Vitalik: True CROPS AI Must Run Across All Hardware!

Ethereum co-founder Vitalik Buterin has shared an update on local AI progress, noting that DeepSeek V4 now has a 2-bit quantized version capable of running within roughly 90 GB of VRAM — hitting approximately 35 tokens per second on Apple hardware and around 7 tokens per second on AMD hardware. The milestone signals meaningful movement toward consumer-grade local inference for large models.

Vitalik argued that genuine "CROPS AI" should be defined by multi-hardware support, not simply rebranded as "decentralized AI." He outlined a "CROPS Ethereum access layer" that overlaps with this vision, encompassing ZK-based paid remote LLM calls and private Ethereum RPC reads — infrastructure that would let AI models interact with Ethereum without leaking user data.