Trends · Bare metal market

Bare Metal Trends 2026: The Great Repatriation and the Return of Owned Hardware

For most of the last fifteen years the infrastructure story ran one direction — into the cloud. In 2026 the current visibly reversed for a meaningful class of workloads, and “bare metal” stopped being a word associated with the past. This is not a nostalgia trend; it is a response to specific, quantifiable pressures that changed the math. This guide covers the bare metal server trends that actually matter this year, what is driving them, and — because no honest trends piece claims one model wins everything — where bare metal still loses.

Is bare metal really making a comeback?

Yes, but selectively, and the data backs it up. Analysts size the bare metal cloud market somewhere between $11.6 and $14.3 billion in 2025, with forecasts reaching the $36–67 billion range by the early 2030s at compound growth rates around 20–25%. North America holds the largest slice, the hardware segment dominates revenue, and — tellingly — small and medium enterprises are the fastest-growing buyers, drawn by high performance without the capital outlay of owning a data centre.

The important nuance is what “comeback” means. Nobody serious is abandoning the cloud wholesale. What is happening is a re-sorting: workloads that are steady-state, data-heavy, latency-sensitive, or regulated are moving to dedicated hardware, while bursty, experimental, and globally-distributed front-ends stay in the cloud where elasticity pays off. Industry estimates put 40–50% of enterprise workloads that need raw compute as genuine bare metal candidates — a large minority, not the whole.

The buyer mix tells the same story. Large enterprises still account for most of the revenue, driven by the high-performance and regulatory demands of finance, healthcare, telecom, and government — sectors where single tenancy simplifies the compliance story. But the fastest growth is coming from small and medium businesses, who can now reach dedicated, high-performance hardware without the capital expense of building their own data centre. Even private equity firms have noticed: several are systematically repatriating the cloud spend of their SaaS portfolio companies to convert unpredictable usage-based bills into fixed costs, reporting infrastructure savings in the 30–50% range and more predictable margins as a result. When the financial engineers start treating repatriation as a portfolio strategy, the trend has moved well past enthusiast territory.

What is driving “the Great Repatriation”?

Three pressures converged at once, which is why the shift feels sudden. The first is software licensing — specifically the Broadcom acquisition of VMware and the pricing shock that followed. The second is cloud egress: the per-gigabyte cost of moving data out of a hyperscaler, which becomes indefensible at scale. The third is the rising, volatile cost of hardware and power, which makes paying a cloud premium on top harder to justify. Individually each is an irritant; together they flipped the business case for a whole category of workloads.

The VMware/Broadcom licensing shock

The single loudest catalyst of 2026 is what practitioners now call the “virtualization tax.” After Broadcom acquired VMware, perpetual licenses gave way to bundled subscriptions, and renewal costs reportedly jumped anywhere from 200% to as much as 1,200% for some customers. The core-based licensing model adds insult to injury: deploy a modern high-density processor to save on hardware, and the per-core licensing eats the saving. For a lot of organisations, the spreadsheet that once favoured virtualization stopped doing so overnight.

The fallout is visible in the market. Pre-acquisition VMware counted more than 350,000 customers; Broadcom has reportedly narrowed its focus to roughly the top 10,000, leaving everyone else looking for an exit. Competitors are catching the runoff — Nutanix reported winning 730 new clients in a single recent quarter, with its CEO noting most came from legacy vendors. Many of those organisations are not just switching hypervisors; they are reconsidering whether they need a hypervisor at all, and bare metal with a lightweight platform like Proxmox is a common landing spot.

Why did cloud egress push data-heavy workloads off the cloud?

Because at scale the bill stops making sense. Hyperscalers typically charge around nine cents per gigabyte to move data out, so transferring a single petabyte can run roughly $90,000 to $120,000 — a recurring, variable cost that scales directly with how much value your data-heavy application produces. This is the “Cloud Paradox”: the cloud offers agility but penalises scale, while dedicated infrastructure does the inverse, getting more cost-efficient per unit the more you grow.

The real-world examples have become reference points. 37signals, the maker of Basecamp, exited the cloud for owned hardware in colocation and projected around $10 million in savings over five years. Smaller shops report the same shape of result moving specific workloads off hyperscalers — a fleet-tracking provider, for instance, cut its customer-facing costs by roughly a quarter to a third after migrating off a public cloud to dedicated infrastructure. The economics are sharpest for streaming, AI training datasets, and large backups, where egress is the dominant line item — and they are exactly why flat, unmetered bare-metal bandwidth is so attractive. We work the full comparison in TCO: dedicated vs cloud, and the broader pattern in cloud repatriation.

What is BMaaS, and how is it different from old dedicated hosting?

BMaaS — bare-metal-as-a-service — is what removed the last good argument against dedicated hardware: that it was slow and manual to operate. The model wraps physical servers in cloud-style provisioning, so you can deploy a single-tenant machine through an API, pay by the hour, and decommission it when the project ends. The control and performance of dedicated hardware now come with the agility people expected only from public cloud.

The operational difference from the dedicated hosting of a decade ago is automation. GitOps, infrastructure-as-code, and BMaaS platforms let teams provision, patch, and replace servers remotely, without on-site engineers, even across distributed locations. That is what makes modern bare metal viable at scale instead of a manual burden.

