Four ThinkAgile Series, Three Hypervisors — and One That Lets You Change Your Mind

Written and maintained by Haink's infrastructure team · Verified against Lenovo Press V4 product guides and Microsoft Azure Local documentation, 29 August 2026 · authorized-channel, MTM and serial verified

Most ThinkAgile selection guides answer a question nobody is actually asking. They compare the appliances. But by the time you are choosing between HX and VX, the decision has already been made somewhere else — you picked a hypervisor, and the series follows from it. The appliance is downstream of the stack.

That was a perfectly good way to buy while hypervisor choices were stable. They are not stable in 2026 — and the hardware side of that instability is covered separately in what leaving VMware actually means for your servers. It changes which of these four series is the right purchase. Three of them commit your hardware to one software stack for its service life. The fourth does not, and Lenovo built it specifically because that commitment turned into a risk.

The short answer

What each series actually commits you to

VX — the VMware line

The 2U VX650 V4 takes one or two Intel Xeon 6700P or 6500P-series processors (Granite Rapids-SP), up to 86 cores, clocks to 4.0 GHz and TDP ratings to 350W. Memory reaches 8 TB across 32 DDR5 DIMMs with eight channels per processor. The front takes up to 24 hot-swap 2.5-inch or E3.S bays on PCIe 5.0 NVMe, and the chassis accepts up to 10 single-wide or 2 double-wide GPUs. Storage is vSAN with ESA. The V4 line also includes the 1U VX630 V4 and the GPU-oriented VX650a V4.

One detail matters more for procurement than any spec on that list. Lenovo states plainly: "Starting with ThinkAgile VX V4 Series, there will no longer be Integrated System/Appliance and Certified Node nomenclature." If your purchasing template, framework agreement or tender text specifies a "VX Certified Node", it will not match anything on a V4 quote. Fix the paperwork before the tender goes out, not during clarifications.

HX — the Nutanix line

The 2U HX650 V4 takes one or two Intel Xeon 6700-series or 6500-series P-core processors, up to 86 cores and 172 threads, to 4 GHz and 350W. It ships as a fully integrated appliance: validated Lenovo hardware and firmware, certified and preloaded with Nutanix software — Cloud Platform, Cloud Infrastructure and Cloud Manager editions. It runs both Nutanix AHV and VMware ESXi, which is worth noting if you want the Nutanix storage layer without committing to AHV on day one.

Memory goes to 8 TB using 32× 256 GB DIMMs at DDR5-6400; the front takes up to 24× 2.5-inch NVMe; GPUs up to 10 single-wide or 2 double-wide. There is an HX650 V4 Storage variant that trades the NVMe front for 8× 3.5-inch SAS/SATA plus 4× 2.5-inch and drops GPU support entirely — a capacity node, not a compute node.

Supported cluster shapes are documented as "1 or 3+ nodes". Single-node deployments are real and used at the edge, but a production cluster starts at three.

MX — the Microsoft line

The 2U MX650 V4 takes one or two Xeon 6700P or 6500P-series processors, again up to 86 cores and 172 threads at up to 4.0 GHz. It arrives preloaded with the Azure Local OS, with Windows Server 2025 Datacenter as an option, and is architected around Hyper-V with Storage Spaces Direct. Memory reaches 8 TB using 32× 256 GB RDIMMs with DDR5 running to 8000 MHz; the front takes up to 28× 2.5-inch hot-swap bays — four more than the VX and HX nodes — and it carries up to 8 single-wide or 4 double-wide GPUs, the highest double-wide count of the four series.

The capability that exists only here is disaggregation. MX compute nodes can attach to external ThinkSystem SAN arrays over Fibre Channel, so compute and storage scale independently instead of storage being trapped inside the HCI nodes. If your organisation already runs SAN, already has the fabric, and has no intention of abandoning either, MX is the only ThinkAgile line that accommodates that position rather than arguing with it.

FX — the line that does not lock you in

Two models: the 1U FX630 V4, a two-socket Xeon 6 node with up to 86 cores and PCIe 5.0, and the 2U FX650 V4, two Xeon 6700 or 6500-series processors, up to 86 cores and 172 threads, 8 TB of memory using 32× 256 GB 3DS RDIMMs with DDR5 to 8000 MHz, up to 24× 2.5-inch NVMe bays, and up to 4 single-wide or 2 double-wide GPUs.

