AI chip startups challenge NVIDIA's dominance with six approaches to data-movement efficiency

ME AI message: Anthropic investor Deedy points out that all startups developing the next generation of AI chips are attempting to break NVIDIA’s dominance in different ways, targeting the fundamental issue of data movement. Six differentiated approaches include: eliminating DRAM (Groq, acquired by NVIDIA for $20 billion), eliminating interconnects (Cerebras, already public with a market cap of ~$48 billion), eliminating the separation of compute and memory (d-Matrix), eliminating server architectures centered on compute (Majestic), eliminating generality (Etched, Taalas, MatX), and eliminating the $400 million lithography machines (Substrate). The IPO and acquisition valuations of Cerebras and Groq have set pricing benchmarks for the entire AI chip sector. This week, Etched’s valuation doubled to $10.3 billion, surging from obscurity in just three weeks, further confirming that capital is rapidly re-evaluating valuations in this space. Among 18 major startups, the combined paper valuations of private companies total approximately $58 billion, while public market valuations reach ~$48 billion, covering multiple sub-sectors including inference, training/new architectures, systems, wafer fabrication, and lithography. Each approach challenges a core assumption of NVIDIA’s GPU architecture—that moving data between compute units and memory consumes both time and energy. Groq eliminates memory hierarchy latency by replacing DRAM with SRAM; Cerebras removes inter-chip bottlenecks with wafer-scale chips; d-Matrix performs computation directly within memory; and Etched abandons generality to design hardware exclusively for the Transformer architecture. The AI chip startup ecosystem has evolved from a single narrative of GPU replacement into a multi-front architectural assault. NVIDIA’s acquisition of Groq signals its deep understanding that the next frontier of competition is no longer raw compute power, but the complete re-architecting of data movement efficiency. (Source: BlockBeats)