AI Research Talent Movement and Inference Cost Analysis
연구/벤치마크 | Sun Jun 21 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 2 sources
Nobel laureate John Jumper moved to Anthropic, and a detailed methodology for calculating LLM inference costs was published.
Analysis
[John Jumper] left DeepMind to join Anthropic [1]
- 2024 Nobel Prize in Chemistry laureate
- Spent approximately 9 years at Google DeepMind
- Led the AlphaFold team
[Noam Shazeer] left DeepMind to join OpenAI [1]
- Co-founder of Character AI
- Announced departure in the same week
- Consecutive departures of key talent from DeepMind
[injuly.in Blog] published napkin math analysis of LLM inference costs [2]
- Calculations based on GPU peak throughput and memory bandwidth
- Assumes FP-8 quantization
- Matrix multiplication cost of 2NMd operations
[LLM Inference Economics] derived cost per token and user processing limits [2]
- Analysis of forward pass cost in auto-regressive models
- Computation reduction via KV-Cache
- Optimization for serving hundreds of concurrent users per GPU