Summary: On August 6, AMD announced that it had signed a definitive agreement to acquire Taalas, a Toronto startup founded in 2023 that etches individual AI model weights directly into silicon. Its HC1 test chip served Llama 3.1 8B at 16,960 tokens per second, though larger model changes require a new chip spin. Terms were undisclosed, and the deal awaited regulatory approval.
On August 6, AMD said it had signed a definitive agreement to buy Taalas, a Toronto startup founded in 2023. Taalas takes one of the more radical routes available in the industry: rather than sell general-purpose GPUs, it etches the weights of one specific model straight into silicon, accepting a rigid design in exchange for speed and cost by building the hardware around the model. Terms were not disclosed, and regulators still have to clear the deal; The Register reports an expected close in Q4. Taalas has raised $219 million in venture funding, according to CNBC.
Viewed from AMD’s side, the purchase adds one more piece to a full-stack build-out that already counts Silo AI ($665 million) and ZT Systems ($4.9 billion). Nor is the timing accidental: it lands months after Nvidia’s roughly $20 billion Groq licensing deal in December 2025. The sequence has a logic to it. Silo AI contributed software talent, ZT Systems contributed rack design, a July partnership pulled Cerebras into the orbit of AMD systems, and Taalas now supplies the most specialized silicon of the lot. Inference hardware built for one job is turning into the freshest battleground in the chip giants’ arms race.
The Deal, in Short
- What is being bought: Taalas, Toronto, founded 2023; AMD calls it “a pioneer in specialized AI inference silicon.”
- No price was given; closing depends on customary conditions and regulatory approvals.
- In the words of Vamsi Boppana, SVP of AMD’s AI Group, customers should get “the flexibility to deploy the right compute solutions for every AI workload.”
- Taalas co-founder and CEO Ljubisa Bajic: “We founded Taalas to rethink AI inference from the ground up by building the hardware around the model.”
How MSICs Work: Printing a Model Into a Chip
Every Taalas part is a model-specific integrated circuit, or MSIC, and each one is split into two regions. A mask-ROM recall fabric carries the weights etched into the hardware; an SRAM recall fabric handles KV caches and fine-tuning adapters. What pushes the design to its extreme is the mask-ROM half: weights are written once at fabrication and cannot be updated, which is why Taalas treats model re-spins as a product cadence rather than a bug. HC1, the test chip, arrived in February, fabricated on TSMC’s 6nm process. Serving Llama 3.1 8B, it reached 16,960 tokens per second — at the time 48x faster than Nvidia GPUs and 8.5x faster than Cerebras accelerators. HC2 follows this summer, targeting 20 billion parameters on a single chip; a trillion-parameter model would need about 50 accelerators chained through pipeline parallelism.
What Speed Costs
What the buyer gives up is spelled out plainly: send a model past LoRA-style adapters and the chip has to be re-spun, although no more than two metal layers need to change. Bajic’s line is that a new model can be “realized in hardware in only two months.” So MSICs belong where weights hold still and throughput is extreme; fast-moving frontier training does not fit. AMD’s plan for stitching this together is hybrid: prompt processing stays on GPUs, token generation moves to Taalas silicon, and the result ships alongside Helios rack-scale systems, EPYC CPUs, and ROCm software. On cost, The Register reports that etching weights into silicon runs about one percent of what training a frontier model costs, and notes that serving the same model on Nvidia’s LPX systems would take a few dozen GPUs plus at least 2,000 Groq LPUs.
Reading the Deal for the Inference Market
Lisa Su put the philosophy plainly: “I’m a big believer that there’s no one-size-fits-all as it comes to chips.” That single sentence captures the industry’s direction of travel: training versus inference, prompt processing versus generation, general versus specialized, each pair splitting apart into its own hardware class. Inference, the press release notes, ranks among the fastest-growing AI segments, and its workloads have grown specialized enough that general-purpose architectures leave compute and memory on the table. AMD also stressed the Canadian angle, promising to retain and grow Taalas’ Toronto engineering base.
Developers who buy inference capacity will feel little change in the short term, but the direction is clear. Fixed-model, high-throughput workloads — recommendations, moderation, agents that run the same pipeline all day — get dedicated silicon options first, and general-purpose GPU pricing gains another source of downward pressure. It is worth recalling, too, how the Groq precedent played out: Nvidia paid for licenses and people rather than a product line, and AMD may look at Taalas the same way — a portfolio of designs plus a team that knows how to hard-wire models, pointed at whichever inference niches grow large enough to justify fixed-function silicon. Iteration speed remains the open question. Two months from trained weights to working hardware is quick by custom-silicon standards, yet beside a GPU that runs a new checkpoint the day it ships, it still amounts to an eternity.