AI reasoning frenzy sweeping the globe, "NVIDIA challenger" Cerebras is coming on strong! Valuation surges 170% to 22 billion US dollars.

Media reports citing informed sources revealed that one of the strongest competitors to the "AI chip superpower" NVIDIA Corporation (NVDA.US) - the artificial intelligence chip supplier Cerebras Systems Inc. - is in discussions to raise a new round of funding of approximately $1 billion to support this AI chip startup's long-term competitive position against NVIDIA Corporation and to significantly strengthen the cost-effectiveness and energy efficiency of its AI computing cluster compared to NVIDIA Corporation's AI GPU cluster. The informed source mentioned that this round of funding would value the startup at $22 billion before the investment, representing a significant 170% expansion in valuation compared to the funding round in September last year. As the discussions are still private, the source requested to remain anonymous. The source also mentioned that the AI chip startup still plans to actively pursue its initial public offering (IPO) on the US stock market. Under the leadership of CEO Andrew Feldman, Cerebras Systems is actively seeking to challenge NVIDIA Corporation's dominant market position in the field of artificial intelligence chips with a market share as high as 90%. Therefore, this round of funding is expected to provide new financial support for this AI chip company seeking to challenge NVIDIA Corporation, which is one of the top artificial intelligence computing infrastructure suppliers globally. Who is Cerebras Systems, the company challenging NVIDIA Corporation? Cerebras Systems has grand ambitions in the field of artificial intelligence. CEO Andrew Feldman stated that the actual efficiency of his company's computational hardware in running large artificial intelligence models is several times that of the NVIDIA Corporation system. In addition to providing physical computational clusters, this AI chip star startup, led by him, also actively provides remote artificial intelligence computing services to large clients like Meta Platforms Inc., the parent company of Facebook, the long-established US technology company IBM referred to as the "blue giant," and the "European OpenAI" Mistral AI. The company's latest valuation has significantly increased from the investment round in September, when Cerebras Systems was valued at approximately $8.1 billion. Shortly after that, NVIDIA Corporation signed an important licensing agreement with Groq, an AI chip startup and one of Cerebras Systems' competitors, and acquired most of the chip design talent from that AI chip company, which also greatly boosted investors' optimism about the field of artificial intelligence chips. The recent $20 billion non-exclusive licensing agreement reached between NVIDIA Corporation and the AI chip startup Groq serves to authorize its AI inference technology to NVIDIA Corporation, with the founder and core research team of Groq joining NVIDIA Corporation after the transaction is completed. This move highlights the imminent surge of the "global AI inference wave" and the increasing competition pressure from Alphabet Inc. Class C TPU AI computing clusters. NVIDIA Corporation aims to maintain its absolute dominance in the field of AI chips with a market share as high as 90% by leveraging "multi-architecture AI computing + consolidating the CUDA ecosystem + bringing in more AI chip design talent." Additionally, NVIDIA Corporation seeks to solidify its position with Groq and the Israeli AI startup AI21 Labs to secure its dominance in the AI full-stack market. Similar to Groq before being "acquired" by NVIDIA Corporation, Cerebras Systems is considered one of NVIDIA Corporation's strongest competitors in the field of AI chips, especially in the rapidly growing subsea market of AI inference. Cerebras' technological route is radically different from NVIDIA Corporation's AI GPU computing system and Alphabet Inc. Class C TPU (AI ASIC technology route). It adopts the "Wafer-Scale Engine" (WSE) architecture, placing the entire AI model on a single giant chip, which greatly enhances inference performance and memory bandwidth, achieves higher energy efficiency in unit inference, and avoids the data segmentation and high-speed communication overhead between GPU clusters. Different from chip giants such as NVIDIA Corporation, Broadcom Inc., and AMD, which focus on smaller high-performance chips and use chiplet advanced packaging exclusively offered by Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR to integrate chips, Cerebras Systems manufactures a super large-scale chip that covers the entire silicon wafer. According to semi-analysis and other semiconductor research institutions, this wafer-level architecture can achieve significantly higher performance density and energy efficiency compared to traditional AI GPU/AI ASIC computing systems when processing large language model inference tasks. Another startup company, Etched, also aspiring to become one of NVIDIA Corporation's strongest competitors, has reportedly raised approximately $500 million in a new round of funding, bringing its