AMD said on Tuesday it has completed its acquisition of Taalas, a Toronto startup that builds specialized chips for AI inference. The cash-and-stock deal is valued at about $2.9 billion, AMD disclosed in a November 4 filing with the US Securities and Exchange Commission.
The Deal
The agreement was signed on August 6, and AMD had expected the deal to close by the end of 2026. The valuation includes up to $1.4 billion in contingent consideration, close to half the headline figure, payable against conditions AMD has not disclosed.
It is AMD’s second AI acquisition to close in two days. The company completed its $8.2 billion all-stock purchase of World Labs on Monday, and Fei-Fei Li joined as executive vice president and chief scientist. Together the deals are worth about $11.1 billion.
What Taalas Built
Taalas designs chips that hard-wire a model’s weights directly into silicon, instead of loading them from memory at run time. Its HC1 demonstrator, built on TSMC’s 6-nanometer process, carries Meta’s Llama 3.1 8B etched into a mask-ROM fabric with on-chip SRAM.
Taalas claims the HC1 serves that model at about 17,000 tokens per second per user, roughly 48 times an Nvidia B200 on its own benchmark, using a tenth of the power. Those figures are the company’s own claims, not independently benchmarked results at production scale.
Our Read
AMD is buying the inference era, not the training era. Training hardware is largely settled, but inference — running finished models cheaply at scale — is where the next money moves, and model-specific silicon is the most radical answer to that problem.
The move mirrors Nvidia’s own inference push: Nvidia structured a roughly $20 billion deal around Groq’s technology last December, and unveiled Rubin HPC for agent-led science this week. AMD’s answer is to own the silicon outright, and pair it with Instinct GPUs and Helios racks.
The trade-off is real. Once a model’s weights are etched, the chip cannot run a different model, and a new model means a new chip. The bet only pays if customers run one model at scale for years — a niche today, though a growing one.
AMD plans to integrate the technology into its accelerator roadmap and build system-level products alongside Instinct GPUs, EPYC CPUs, ROCm software and Helios racks. Taalas is also developing its second-generation HC2 chip, which tops out at 20-billion-parameter models, still far below the largest frontier systems.
Why It Matters
Why it matters: in one week AMD has bought both ends of the AI stack — world models from World Labs and inference silicon from Taalas. That is a full-stack challenge to Nvidia, which opened four Cosmos world models of its own earlier this month.
What to watch is integration: whether Taalas silicon shows up in Helios racks, and whether AMD names customers running single models at scale. NVIDIA’s earnings tonight may also say how the inference race looks from the other side of the market.
Background
Background: Taalas was founded in 2023 by Ljubisa Bajic, a former architect at AMD and Nvidia and a Tenstorrent co-founder, with engineers Drago Ignjatovic and Lejla Bajic. The startup raised $219 million in venture funding, including $169 million in February 2026.
Source: AMD’s acquisition announcement and AMD’s original agreement release.


0 Comments