\\DeepSeek quietly testing their new "v4-lite" model in their app this week with a massive 1-Million token context window and lightning-fast speeds.
A few clues strongly suggest this new model is based on a new architecture (i.e not Nvidia)
1. This model has 1M context, which was too expensive to do in previous-gen Deepseeks trained on Nvidia GPUs.
2. The attention quality is extremely high - many early testers expressed surprise DeepSeek can now find a specific passage in 500k-word novel. This level of attention was previously only possible with Gemini.
3. Faster inference - people deduced that this new model is smaller than previous iterations - likely a compromise they had to make as they build up their Ascend cluster.
# The Catalyst 🧨
DeepSeek's v4 technical report is imminent (Lunar New Year timing). When the paper confirms v4-lite was natively trained end-to-end on Huawei Ascend CM384 clusters... the psychological "CUDA Moat" instantly evaporates. NVDA goes from an "invincible software monopoly" to a "commoditized hardware vendor." US Hyperscalers (Zuck, Sundar, Satya) will realize they don't need to pay Jensen's 75% gross margin tax anymore. We are looking at a 15-20% gap down overnight.
Do I know this for sure? I am not an insider and I cannot forsee the future. But the risk/reward looked good enough to me.
# Positions:
# NVDA 2/20 165p
**TL;DR:** DeepSeek possibly built their new 1M-context model natively for Huawei chips and only used Nvidia as an expensive backup plan. CUDA moat is no moat.