PRACTICE AREAS

The future of AI-native is open source

Companies worth trillions of dollars have spent the last 18 months giving away the software that makes their hardware worth buying. That is a bet on ecosystem control, and its consequences land on platform teams, who now own the layer every AI initiative in the company runs through. This report, produced by Weave Intelligence and commissioned by Vultr, sets out what that openness hands platform teams, what it leaves them to finish, and how to evaluate the curated stacks now emerging on top of it.

About the report

The report examines this phenomenon with examples from twin giants NVIDIA and AMD, who opened their AI software layers from opposite competitive positions. It details what that shift means for platform teams, and how to judge an opinionated stack against your own constraints. It includes a direct comparison of two production stacks published in 2026 and a five-point framework for evaluating whatever is put in front of you.

Key takeaways

Openness hands platform teams capability rather than product, and the integration work it creates is now their work, whether or not anyone has said so out loud.
Accelerator choice has become reversible, which turns it from an inherited procurement default into a decision the platform team can actually make.
Not all open is equally open: free software tuned for a single vendor's hardware sits a long way from permissive licensing with full modification rights.
A side-by-side comparison of an NVIDIA-based and an AMD-based stack shows the accelerator changing while everything above it stays constant, which is the composability property worth testing for.
Five criteria narrow the shortlist fast: accelerator flexibility, regional autonomy, Kubernetes governance integration, position on the openness spectrum, and operational investment profile.

Key takeaways

Openness hands platform teams capability rather than product, and the integration work it creates is now their work, whether or not anyone has said so out loud.
Accelerator choice has become reversible, which turns it from an inherited procurement default into a decision the platform team can actually make.
Not all open is equally open: free software tuned for a single vendor's hardware sits a long way from permissive licensing with full modification rights.
A side-by-side comparison of an NVIDIA-based and an AMD-based stack shows the accelerator changing while everything above it stays constant, which is the composability property worth testing for.
Five criteria narrow the shortlist fast: accelerator flexibility, regional autonomy, Kubernetes governance integration, position on the openness spectrum, and operational investment profile.