I came across a company called Taalas, an AI chip company with a team of just over twenty people, yet they have already raised over $200 million in funding.

Their first product, HC1, took an extreme approach: using Mask ROM technology to directly embed the weights of Llama3.18B into the silicon's metal layer. The compute units and model parameters are on the same silicon, almost abandoning programmability to achieve extreme optimization in throughput, latency, and energy efficiency.
Currently, they have embedded Llama3.18B; I tried it out, and the responses are unreliable and quite unstable.
But the remarkable thing is how fast it is—it's incredibly counterintuitive in terms of user experience. It can crush Groq in performance, with a single chip capable of 17,000 tokens per second output, generating tens of thousands of words in the blink of an eye—this surpasses even database query capabilities.
If in the future, large models are truly limited to just a few top players iterating, with model structures gradually stabilizing and weight update frequencies slowing down, then designing a dedicated chip for a specific model might not be crazy at all.
Right now, we assume models will continue to change rapidly, so computing power needs to be versatile.
But what if models start to standardize?
Embedding weights directly into silicon and replacing throughput with highly specialized architecture to cut costs.
It seems the landscape of models is beginning to centralize. Once the structure of top models stabilizes, it’s definitely worth developing dedicated chips tailored to their architecture.
The explosive potential of this approach could be very significant.
If that’s the case, a counterintuitive question arises: will GPU-based solutions really be the ultimate endgame forever?