reports that deal was Nvidia's largest transaction on record, and it closed roughly seven months before AMD's Taalas announcement.

AMD is also stacking up partnerships beyond this acquisition. In July, the company announced it would integrate AI chips from Cerebras into its systems later this year. AMD says it plans to fold Taalas' chips and underlying technology into its broader product roadmap, pairing them with its central processors and Instinct GPUs in future systems.

The company has also started shipping Helios, its first rack-scale server system built to compete directly with Nvidia's integrated server racks. Customers already receiving Helios include Meta and Microsoft, according to CNBC.

What this signals

The leading AI chipmakers no longer think a single type of processor can carry the entire AI workload. GPUs remain the default because they're adaptable. But adaptability costs something: extra silicon, extra power draw, extra latency. For companies running the same model over and over at massive scale, that flexibility is dead weight.

Taalas and Groq represent a bet that specialization wins for high-volume, repetitive inference tasks, while GPUs keep their grip on training and anything requiring adaptability. Whether that specialization pays off depends on something nobody can fully answer yet: how often AI companies actually swap out their production models. If Meta, OpenAI or Google update their flagship models every few months, a chip hardwired to one specific model version could become outdated fast. Bajic's claim of a two-month turnaround to bake a new model into silicon is the company's answer to that risk, but it hasn't been tested yet at the scale AMD will now be pushing it toward.

No financial terms have been disclosed, and AMD has not said when the deal is expected to close or how Taalas' roughly three-year-old team will be integrated into AMD's existing product groups. Those specifics will matter for anyone trying to gauge whether AMD paid a reasonable price relative to Nvidia's $20 billion Groq deal, or whether investors should expect similar consolidation moves from other chipmakers chasing the same inference bottleneck.