@SaitoWu: A group at Baidu Research US predicted ten years ago: Don't bet all AI compute on NVIDIA. So they actually invested in a 'wafer-scale' chip company — Cerebras. In 2016, Zhou Nan left investment banking for Baidu's US AI research institute. Andrew Ng was leading the team, budgets were ample, GPUs were bought freely. Dario (An…

A group at Baidu Research US predicted ten years ago: Don't bet all AI compute on NVIDIA. So they actually invested in a 'wafer-scale' chip company — Cerebras. In 2016, Zhou Nan left investment banking for Baidu's US artificial intelligence research institute. Andrew Ng was leading the team, budgets were ample, GPUs could be bought freely. Dario (Anthropic founder) and Greg Diamond were there too. The team had already vaguely sensed the shape of scaling laws: the larger the model, the more data, the more compute, the better the results. The DeepSpeech 2 paper was one of the earliest signals. Zhou Nan's first project was Cerebras. At the time, he searched globally for training chips 'not made by NVIDIA,' looked at Graphcore and Wave Computing, and ultimately chose Cerebras. Because this company was the most aggressive — they aimed to build a wafer-scale engine, turning an entire wafer into a massive AI compute engine, placing compute units and memory extremely close together to drastically reduce communication costs. When invested, Cerebras hadn't even taped out; they only had a signature. Baidu Research's researchers directly verified it, running it on the world's largest language model at the time, and the signals were decent. The investment decision faced almost no resistance — Robin, Lu Qi, and Robin Li quickly approved. The valuation was already over $700 million, expensive for 2017, but Zhou Nan was betting on 'non-consensus.' The truly hard part wasn't the idea, but building the wafer-scale. Yield, heat dissipation, power, compiler — each was a hard nut to crack. 2017 to 2019 were almost the darkest hours, with tapeouts delayed repeatedly. But early investors like Benchmark, Foundation, and Eclipse stayed with them, offering plenty of patience. Chips truly take ten years to show results. Now Cerebras's opportunity lies in inference, not training. OpenAI signed a deal worth at least $20 billion precisely for its low latency and high throughput. NVIDIA's CUDA ecosystem is too strong, making migration costs extremely high for training, but in inference scenarios, Cerebras's architecture advantages begin to shine. In fact, Sam Altman had personally invested in Cerebras in 2016, even earlier than Baidu. Zhou Nan reflects that Baidu Research US was like the 'Whampoa Military Academy' of Silicon Valley. Too many people later left to start companies: Inflection, Adept, Anthropic, Cohere… Unfortunately, due to geopolitical issues, Baidu later tried to raise a dedicated AI fund but didn't succeed. OpenAI and Databricks were on the list at the time, but ultimately no investments were made. Everyone worried ten years ago about relying solely on NVIDIA, yet NVIDIA still became a de facto monopoly. But now the explosion of inference demand has opened a new window for heterogeneous chips. Cerebras is just the beginning; more inference chips with new architectures may follow. Now Zhou Nan approaches AI investment: consensus comes too fast, early windows become shorter and shorter. Instead, he focuses on directions without consensus yet, such as Physical AI (robotics) and new inference chip architectures. What moved me most in this episode is how it recreates the atmosphere of that 'clash of titans' at Baidu Research US when scaling laws were just emerging. Many things were seen early on, but truly coming to fruition took ten years. Cerebras's story is just one example.