Self-Hosted Deployment Maintenance Mode

tl;dr We have decided to place our self-hosted offering in maintenance mode, and offer a new single-tenant hosted offering that can expose our agentic capabilities.

A little bit of history and context first.

Two years ago, we released the first enterprise offering we ever had, which was a self-hosted deployment. We released this even before we had a hosted cloud solution. Why? Well, our background as a GPU infrastructure company meant that we could uniquely build such an offering, and at a time when that was the entirety of our expertise, we saw an opportunity to work with large enterprises that were in regulated industries or had other security-related constraints. It was also early in the generative AI era, and it would have been harder at the time, as a tiny startup with essentially zero security certifications, to gain much traction with a hosted solution, even with enterprises that weren’t strictly regulated. It made sense, and we were, by most metrics, highly successful with this offering.

However, two years in the generative AI space might as well be two decades in normal software development. A number of different factors have progressed simultaneously:

• We became a lot better at creating offerings that were intermediate between fully commercial-cloud hosted and fully self-hosted, and our security posture massively improved (not to mention size and stability of the company). We created a Hybrid deployment to solve the persisted data concerns, we became the first AI code assistance platform to receive FedRAMP High authorization to support highly regulated industries (not just SOC2 Type 2!), and we launched a cluster in Germany to satisfy European regulations around data residency.
• The underlying technology continued to progress. The frontier models of today dwarf the largest foundation models two years ago in size, context length, and practically any other metric you can think of. It has come to the point that hosting even our best Tab model would require a full rack of frontier GPUs to serve, and those are our smaller models. This stops making sense from a TCO perspective for almost all companies. Connecting to a third-party endpoint for models from the foundation labs has essentially become a requirement for our self-hosted deployment. And as users try these tools in their spare time on personal machines, they start noticing the delta in performance from using smaller, previous generation models under the hood.
• The fork in capabilities that we could offer on self-hosted and hybrid/cloud grew. For quite a while, even if it was with different sized models, we could provide similar sets of capabilities between our hosted Cloud/Hybrid solution and our self-hosted offering; However, with the advent of Cascade and the new cutting edge agentic capabilities, which simply cannot be run entirely in a self-hosted environment, there is a very meaningful delta in what we can offer our customers between our different deployments. We no longer can say honestly that we are providing our self-hosted customers even close to the best of what we have to offer.
• Companies have become more comfortable with Cloud/Hybrid solutions that provide guarantees around zero-data retention. Maybe it is just time and familiarity with the technology, maybe it is our new certifications and more stability as a company, or maybe it is an understanding of the kinds of agentic capabilities that are restricted to these offerings. Likely some combination of all three.

Combining all of these together, we had to make a tough decision: in a world where self-hosted will, within time, not be the answer for over 99% of developers, do we (a) continue to invest in building on the self-hosted deployment knowing that we will likely lose almost all such customers over time, or (b) double down all of our resources to maximize the value we can drive to the vast majority of developers at the vast majority of our current and future customers.

While we stuck with (a) for a while, partially because a meaningful chunk of our customers have trusted us with that offering over the last couple of years, we eventually came to the realization that we would actually be doing our enterprise customers a disservice if we did not adapt to the times and adopt option (b). Given our certifications, security posture, GTM and partner motions, and product-level bets (see Wave 8), we hope that it is still clear that we are still primarily focused on the enterprise.

And we are not stopping in iterating on our security posture. Today, we are also simultaneously rolling out a single-tenant hosted offering as an additional security guarantee to our largest customers. This does require a minimum commitment given the hardware costs on our end, but we recognize that multitenancy is the second most common objection to Cloud solutions after data retention.

By maintenance mode, we will of course continue to support our current self-hosted customers until the end of their term and provide great financial incentives to switch to and adopt our other deployment offerings, but we are no longer investing in feature development or bringing on new customers to the self-hosted platform.

We will also never say never. Maybe one day in the future, there are new changes in the tech in a way where self-hosting becomes possible and more meaningfully equivalent to the hosted deployment options. Our customers are not just betting on our offerings today, but on the ability of our team to adapt to the rapidly-changing landscape of this space to consistently deliver the frontier.

We are grateful for the success that the self-hosted offering enabled us up to today, but are even more excited for what the future holds.