Today, most of my Chinese lines on X were discussing OpenUSD developed by Stripe and Visa, while there were more English lines discussing @ Etched. This thing is a bit interesting, it has the same feeling as when I first saw CBRS. One sentence introduction: Etched is the craziest and most extreme bet in the AI chip industry. Without making a better GPU, simply solder the Transformer into the chip - it doesn't support CNN, graphics, or the useless general-purpose computing in CUDA. Just like a GPU is like a Swiss Army knife, it can cut anything; Etched is a sushi knife that only cuts fish, but it is the best at cutting fish in the world. Of course, there is a big risk involved here. If in the future everyone switches from eating fish to eating meat, this project will go straight to GG! Back then, we also went through the process of Bitcoin mining - CPU mining, GPU mining, and then ASIC mining. The logic of ASIC mining machines is that since the whole world is computing SHA-256 as a hash function, there is no need to keep the pile of general-purpose computing circuits in GPUs and directly solder SHA-256 into silicon wafers, which can greatly improve efficiency. Etched is doing the same thing now, just replacing "SHA-256" with "Transformer". But the key difference, and also the most fatal aspect of this bet, is that the algorithm for BTC mining is fixed and will not change, SHA-256 is soldered dead, and the Bitcoin protocol does not change at will. But AI is not - Transformer is just an architecture that has won in the past few years, not necessarily the end result. So Etched's entire company is essentially betting on one thing: Transformers will not be replaced by more advanced architectures. If in the future there is a new architecture like Mamba, RWKV, SSM, or something we haven't seen yet, and Transformer is taken down, then Etched, this chip soldered to a silicon wafer, will be no different from the mining machines that were eliminated by algorithm upgrades in the past - it will directly become a pile of useless silicon. At least GPU can still run other models, ASIC has no way out - bet on the right order of magnitude advantage, bet wrong and instantly reset to zero, there is no gray area in between. There are two important points to note in this official announcement: one is A0 tapeout success, which means that the first version of the silicon wafer can be lit up and run through after being taped out, which is already halfway through the chip industry; The other is $1B customer contracts, which, although not equal to revenue, at least indicate that this is not a pure story and that there is real money betting on it. And what I really want to talk about is the relationship between this thing and the Memory line. That's also why I think of CBRS. Many people's first reaction is: ASIC is more efficient → requires less HBM → bearish for Micron/Hynix. I think this judgment is reversed. Etched's Sohu still comes with 144GB HBM3E per chip, without cutting HBM, but maximizing HBM utilization. Today's big model inference (especially in the decode stage) is becoming more and more bottleneck oriented towards KV Cache read bandwidth. It is betting that 'HBM will always be the bottleneck', but it doesn't want the general-purpose computing shell of GPU anymore. This leads me to the most interesting part - three possible parallel worlds in the future: Speaking of which, when I wrote CBRS in May, I also wrote three different timelines, but at that time it was based on branch inference of future model size, while this time it was based on another variable - memory architecture World A: NVIDIA wins. GPU+HBM+CUDA continue to dominate. MU/Hynix explosion, CBRS average. World B: Etched wins (provided that Transformer continues to dominate). Transformer ASIC+HBM+Inference Rack has become a new paradigm. MU/Hynix still explodes (even more so, as ASIC relies more on high-capacity HBM), while CBRS does not have an advantage. World C: Cerebras wins. Wafer Scale+wins the SRAM route on large chips and tries to avoid HBM as much as possible. In this situation, MU/Hynix is the most under pressure, and CBRS is the one that really skyrockets. So the companies that are really competing with CBRS for territory are Etched, Groq, Taalas, and others who are betting on different memory routes. NVIDIA and Etched are actually on the same side - both are betting on HBM to continue dominating AI computing. Of course, it should also be noted that all performance figures of Etched currently come from the official source, which is 10 to 20 times faster than H100, and there is no large-scale third-party benchmark verification for the term 'SOTA Throughput'. Graphcore, Cerebras, SambaNova, Habana, Groq have almost all said the same thing, whether they can be bought by the market in the end is another matter. What's even more extreme about Etched than these predecessors is that it voluntarily gave up the option of "universality", which is a gamble without a Plan B