# Is Open-Source AI Shifting the Primary Economic Surplus Directly to End Users?: 36 Sourced Claims, Contested Facts & Evidence Rapid commoditization of model weights has eroded frontier tech rents, distributing cost savings directly to enterprise adopters. Authors: djwide HTML: https://moyomap.com/maps/open-source-ai-shifting Markdown: https://moyomap.com/maps/open-source-ai-shifting.md License: HTML and markdown are for reading and citation. Bulk copying and AI training need a written license. See /terms and /developers#maps-data. Cite a claim as: MoyoMap, "Is Open-Source AI Shifting the Primary Economic Surplus Directly to End Users?: 36 Sourced Claims, Contested Facts & Evidence", claim {id}, https://moyomap.com/maps/open-source-ai-shifting.md#{id}. “Moyo, a product of SenTeGuard Inc.” ### A leaked May 2023 internal Google memo attributed to Luke Sernau argued open-source models would erase the co… Id: claim_08d10e90e25a81585f220b29 Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 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. Labels: Source-linked, Low, known ### Epoch AI data indicates model token price deflation varied by benchmark, ranging from 9x per year at the slow… Id: claim_0ea809ba907a9d05894ac06b Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 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. Labels: Source-linked, Low, known ### Mark Zuckerberg stated in a July 2024 letter that Meta open-sources Llama partly to commoditize the model lay… Id: claim_1711ae8dc6d0a8f229f39f91 Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 Mark Zuckerberg stated in a July 2024 letter that Meta open-sources Llama partly to commoditize the model layer and prevent competitors from taxing Meta's core business. Labels: Source-linked, Low, known ### Hugging Face surpassed 1 million public model repositories on its platform in 2024. Id: claim_1747e54b53967ffea17f33ee Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 Hugging Face surpassed 1 million public model repositories on its platform in 2024. Labels: Source-linked, Low, known ### A 2022 controlled experiment reported by GitHub found software developers completed tasks up to 55% faster wh… Id: claim_1daaf28781b16e6160867923 Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 A 2022 controlled experiment reported by GitHub found software developers completed tasks up to 55% faster when using GitHub Copilot. Labels: Source-linked, Low, known ### An OECD 2026 analysis reported that cloud-available open-weight text models reached approximately 90% of clos… Id: claim_5b0112a2b20ed8dd660e4b20 Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 An OECD 2026 analysis reported that cloud-available open-weight text models reached approximately 90% of closed-model quality while costing roughly 20% as much. Labels: Inferred, Low, known ### Andreessen Horowitz found enterprise LLM budgets grew roughly 2 to 5 times from 2023 to 2024 despite sharp dr… Id: claim_5ff6fb2f09024cacfad35cc9 Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 Andreessen Horowitz found enterprise LLM budgets grew roughly 2 to 5 times from 2023 to 2024 despite sharp drops in per-token inference prices. Labels: Source-linked, Low, known ### A 2023 NBER field study by Brynjolfsson, Li, and Raymond found generative AI assistance boosted customer-supp… Id: claim_659b28cc89f04ff3bb51227b Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 A 2023 NBER field study by Brynjolfsson, Li, and Raymond found generative AI assistance boosted customer-support worker productivity by 14% on average and by 34% for novice workers. Labels: Source-linked, Low, known ### AT&T used LiteLLM to implement complexity-based query routing, reducing costs on advanced coding tasks by up… Id: claim_7254075b5b7541a8c51a41bf Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 AT&T used LiteLLM to implement complexity-based query routing, reducing costs on advanced coding tasks by up to 56% with a 2% drop in quality. Labels: Inferred, Low, known ### Vercel's July 2026 AI Gateway Production Index showed open-weight models executed 29% of June gateway tokens… Id: claim_8523836fe1d00d211697d3d0 Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 Vercel's July 2026 AI Gateway Production Index showed open-weight models executed 29% of June gateway tokens while representing under 4% of total customer spend. Labels: Inferred, Low, known ### An arXiv analysis of 2024–2025 LLM pricing found that algorithmic efficiency progress drove price declines ro… Id: claim_8906138621429b1f64c898ed Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 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. Labels: Source-linked, Low, known ### Stanford AI Index 2024 estimated the training costs for GPT-4 and Gemini Ultra at approximately $78 million a… Id: claim_a7ac6790ae1e95b5d1f1aa55 Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 Stanford AI Index 2024 estimated the training costs for GPT-4 and Gemini Ultra at approximately $78 million and $191 million, respectively. Labels: Source-linked, Low, known ### Menlo Ventures' survey of 495 enterprise AI decision-makers in November 2025 revealed open-source models' sha… Id: claim_ba867986dda4afaf149c24b5 Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 Menlo Ventures' survey of 495 enterprise AI decision-makers in November 2025 revealed open-source models' share of enterprise LLM API spend dropped from 19% to 11% year-over-year. Labels: Inferred, Low, known ### McKinsey's 2024–2025 surveys reported over 65% of organizations regularly using generative AI, with cost redu… Id: claim_e52f93bb4efd016646b09ab5 Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 McKinsey's 2024–2025 surveys reported over 65% of organizations regularly using generative AI, with cost reductions concentrated in HR, supply chain, and service operations. Labels: Source-linked, Low, known ### Andreessen Horowitz's 2024 enterprise survey found approximately 46% of respondents preferred or used open-so… Id: claim_ea19c683ad6cf177cec69a04 Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 Andreessen Horowitz's 2024 enterprise survey found approximately 46% of respondents preferred or used open-source models in production deployments. Labels: Source-linked, Low, known ### Reasoning Models and Inference Evaluation Scaling Id: section_3fddb728f7beedd5b804872b Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 Reasoning Models and Inference Evaluation Scaling Labels: Inferred, Low ### DeepSeek Training Costs and Technical Metrics Id: section_563005c70d5afc4a51942ed4 Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 DeepSeek Training Costs and Technical Metrics Labels: Inferred, Low ### Market Impact of DeepSeek-R1 Release Id: section_7418bbad917ed775d8884fb6 Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 Market Impact of DeepSeek-R1 Release Labels: Inferred, Low ### Brynjolfsson Study on Generative AI Consumer Surplus Id: section_a21281db66be9894207bd7c2 Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 Brynjolfsson Study on Generative AI Consumer Surplus Labels: Inferred, Low ### Meta Llama Releases and Adoption Id: section_b4de42b4a807ed808e798b62 Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 Meta Llama Releases and Adoption Labels: Inferred, Low ### Economics of Enterprise Open LLM Self-Hosting Id: section_bc835a0eaed69e86335fe081 Parent: topic_wtBApG3xpZ5YzdAgcf5P Depth: 1 Economics of Enterprise Open LLM Self-Hosting Labels: Inferred, Low #### A 2026 Stanford working paper by Brynjolfsson et al. Id: claim_11d7cb04d4c3e9fab31a1734 Parent: section_a21281db66be9894207bd7c2 Depth: 2 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. Labels: Source-linked, Low, known #### Predibase reported that fine-tuned Llama-2 matched or outperformed GPT-4 on narrow classification benchmarks… Id: claim_1b62d83ac40b217eb7dae639 Parent: claim_5b0112a2b20ed8dd660e4b20 Depth: 2 Predibase reported that fine-tuned Llama-2 matched or outperformed GPT-4 on narrow classification benchmarks at approximately 1/100th the inference cost. Labels: Source-linked, Low, known #### Meta released LLaMA weights for research use in February 2023, and the weights leaked on 4chan in early March… Id: claim_1c36b3deed33f878a039392a Parent: section_b4de42b4a807ed808e798b62 Depth: 2 Meta released LLaMA weights for research use in February 