Taalas HC1 AI chip outed, runs Meta’s Llama 3.1 with 10x less power

AI hardware startup Taalas has unveiled the HC1, a specialized chip that hardcodes Meta’s Llama 3.1 model into silicon. This promises massive speed and efficiency gains over general-purpose GPUs.

The company claims throughput of around 17,000 tokens per second, putting it roughly 40x faster than Nvidia’s flagship B200. This also includes other data center accelerators in internal comparisons.

Beyond raw performance, the HC1 is designed to be cheaper and leaner. Taalas reports build costs up to 20x lower and power consumption approximately 10x less than leading alternatives.
Of course, this is aided by a design that tightly merges memory and compute on a single chip. By removing the software–hardware abstraction layers, the LLM effectively ‘lives’ inside the chip and can respond almost in real-time.

However, the approach has clear trade-offs. Because the HC1 is optimized for Llama 3.1, it cannot simply swap in its competitors without designing a new chip. For reference, this consists of the likes of DeepSeek or Qwen.

Even so, it points to a future where high‑volume AI models ship as dedicated hardware. Its custom silicon-built design around a single model architecture offers best‑in‑class latency and efficiency.

Taalas has published additional technical details in a release page for early adopters and researchers. Interested readers can check it out here.