Taalas

Taalas develops custom silicon for individual AI models. Its first product, HC1 hard-wired with Llama 3.1 8B, debuted in February 2026 as a beta chatbot demo and inference API. On August 6, 2026, AMD announced a definitive agreement to acquire the Toronto-based company, subject to customary closing conditions and regulatory approvals.[2][5]

Taalas specializes its chips for particular models rather than executing every model on general-purpose hardware. In its February 2026 account, the company described merging storage and computation on one chip at DRAM-level density, removing the need for HBM, advanced packaging, 3D stacking, liquid cooling and high-speed I/O. It claimed that a previously unseen model could be realized in hardware within two months of receipt.[2]

CEO Ljubisa Bajic gave Reuters a concrete manufacturing explanation: Taalas starts with a nearly complete chip of roughly 100 layers, then customizes two metal layers for the model. He said its manufacturer, TSMC, takes about two months to finish fabricating a customized chip. “This hardwiring is partly what gives us the speed,” Bajic said.[4]

Taalas chose Llama 3.1 8B because its small size and open-source availability minimized the logistical effort of implementing it in silicon. Although largely hard-wired, the product supports configurable context-window size and fine-tuning through low-rank adapters, or LoRAs. Bajic said the company released it as a beta despite its distance from frontier capability so developers could explore applications enabled by very low inference latency.[2]

Latent Space’s February 21, 2026 account reported 16,960 tokens per second per user for HC1 running Llama 3.1 8B. Taalas rounded the result to 17K and claimed nearly 10 times faster execution, 20 times lower build cost and 10 times lower power consumption. Its performance caption specified “Input sequence length 1k/1k” and combined an Nvidia H200 baseline, Taalas-measured B200 results, Artificial Analysis results for Groq, SambaNova and Cerebras, and Taalas-lab results for HC1.[1][2]

Those comparisons came with a quality tradeoff: HC1 used a custom 3-bit base format and combined 3-bit and 6-bit parameters, which Taalas acknowledged degraded quality relative to GPU benchmarks. Its second-generation platform, HC2, was designed to adopt standard 4-bit floating-point formats to address those limitations. The February 2026 roadmap called for a mid-sized reasoning model on HC1 to reach its labs in spring, followed by service integration, and a frontier model on HC2 with deployment planned for winter.[2]

The model-specific approach preceded HC1’s debut: Taalas had publicly described an automated model-to-silicon development flow when it emerged from stealth in March 2024. Its founders’ earlier processor work supplies the background to that effort.[3]

Taalas was founded in 2023 by Ljubisa Bajic, Drago Ignjatovic and Lejla Bajic. Ljubisa had founded Tenstorrent in 2016; Drago and Lejla joined soon afterward as early engineering leaders. The company’s March 2024 announcement described the team’s collective experience developing AI processors, GPUs and CPUs across Tenstorrent, AMD and Nvidia.[5][3]

On March 5, 2024, Taalas announced its exit from stealth and $50 million raised across two rounds led by Pierre Lamond and Quiet Capital. Its proposed automated flow covered Transformers, state-space models, diffusion models and mixture-of-experts models. At that point, it planned its first LLM chip tape-out for the third quarter of 2024 and early-customer availability for the first quarter of 2025.[3]

On February 19, 2026, Reuters reported that Taalas had announced $169 million in additional financing, bringing cumulative funding to $219 million, with Quiet Capital, Fidelity and Pierre Lamond among its investors. Separately, Bajic’s February launch account said the first product had been delivered by 24 team members with $30 million spent. He described a core of long-time collaborators, many together for more than twenty years, supported by external partners.[4][2]

Under the agreement announced August 6, 2026, AMD planned to integrate Taalas technology into its accelerator roadmap and develop system-level solutions alongside AMD Instinct GPUs. Bajic, identified in the announcement as Taalas co-founder and CEO, said joining AMD would provide scale, engineering resources and global reach. The announced transaction remained subject to closing conditions and regulatory approvals.[5]