Next-generation AI chip companies challenge Nvidia's dominance in six ways, shifting the core from peak computing power to restructuring data mobility efficiency
Comparing news, Anthropic investor Deedy pointed out that all startups that are developing next-generation AI chips are trying to break Nvidia's dominance in different ways and want to attack the fundamental problem of data mobility. The six ways to differentiate include eliminating DRAM (Groq, which has been acquired by Nvidia for $20 billion), eliminating interconnection (Cerebras, which has already been listed at a market value of about $48 billion), eliminating computation/memory separation (D-matrix), and eliminating compute-centered servers Architecture (Majestic), elimination of generality (Etched, Taalas, matX), and elimination of $400 million worth of lithographs (Substrates).
The listing and acquisition prices of Cerebras and Groq set the pricing benchmark for the entire AI chip circuit, while Etched's valuation doubled to $10.3 billion this week, breaking out of a low profile in just three weeks, further confirming that capital is rapidly revaluing its pricing for this track. Among the 18 major startups, private companies have a total paper value of about US$58 billion and an open market market capitalization of about US$48 billion, covering various segments such as reasoning, training/new architectures, systems, and fabs and lithography.
Each route challenges the core assumption of Nvidia's GPU architecture—moving data between computing units and memory consumes both time and energy. Groq uses SRAM instead of DRAM to completely eliminate memory-level delays, Cerebras uses wafer-level chips to eliminate inter-chip connectivity bottlenecks, D-Matrix directly performs calculations in memory, and Etched abandons generality and only designs hardware for the Transformer architecture. The AI chip startup ecosystem has evolved from a narrative of a single GPU replacement to a multi-point attack at the architecture level. Nvidia actively participated in this process by acquiring Groq, which also shows that the core of next-generation competition is no longer a peak in computing power, but a complete restructuring of data movement efficiency.
