There is a massive shift in the world of AI infrastructure that people aren't appreciating enough. Every major AI infrastructure player is open-sourcing software that builds on top of, runs, and makes the most of their infrastructure.
Almost every single AI conversation in the world right now is focused on models and the specs of new metal. But the most important shift in AI infrastructure this year has nothing to do with which model is winning. It’s that the companies building the chips are giving away the software that runs them. And that hands platform engineering a new decision, one that most teams are not prepared at all to answer.
Why the chip giants are opening up
For two decades, the assumption was simple. The chip was the product, the software stack on top of it was the moat, and you paid for both. NVIDIA built the most valuable company on earth on exactly that logic. So when NVIDIA ships Dynamo, its inference engine, as free and open source software, and open sources NeMo across training, data curation, guardrails, and agents, something structural is happening. And since Nvidia is the world's most valuable company, something tells me it's more than just philanthropy.
At the same time, AMD has built its entire competitive identity around the same move. ROCm 7, the AMD Enterprise AI Suite, an open-source Rocky Linux build for AI, all of it open. What’s happening here? Well, Raimine Roane, who leads AMD's AI Solutions Group, put the logic perfectly,
"It's whoever is going to write the new kernels faster who is going to win. And we're going to win because we work with open-source." The chip vendors have worked out that ecosystem control now matters more than a proprietary kernel. You do not win the next architecture by locking the door. You win it by being the default everyone already builds on.
The industry has a name for this, and it is not new. Jim Zemlin, who runs the Linux Foundation, has been blunt about it, "AI is having that moment right now." He means the Linux moment, the point where an open stack produced by a community out-competes the biggest proprietary vendor in the category. Jensen Huang has started saying the quiet part out loud too. At GTC this year he told the room, "For the CEOs, the question is, what's your OpenClaw strategy? We all have a Linux strategy." When the CEO of the most powerful chip company on the planet and the boss of one of the world's most important open source foundations start saying the same stuff, the direction of travel is set.
The misread
Here is where the shift is easy to misinterpret. The temptation is to treat open source at the silicon layer as a cost story, or as a license to assemble everything yourself. Every component is open, every layer swappable, so why not build the bespoke stack from parts?
The catch is that optionality carries a cost. Once your platform team owns the integration of an open inference engine, orchestration layer, model runtime, and GPU drivers, that becomes a full-time project competing with the actual work. Open source may remove license cost, but it doesn't remove the integration burden and the risks that come with that. We are very familiar with that story in platform engineering.
The platform engineering read
This is a platform engineering problem at heart, and the discipline has seen many versions of it before.
Composable AI infrastructure is the Internal Developer Platform pattern, refined for years on cloud-native, applied to AI workloads. The point of an IDP was never infinite choice. It was the opposite. Centralize the cross-cutting concerns, settle on a few paved roads, and hand developers an opinionated starting point so they ship instead of assembling. Golden paths, not a parts bin.
The State of Platform Engineering Report backs this up. Across 100s of platform teams, 68% prioritized standardizing infrastructure provisioning through an Internal Developer Platform and 65% improving developer experience. Teams want fewer, better knobs. The chip vendors going open is what finally makes an opinionated, swappable AI stack possible.
Kevin Cochrane, from Vultr frames the destination well, "It's a new stack but it's an open stack." Both halves matter. New, because AI workloads too frequently break the abstractions cloud-native taught us to rely on. Open, because a closed equivalent risks trading one lock-in for another.
The risks
None of this comes without strings. Open source from a single dominant vendor is still shaped by that vendor, who steers the roadmap and can stop maintaining it if priorities change. The openness also serves them, since it is a route into their ecosystem, and an opinionated stack can become its own soft lock-in. That does not make it a bad bet, but the word open is worth reading closely, and it is fair to ask who controls a project before you build on it.
This warning does not mean we should discount this trend. There are ways to take advantage of the open source, while minimising risk. Vultr's Cloud Alliance is one example, with pre-validated, composable stacks for NVIDIA and AMD workloads. The idea is that an opinionated, supported starting point keeps much of the open-source upside while taking some of the integration and abandonment risk off the platform team.
Why this is interesting to platform teams
Two things in particular are drawing teams in.
The first is sovereignty. When the inference path is open source, a regulated buyer in healthcare or finance can inspect it, and a sovereignty requirement becomes a placement decision rather than a blocker, which is harder with a fully proprietary alternative.
The second is the skills gap. Our research finds that 57% of organizations cite skills gaps as their top barrier to going AI-native. That helps explain the pull toward curated stacks. A supported path asks less of developers than a pile of components and good luck. Whether the trade is worth it is for each team to weigh. But something tells me this won’t be the last time you are hearing about this story.
We are pulling the full picture together, the chip-vendor playbook, the composable-stack pattern, and what it means for teams making real infrastructure decisions, into a report landing soon. Subscribe to hear when that research comes out.
This article is sponsored by Vultr as part of PlatformCon 2026
