The Twenty Minute VCWhy OpenAI and Anthropic Won't Win the App Layer | Glean Founder
At a glance
WHAT IT’S REALLY ABOUT
Why enterprises shift to open models and context-first AI apps
- Enterprises fear frontier model providers not just for data exposure but for operational dependence and loss of long-term “institutional learning” that agents accumulate over time.
- Open-source models are hitting an inflection point where most enterprise use cases can be served by multiple models, pushing commoditization and strong price pressure on the model layer.
- In enterprise AI, ROI is uneven: some functions like customer support show measurable gains, while broad areas like software shipping speed remain hard to prove despite near-universal AI-assisted coding.
- Most enterprise AI failures come from poor context assembly and “brute force” agent approaches that waste tokens, slow execution, and raise costs rather than delivering throughput gains.
- Workforce impact will be shaped less by full job replacement and more by power-law adoption, composite roles, and the disappearance of narrow execution-focused analyst and recruiting sourcer tasks.
IDEAS WORTH REMEMBERING
5 ideasThe real enterprise risk is losing agent-built institutional learning, not just leaking data.
Jain argues that as agents do more work, the compounding process knowledge lives inside the agent; if a vendor controls that agent, the enterprise becomes operationally dependent and may lose ownership of how work is done.
Open source is becoming “good enough” for most enterprise workloads.
He claims 90%+ of enterprise use cases can be handled by many models (including open source), implying the differentiator moves up-stack to context, workflow design, governance, and product experience.
Open source pull is currently more about cost than privacy.
Early enterprise anxiety about model providers training on their data has eased with contracts; today, runaway AI budgets and per-token expense are the main catalysts for considering open-source inference.
The biggest blocker to enterprise ROI is context throughput, not model intelligence.
Many deployments connect tools crudely and let agents “brute force” retrieval, burning tokens to assemble context; investing in structured context (permissions, relevance, workflows) makes agents faster and cheaper.
Model pricing power is likely weaker than hype suggests.
Competition among labs plus open source creates pricing pressure; Jain expects inference costs to fall by orders of magnitude, making pure model businesses less lucrative than many assume.
WORDS WORTH SAVING
5 quotesIn the future, all of that institutional learning, uh, is actually going to accumulate in that agent that is doing that work.
— Arvind Jain
I think right now the open source drive is coming from the cost point of view.
— Arvind Jain
Once you move towards consumption, there's no inherent bundling advantage.
— Arvind Jain
It also becomes very, very costly because most of the tokens are being burnt, um, just trying to assemble the right context for that given task.
— Arvind Jain
I think, uh, it's a, it's a wrong goal in my opinion to say that, "Hey, like, replace yourself with AI."
— Arvind Jain
High quality AI-generated summary created from speaker-labeled transcript.
