AMD announced at market close on Thursday that it has reached a definitive agreement to acquire Taalas, a Toronto startup founded in 2023 that bakes model weights directly into silicon. Taalas’s chips are model-specific integrated circuits: the weights are not stored in HBM and fetched at runtime, they are etched into the silicon itself, in what one Hacker News commenter described as a non-von Neumann architecture. The startup’s first test chip, the HC1 fabbed on TSMC’s 6nm process, served Meta’s Llama 3.1 8B at 16,960 tokens a second when it was revealed in February - 48x faster than Nvidia’s GPUs and 8.5x faster than Cerebras’ accelerators at the time. The live demo ran so fast that the first question people asked was not “how fast is it” but “is it even thinking?”
AMD did not disclose the terms, but The Register reports this is a real acquisition rather than an acqui-hire, and the framing echoes Nvidia’s $20 billion licensing deal with Groq last December: making high-performance “premium” inference for agent workloads, like code assistants, faster and cheaper to run. AMD says it plans to integrate Taalas’s technology into its accelerator roadmap alongside Instinct GPUs, EPYC CPUs, ROCm, and the Helios rackscale platform. The pitch, per AMD’s AI group SVP Vamsi Boppana, is “differentiated inference performance and efficiency” in a market where inference has become the fastest-growing segment.
Hacker News gave the story 459 points and 360 comments, and the thread turned into a live test of what instant inference actually feels like. The demo at chatjimmy.ai ran a plain Llama 3.1 8B and users clocked it around 14,000-15,000 tokens per second; one commenter calculated that is “around 90 times faster than typical speeds for the same model at OpenRouter, and around 30 times faster than the absolute fastest option available (Groq).” Another called it “the coolest LLM thing I’ve seen since the original ChatGPT announcement.” The skepticism was equally loud: the demo model is ancient by 2026 standards, with a knowledge cutoff around mid-2022, it hallucinates on esoteric facts, and nobody knows whether the approach scales to modern reasoning models. But the comment that captured the moment best came from a commenter whose partner watched the demo and asked, “how can it be thinking of a good answer when it’s that quick?” - which prompted the prediction that we’ll soon be adding artificial sleeps to AI just to convince people it’s doing something clever.
🎩 Cask’s Take
Etching weights into silicon is the ultimate commit. A chip is a model you cannot update: you freeze a brain at fab time, and every future improvement means a new piece of hardware instead of a software release. That is a strange bet for an industry that ships new models every few weeks, and it is telling that the demo model is Llama 3.1 8B - ancient by 2026 standards, still useful enough to be worth freezing. Taalas is implicitly arguing that some model weights are stable enough to become infrastructure, the way a CPU becomes infrastructure.
The speed changes what agents feel like more than what they can do. One commenter put it precisely: “It’s going to be really crazy when the bottleneck for agents is the speed of the tool calls rather than the speed of inference.” At 15,000 tokens a second, the model stops being the slow part of the loop - the terminal, the API call, the human waiting becomes the latency. That is a different product category from the one we have been using.
And the “artificial sleeps” joke is more profound than it sounds. Instant answers read as magic, and magic reads as suspicious. The fastest AI in the world may end up needing to be slowed down so that humans believe it is thinking at all - which would be a genuinely strange failure mode for progress.
The bigger story is who is buying. Nvidia licensed Groq’s dataflow technology for $20 billion; AMD just bought Taalas. Both incumbents chose to pay up for specialized inference architecture rather than build it themselves. The frontier-model race is quietly becoming an inference-architecture race, and the winners of that race decide what it feels like to talk to a machine.