I remember during the three years of Defi Summer from 2020 to 2022, a phrase that many people often lamented was' One day in the cryptocurrency industry, one year in the human world ' Because technology iteration is too fast, there are countless new things coming out every day, and if you don't watch them for a few days, you can't keep up with the pace Now this feeling has completely disappeared in the cryptocurrency circle and appeared on the AI side. I can't finish watching new things every day, I really can't finish watching them Coincidentally, I have seen more posts discussing Cerebras in the circle these days. Yesterday, I also saw a lot of roadshow information, after all, it will be listed for trading on the 14th. I have previously talked about this company separately when sharing with Amber (the PPT used at the end of the article), so let me briefly state my judgment on Cerebras Let's start with the conclusion: This company is very interesting, but the core variable may not necessarily be the chips they produce, but rather what the future form of AI workloads will be. First, let's talk about what it is for to prevent the old iron from watching Cerebras Its core technology is wafer scale chip, which simply means that others are still using "stamp sized" chips. It directly turns the entire wafer into a super large processor, coupled with super large S RAM, to keep a large amount of data locally for high-speed processing, reducing the most headache inducing memory bottleneck of traditional GPUs. Now the rise of Hynix Micron is due to the high demand for HBM, and Cerebras has directly avoided relying on HBM Many people see Cerebras' most impressive benchmark: inference speed 10-15 times faster than GPU, and their first reaction is the next Nvidia?! Don't rush for now. The biggest problem with this benchmark is that the default core requirement for AI is always to 'dispense tokens faster'. If it's just humans staring at ChatGPT chatting, this story isn't really that sexy. You spit out 30 tokens per second, I can hardly read them anymore. Ten times more, the marginal experience improvement is almost zero. What's really interesting is the agent. Agent does not read words, Agent consumes tokens. Speed directly equals productivity. An OpenClaw/Hermes agent may require dozens of inference calls to read web pages, write plans, call APIs, run code, retry errors, and continue execution. Every 2 seconds, the task is a minute level experience. Every 200ms, it's another world. So what deserves more attention in Cerebras is the AI worker line, rather than just chatbot acceleration. But the problem is - its magic comes from wafer scale+ultra large SRAM, and local access is extremely fast. But SRAM has a natural trade-off, fast speed, expensive capacity, and if a large model cannot fit it, it must be split. And once split, chip to chip communication comes up. The most feared aspect of communication in LLM inference is precisely the decode phase. Tokens are ejected one by one, and with each additional hop, the delay is forcibly added and cannot be hidden. So whether Cerebras can succeed or not depends not on how many times faster it is than GPUs, but on what the mainstream computing form of AI will be in the future. 1. Timeline 1- In the next few years, cutting-edge super models will dominate the world, with billions or even trillions of parameters, and all requests will be handled by the super models themselves. Nvidia's distributed infrared will still be the most comfortable, and Cerebras' speed advantage will be greatly eroded by communication losses. 2. Timeline 2- If technologies such as MoE, distillation, and quantification continue to make rapid progress, I am not surprised that a model of around 70B in the next two years can achieve 80% -90% of the performance of today's 700B model. I would like to express my strong gratitude to Deepseek for their great success! ) If the world goes in this direction, the story will change. The big model is responsible for planning/judgment/orchestration. The large number of worker models that actually perform tasks fall within the range of 30B-70B. These models are smart enough to reap the benefits of high-speed local inference. In the world of agents, most tokens do not require the smartest brain at all. Many jobs are essentially manual labor at the execution level: browsing web pages, modifying code, debugging tools retry、 Continue with the process. Once this topic is established, Cerebras will directly enter its sweet zone 3. Timeline 3- Future inference is mainly based on the end side, using small models such as 8B and 14B. The GPU can also run well, and even dedicated ASIC chips have higher efficiency. In this scenario, Cerebras' advantages and moat are not high In other words, the advantages of Cerebras in the two parallel universes of cloud based inference for ultra large models or end-to-end inference for ultra small models are not obvious enough. Only mainstream inference falls within the size of the 32B-70B medium model, which happens to be the world where Cerebras can fully demonstrate its abilities, with "Big enough to stress GPU memory, Small enough to fit locally" So my assessment of Cerebras is that with a market value of over 30 billion, looking at short-term data such as orders and financial statements, the long-term bet is on which parallel universe timeline the computing paradigm of the future Agent era will fall on