AI inference costs are falling faster than Moore's Law
The cost of running AI models is dropping 70% annually due to hardware improvements, algorithmic efficiency, and competition. This will unlock new use cases and margin expansion for AI companies.
Bull Case
- +
NVIDIA H100 inference costs down 60% year-over-year while performance doubled
NVIDIA Q4 2024 Investor Presentation
- +
New architectures like mixture-of-experts reduce compute by 5-10x with minimal accuracy loss
Google DeepMind Technical Report, Dec 2024
- +
Hyperscaler competition driving aggressive pricing - AWS Bedrock prices down 40% in 2024
AWS Re:Invent 2024 Announcements
Bear Case
- -
Training costs remain high and rising, limiting model improvements
OpenAI Economics Paper, Nov 2024
- -
Energy constraints may limit datacenter expansion and increase inference costs
Goldman Sachs Energy Infrastructure Report
- -
Diminishing returns on hardware improvements as we approach physical limits
IEEE Spectrum: The End of Moore's Law
Unlock Deeper Insights
Join the waitlist to access 2nd & 3rd order effects when we launch
Related Companies
NVDA
NVIDIA Corporation
Dominant AI chip provider, 80%+ inference market share
GOOGL
Alphabet Inc.
Major cloud provider with TPU chips and large AI inference workloads
META
Meta Platforms
Heavy AI inference user for content recommendations and Llama models
AMD
Advanced Micro Devices
Challenger in AI chips with MI300 series, growing datacenter share
Key Catalysts
Mar 15, 2025
NVIDIA GTC Conference - Expected H200 and B100 announcements
Jun 1, 2025
Google I/O - TPU v6 and Gemini inference pricing updates
Sep 30, 2025
Major hyperscaler capex reports for Q3
Disclaimer: For informational purposes only. Not investment advice. ThesisSwipe provides research and analysis but does not recommend any specific investment decisions. Always conduct your own research and consult with a qualified financial advisor before investing.
What's your take?
Your reaction is anonymous and helps us improve our theses
Get Early Access
Hundreds of AI-generated investment theses across 12 categories. Launching soon.
Save This Thesis
Join the waitlist to save theses, build your library, and get notified about new ideas.
Free to use. No credit card required.