According to BlockBeats, on July 26, Anthropic investor Deedy pointed out that all startups developing next-generation AI chips are trying to break Nvidia's dominance in different ways, aiming to tackle the fundamental problem of "data movement." Six differentiated challenge paths include: eliminating DRAM (Groq, acquired by Nvidia for $20 billion), eliminating interconnect (Cerebras, listed with a market capitalization of approximately $48 billion), eliminating compute/memory separation (d-Matrix), eliminating compute-centric server architectures (Majestic), eliminating versatility (Etched, Taalas, MatX), and eliminating the $400 million lithography machine (Substrate).
The IPOs and acquisitions of Cerebras and Groq set a benchmark for the entire AI chip sector, while Etched's valuation doubled to $10.3 billion this week, a dramatic surge in just three weeks, further confirming the rapid revaluation of this sector by investors. Among the 18 major startups, the combined paper value of private companies is approximately $58 billion, and the market capitalization is approximately $48 billion, covering multiple sub-sectors including inference, training/new architectures, systems, and wafer fabs and lithography.
Each approach challenges a core assumption of NVIDIA's GPU architecture—that moving data between computing units and memory is both time-consuming and energy-intensive. Groq completely eliminates memory-level latency by replacing DRAM with SRAM, Cerebras eliminates inter-chip interconnect bottlenecks with wafer-level chips, d-Matrix performs computations directly in memory, and Etched abandons versatility to design hardware solely for the Transformer architecture. The AI chip startup ecosystem has evolved from a narrative of simply replacing GPUs to a multi-pronged attack at the architecture level. NVIDIA's acquisition of Groq demonstrates its active participation in this process, indicating its understanding that the core of the next generation of competition is no longer peak computing power, but a complete reconstruction of data movement efficiency.