Lunch with CTO of The Linux Foundation & PyTorch Foundation Matt White

Last month at MLSys 2026 in Bellevue, I had lunch with Matt White — Global CTO of AI at the Linux Foundation, CTO of the PyTorch Foundation, and CTO of the newer Agentic AI Foundation (among a long list of open-source roles, from LF AI & Data to the Generative AI Commons). It was the kind of conversation that only happens in person, and it had been a couple of years in the making.

With Matt White at MLSys 2026, Bellevue, WA.

How we (finally) met

Matt and I first connected on LinkedIn back in 2024, when he was CTO of LF AI & Data and the large-model era was just beginning (ChatGPT had only launched in November 2022). That year we were both at the PyTorch Conference in San Francisco and somehow missed each other. So I kept up the other way: reading what Matt writes and shares, most recently his excellent field report, "Eight Days in China: A Field Report from Inside the Labs."

Because I've spent the last few years contributing to the Linux Foundation community, and we already knew each other, lining up a meeting during MLSys week came together naturally. We bumped into each other at the venue on 5/18, and set a lunch for 5/19.

The booth and the brunch

When I found Matt on the 19th, he was still at the PyTorch booth — hands-on, debugging a stubborn USB connection on the demo screen. Very on-brand for someone who runs foundations and still fixes the hardware. I pulled him away and we walked over to Cafe Hagen, a cozy brunch spot nearby (avocado toast included), where we traded notes on what each of us has been building lately.

Every foundation needs an "anchor project"

Matt talked about his work at the Agentic AI Foundation and his earlier years at IBM. I asked the question I'd been saving: how does he see the Silicon-Valley-hot space of world models and robotics — the kind of work Fei-Fei Li and others are pushing into multimodal, spatial intelligence — and does the Linux Foundation have plans there?

He's studying it, he said, but the honest answer is that a credible open movement needs an anchor project to rally around first. The pattern repeats across the open ecosystem:

A movement coalesces around one indispensable project. World models and robotics don't have their open anchor yet — finding (or seeding) it is the hard part.

He shared sharp observations about which companies are quietly building these capabilities in-house versus leaning on open foundations — and where the gaps are.

Open-weights frontier models & the supply chain

That led us into open-weights frontier models. Matt's read: the U.S. doesn't really have an open-weights frontier model right now — Google's open releases being the notable exception — and the country still needs to build up the supply chain around open models (data, weights, tooling, evals, distribution). His point was that China is doing this remarkably well: a dense, fast-moving open-weights ecosystem. It reframed "open source AI" for me as less of a licensing question and more of an industrial one — a supply chain and a community, not just a repo.

Eight days in China

Naturally I asked about his China trip. He lit up — visiting labs and companies across Hangzhou, Shenzhen, and beyond, and (of course) the food. If you haven't read his field reports, they're the most grounded "inside the labs" accounts I've seen this year — see also "From Code to Cognition: How LLMs & VLAs…". He also mentioned he'd be heading back to China soon after MLSys, for BEYOND Expo. Afterward we walked to the Starbucks near the conference hotel and grabbed drinks.

A French grandmother, a phone, and Gemini

The best moment wasn't on the agenda. On our way to lunch, an elderly French lady stopped us — her English wasn't fluent — holding up her phone with Gemini open, asking how to get to the Old Navy at Factoria. The mall turned out to be far from where we were standing. I don't speak French; Matt did, and helped her sort it out.

We laughed about it afterward: AI has genuinely reached every generation and woven itself into daily life — a French grandmother navigating a foreign city with a model in her pocket. That's the adoption curve no benchmark captures.

Again, at the Together AI meetup

That evening I ran into Matt once more — at the Together AI meetup. We picked the thread back up on open-source AI community building: how to grow contributors, how to keep neutral governance, and how to give a new direction its anchor.

Two takeaways I'm still chewing on: open source AI is a supply chain, not a checkbox; and every new direction — world models included — waits on someone willing to build its anchor. Grateful to Matt for the time, the French translation, and the perspective.

Build in the open. Then bring people along.