The TFD Experience
About a week ago I attended my first Tech Field Day event, specifically AI Field Day 8, and I’m still processing it. I’ve been to many technical conferences: I’m a regular at Cisco Live, attended VMware events, Microsoft TechEd back when that was a thing, etc. etc. Those events serve a real purpose and I always come home with something useful, but there’s an undercurrent at vendor-sponsored conferences that you can’t really escape: the programming is ultimately structured around making the primary sponsor look good. Everything is shaped a little in that direction, and that’s fine, it’s just the nature of the format.
AIFD8 was a different animal entirely. The format is closer to a very technical user group, with a small group of independent delegates sitting down with vendors who are there to be challenged, not applauded. The Q&A is genuine, the dialogue is real, and the follow-up questions get pointed. The intimacy of the setting creates space for the kind of technical depth that is really hard to achieve any other way, and I have to commend the TFD team for cultivating this environment.
What I wasn’t fully ready for was the caliber of the other delegates. I was sharing a table with people who have been at the frontier of data, AI/ML and bleeding-edge research for decades. The quality of the questions being asked, the frameworks people brought to the discussions, and the genuine intellectual generosity in how they shared their perspectives – it was, truly, humbling, and the opportunity to learn from them gave me a much wider context on which to form my own opinions. I can say with confidence that this is perhaps the best technical networking event I have attended, and I’m already looking forward to attending TFD again in the future.
The Technology
Six vendors presented, and I’m going to try to give each of them an honest summary without turning this into a book report.
Selector opened the event with a presentation on ML-driven network observability. The concept is that traditional monitoring tools produce alert fatigue and slow MTTR, and their platform uses autocorrelation across the full telemetry stack (network, application, infrastructure, cloud) to surface root cause in seconds rather than hours. The architecture they walked through was genuinely interesting: a multi-agent system built on MCP with specialized domain agents for things like firewalls, load balancers, and cloud observability, all orchestrated through a general contractor/subcontractor model that I found pretty intuitive as a mental model. The target market here is Fortune 1000 scale orgs, so your mileage will vary significantly depending on environmental complexity – that said, this approach to ops and observability has a lot of validity across the entire market.
Cisco presented on intelligent observability and agentic ops for data center networking, and I want to be deliberate here because I’m planning a dedicated recap on what they showed. The short version: the AI Canvas work is the most interesting vision I’ve seen for how AI can actually improve operational UX in a meaningful way, not just add a chatbot on top of existing tooling. More on that soon.
HPE covered their compute portfolio and the ProLiant roadmap for AI workloads. The story they’re telling points toward something real: datacenters can only scale so far, token costs are a genuine constraint, and local LLM hosting at the edge and on-premise is increasingly where the puck is going. HPE is clearly positioning to answer that need, and for organizations that are thinking ahead about where AI inference lives long-term, that positioning is worth paying attention to.
Hammerspace caught my attention with a problem framing I don’t hear articulated well very often: the data isn’t in one place, and moving it is expensive and slow, so what if you didn’t have to? Their software-defined platform creates a global namespace across existing storage (NAS, object, even local NVMe on GPU nodes they call “tier zero”) without requiring data migration. The value proposition for AI workloads is that you can feed your GPU clusters from wherever the data already lives, using standard protocols, without buying a new box and reshuffling everything into it first. For organizations sitting on distributed legacy storage who are trying to get AI pipelines running without a rip-and-replace project, that’s a practical answer to a practical problem.
Scality announced their ADI (Autonomous Data Infrastructure) product at the event, which is their answer to the storage software challenge at very large scale. They’re positioning it as hardware-independent, open-code (inspectable rather than fully open source), and built around AI-assisted operations through what they call Guardian. The idea is that storage teams aren’t growing but storage volumes are, and you need the platform itself to help close that gap. The data sovereignty angle they kept returning to is also real, especially for GDPR locales and regulated industries.
Solidigm, like Cisco, is getting its own recap post soon. Their session on the anatomy of a prompt and what that means for storage, software, and overall system design was the most rigorous bottom-up technical content I’ve seen on the subject. If you think storage is a procurement afterthought in AI infrastructure planning, Kapil Karkra will fix that belief in about forty-five minutes.
The Takeaways
Stepping back from the vendor content: the thing that struck me most across all six sessions was the degree to which the same set of concerns kept surfacing in Q&A regardless of what was being presented:
- How do we keep humans in the loop?
- How do we implement and enforce RBAC in a world where agents are acting on behalf of users?
- How do we think about identity when we have a new class of non-human interaction that goes well beyond traditional service accounts?
These aren’t hypothetical governance questions anymore, they’re design requirements that practitioners are being asked to solve right now, and the honest answer from the room was that the tooling and the frameworks are still catching up to the adoption curve.
I came into this event with AI as a surface-level area of expertise for me personally, and I left with a clearer sense of a few things. The pace of change is real and it isn’t slowing down. Tokenomics are a genuine architectural constraint, not just a cost concern, and agentic workloads are making that more acute. The “just build more datacenter” answer is running into physics and power grid limitations in ways that are forcing more creative thinking about efficiency, local inference, and smarter use of existing infrastructure. And for anyone who spent the last decade migrating workloads to the public cloud, the question of whether you’ve just traded on-prem CapEx for token spend is one worth sitting with.
Data truly is the currency of AI. That sounds like a slogan, but it kept proving itself true in session after session. Where the data lives, how fast you can access it, who is authorized to see it, and how you prove that, those are the real infrastructure questions right now. I’m glad I was in the room for this one and am looking forward to the next.
CCIE #TBD 