TL;DR — VCF 9.1 is Broadcom's focused release. Three pillars: cheaper memory & storage, faster lifecycle, and production AI on your own metal. If you run large VMs or GPUs, the NVMe tiering + Live Patching + ZTP trio alone is worth an evaluation this quarter.
VCF 9.1 landed as what Broadcom calls a focused release — and after going through the release notes end to end, I agree. The theme is unmistakable:
Core thesis: run production AI and demanding workloads on your own hardware, at the lowest possible cost per workload, without giving up security or velocity.
Here's the full breakdown of what actually matters — with the noise removed.
At a Glance
| Pillar | Headline Feature | Why It Matters |
|---|---|---|
| Economics | NVMe Memory Tiering | 2–3× effective memory without new DIMMs |
| Storage | vSAN Dedupe + Native S3 | More TB per raw TB + S3 without a separate stack |
| Lifecycle | Zero-Touch Provisioning | PXE-less bare-metal bring-up at scale |
| Uptime | Live Patching for ESX | 80% of patches with zero VM downtime |
| AI | GPU metrics + topology-aware scheduling | Private AI you can actually operate |
Memory & Storage Economics
Enhanced NVMe Memory Tiering
ESX now tiers hot pages → DRAM and cold pages → local NVMe, expanding effective memory capacity without buying more DIMMs. VCF 9.1 adds software-based mirroring and a cost-savings analysis view on top.
When to use it: AI inference, large in-memory DBs (SAP HANA, Redis), and any VM where
mem.active≪mem.configured. Start with a 1-host pilot and compareesxtopMEM stats before/after.
| Before | After (with tiering) |
|---|---|
| 512 GB DRAM per host | 512 GB DRAM + 2 TB NVMe tier |
| Over-provision or cap VMs | Right-size with headroom |
| DIMM purchase for growth | NVMe for growth (5–10× cheaper/GB) |
Extended vSAN Deduplication & Compression
Inline data reduction now covers more cluster types and workload profiles, including dedupe for data-at-rest encryption — previously a dealbreaker. Compression ratios improved as well, directly lowering raw storage per logical TB.
Logical TB ──► Deduped ──► Compressed ──► Raw TB on disk
10 TB 6.2 TB 3.8 TB 3.8 TB
Native Object Storage TECH PREVIEW
Self-service S3-compatible object storage built into the platform. Developers get an endpoint; IT keeps the guardrails. Provisioned, scaled, and managed with the same workflows as block and file.
- Use for: backups, ML datasets, artifact registries
- Not yet for: production primary storage (tech preview)
Compute & Lifecycle
Zero-Touch Provisioning (ZTP)
vSphere Elastic Provisioning bootstraps ESX on bare metal via secure UEFI + HTTPS network boot — no PXE, no TFTP, no manual imaging. Boot multiple servers from a single URL with static network config.
This is the official forward-looking replacement for Auto Deploy (deprecated since 9.0).
# Example: ZTP boot URL serves per-host config via HTTPS
# Host iDRAC → UEFI HTTPS Boot → https://ztp.techen.in/esx/9.1/profile.json
{
"host": "esx-07.techen.in",
"mgmt_ip": "10.10.30.7/24",
"vlan": 30,
"image": "esx-9.1-nvme-tiering"
}
Migration note: If you still run Auto Deploy, plan the ZTP cutover now. PXE/TFTP won't be enhanced going forward — ZTP is where the investment is.
Live Patching for ESX
Patches applied to running kernel memory on TPM-enabled hosts. VMs keep running — no maintenance window — covering up to 80% of patches. Combined with Quick Boot (now supports confidential VMs), patch cycles get dramatically shorter.
| Patch type | VM downtime? | Host reboot? |
|---|---|---|
| Live Patch (kernel) | No | No |
| Quick Boot (most others) | No (vMotion first) | No (kernel reload) |
| Full reboot | Yes (maintenance) | Yes |
VKS & VM Fast-Deploy
Kubernetes and VM cluster deployment now uses linked-clone technology, cutting time-to-running significantly. Upgrades benefit too — less clone, less copy.
VCF Management Services
A new common runtime unifying lifecycle and operational capabilities across the platform — the beginning of a cleaner, more consistent management plane. Less "which SDDC Manager does what" and more single pane.
Security & Resilience
| Capability | What Changed in 9.1 |
|---|---|
| Live Patching | Eliminates the #1 source of unplanned downtime — patching |
| Continuous Compliance | Cyber Compliance moves from quarterly audits to runtime remediation |
| On-prem Ransomware Recovery | Cyber recovery into isolated clean rooms inside your own VCF estate |
| File Integrity Monitoring | vCenter FIM on by default, NIST-aligned |
| Identity | OAuth 2.0 API tokens for vCenter + streamlined OIDC brokering |
Ops win: FIM on by default means you get tamper detection without a separate hardening project. Check vCenter → Administration → File Integrity on day 1.
AI on Your Own Hardware
This is where 9.1 aims squarely — and finally feels concrete:
- Private AI Model & GPU Metrics — utilization, memory pressure, and model-level telemetry on the same console as the rest of the fleet
- Enhanced DirectPath I/O for AMD MI350 and NVIDIA ConnectX-7 / BlueField-3 — near-bare-metal accelerator performance with vMotion still in play
- Topology-Aware Scheduling — places workloads with NUMA and accelerator locality in mind (less cross-socket chatter, more consistent latency)
- Ubuntu OS Enterprise Support — broadening the AI workload ecosystem beyond Photon
If Private AI felt like "someday" in 9.0, 9.1 is the release where the ops story catches up — GPU metrics in the same pane as your VM fleet is what teams have asked for since day one.
Developer Experience
| Area | What's New |
|---|---|
| CaaS | Self-service namespace provisioning — registry, ingress, quotas, and identity inherited from VCF constructs |
| Metrics | Real-Time Metrics APIs — one call replaces the legacy vStats config model |
| Federation | vCenter Group Federated API — one endpoint for inventory across an entire vCenter group |
| Data | SQL Server DBaaS + Tanzu Marketplace for certified middleware |
Upgrading
Path: VCF 9.0 → in-place upgrade to 9.1 is supported.
Before you touch the management domain:
- Read the VCF 9.1 Upgrade Guide + compatibility matrix on Broadcom Support Portal
- Run the new prechecks (now native VCF capabilities, exports to CSV)
- Snapshot and validate backups — then upgrade the management domain first
My Take
The memory tiering + live patching + ZTP trio alone justifies planning a 9.1 evaluation this quarter. The AI story is finally concrete rather than aspirational — GPU metrics in the same pane of glass as the rest of the estate is what ops teams have been asking for since Private AI was announced.
If you run VCF today, put tiering on your lab backlog and run the prechecks. The economics will do the rest of the convincing.
Full release notes and component-level details are on the Broadcom TechDocs portal.
Planning a 9.1 upgrade? I'll be publishing a step-by-step upgrade runbook and a memory tiering benchmark series soon — watch the archive.