2023, and the weights leaked on 4chan in early March 2023. Labels: Source-linked, Low, known #### Meta reported that Llama models surpassed 1 billion cumulative downloads by March 2025, rising from approxima… Id: claim_49cebb1174b7cc7cecda1fd3 Parent: section_b4de42b4a807ed808e798b62 Depth: 2 Meta reported that Llama models surpassed 1 billion cumulative downloads by March 2025, rising from approximately 650 million in December 2024. Labels: Source-linked, Low, known #### On January 27, 2025, the release of DeepSeek-R1 triggered a roughly $1 trillion intraday drop in US tech stoc… Id: claim_6154257d1d5403063edb5fff Parent: section_7418bbad917ed775d8884fb6 Depth: 2 On January 27, 2025, the release of DeepSeek-R1 triggered a roughly $1 trillion intraday drop in US tech stocks, including an approximate 17% ($589 billion) single-day decline for Nvidia. Labels: Inferred, Low, known #### Inference costs for GPT-3.5-level systems fell from $20.00 per million tokens in November 2022 to $0.07 in Oc… Id: claim_70615df8def21ef3c87166cc Parent: claim_0ea809ba907a9d05894ac06b Depth: 2 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. Labels: Source-linked, Low, known #### The cost of evaluating frontier LLMs on GPQA-Diamond, SWE-bench Verified, and AIME has increased exponentiall… Id: claim_829f155c5655b533448a71ef Parent: section_3fddb728f7beedd5b804872b Depth: 2 The cost of evaluating frontier LLMs on GPQA-Diamond, SWE-bench Verified, and AIME has increased exponentially at 3x to 18x per year due to increased inference requirements per task. Labels: Inferred, Low, known #### DeepSeek-R1 was released on January 20, 2025, under an MIT license, with DeepSeek claiming a base V3 model tr… Id: claim_8a48ad90d70833ba60db82bf Parent: section_563005c70d5afc4a51942ed4 Depth: 2 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. Labels: Source-linked, Low, known #### DeepSeek's consumer application reached the #1 spot on the US Apple App Store in late January 2025. Id: claim_de5d4828186447a961eb093d Parent: section_7418bbad917ed775d8884fb6 Depth: 2 DeepSeek's consumer application reached the #1 spot on the US Apple App Store in late January 2025. Labels: Inferred, Low, known #### Enterprise self-hosting of open LLMs becomes cost-effective relative to proprietary APIs above 8 to 30 millio… Id: claim_fc3134be282f1d7e2e10a29d Parent: section_bc835a0eaed69e86335fe081 Depth: 2 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. Labels: Contested, Low, known ##### Brynjolfsson et al. Id: claim_111873940e54c86369b2f1d6 Parent: claim_11d7cb04d4c3e9fab31a1734 Depth: 3 Brynjolfsson et al. concluded that generative AI consumer surplus substantially exceeds estimated producer revenues in the United States. Labels: Inferred, Low, known ##### Self-hosting open-weight models imposes an estimated 15% to 20% operational overhead for GPU provisioning, se… Id: claim_57820d13e1523337ade7d6ce Parent: claim_fc3134be282f1d7e2e10a29d Depth: 3 Self-hosting open-weight models imposes an estimated 15% to 20% operational overhead for GPU provisioning, serving infrastructure, monitoring, and prompt maintenance. Labels: Inferred, Low, known ##### DeepSeek reported consuming 2.788 million H800 GPU-hours at a cost of $5.576 million for the final training r… Id: claim_84373d653f710a77a61547b6 Parent: claim_8a48ad90d70833ba60db82bf Depth: 3 DeepSeek reported consuming 2.788 million H800 GPU-hours at a cost of $5.576 million for the final training run of its model. Labels: Inferred, Low, known ##### Reasoning models' average token output length has grown 5x per year compared to 2.2x per year for traditional… Id: claim_87320b66803c5e8d3440d9fe Parent: claim_829f155c5655b533448a71ef Depth: 3 Reasoning models' average token output length has grown 5x per year compared to 2.2x per year for traditional LLMs. Labels: Inferred, Low, known ##### On January 27, 2025, Larry Ellison's net worth declined by $24.9 billion and Jensen Huang's fell by $19.8 bil… Id: claim_90f63fcedbbdfcadf4670aa0 Parent: claim_6154257d1d5403063edb5fff Depth: 3 On January 27, 2025, Larry Ellison's net worth declined by $24.9 billion and Jensen Huang's fell by $19.8 billion. Labels: Inferred, Low, known