Open Models Change The Economics of AI

Y Combinator · 2026-09-04 · 57 分鐘
https://www.youtube.com/watch?v=rY0wnfFHYbs影片總結
Ollama sees open models capture 80–90% of enterprise tokens as usage rises 150x
Ollama co-founder and CEO Jeffrey Morgan says enterprise AI is shifting rapidly toward open models, driven by coding agents, customization, and lower costs. AT&T has already moved 40% of token consumption to open models, while Ollama Cloud token usage has grown 150x since the start of the year. Chinese models dominate cloud usage globally, although local deployments show a closer US-China-Europe mix. Morgan expects open models to handle 80–90% of enterprise tokens, with frontier closed models reserved for the hardest tasks. He also highlights hybrid execution: local models for cheaper, lower-latency document workflows and cloud models for complex coding. Models such as DeepSeek Flash are approaching “unlimited token” economics, while Ollama increasingly acts as an operating-system layer connecting models, hardware, inference providers, and application harnesses.
章節
- 0:00Intro: Ollama’s 9 Million Developers and Enterprise Shift to Open Models
- 2:13Is It All About Cost?: AT&T Moves 40% of Tokens to Open Models
- 3:31The Token Usage Explosion: Ollama Cloud Grows 150x Since January
- 5:31Fine Tuning: DeepSeek Flash Iterations and Open Models for Security
- 8:26Launching Models at Scale: DeepSeek Multimodal Releases Require a 24-Hour Fire Drill
- 11:31Ollama as the OS Layer: Matching Any Harness to Any Model
- 13:58Hidden Layers Between Model and App: Knowledge, Coordination, and Execution
- 17:28Open vs. Closed: The Steady State Is 80–90% Open Tokens
- 20:41Local vs. Cloud Models: Quinn 3.8 38B Matches Opus 4.6 for Coding
- 24:00NVIDIA’s Open Source Play: DGX Spark Brings 128GB Models to the Desk
- 26:03The Return to Local: GB300 Brings Cloud-Scale Coding Loops Back to Desktops
- 27:15Getting GPUs Is Hard: B200 and B300 Access
- 29:16The Flash Model Revolution: DeepSeek Flash Enables Near-Unlimited Tokens
- 31:35God Model vs. Orchestration: Specialized Models Win Everyday Tasks
- 33:37The Geopolitics Question: Model Origin and Power-Plant Safety
- 37:12Applied to YC With the Wrong Idea: From Kubernetes SSO to Ollama
- 40:36Lost in the Wilderness for Two Years: Searching for a Developer Problem
- 42:15The Pivot That Changed Everything: Launching Local Llama in Two Weeks
- 44:49100K GitHub Stars, No Revenue: From Hobbyists to Fortune 500
- 47:40How Do You Monetize Open Source?: Privacy, Security, and Cloud Access
- 49:43Why Do YC as a Second-Time Founder?: Peer Support and Avoiding Mistakes
- 51:49What Seeing “Good” Actually Does for You: Learning Product Excellence
- 53:38Old DevOps Rules That Don't Apply Anymore: AI Makes Non-Determinism and Leaner Teams Normal
這是 Tier 1 公開摘要
每章重點、段落總結、心智圖由分享者控制是否公開。想看完整分析?自己提交一支。
同頻道的其他分析
Jeff Dean: The 1% Rule for Building in AIY CombinatorJeff Dean links Google Search, TPUs, Gemini agents, and founder strategy to 1,000x energy gaps and 0%-success niches.
Patrick Collison: Is AI Breaking the Lean Startup Playbook?Y CombinatorPatrick Collison says Stripe took two years to launch, while Stripe data shows AI-era startups nearly doubled year over year.
Waymo Co-CEO Dmitri Dolgov: The Demo Is Only 1% Of The WorkY CombinatorWaymo’s demo took 18 months, its product 15 years, now reaching 17× human safety across 220 million autonomous miles.
Garry Tan: Own Your IntelligenceY CombinatorGarry Tan argues personal AGI, G Brain, and Markdown skills let one founder achieve 400x leverage while owning their intelligence.
Max Hodak: Average Is Not Good EnoughY CombinatorMax Hodak explains how Science’s Helix infrastructure, hiring system, and retinal implant work turn speed into startup advantage.
相關主題的分析
Jensen Huang: The Doomer Hoax, Superintelligence is Here, and The Future of AI (ft. President Trump)All-In PodcastJensen Huang dismisses AI doomer predictions, backs open models, and says superintelligence already exists in narrow domains
EP316. Claude Sonnet 5、Meta 也要賣算力、PLTR 合作 NVDA | M觀點M觀點Anthropic 推出 Claude Sonnet 5,Meta 評估出租 GPU,Palantir 與 NVIDIA 搶攻企業私有 AI
EP323. Opus 5 正式發表、星艦第十三飛達標、開放模型的公開信 | M觀點M觀點Claude Opus 5成本降26%、Starship Flight 13完成九成任務,輝達等50家力挺開放權重模型
Peter Steinberger: "Fun Is Velocity"Y CombinatorOpenClaw grew from a WhatsApp relay into a 4.7-million-download open-source storm—and taught Peter Steinberger that fun drives velocity.
EP330. 輝達 GPU 要漲價、中國機器人跑很快、特斯拉新消息更新 | M觀點M觀點輝達GPU喊漲15%,中國機器人百米破博爾特,特斯拉更新Cybercab與Optimus進度