---
title: This Chip Startup Builds AI Models Directly Into Its Silicon
description: Toronto startup Taalas hard-wired Llama 3.1 8B into its HC1 chip and claims 17,000 tokens a second per user at a tenth of the power of a GPU.
author: Darie Nani (Editor-in-Chief)
date: 2026-07-28T11:55:33.914Z
updated: 2026-07-28T11:58:22.435Z
canonical: https://www.sovereignmagazine.com/article/taalas-hc1-ai-chip
image: https://cdn.nanimediahouse.com/taalas-hc1-card.webp
categories: Artificial Intelligence, Startups
content_type: News
region: Global
publication: Sovereign Magazine
about:
  - type: Organization
    name: Taalas
    description: Taalas is a Toronto AI chip startup founded in 2023 by Ljubisa Bajic, Lejla Bajic and Drago Ignjatovic. It makes processors with AI models hard-wired into the silicon rather than loaded from memory, trading the flexibility of a GPU for speed and lower power. Its first chip, the HC1, carries Meta's Llama 3.1 8B in its circuits, and the company has raised $219 million.
    url: https://taalas.com
    foundingDate: 2023-09-01T00:00:00.000Z
    industry: Semiconductors
    sameAs:
      - https://www.linkedin.com/company/taalas
---

Taalas, a Toronto chip startup, has fixed Meta's Llama 3.1 8B into the transistors of its first processor. The HC1 cannot be reprogrammed to run any other base model. The weights are part of the circuit rather than something the chip fetches from memory while it runs.

Underneath that is a bet against the assumption the AI hardware industry has worked on for a decade. Flexible silicon exists because models were changing too fast to commit to any one of them. Taalas is wagering that the assumption has expired: that a model like Llama 3.1 8B is now good enough for a large share of real work, and that once it is, the speed and power saved by freezing it beat the freedom to swap it later.

The company has raised $219 million, most of it in a round announced in February 2026, after coming out of stealth in March 2024 with $50 million. About 25 people were on staff in February, hired from AMD, Apple, Google, Nvidia and Tenstorrent.

## Only the Top Two Metal Layers of the Die Change Per Model

The HC1 is made by [TSMC](https://www.sovereignmagazine.com/article/tsmc-s-ai-chip-leadership-drives-taiwan-s-semiconductor-dominance) on a 6nm process, an 815mm² die carrying 53 billion transistors. Of its roughly 100 layers, only the top two metal layers are customized for a given model, which is what lets Taalas put a new model into silicon in about two months against roughly six for a full custom chip.

Chief executive Ljubisa Bajic founded Tenstorrent in 2016 and ran it until December 2022, when he and Jim Keller swapped roles. He left in March 2023 and started Taalas that September with Lejla Bajic, now chief operating officer, and Drago Ignjatovic, Tenstorrent's former vice president of hardware. Bajic will not say how the weights are encoded, calling it "a clever trick that we don't want to broadcast".

> "Nobody went into this [flexibility-performance] corner because everybody felt AI was changing so rapidly that it would be a massively risky thing to do… which it is, to some extent."
> — Ljubisa Bajic

"Essentially, we looked for ways to make painful tradeoffs in flexibility for the sake of economics and speed," Bajic says of the design. Taalas wanted to see "what's hiding in that corner, what we could get if we explored it, and you can get a lot".

## A Tenth of the Power, If the Numbers Hold

[The expensive part of AI is now running the models](https://www.sovereignmagazine.com/article/femtoai-5x-growth-spu-sparsity-2026). Gartner expects inference to take 55% of AI-optimized infrastructure spending in 2026, rising past 65% by 2029, and the International Energy Agency projects data-center electricity roughly doubling to about 945 TWh by 2030, with AI accounting for most of the increase. A GPU can run any model, and that flexibility costs power on every query. Taalas removes it so that running one model costs a fraction as much.

