BuffConf 2025: AI at the Edge, Video in Real Time

Where GenAI, monetization, and metadata are reshaping video infrastructure

This week I joined a few hundred engineers and architects at BuffConf, a new technical summit hosted at T-Mobile Park and SURF Incubator in Seattle by Momento, along with multiple sponsors. This wasn’t a vision fest or a “what if” retreat. It was a pragmatic, detailed look at what’s already working (and sometimes breaking) across the video stack as media goes real-time, AI-native, and monetizable at global scale.

Years ago, when I was deep in the weeds building encoders, I started thinking of myself as a media plumber, laying new pipes to move the water through the system. BuffConf was full of media plumbers. These are the folks who understand that getting video to work at scale isn’t about prompts or pixels. It’s about flow, pushing bits through real pipes with real constraints. I always seem to return to this metaphor because without great plumbing, the whole thing backs up, leaks, or fails in ways the audience never sees but always feels.

Generative Video Breaks the Stack (In All the Right Ways)

The standout lightning talk for me came late in Day 1 from Rushaan Mahajan and Sophie Raniwala of Sima Labs, who offered a surgical breakdown of how GenAI video fundamentally shatters traditional video infrastructure.

• Creation to Encoding: Diffusion models don’t obey physics, so the temporal coherence we’ve baked into codecs becomes a liability. We’re wasting bandwidth encoding “motion” that only exists in latent space. A fascinating shift for an old compressionist to consider.
  

• Encoding to Delivery: Personalized, ephemeral video breaks the assumptions behind GOP-based caching. What used to be “one-to-many” becomes “many-to-many,” and the infrastructure groans under the weight.
  

• Delivery to Experience: When generation is user-driven—think sub-200ms from prompt to pixel—you need model partitioning, adaptive bitrate in latent space, and watermarking at runtime. Post-production no longer applies.
  

Maybe the part I found most compelling was how they didn’t pitch a product. They offered a mental model for rethinking encoding, delivery, and experience when content is no longer captured, but synthesized, personalized, and distributed on demand.

Monetization Is Infrastructure Now

On Day 2, I had the opportunity to moderate a panel on Engineering Monetization and Personalization at Scale and what emerged was just how operational ad tech has become. FAST platforms, AVOD-SVOD hybrids, and personalized ad pipelines are no longer bolt-ons. They’re foundational.

From SSAI vs SGAI tradeoffs to CDN edge prefetching during live spikes, panelists Spencer Shanson (Paramount), Josh Lamb (Red Bull), and Phil Harrison (AWS Elemental) walked through the hard-earned lessons of keeping monetization stable under load, maintaining UX quality while targeting dynamically, and building redundancy into real-time ad decisioning systems.

Metadata and Modularity Win Everywhere

Across both days, one theme kept surfacing: video is becoming modular, and metadata is the protocol holding it all together.

From Zach Cava’s (Disney) talk on content fluidity to Matthew Scharr’s (AWS) breakdown of ad prefetching, the clearest path forward involves decomposing the video experience into addressable parts, treating assets like code, and orchestrating them with real-time signals. If your infrastructure doesn’t treat metadata as first-class or worse, treats it like an afterthought then you’re already behind.

What to Watch Next

BuffConf reinforced a shift I’ve been thinking about for a while. We’re moving from video as a file or signal to video as a programmable object. Distribution isn’t a fixed pipeline anymore. It’s just another form of orchestration.

And that shift breaks the old plumbing metaphor.

The pipes we built were designed for fixed flows—linear channels, pre-encoded VOD libraries, predictable ad breaks. But now, the water itself is dynamic. It’s changing shape in real time, reacting to every user prompt, every signal, every edge condition.

What does plumbing even look like when the liquid you’re routing is alive?

The engineers, product managers, and architects at BuffConf aren’t just maintaining infrastructure. They’re inventing. Testing. Prototyping a world where media behaves more like a system than a payload, where every stream is assembled on request, every frame might be modified mid-flow, and every viewer can trigger a different path through the pipe at any time.

None of this is the sexy side of AI-powered storytelling. But it’s exactly what makes the magic work. And the next wave of media innovation will depend entirely on how well these programmable, personalized streams can be routed, monetized, and delivered at speed, at scale, and without bursting the pipes.

After I left Microsoft, I started this Substack, which I’ve genuinely loved using to explore the intersection of AI and media tech. I also set up a consulting shop called Alchemy Creations.

It wasn’t my plan to become a consultant, but enough people kept asking to work together that I needed a way to say yes. What surprised me is how much I’ve enjoyed being more self-contained. It’s let me spend time with smaller groups and organizations that a company like Microsoft would never have let me focus on.

There’s a website, but it’s super generic and probably always will be. There’s also a LinkedIn page for Alchemy. Josh Schwerin runs it, so it’s more likely to be updated than the site. If you want to keep up with what I’m working on, or just see where this all goes, you might want to follow it.

Thanks to the team at Momento for making BuffConf happen and for keeping it focused on substance over spectacle. Huge appreciation to Khawaja Shams, Lionel Bringuier, and Mike Callahan and all of the Momento team for bringing this event to life. Hopefully we’ve just established a new, BS-free checkpoint where the people building the future of media can keep learning from each other.