The agreement was announced on 6 August, but the deal hasn't closed, and no price has been disclosed. Taalas is betting on an extreme idea: the model shouldn't run on a chip - it should be the chip.
- A definitive agreement has been signed, but closing awaits regulatory approvals. No price, no timeline, no headcount has been disclosed.
- Taalas was founded in 2023 in Toronto. It claims models built this way are 1000 times more efficient than software ones - with no published methodology.
- Telling detail: AMD is switching its unit of measure - it now counts its edge in output tokens per dollar, not raw compute.
Taalas's idea is absurdly simple, and that's what makes it unsettling: either it flips the industry's whole cost equation, or it hangs around as a pretty thesis on the calendar. For now, I don't know which.
Their thesis, put simply: today the model is a program that runs on a general-purpose chip. Taalas wants the model to be the chip. The gain is clear - moving weights between memory and processor disappears, and that's exactly what burns the power. The loss is just as clear: a chip can't be rewritten, and models change every few months.
There's one number, though, that says why AMD is in a hurry - and it's their own, from a separate announcement two weeks earlier. In November 2025 they talked about a market opportunity above one trillion dollars by 2030. In July 2026 the same company raised that figure to around two trillion for the same year. Their own forecast doubled in nine months. And that same announcement says they now measure their edge in output tokens per dollar spent - not raw compute.
The weight in AI is shifting from training to inference, and once you're measuring cost per answer, specialized hardware always beats general-purpose. Whether Taalas's claim holds up on flexibility - the market will show that, not the press release. Today there's simply no primary source for that answer.