Qilin Labs publishes research for educational and informational purposes only. Not investment, financial, or tax advice. The firm holds positions in many assets discussed and may transact at any time without notice.
Meta's Muse puts agentic AI in front of 3.6B users. The obvious winners get the headlines. The real bottleneck sits somewhere nobody's talking about.
Pre-training compute at Anthropic and OpenAI collapsed from 60% to 7% in two years. Post-training now dominates — and it demands a completely different hardware stack. Here's where I'm looking.
DeepSeek and Kimi just slashed KV cache requirements by orders of magnitude. Anthropic and OpenAI seem to be taking a different path. The memory stack winners and losers are about to reshuffle.