Summary of Key Points
While the AI industry is competing for high-bandwidth memory (HBM) to enhance the computing power of existing chips, Cerebras has chosen a different approach by “redesigning dedicated AI computers.” The company went public on NASDAQ in May 2026, with its stock price soaring 68% on the first day of trading, and its market value briefly approaching $100 billion. However, ten years prior (in 2015), Cerebras was just a “crazy bet” by five engineers. At that time, AI had not yet exploded, but they decided to build an AI computer from scratch. Around the same period, Baidu’s AI lab discovered the “Scaling Law”—the larger the model, the better the performance, but the demand for computing power grows exponentially—recognizing that AI would soon face a bottleneck in computational resources. Thus, this bold bet on the future of AI began.
Detailed Analysis
1. Cerebras’ “Anti-mainstream” Approach: Focusing on AI-Dedicated Computers Instead of HBM
The current trend in the AI industry is to equip existing chips (such as GPUs) with larger and faster memory to process more data. However, Cerebras has taken a different path by designing computers specifically for AI. To put it simply, while everyone is upgrading ordinary cars with bigger fuel tanks and faster engines, Cerebras created a racing car from the ground up. Its chips may be larger (integrating more computing units) and have architectures that are better suited to the parallel computing needs of AI, which requires handling vast amounts of data simultaneously. This “customization” is what sets it apart from the mainstream approach.
2. The 68% Stock Price Surge on Market Launch
The success of Cerebras’ stock launch was largely due to its addressing the most pressing issue in the AI industry: the shortage of computing power. As large models (like GPT-4 and Wenxin Yiyuan) continue to grow, the demand for computing power is increasing exponentially, outpacing the capabilities of traditional chips with added HBM. Cerebras’ dedicated computers can solve this problem; for example, its chips may be able to process more data at once or operate several times more efficiently than conventional GPUs. Investors saw that for AI to advance, more efficient computing power is essential, and Cerebras’ approach seemed like the solution, leading them to value the company at nearly $100 billion.
3. The “Crazy Bet” of Ten Years Ago
In 2015, AI was not as popular as it is now, and large language models had not yet emerged. The idea of redesigning an AI computer from scratch sounded absurd to many:
- High investment: Creating new computers required designing chips and architectures from scratch, which was a significant financial commitment.
- High risk: No one knew how quickly AI would develop; if the investment failed, the money would be lost.
- Different industry consensus: The industry was using general-purpose chips (like GPUs) for AI, believing that modifying existing ones would suffice. Therefore, Cerebras’ approach was seen as a bold gamble that few were willing to take on.
4. Baidu’s Scaling Law: Unveiling the Limitations of AI Computing Power
Baidu’s discovery of the Scaling Law provided theoretical support for Cerebras’ approach. This principle states that the performance of AI models is directly proportional to their size, the amount of training data used, and the computing power invested. The larger the model and the more data, the better the results, but the demand for computing power grows exponentially. Baidu realized that future large models would face a shortage of computing resources, and existing chips and architectures would not be sufficient. Cerebras’ vision became a potential solution to this issue.
5. The Competition Between Two Approaches to AI Computing Power
Cerebras’ success represents the competition between two approaches to developing AI computing power:
- Approach One: Making incremental improvements by adding HBM to existing chips and continuing to use general-purpose hardware.
- Approach Two: Designing dedicated hardware for AI (such as Cerebras’ computers) to optimize computing power from the ground up.
The outcome of this competition will determine the speed of AI’s development. If Approach Two prevails, there could be a dramatic increase in AI computing power, making large models more intelligent and affordable. If Approach One remains dominant, the bottleneck in computing power may slow down AI’s progress. Cerebras’ public offering is a vote of confidence from the market in Approach Two.
Conclusion
Cerebras’ story illustrates the spirit of innovation in the AI industry: taking a harder but potentially more promising path when others are following the crowd. Baidu’s Scaling Law provided the rationale for this approach. The market’s enthusiastic response to Cerebras shows that companies that address the limitations of computing power will be highly sought after in the future. This also highlights that in times of technological change, “anti-mainstream” bets can lead to significant success.