DEEPSEEK

A Chinese AI lab spun out of hedge fund High-Flyer in 2023. In early 2025 it triggered an 'AI Sputnik moment' by matching GPT-class performance at a fraction of the cost. With MoE architecture and distillation driving rock-bottom inference prices, its open-source models are gaining global share fast.

CEO Liang Wenfeng · Founded 2023 · HQ Hangzhou, China · Exchange 비상장 · Website deepseek.com

Predictions & track record (5)

• Awaiting grading SemiAnalysis, long tough on AMD, now says 'great chance.' Can it really break the CUDA moat?Check whether AMD actually reaches SemiAnalysis's revised October 2026 target for CI test parity with Nvidia's CUDA on vLLM and SGLang gating tests.
• Awaiting grading Nvidia's real moat is software, not silicon. Can AI coding tools crack it in a year?Grade around mid-2027 whether Huawei's chip software ecosystem (DeepSeek's compiler and TileLang tooling) has closed the gap with CUDA as Wenfeng predicted, and whether Nvidia's China-exposed growth underperformed as a result.
• Awaiting grading SemiAnalysis: Nvidia's next-gen Rubin chip shows 5x better performance per dollar than today's Blackwell, per early dataGrade whether Nvidia submits verifiable Rubin benchmark numbers to InferenceX by Q3 2026, and whether they roughly match today's early 5x per-dollar claim.
• Awaiting grading China weighs restricting overseas transfer of AI training data and model weights at Alibaba, ByteDance, Zhipu AIWatch whether China's Ministry of Commerce formalizes restrictions on overseas AI data and model-weight transfers in the coming months, and whether it affects future open-weight releases from Alibaba, ByteDance, or Zhipu AI.
• Hit The 'Kimi moment': a $1.3B open-weight model that could reprice the whole US AI tradeIndependent benchmarking firm Artificial Analysis ranked Kimi K3 (full weights released July 27, 2026) #3 on its Intelligence Index (score 57), matching or surpassing GPT-5.5 and Claude Opus 4.8 -- graded hit.

6 related posts

• Bull SemiAnalysis, long tough on AMD, now says 'great chance.' Can it really break the CUDA moat? SemiAnalysis · 2026-07-25
• Bear Nvidia's real moat is software, not silicon. Can AI coding tools crack it in a year? Jukan (@jukan05) · 2026-07-23
• Bull SemiAnalysis: Nvidia's next-gen Rubin chip shows 5x better performance per dollar than today's Blackwell, per early data SemiAnalysis · 2026-07-23
• Watch China weighs restricting overseas transfer of AI training data and model weights at Alibaba, ByteDance, Zhipu AI Jukan (@jukan05) · 2026-07-21
• Watch The 'Kimi moment': a $1.3B open-weight model that could reprice the whole US AI trade 메르 (ranto28) · 2026-07-18
• Half of America is quietly using Chinese AI: what OpenRouter's real usage data shows 메르 (ranto28) · 2026-07-17