"cost is by far the largest pain point that open models can jump in and solve but you know every business has a vision of getting better control over AI and customizing it for their business and that's really their north star" - Jeffrey Morgan [00:00:00]
"seeing out of the box open models being served that really only took off at the start of this year" - Jeffrey Morgan [00:05:24]
Disclaimer: Orignal content owned by or sourced from third parties. It does not represent the views of 'Nuggets' platform or it's team. AI is used extensively across this platform including for summaries. Accuracy is not guaranteed, there can be mistakes. Any info or content on this platform is not a financial, legal, or investment advice. Do your own research. Refer for complete disclosures:- Terms of Use · Full Disclaimer
"these LMS are never perfect and like in the systems world you want everything to be exactly as it's designed to run it's tested it's validated but LMS by definition are not that's a feature not a bug" - Jeffrey Morgan [00:54:15]
"there's a scarcity in bringing that together into something that works" - Jeffrey Morgan [00:56:55]
Speakers & Credentials
Jeffrey Morgan: Co-founder and CEO of Ollama, an open-source framework and cloud service used by over 9 million developers and 85% of Fortune 500 companies to run open AI models locally and in the cloud.
Hosts (Y Combinator - The Lightcone): YC Partners leading discussions on frontier AI tech, ecosystem shifts, and startup economics.
1. Executive Summary
Enterprise AI adoption is experiencing a dramatic structural pivot from closed API monoliths toward open-weight models driven primarily by immediate cost reduction and long-term data control [00:01:23].
Developer behavior reveals two massive inflection points in token consumption: an initial surge driven by coding agents (Kimi, GLM, MiniMax) followed by a massive secondary surge powered by long-running task automation agents like OpenClaw and Hermes [00:04:08].
Per-developer token volume on Ollama Cloud grew 5x following agentic workflow launches, while aggregate token platform volume surged 150x year-to-date [00:04:46].
Large enterprises like AT&T have already shifted up to 40% of their total token consumption to open models, highlighting enterprise willingness to replace proprietary endpoints [00:02:50].
Open source model release cycles have compressed from six-month intervals down to weeks, forcing a paradigm shift away from traditional systems engineering toward accepting non-determinism as a core feature [00:06:11].
Value capture in the open AI ecosystem is migrating toward curation layers that unify fragmented models, hardware architectures, and unstable inference APIs into dependable developer experiences [00:56:03].
2. Chronological Table of Contents
[00:00:00] Macro Drivers for Open Model Adoption: Cost vs. Customization
[00:01:14] State-of-the-Art Open Models & Enterprise Token Flows
[00:02:44] Case Study: Enterprise Token Migration at Scale (AT&T)
[00:55:58] Curation as the Primary Value Layer in Abundant AI Infrastructure
3. Detailed Thematic Summary
Enterprise Token Economics & Open Model Migration
High operational costs of proprietary closed APIs act as the initial trigger forcing enterprises to migrate workloads to open-weight models [00:00:00].
Enterprise "North Star" strategic vision centers around getting granular control over model weights, data privacy, and custom alignment rather than relying solely on third-party API providers [00:00:06].
Major enterprise players such as AT&T have already successfully transitioned 40% of their production token consumption to open-weight models [00:02:50].
Open model adoption is heavily driven by both US-developed architectures and Chinese-origin models (e.g., DeepSeek, Kimi, GLM, MiniMax), with heavy global consumption concentrated across the US and Germany [00:01:23].
The Agentic Token Explosion: Coding & Co-Worker Workflows
Agentic workflows represent a fundamental step-function shift in developer token consumption compared to standard chat interfaces [00:03:05].
The first major inflection point in per-developer token usage was triggered by coding agents leveraging open models like Kimi, GLM, and MiniMax [00:04:08].
The second massive inflection point occurred with autonomous co-worker agents such as OpenClaw and the Hermes project, which automate extended multi-step workflows for non-technical departments (finance, support, sales, marketing) [00:03:19].
Autonomous agents expanded context usage from legacy 128K windows to over 1M+ tokens while performing tool calls and fetching external data, driving a 5x jump in individual user usage (from ~15M weekly tokens) and a 150x total platform growth year-to-date [00:04:33].
Model Iteration Speed & The Fine-Tuning Cycle
Model development cadences have drastically compressed; open-source release cycles shifted from 6-month gaps down to iterations arriving every few weeks (e.g., DeepSeek Flash iterating 3 times over a single summer) [00:06:11].
