Skip to main content

Taalas

Category: AI Chips / Semiconductors

A platform for rapidly turning any AI model into custom silicon, embedding entire deep learning models directly into specialized chips that achieve 1000x efficiency improvements over traditional GPU-based approaches. Taalas was founded in 2023. The company is led by Ljubisa Bajic. Based in Toronto, Canada. Team size: 24-25. Total funding raised: $219.0M. Latest round: Series B ($169.0M, Feb 2026). Key investors include ["Quiet Capital","Pierre Lamond","Fidelity"].

Founded
2023
Headquarters
Toronto, Canada
Team size
24-25
Total funding
$219.0M

Value proposition

Delivers 1000x improvement in AI efficiency by hard-wiring entire AI models directly into silicon using mask ROM and SRAM, eliminating the memory-compute bottleneck. Single chip can outperform a small GPU data center while consuming 10x less power, enabling faster inference (17K+ tokens/sec on Llama 3.1 8B) at dramatically lower cost.

Products and solutions

["HC1 Technology Demonstrator Chip (Llama 3.1 8B hard-wired inference)","Taalas Foundry Platform (automated AI-model-to-silicon conversion)","Hardcore Model Inference Service (API access)","PCI-Express Cards for Deployable Inference","HC2 Chip (upcoming - for frontier-scale models)"]

Unique value

Pioneers 'Hard Coded Inference' technology that embeds entire AI models (weights and architecture) directly into silicon using mask ROM and on-chip SRAM, rather than simulating models on general-purpose GPUs. Creates custom model-specific chips in just 2 months via foundry-optimized workflow with TSMC, assembling 100-layer chips with final customization on 2 metal layers.

Target customer

Tech companies involved in AI and deep learning, enterprises requiring high-performance AI inference at scale, organizations with long-lived AI models (1+ year production use cases), companies seeking alternatives to GPU-centric infrastructure

Industries served

["AI Infrastructure & Semiconductors","Large Language Models (LLMs)","Machine Learning & Deep Learning","Data Center & Cloud Computing","Natural Language Processing","Computer Vision & Machine Vision","Transformer-based AI Applications"]

Technology advantage

Achieves 17K+ tokens/sec per user on Llama 3.1 8B (nearly 10x faster than current state-of-the-art), 73x more tokens than Nvidia H200, using only one-tenth the power. First chip (53B transistors, 815mm², TSMC 6nm, 200W) developed in just 2.5 years with only 24 employees and $30M spent. Supports fine-tuning of hard-wired models and enables running powerful AI locally on consumer devices. Eliminates redundant components found in general-purpose GPUs, replacing unused RAM with efficiency-boosting transistors.

How they differentiate

Model-specific silicon approach that embeds portions of individual AI models directly into chips via final metal layer customization on ~100-layer chips, versus competitors' general-purpose AI accelerators. Delivers 2-month production cycle via TSMC vs 6-month industry standard.

Main competitors

["Groq","Cerebras Systems","Etched"]

Key partnerships

["TSMC (Taiwan Semiconductor Manufacturing Co.) - exclusive manufacturing partner for foundry-optimized workflow","Quiet Capital (lead investor)","Pierre Lamond (independent investor, Eclipse Ventures advisor)","Fidelity (investor)"]

Notable customers

["Tech companies involved in AI and deep learning","Enterprises with long-lived AI model deployments","Organizations seeking GPU alternatives"]

Major milestones

["Founded in August 2023 by Tenstorrent founder Ljubisa Bajic","Exited stealth mode with $50M funding over two rounds (Mar 2024)","First HC1 chip developed in 2.5 years with 24 employees","Raised $169M Series B funding (Feb 19, 2026)","Filed 14 patents covering hard-coded inference technology","Achieved working chips for less complex AI models with plans for GPT 5.2-scale by year-end 2026"]

Growth metrics

Developed HC1 chip with 24 employees and $30M spent; achieved 17K tokens/sec on Llama 3.1 8B (nearly 10x faster than current state-of-the-art); 73x more tokens than Nvidia H200 with one-tenth power consumption

Market positioning

Ultra-specialized AI inference chips targeting long-lived production models (1+ year deployments) with claims of 1000x efficiency improvement and 10x cost reduction versus GPU-based approaches. Single chip purportedly outperforms small GPU data centers.

Geographic focus

North America (Toronto, Canada headquarters), with global manufacturing partnership via TSMC in Taiwan. Competes directly with US-based AI chip startups like Groq and Cerebras.

Patents and IP

14 patents filed under CEO Ljubisa Bajic covering Taalas' unique hard-coded inference technology and direct-to-silicon foundry methodology

About Ljubisa Bajic

Founder of Tenstorrent (2016), former Director of IC Design and Architecture at AMD with over a decade of experience developing GPUs, CPUs, and hybrid CPU-GPU chip designs. Previously worked as senior architect at Nvidia for one year, and spent time at Teralogic and Oak Technology designing video encoders post-Dot Com Boom. Veteran semiconductor industry expert with extensive experience in VLSI design, accelerator architecture, and AI hardware. Founded Taalas in August 2023 after leaving Tenstorrent in March 2023.

Official website: