Is Open-Source AI Shifting the Primary Economic Surplus Directly to End Users?36 affirmations sourcées, faits contestés et éléments de preuve

Topic summary

Rapid commoditization of model weights has eroded frontier tech rents, distributing cost savings directly to enterprise adopters.

Key research questions

• Investigate "Who Actually Benefits From AI?" from this distinct point of divergence: Open-source commoditization and end-user surplus.

Important claims and supporting citations

• An arXiv analysis of 2024–2025 LLM pricing found that algorithmic efficiency progress drove price declines roughly 3x more strongly than hardware progress or estimated competitive pricing effects.

  • https://arxiv.org/html/2601.12488v1
  • https://arxiv.org/pdf/2403.12004

• DeepSeek-R1 was released on January 20, 2025, under an MIT license, with DeepSeek claiming a base V3 model training cost of approximately $5.6 million.

  • https://mitsloan.mit.edu/press/calculating-economic-well-being-information-age-a-new-study-researchers-mit-sloan-measures-value-digital-goods-and-services-create

• A 2026 Stanford working paper by Brynjolfsson et al. found mean consumer willingness-to-accept for generative AI rose from $98 in 2025 to $124.50 in 2026, with median rising from $3.40 to $11.40, and aggregate US consumer surplus expanding from $116 billion to $172 billion.

  • https://digitaleconomy.stanford.edu/publication/what-is-generative-ai-worth/
  • https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6569938

• A leaked May 2023 internal Google memo attributed to Luke Sernau argued open-source models would erase the competitive moats of both Google and OpenAI.

  • https://openai.com/index/hello-gpt-4o/

• Epoch AI data indicates model token price deflation varied by benchmark, ranging from 9x per year at the slowest end to 900x per year at the fastest during 2024.

  • https://dontpaniclabs.com/blog/post/2025/12/02/the-price-per-token-is-dropping-will-it-stay-that-way/

• Inference costs for GPT-3.5-level systems fell from $20.00 per million tokens in November 2022 to $0.07 in October 2024 (Gemini-1.5-Flash-8B), a deflation factor exceeding 280x.

  • https://report-ai.org/indexes/ai-economics/llm-token-price-index/
  • https://digitaleconomy.stanford.edu/publication/what-is-generative-ai-worth/
  • https://aiindex.stanford.edu/report/

• Andreessen Horowitz found enterprise LLM budgets grew roughly 2 to 5 times from 2023 to 2024 despite sharp drops in per-token inference prices.

  • https://a16z.com/generative-ai-enterprise-2024/

• Stanford AI Index 2024 estimated the training costs for GPT-4 and Gemini Ultra at approximately $78 million and $191 million, respectively.

  • https://aiindex.stanford.edu/report/
  • https://hai.stanford.edu/ai-index/2025-ai-index-report

Competing explanations

• Enterprise self-hosting of open LLMs becomes cost-effective relative to proprietary APIs above 8 to 30 million daily tokens, with 100M+ token volumes achieving 60% to 70% cost savings.