Key points
- Arvind Jain co-founded Rubrik and now leads Glean, which crossed $300m in ARR and was last valued at $7.2bn.
- Jain says open source models already handle 80-90% of enterprise AI tasks and will keep closing the gap with frontier models.
- Jain says Glean underinvested in marketing and would be much bigger today despite training transformer models per customer in 2019.
Glean bets the context layer outlasts the models
Glean's bet is that enterprise AI is won on context rather than model quality. Arvind Jain started the company after watching productivity fall at Rubrik because employees could not find information or the right colleague. Glean connects to a company's systems, assembles context for each task and verifies answers against it. Jain expects open source models to handle most of that work, which leaves the context layer as the durable asset.
Takeaways
- Glean deploys a coworker across the whole company. Connected to a company's systems, it understands the business, its data, people and customers, and is deployed for every employee. It draws on over 100 models and offers its context layer to other AI tools.
- Every answer is verified against enterprise context. Glean assembles the context for a task, then checks each line of the model's response against it, redoing work that has no evidence and attaching references. Jain says models remain unpredictable and still make many mistakes.
- Confidentiality is enforced before the model runs. Glean treats all enterprise information as confidential, offers zero data retention with model providers, and adds runtime security, obfuscation and PII remapping. Customers choose from a hub of over 100 models, including self-hosted ones inside their firewall.
- Open source wins on enterprise cost and control. Jain says open source models already handle 80-90% of enterprise AI tasks, that AI has become expensive enough that open source is critical for cost control, and that adoption will catch up over the next 18 months.
- A router matches each task to a model. Glean picks models by quality and cost using customer criteria, which it can do because it sees tasks across thousands of enterprises and millions of tasks a day. Feedback is mostly implicit, from follow-ups and copy-pasting.
- AI will not shorten the work week. Jain says succeeding with AI makes people do more work, and that prosperity, culture and social norms change working hours rather than technology. He calls today's AI basic and no replacement for human judgment.
What Jain predicts next
- Four or five labs can build comparable frontier models, and serving models at scale stays hard enough to keep model providers valuable.
- Models become more specialised, with small distilled models matching frontier quality on a single task while running faster than a model built to do everything.
- Google looks behind on frontier models now, but Jain expects it to catch up, and the top labs' models will end up indistinguishable in capability and cost.
- Companies will design separate access modes for humans and AI agents, and the bigger shift is a flood of AI-generated content that is hard to authenticate.