Matthew Berman: 'DeepSeek Is a Problem' — 123K Views on the Competitive Threat and Cost Disruption

Matthew Berman·medium signal

Matthew Berman's analysis of DeepSeek's competitive positioning reached 123K views, covering DeepSeek R2's 1.2T-parameter MoE architecture (78B active per token) running on Huawei Ascend chips and claiming 97% cost reduction versus GPT-4 Turbo. The video frames DeepSeek as a structural problem for Western AI labs due to hardware independence from NVIDIA and pricing that undercuts by 140x on reasoning tasks.