Hyper-fast base model iteration creates a strategic dilemma for fine-tuning: custom fine-tuned weights risk being rendered obsolete ("stomped") by the next base model release unless modern tooling allows instant retraining [00:00:27].
Off-the-shelf open models are increasingly capable out-of-the-box for complex tasks without requiring immediate custom fine-tuning [00:05:24].
Engineering Mindsets & Curation as Value
Software engineering rules from Infrastructure 1.0 (2000s) and Cloud 2.0 (2010s) break down in the AI era, requiring engineers to embrace non-determinism as a core system feature rather than a bug [00:54:15].
Moving up the tech stack closer to end-developers creates massive enterprise value, dismantling old infrastructure assumptions about application-layer vulnerabilities [00:54:03].
In an era of model and hardware abundance, true scarcity lies in curation—shielding developers from inference provider errors, broken JSON parameters, and fragmented APIs by offering unified, seamless developer tooling [00:56:03].
The Tactical-to-Strategic Open Model Pipeline: Organizations initially adopt open models as a tactical cost-cutting measure for immediate token margin relief. However, this serves as an onboarding bridge toward their ultimate "North Star": establishing full control over weight ownership, data privacy, and custom domain alignment [00:00:00].
Non-Determinism as a Feature (Systems Engineering 3.0): Traditional systems engineering prioritizes absolute validation, deterministic paths, and strict unit testing. AI systems flip this paradigm—non-deterministic outputs are the exact mechanism that enables creative problem-solving and reasoning in agentic tasks [00:54:15].
The Abundance/Scarcity Inversion in AI Infrastructure: As base models, raw compute, and underlying inference APIs proliferate into an abundant commodity, developer friction increases due to API inconsistencies and provider bugs. Value shifts decisively from raw model supply to curation layers that synthesize fragmented infrastructure into clean endpoints [00:56:55].
6. Anecdotes
AT&T's 40% Open Token Shift: Featured in The Information, AT&T quietly migrated 40% of its total company token consumption away from proprietary models to open models for developer automation and coding agents, demonstrating that large legacy enterprises are actively proving out open model unit economics at scale [00:02:44].
The OpenClaw Step-Function Explosion: Ollama tracked individual developer usage graphs expecting aggregate curves, only to discover single-developer weekly token graphs jumping exponentially by 5x (from ~15M tokens) as non-developers adopted OpenClaw and Hermes agents for multi-hour autonomous tasks [00:03:39].
The Inference API Minefield: Developers attempting to query raw inference providers directly encounter obscure undocumented errors (e.g., rigid JSON formatting bugs), causing hours of debugging until abstraction layers smooth out the underlying interface [00:55:26].
7. References & Recommendations
Companies & Platforms
Ollama: Open-source infrastructure and cloud service for running and deploying open-weight AI models locally and at scale [00:00:50].
AT&T: Major telecommunications conglomerate cited for shifting 40% of its token load to open models [00:02:44].
NVIDIA / AMD / Intel: Hardware chipmakers whose architectures Ollama unifies for local model execution [00:01:49].
OpenRouter: Model routing platform referenced as a key consolidation layer allowing developers to access disparate models under a unified billing umbrella [00:56:29].
Cursor: AI-native IDE referenced for fine-tuning open-weight models specifically for custom code generation [00:05:22].
AI Models & Open Source Projects
OpenClaw & Hermes Project: Autonomous agentic frameworks capable of performing multi-step tasks across extended durations for technical and non-technical business functions [00:01:30].
DeepSeek / DeepSeek Flash: High-performance open-weight model family highlighting rapid release cadences [00:05:16].
Kimi, GLM, MiniMax: Chinese-origin open models that significantly catalyzed the coding agent boom [00:04:08].
Media & Publications
The Information: Tech journalism outlet that published the enterprise case study detailing AT&T's adoption of open AI models [00:02:44].
Sep 11, 2026
How Open-Source is Reshaping the AI Infrastructure Stack
1. Executive Briefing TL;DR Open Source AI Trade offs & Open Weights vs. Open Source: Open weights models do not equal true open source. True open source AI requires open data, full infrastructure stacks, and reproducible training pipeline…
Weekly per-developer token consumption prior to agent boom