News › ai-infrastructure
By Zayden R., August 9, 2026
AMD has acquired AI chip startup Taalas, aiming to improve inference performance by embedding model weights directly into silicon. This approach could significantly disrupt Nvidia's hold on AI hardware by offering faster and more efficient inference solutions.
AMD has made a strategic move in the AI hardware sector by acquiring Taalas, an AI chip startup known for embedding model weights directly into silicon. This acquisition, announced at the market close on Thursday, is seen as AMD's latest effort to challenge Nvidia's dominance in AI inference performance.
Taalas, founded in 2023 and based in Toronto, has developed a unique approach to inference that sets it apart from conventional GPUs and other AI accelerators. Instead of relying on high-bandwidth memory (HBM) to store model weights, Taalas' technology etches these weights directly into the silicon, creating what are essentially model-specific integrated circuits (MSICs).
This technique is not just theoretical. Earlier this year, Taalas revealed its first test chip, the HC1, fabricated on TSMC's 6nm process technology. In initial benchmarks, the chip demonstrated impressive performance, processing Meta's Llama 3.1 8B model at a rate of 16,960 tokens per second. Compared to Nvidia's GPUs and Cerebras' accelerators, this represents a 48x and 8.5x speed increase, respectively.
The technical details of Taalas' chips remain somewhat under wraps, but we know they consist of two main regions: the mask-ROM recall fabric for etched model weights and the SRAM recall fabric for storing KV caches and fine-tuning adapters. Taalas plans to release a second-generation chip, the HC2, this summer, which will increase the parameter count to 20 billion per chip. This could potentially allow for pipeline parallelism across multiple accelerators, supporting trillion-parameter models with just 50 accelerators.
For engineers and sysadmins, this acquisition could mean more efficient and cost-effective AI inference capabilities, making it a development worth watching closely. AMD's integration of Taalas' technology could lead to significant shifts in AI infrastructure and performance optimization strategies.
The Linux Camp teaches these topics as hands-on labs on real virtual machines, verified as you type.