valuation to around $5 billion. It is worth noting that Cerebras Systems relies heavily on the business of the Abu Dhabi-based artificial intelligence company G42, and this deep cooperation has attracted scrutiny from the Committee on Foreign Investment in the US (CFIUS), possibly leading to continued obstacles in the IPO process. As the wave of AI inference sweeps in, Cerebras Systems aims to seize this trend. Recent market trends indicate that Cerebras Systems is actively positioning itself in the wave of AI inference and enhancing its competitive position through funding and advancing its IPO to continue to erode NVIDIA Corporation's significant market share as high as 90%. The ongoing fundraising process and IPO promotion of Cerebras reflect the company's intentions to expand its market influence, enhance competitiveness, and challenge NVIDIA Corporation through the AI inference wave. The latest CS-3 system from Cerebras (equipped with the WSE-3 chip) has been reported to outperform NVIDIA Corporation's latest GPU system - the Blackwell architecture AI GPU - in large language model inference scenarios. According to Cerebras' own comparison data, the CS-3 is approximately 21 times faster than the B200 AI GPU system based on the Blackwell architecture when running tasks like Llama 3 70B inference, while also offering lower overall costs and energy consumption (including hardware and power consumption costs). Third-party analyses have indicated that in certain high-parameter inference benchmark tasks, Cerebras' output throughput (tokens/sec) is multiple times higher than traditional GPUs, with reports mentioning speeds up to 20 times faster or more in some model inferences. In the field of large-scale inference, especially in handling large LLMs, Cerebras' WSE-based architecture demonstrates significant advantages in terms of cost-effectiveness (cost/performance ratio) and energy efficiency (energy consumption/inference output ratio) compared to NVIDIA Corporation's GPU computing clusters. These advantages mainly stem from its wafer-level single-chip design, high bandwidth, and inference throughput capabilities. However, these advantages are more pronounced in specific inference scenarios rather than covering all AI computing tasks. NVIDIA Corporation still holds a significant advantage in general computing task deployment, AI training operators, and CUDA ecosystem compatibility. The AI training side, which is almost monopolized by NVIDIA Corporation's AI GPU, requires a more powerful and versatile AI computing cluster and rapid iteration capabilities for the entire computing system. On the other hand, the AI inference side prioritizes unit token cost, latency, and energy efficiency after the scaling deployment of cutting-edge AI technology. Alphabet Inc. Class C clearly positions Ironwood as the next generation of TPU designed for the "AI inference era," emphasizing the cost/performance/energy efficiency of the computing cluster and scalability. As AI inference computing becomes a long-term cash cost center for global technology companies, customers are increasingly inclined to choose more cost-effective and competitive AI ASIC accelerators in the cloud. Reports have mentioned that OpenAI rents TPUs (Alphabet Inc. Class C TPU falls under the AI ASIC technology route) in large quantities through Google Cloud, a platform of Alphabet Inc. Class C, as one of the core motivations is to reduce AI inference costs. This is a typical case of rising competitive pressure from TPUs. Based on the data analyzed by semi-analysis, Alphabet Inc. Class C's latest TPU v7 (Ironwood) shows remarkable generational leaps, with the BF16 computing power of TPU v7 reaching 4614 TFLOPS, while the widely used previous generation TPU v5p only reached 459 TFLOPS. This represents a significant leap in performance. Furthermore, TPU v7 is directly comparable to NVIDIA Corporation's Blackwell architecture B200, and for specific AI applications, AI ASICs with superior cost-effectiveness and energy efficiency advantages can more easily handle mainstream inference end computing loads. For example, TPUs can provide up to 1.4 times higher performance per dollar compared to Blackwell. The current rapid growth trend in large-scale AI inference demand, with a doubling every six months, combined with the increasing competition pressure brought by the AI inference wave and the powerful computing capabilities of Alphabet Inc. Class C TPU, has led NVIDIA Corporation to secure its defense/counterattack strategy by acquiring Groq for inference chip technology and top talent, as well as supplementing software and modeling capabilities through AI21. The transaction between NVIDIA Corporation and the AI chip startup Groq essentially involves non-exclusive technology licensing for inference-based AI chips and the absorption of Groq's founder/CEO Jonathan Ross, senior executives, and some core engineering teams. Some semiconductor industry analysts have emphasized that Groq's exclusive chip technology focuses on inference and reduces data movement bottlenecks through on-chip SRAM, directly addressing the cost/latency pain points in the inference stage.