Thank you all for joining the webinar today. We're very excited to have with us a discussion between our CEO Ali Ghodsi and Sam Altman, CEO of OpenAI. So thank you very much both for being here.
Thank you. Yeah, excited.
Yeah. Well, let's get started and I hope we can have a fun, thoughtful discussion. You know, Databricks and OpenAI recently announced a first-of-its-kind partnership to make the OpenAI models natively available within Databricks and Agent Bricks. Maybe starting with Ali first: what's the importance of partnering with OpenAI on this endeavor?
Yeah, I mean, every one of our enterprise customers wants to use OpenAI. They all want to have the models available, use them on their enterprise data. And getting those two things working together is non-trivial because the data is sensitive. They want privacy, auditing, GDPR rights, but they also want to use the models to build agents and get insights. So it's really customer demand that has been overwhelming.
Yeah, we're thrilled to do this. Enterprise is becoming one of our biggest focuses. We've had 7x enterprise growth this year. We really think we're heading into the phase of AI where the models are getting so good that enterprises will need to use them, want to use them, bring them into their whole ecosystem. We cannot imagine a better partner than Databricks to make that happen. We're also very happy to be a Databricks customer.
Yeah, that's amazing. As you said, the partnership not only brings OpenAI models to all of our enterprise customers, but also OpenAI has been using Databricks for data analytics for some time. Maybe starting with Ali: how has that partnership and work been going?
It's awesome working with extremely technical teams that push us to our boundaries and demand more. It's nice that the two companies are so close geographically, so it's easy to walk over and iterate on the product. It's one of our most demanding customers in a very good way that made us much better as a company.
Maybe moving separately a little bit. Sam, you and OpenAI have really revolutionized AI for consumers. How do you think in this new chapter about going to enterprise, Databricks and OpenAI together can revolutionize AI for enterprise as well?
Yeah, we always planned to do enterprise too, but the models started with a lot of problems. They started fairly weak. It was easier to get consumers to adopt them. They're coming to the point in their maturity where we are clearly seeing huge enterprise demand. The need is to figure out how to bring this technology to enterprises in a way they can use it with all their data, constraints, and concerns about security and safety. But also to jump into this new world where AI will do an increasing amount of intellectual work in an enterprise. This feels like it's really the time. I think in 2026 and 2027, we'll see a huge transformation of how enterprises think about this. The example we can look at from 2025 is what's happened with how code gets written. Now imagine that for every other function of the enterprise. That would be a big deal, and I'm thrilled to go do that together.
Yeah, I'm super excited. There's a lot of context in the data that enterprises have, and it's very proprietary. The consumer side initially started with all the public data mankind has created over thousands of years. Now there's this data not available to the LMs, which we can bring together. Providing the proprietary data in the enterprise as context to the agents is going to be the big unlock. I'm excited to build that together.
Yeah. On that point, both Sam and Ali, on model capability reaching a point where enterprises make a lot of sense, the other piece is the age of agents arising. I want to ask both of you more on the research side. With AI research moving from building great frontier models to including context management, agents, and multi-agentic patterns and systems, where do you think future innovations across these research buckets will drive the move into enterprise?
I think there will be all these different research buckets. We'll keep pushing on pre-training. We'll get these multi-agent systems doing complicated things. Models will get smarter across the board. But there are maybe two axes to think about here. One has nothing to do with model intelligence—it's about how tightly you can integrate into an enterprise: all the knowledge, data sources, business processes. You could bring a brilliant physicist and drop them into Databricks, and maybe they wouldn't get much done on day one because they'd be missing that understanding. So this ecosystem layer—not just our companies but the whole world—will have to build integration into these models. That's a lot of work to teach a model to be a productive employee in a particular enterprise. The other category is that we have seen a dramatic increase in how long a horizon a model can work on a task. For a particular task with enterprise connection, how long until a model has a 50% chance of success? In coding, we've gone from 5-second tasks at the launch of GPT-3.5 to 5-minute tasks with GPT-4 iterations to 5-hour tasks with GPT-5. That's remarkable. But a lot of what an enterprise does requires months or years. Also, we can only currently do this in one vertical. Lengthening the horizon, giving models all the context inside an enterprise or the world, and adding more verticals will be the important thrust.
Yeah, the 50th percentile task completion time horizon is super interesting. With more context, you can lengthen that horizon. Something that would take a human a day, agents can now do. One cool thing is optimizing the context automatically. We developed a technique called JAU, inspired by genetic algorithms, that automatically gets enterprise context into the model without humans sitting there optimizing it. So you can live-feed the relevant context from different docs inside the enterprise. That's a really good metric, and the progression has been insane—from seconds to many hours now.
It always strikes me that model capability today is the dumbest it will ever be for the rest of my lifetime. That's an incredible thought when you think about the potential applications. As we mentioned, agents getting deeper into enterprise, operating for hours and potentially taking actions—what are the main things to consider for ensuring responsible AI and governance over the tools, data, and actions?
Yeah, we've worked very closely on that. We've built with privacy and security from the ground up. The whole partnership and integration builds on having guardrails in place: audit logging, access control, making sure the model is on-brand and not recommending competitors. You can build those guardrails in from the beginning. If you use GPT-5 inside Databricks, you get all that out of the box. I think this will become the fundamental limiter to AI adoption in the enterprise. It won't be about intelligence or price. Research and infrastructure teams will figure that out, but enterprises are starting to realize how critical this is. It will be the limiting reagent.
Yeah, 100%. I agree. Maybe this is a question for Sam: as Databricks can attest, the open-weight model GPTs have been a game changer for many companies in terms of capabilities, use cases, and customizability. OpenAI just released the GPT OS safeguards version as well. What's the future for open source and open weights from OpenAI?
There's clearly demand for it. Honestly, there's much less demand than for the most capable models we can run in a cloud. But we're a big-ish company, and we should do both. People want a model they control and can run on their device or system. There needs to be some way we support them.
Yeah. We're seeing huge demand for GPT-2 OS. People really want an American open-source model at the frontier. What are OpenAI's plans going forward? Is GPT-2 OS the last?
No, I hope not. We're trying to figure out how we can someday do a model of GPT-5 quality running on one device as an open-source weight model. We don't know how to do that yet, but we didn't think we'd get a model as good as GPT-OS running at 120 billion parameters either. I don't think this will be what most people want, but the people who want it really want it. We'll try to find ways to offer incredible open-source models.
Wow, that's pretty amazing. So you could run it on your laptop if you got that form factor down.
