The chip is called HC1 (Hardcore 1) by a Toronto startup called Taalas. It doesn’t run an AI model. It literally IS the model — Llama 3.1, hardwired into the metal layers of silicon itself. No HBM. No exotic cooling. $30M to build. Two months from model weights to working chip.
The numbers are hard to believe until you line them up:
Nvidia H200 — 230 tokens/sec
Groq — 600
Cerebras (fastest until last week) — 1,900
Taalas HC1 — 17,000
Not with a bigger chip. Not with more power. With a completely different idea about what a chip even is.
That’s not an incremental improvement. That’s a different category.
Most people are reacting to the speed numbers. I get it. They’re insane.
But here’s what I keep thinking about —
This is genuinely great news for verification engineers. And I don’t think enough people are saying that.
We’ve spent careers being called the bottleneck. The team that holds up tapeout. The necessary evil between RTL and silicon. I’ve sat in program reviews where people forget that verification isn’t the risk — it’s the function that stops risk from reaching silicon.
A 2-month design cycle changes that conversation overnight.
You can’t throw more sims at it. You can’t just hire 10 more people in January. When the timeline is that compressed, verification isn’t the friction anymore — it’s the whole game. The thing that either makes or breaks the chip.
And yes, the verification problem for a hardwired AI chip is genuinely different. The weights ARE the logic. There’s no instruction path to cover. Coverage means something completely new. Formal methods have to be rethought for fixed-function silicon.
That’s not a threat. That’s an open field.
Whoever figures out how to verify AI-native ASICs — the methodology, the tooling, the coverage model — writes the playbook for the next decade of the industry.
I’d rather be the person writing that playbook than waiting to see who does.
Are any of you already thinking about this? Would love to know what teams working on AI silicon are actually doing for verification today.