The Pendulum Is Swinging Back Toward Three-Tier Infrastructure

I have been having a lot of conversations lately with customers about the future of their data center infrastructure. VMware’s licensing and program changes continue to dominate those discussions.

Many organizations are frustrated with the direction things have gone. Some feel priced out. Others feel the platform’s future no longer aligns with their long-term strategy.

But this is not a post about VMware or Broadcom. It is about something more interesting: the economics of infrastructure design have shifted, and that shift is changing how organizations approach their next refresh cycle.

Two years ago, the math favored hyperconverged

If you had asked me then what the most cost-effective architecture was for a typical mid-market customer, I would have said hyperconverged infrastructure.

Every environment is different, of course. Team size, workload requirements, growth projections, and operational maturity all matter. But for organizations with lean IT teams made up of talented generalists, hyperconverged offered a compelling combination of simplicity and economics.

The math worked. vSAN came bundled with the license. NVMe and SSD storage were relatively affordable. As clusters grew, organizations could use storage efficiency technologies and erasure coding to maximize capacity without sacrificing performance. The result was a platform that was easy to manage, easy to scale, and often cheaper than a traditional architecture.

Two things have happened since. Both of them attack the same assumption.

Flash stopped being cheap

NAND contract prices rose 55 to 60 percent quarter over quarter in Q1 2026, then climbed another 70 to 75 percent in Q2, according to TrendForce. Enterprise SSD contract pricing has moved even harder than the raw flash underneath it. Dell, Lenovo, HP, and HPE have all pushed server pricing up a minimum of 15 percent (sometimes much more) to absorb what they are paying for components.

This is not a blip in a quarterly forecast. As you have probably heard by now Samsung, SK Hynix, Kioxia, and Micron cut NAND output in the back half of 2025 and then redirected capacity toward AI demand. Most 2026 supply was committed before the year started. Nobody credible is calling for meaningful relief before 2027.

Now apply that to an architecture where every node carries its own storage.

In a hyperconverged cluster, capacity is distributed across every host, and protection policies multiply the raw footprint on top of that. A mirrored policy tolerating one failure consumes roughly twice the raw capacity of the data it protects. When flash was cheap, that overhead was an acceptable trade for operational simplicity. At current pricing, it is the largest line item in the quote.

A dedicated array centralizes those drives and shares them across the environment. Fewer drives, higher utilization, and the protection overhead is absorbed by the array rather than replicated node by node.

If that argument sounds familiar, it should. It is the same case that made shared storage the default a decade ago.

vSAN is no longer simply included

This is the part I find most customers have not fully modeled yet.

Under the current entitlement structure, vSphere Foundation includes 0.25 TiB of raw vSAN capacity per licensed core. Cloud Foundation includes 1 TiB per core. Anything beyond that is an add-on purchase, sold per TiB, on top of the license you already bought. Core counting carries a 16 core per populated socket floor regardless of what the chip actually ships with.

Run the arithmetic on a typical mid-market build. A four node cluster with two 16 core CPUs per host is 128 licensed cores, which entitles you to 32 TiB of raw vSAN capacity under VVF. Ask that same cluster to hold 100 TB usable with a mirrored protection policy and you need roughly 200+ TB raw. Everything past the entitlement is a separate line item.

The entitlement is measured against raw contributed capacity, not the usable datastore figure your console reports. That distinction has caught more than one customer at renewal, and it’s an additional tax on the already astronomical pricing we see today.

This is an economics story, not a technology one

Hyperconverged has not gotten worse. It remains a strong architectural approach, and for smaller clusters, remote sites, and teams with no dedicated storage administrator, it is still frequently the right call. What changed is the price of the inputs and the shape of the license.

We are also seeing customers pair these architectural decisions with consumption-based financing. Instead of a large capital investment upfront, costs align more closely with actual usage, and storage, compute, and virtualization can be bundled into a single scalable monthly expense.

In a market where the same drive costs materially more this quarter than last, that matters more than it used to. For organizations looking to preserve cash while maintaining flexibility, those programs have become increasingly attractive. Depending on workload and growth profile, we have seen the overall cost compare favorably to a traditional capital purchase while providing considerably more flexibility over the life of the solution.

None of this is permanent

Technology trends rarely move in a straight line. They behave more like a pendulum.

For several years it swung heavily toward hyperconverged. Today it is moving back toward three-tier for many workloads. When NAND supply catches up, or the licensing model shifts again, it will swing back.

That is what makes this industry interesting. The right answer keeps moving, and the organizations that regularly challenge old assumptions tend to make the best decisions.

So I am curious: are you reconsidering hyperconverged, or doubling down on it? And what is actually driving your refresh decisions right now?

Sources