In Post 6 I wrote that a keynote you skip is gone for good, and that no amount of reading about it afterward puts you in the room. I'll stand by the first half. The second half needs an asterisk. The replay doesn't put you in the room either, but it does let you pause.
The plenary was Monday at four, the hour I spent on Level 3 chasing Odyssey points. On replay it runs under its official title, Shaping the Future of Private AI Cloud and Agentic Innovation. Most of it is what the press already wrote up: hardware costs, private AI on infrastructure you already own, and an AI assistant that patches ESXi without restarting a VM. One segment is the reason this post exists.
Acme Insurance
My day job sits in the risk management sector, so the Acme Insurance segment was the one I leaned in for.
Acme is fictional, the company you call when your car is in a ditch or your basement is underwater. Sabina Anja, the chief technologist for VMware Cloud Foundation, plays its platform admin. Four groups at Acme run on AI: claims, underwriting, finance and product engineering. They share one inference platform. Her job for the next few minutes is to hand those teams a new model without giving up her guardrails.
Watch what she does, because it's the whole job in miniature.
She imports NVIDIA's new open-weight Nemotron 3.5 from a Hugging Face link, and it lands in Acme's own registry, versioned and scanned like any container image. Every copy that serves traffic comes from Acme, not from the internet. She places it on a cluster the company already owns, on the same GPUs already serving claims and underwriting. Placement is just a scheduling choice.
She gives it a name, and that name is all an application ever asks for. Never a model file, never a server. Then she points a model route at it: a permanent address that sits in front of whatever model answers. "This isn't networking," she says, "although it'll look a lot like it." A route can serve a single model, fail over to a second, or split traffic between them, and changes go live the moment you save. Her words: handle with care.
The guardrails live on the route, not in the application. Every request gets inspected twice. On the way in, personal data is masked, and jailbreak attempts and unsafe language are blocked before they reach the model. On the way out, personal data is masked again and internal information is stripped. An application built two years ago gets the same enforcement as one a team ships this week, and when the model underneath changes, none of that does.
Each team gets a token budget and a rate limit. Then comes the observability: who's burning the tokens, how long a user waits for the first one to appear (under half a second is excellent, she says), and how fast the rest stream in, which is how she decides whether to add a replica.
"All of that took me about four minutes."
What stayed with me isn't the model. Models come and go, and Acme already had four serving traffic. It's two things. The admin never gave up control: applications ask for names instead of servers, policy sits at the route instead of in a hundred codebases, and every team has a budget. And the whole job, from a Hugging Face link to a governed, monitored model, took four minutes, start to finish. I spent my session time at Explore on infrastructure management, hardware and software, and this was the same subject wearing an AI badge.
Grading the wishlist
In June, before the catalog dropped, I wrote down five things I wanted from the platform and promised to come back with grades. One confession first: Broadcom announced VCF 9.1 on May 5, six weeks before I wrote that list, so some of what I wished for was already in a press release. I'll grade the keynote anyway, and own the timing.
NSX, a simpler firewall story: miss. The keynote had nothing for NSX's distributed firewall.
vSAN ESA, an on-ramp for smaller shops: miss. vSAN appeared in a customer story about racks built and cabled at the factory. That's the far end of the market from a team of two running a mixed fleet.
VCF Operations, an on-ramp for teams flying blind: hit. The VCF AI Assistant takes questions in plain language. In the demo it found a cluster where DRS had been switched off by mistake. A team that hasn't built a single dashboard can still ask the question. That's the on-ramp I asked for.
SDDC Manager, more automation and less clicking: hit. The same demo ran its pre-checks and patched ESXi live, with no VM restarts. Broadcom says live patching in 9.1 covers up to 80 percent of patches without evacuating a host. That's the patch day I asked for, closer to a pipeline trigger than a checklist.
Multi-workload domains and brownfield: hit. The keynote never touched it, but the catalog did. In June I promised that any session with "brownfield" in the title went straight onto my schedule. Choosing the Right Path to VCF 9: Upgrade, Converge, Import, or Rebuild did, and it delivered: four ways into VCF 9 for an environment that didn't start greenfield.
Final tally: three hits, two misses, for a list written before the catalog dropped.
Where the hours went
I spent my session time on infrastructure management, from the hardware up through the software that runs on it.
Two of them earned their slot: Fleet Commander: Full-Stack VMware Cloud Foundation Operations and Modernization, and Full-Stack Troubleshooting with VMware Cloud Foundation Operations. Both lived in VCF Operations, the same ground that turned out to be the biggest block of Thursday's exam.
What I'm carrying to 2027
Post 6 ended with a rule: put the immovable content on the board first, and let the game route around it. Next May, at Resorts World, the plenary goes on the board first. I'll still play the game. It just gets the hours the keynote doesn't need.
That's the last loose end from 2026. Next post is the road to Explore 2027.