What makes them different is not the silicon. Both are validated for all three stacks on the same hardware model:

• VMware — Cloud Foundation (VCF) and vSphere Foundation (VVF)
• Microsoft — Windows Server HCI with Storage Spaces Direct, and Azure Local
• Nutanix — Cloud Platform (NCP) with AHV

Lenovo calls this "unique, industry first flexibility" and describes an "any to any path to migrate" between the three, with professional services to support the transition. Read plainly: FX is the series you buy when you are not certain the hypervisor you are standardising on today is the one you will be running in four years.

SX — do not design on it

ThinkAgile SX and SXM for Microsoft Azure Stack Hub are marked as withdrawn products on Lenovo Press, alongside the Xeon SP Gen 3 HX 1U appliances — the full platform status and successor map covers both. If you are running Azure Stack Hub today, the conversation is not which ThinkAgile to refresh onto — it is a platform migration, and Azure Local is the successor discussion. Budget it as a migration project, not a hardware line item.

The comparison table nobody publishes cleanly

Cluster minimums set your entry price, and they are not the same

This is where HCI budgets most often go wrong. The smallest cluster you can buy is decided by the software stack, not by the appliance, and the four options do not agree with each other:

The practical effect: identical hardware, different floor. A VCF management domain sets a four-node minimum before a single production workload runs. The same estate on vSphere Foundation or Nutanix starts at three. On a two-socket Xeon 6 node with meaningful memory and NVMe, one node is frequently the difference between two budget brackets — and it is the single most common reason a "small" HCI project comes back over budget after the first quote.

How to get the node count right before you ask for a price

1. Count committed vCPU and RAM, not allocated. Allocation figures from an existing vCenter overstate real demand, often by a wide margin. Consolidation ratios built on allocation buy nodes you do not need.
2. Divide by usable per-node capacity, not raw. Erasure coding or replication factor, the reserve for rebuild, and the storage controller's own footprint all come off the top before workloads see anything.
3. Add the resiliency node. N+1 is the floor for anything with an availability commitment; N+2 if a rebuild window has to survive a second failure.
4. Add the stack's own minimum from the table above. This is the step that gets skipped.
5. Confirm the licensing unit and any per-CPU core minimum before fixing the CPU model. On core-based licensing a low-core processor can cost the same to licence as a high-core one, which changes the optimal node size — sometimes toward fewer, larger nodes than the hardware quote assumed.

Steps one and five are where the money is. Steps two through four are arithmetic; those two are judgement, and they are the reason two vendors can quote the same workload with a three-node difference between them. The full arithmetic, with the raw-capacity factor for each resiliency scheme and a worked example, is in sizing an HCI cluster.

When a ThinkAgile appliance is the wrong answer

We supply all four series and we also supply the alternative, so there is no reason to pretend the appliance always wins:

• If you want configuration freedom, build on plain ThinkSystem. An SR630 V4 or SR650 V4 cluster with the stack licensed separately gives you component choice the appliance line will not. What the appliance premium buys is validation, firmware matching, preload and one support path to call — real value on a first HCI deployment, much less obvious on your fifth.
• If GPU density is the point, none of these are the platform. The ThinkAgile ceiling is two to four double-wide GPUs per node. Purpose-built platforms go considerably further, and forcing the requirement into an HCI appliance produces an expensive compromise in both directions.
• If you are already standardised on VxRail or Nutanix-direct, changing appliance vendor mid-estate rarely pays for itself. Mixed-appliance estates carry two firmware matrices, two support relationships and two upgrade calendars.
• If you are on Azure Stack Hub, SX is withdrawn and the successor path is a migration, not a refresh.
• And the honest caveat on FX: its flexibility is a hardware property, not a promise that migration is free. Changing stacks is still a project — data moves, licences are repurchased, the operational model changes and staff retrain. What FX removes from that bill is the hardware write-off. It does not remove the bill.

Whether HCI is the right model at all — against three-tier with external storage — is a separate question we cover in choosing an HCI platform.

What this looks like as a purchase

ThinkAgile VX, HX and MX are to-order lines, not stock lines: our published lead time is 3–6 weeks as of August 2026, against days to a week for standard ThinkSystem SR630 and SR650 V3 from Hong Kong stock. That gap is the reason node count needs settling early — a four-node cluster and a six-node cluster have the same lead time, but discovering you need the sixth node after the order is placed costs another full cycle.

Two procurement mechanics change the number on a multi-node order. The first is the route: stock versus BTO and CTO determines both price and delivery date, and appliance lines are configure-to-order by nature. The second is volume: at cluster scale the pricing mechanism itself changes, and special bid and project pricing is not the same thing as a list discount — it is registered against the specific opportunity through deal registration, which is why the deal has to be structured before it is quoted, not after.