Taalas says the HC1 reaches up to 17,000 tokens a second per user on Llama 3.1 8B, which it puts at 73 times an Nvidia H200 at about a tenth of the power. Those are the company's own measurements. Karl Freund of Cambrian AI, writing in Forbes, measured 14,357 tokens a second in his own test and reports HC1 racks drawing 12 to 15kW where [GPU racks draw 120 to 600kW](https://www.sovereignmagazine.com/article/the-swiss-startup-cooling-down-ai-s-21-billion-overheating-problem). Freund calls the HC1 a prototype rather than a product generating revenue, and no paying customers have been confirmed. API access is granted on request, and there is a public demo chatbot at [chatjimmy.ai](https://chatjimmy.ai).

LoRA adapters load into on-chip SRAM and the context window is configurable, so a customer can adapt the chip's behavior within limits, but the base model cannot be swapped. The HC1 runs at roughly 3-bit quantization with quality trade-offs the company acknowledges, and HC2 moves to standard 4-bit MXFP4.

## The Model May Be Old Before the Data Center Is Live

Whether the bet is early, on time or wrong turns on how long a frozen model stays useful. Models are still updated several times a year, and kitting out a data center with thousands of Llama-frozen chips takes time, so by the time the fleet is running, Llama is several versions ahead and none of those chips can be repurposed. The German technology publication Heise and others put the question directly: does model-specific silicon work economically when models turn over every few months?

A lot of high-volume work does not need the frontier model, which is the half of the bet that looks strongest. If Llama 3.1 8B answers the query well enough, running it in a rack drawing 12 to 15kW instead of 120 to 600kW is a rational deployment. Cheap, fast inference on a known model is a product plenty of buyers can price.

The two metal layers are Taalas's answer to the other half: a new model reaches silicon in about two months, so a fleet can be refreshed rather than written off. A technical write-up on zach.be reports that the company's economics assume a one-year life for each Taalas data center.

A mid-sized reasoning model is due in spring 2026, HC2 with a 20 billion parameter model in the summer, and a terabyte-class frontier model across multiple cards in the winter. The original plan put the first chip with customers in early 2025, so Taalas is running about a year behind it.

## Rivals Are Betting on Architectures, Not on One Model

Etched closed a $300 million Series C led by Sequoia on 23 July 2026 at a $10.3 billion valuation, for a chip fixed to the transformer architecture but able to run any transformer model, a looser bet than Taalas has taken. Nvidia agreed a licensing deal worth roughly $20 billion for Groq's inference technology in December 2025, and Cerebras went public in May 2026. [OpenAI is designing its own AI chips with Broadcom](https://www.sovereignmagazine.com/article/the-ai-chip-wars-heat-up-how-openai-and-broadcom-are-reshaping-manufacturing-s-future), another route around the general-purpose GPU.

## FAQ

**Q: What is Taalas HC1?**
The HC1 is Taalas's first chip, an 815mm² die made by TSMC on a 6nm process with 53 billion transistors. Meta's Llama 3.1 8B is physically encoded in the circuit rather than loaded from memory, so the chip runs that one model and nothing else. Taalas claims up to 17,000 tokens a second per user from it, and the analyst Karl Freund measured 14,357 in an independent test.

**Q: How does Taalas work?**
Taalas etches a model's weights into the transistors of a chip, trading the flexibility of a GPU for speed and lower power. Only the top two metal layers of the roughly 100-layer die are specific to the model, so a new model can be put into silicon in about two months rather than the six a full custom chip takes. LoRA adapters load into on-chip memory and the context window is adjustable, but the base model is permanent.

**Q: Who is the CEO of Taalas?**
Ljubisa Bajic, who founded Tenstorrent in 2016 and led it until December 2022, when he and Jim Keller exchanged roles. He left Tenstorrent in March 2023 and started Taalas in September of that year with Lejla Bajic and Drago Ignjatovic, both also from Tenstorrent.