For estates crossing borders, node counts also interact with delivery: a twelve-node cluster does not have to arrive as one shipment, and staging by rack against a phased cutover is usually cheaper than storing hardware on site waiting for a data-hall window.

Frequently asked questions

Can I run Nutanix on a VX node, or VMware on an HX node?

HX is validated for both Nutanix AHV and VMware ESXi, so that direction works. VX is the VMware line. If you want the option to move in either direction later, FX is the series built for exactly that — it is validated for VMware, Microsoft and Nutanix on the same hardware model.

What is the smallest ThinkAgile cluster worth buying?

It depends entirely on the stack, not the hardware. VMware Cloud Foundation needs four nodes for a management domain, vSphere Foundation three hosts, Nutanix three for production, and Azure Local will run on a single machine. Add one more node for N+1 resiliency on anything carrying an availability commitment.

Is ThinkAgile just ThinkSystem with different software?

The hardware is ThinkSystem-based, yes. What you are buying on top is validation of the specific hardware and firmware combination, factory preload of the stack, and a single support path instead of arguing about whether a fault is hardware or software. On a first HCI deployment that is worth real money. On an estate where your team already runs the stack daily, it is worth less — and building on plain SR nodes becomes a defensible choice.

Does FX genuinely let me change hypervisor later?

At the hardware level, yes — FX630 V4 and FX650 V4 are validated for VMware Cloud Foundation and vSphere Foundation, Microsoft Azure Local and Windows Server with Storage Spaces Direct, and Nutanix Cloud Platform on AHV, and Lenovo documents an any-to-any migration path with professional services. The transition itself is still a project with data movement, new licences and retraining. FX removes the hardware from that cost, not the rest of it.

What happened to ThinkAgile SX?

ThinkAgile SX and SXM for Microsoft Azure Stack Hub are withdrawn products. Anyone running Azure Stack Hub today should be planning a platform migration rather than a hardware refresh, with Azure Local as the successor conversation.

Can a ThinkAgile cluster use our existing SAN?

MX can. The MX series supports a disaggregated deployment where compute nodes attach to external ThinkSystem SAN arrays over Fibre Channel, letting compute and storage scale independently. The other series keep storage inside the nodes.

Which series handles GPU-backed virtual desktops best?

MX650 V4 carries the highest double-wide GPU count of the four at up to four per node, against two on VX, HX and FX. For heavier GPU requirements than that, an HCI appliance is the wrong chassis and a purpose-built platform is the cheaper answer.

Related

• How to choose Lenovo ThinkSystem servers — workload-to-model selection and host sizing
• ThinkSystem V3 vs V4 — the generation question if you build HCI on standard rack servers instead
• VMware exit — the hardware decision — what you can carry across, what has to be rebought, and how the budget really breaks down
• VxRail vs Nutanix — the platform comparison behind the appliance choice
• Choosing an HCI platform — when hyperconverged is and is not the right model
• Sizing an HCI cluster — the resiliency factor, the rebuild reserve and the platform floor, worked through
• Enterprise storage selection — relevant if you are considering the disaggregated MX route
• Lenovo stock, lead times and pricing · stock vs BTO/CTO · project pricing
• Sovereign cloud infrastructure — 40 nodes, 7,680 vCPUs, 96 TB NVMe, national data residency

Sources

• Lenovo Press — ThinkAgile FX630 V4 product guide (1U, Xeon 6, three-stack validation, "industry first flexibility", "any to any path to migrate")
• Lenovo Press — ThinkAgile FX650 V4 product guide (2U, 8 TB, 24 NVMe bays, GPU limits)
• Lenovo Press — ThinkAgile HX650 V4 product guide (Nutanix preload, AHV and ESXi, "1 or 3+ nodes", Storage variant)
• Lenovo Press — ThinkAgile VX650 V4 product guide (vSAN ESA, VCF 4-node management domain, VVF 3 hosts, nomenclature change)
• Lenovo Press — ThinkAgile MX650 V4 product guide (Azure Local OS preload, disaggregated FC to ThinkSystem SAN, 28 bays, 4 double-wide GPUs)
• Lenovo Press — Introducing the ThinkSystem V4 servers with Intel Xeon 6
• Lenovo Press — ThinkAgile SX for Azure Stack Hub datasheet (marked withdrawn product)
• Microsoft Learn — Azure Local system requirements (1–16 machines standard, stretched 4–16 across two sites)