# BMaaS: a single-tenant physical server, provisioned like a cloud instance
$ curl -s -X POST https://api.metal.example.net/v1/servers \
-H “Authorization: Bearer $METAL_TOKEN” \
-d ’{“plan”:“epyc-7003-128g”,“region”:“ca-central”,“os”:“debian-13”}’
{ “id”: “srv_9f12”, “state”: “provisioning”, “eta_minutes”: 7 }
 
# Driven from infrastructure-as-code, not a ticket and a phone call
$ terraform apply -target=metal_server.db_primary
metal_server.db_primary: Creation complete after 6m41s
# API + IaC = cloud-like agility on dedicated hardware. That’s BMaaS.

Confidential computing and the regulated-sector pull

Single tenancy has always been bare metal’s security advantage — when you are the only tenant on a physical machine, entire categories of noisy-neighbour and multi-tenant escape risks simply disappear, which is why finance, healthcare, and government have long favoured it for simpler compliance audits. What is new in 2026 is that hardware-level confidential computing has matured past the proof-of-concept stage. Technologies like Intel TDX let you run workloads in encrypted, attested enclaves on dedicated hardware, so even sensitive data such as protected health information can be processed without exposing it to the broader platform.

Combined with single tenancy, this makes bare metal a credible home for regulated AI and analytics that organisations were previously nervous about running anywhere shared. It also dovetails with the sovereignty trend: a single-tenant machine in a jurisdiction you choose, processing encrypted workloads, is a strong answer to data-residency obligations. We cover that angle in depth in data sovereignty trends.

How AI demand is reshaping bare metal

AI is pulling bare metal in two directions at once. At the top end, the largest operators are locking in long-term dedicated GPU capacity — agreements like the multi-year, multi-billion-dollar CoreWeave–Meta deal exist precisely because frontier AI needs guaranteed, non-shared compute, and Oracle and others have built GPU superclusters on the same logic. That is bare metal at hyperscale, and it is a major reason the market’s revenue figures look the way they do.

For everyone else, AI raises the floor on what a “normal” server needs — faster NVMe, more memory bandwidth, and often a modest GPU for inference or fine-tuning. The supply pressure from the AI boom also has a downside that deserves its own treatment: it is squeezing availability and pricing for ordinary hardware, which we cover in GPU shortage 2026 and the wider AI infrastructure trends.

The hardware itself is moving fast

One reason the bare metal conversation feels different in 2026 is that the hardware underneath it took a visible generational leap, and on dedicated servers you get that leap directly rather than diluted through a hypervisor. Storage is the clearest example: PCIe Gen5 NVMe drives push sequential throughput and IOPS well past the previous generation, and with no virtualization layer between the operating system and the disk, a single-tenant box turns that raw I/O into real database and analytics performance instead of shared-tenant contention.

The CPU and memory side moved in step. High-core-count parts — AMD’s latest EPYC generations and Intel’s Xeon 6 (Granite Rapids) — pair with DDR5 memory running at 6400 MT/s and beyond, which is exactly the combination that data-heavy and inference workloads are bandwidth-starved without. Networking is climbing too: 10 and 25 gigabit ports are now routine on dedicated servers, with 40 and 100 gigabit at the high end, and NVMe-over-fabrics and composable approaches letting providers attach pools of storage and accelerators to bare metal on demand. The net effect is that the performance gap between “a server you rent” and “the fastest hardware money can buy” has narrowed sharply — buying bare metal in 2026 means buying genuinely current silicon, not last cycle’s leftovers. The same is true of accelerators: providers increasingly offer GPUs and, for specialised workloads, FPGAs as configurable options on dedicated machines, so a team can match the exact accelerator to the job — inference, transcoding, or fixed-function processing — instead of accepting whatever a virtual instance type happens to bundle.

Where bare metal still loses

An honest trends piece has to name the trade-offs, because bare metal is not the right answer everywhere. Its core weakness is elasticity: a physical server has to be allocated, configured, and provisioned, which takes minutes at best — it cannot spin up and tear down in seconds the way a cloud VM can. For genuinely bursty or unpredictable workloads, that rigidity is a real cost, and the cloud’s instant scaling is worth paying for.

There is more ownership, too. With bare metal you take on more of the operational responsibility — capacity planning, hardware lifecycle, and the discipline of running your own platform — even when BMaaS automates the provisioning. And one practical 2026 caution: the provider landscape is shifting under your feet, with even established services reaching end-of-life (Equinix Metal is winding down in mid-2026), so choosing a stable provider matters as much as choosing the architecture. Bare metal rewards teams that want control and have steady, predictable demand; it punishes those who needed elasticity and underestimated the operational load.

What these trends mean for your 2026 choices

The practical reading is to stop treating “cloud versus bare metal” as an identity and start treating it as a per-workload allocation. Audit where you are paying the cloud premium without using its elasticity — the steady database, the data-heavy pipeline, the regulated dataset, the high-egress service — and those are your repatriation candidates. Keep the bursty front-ends and the experiments where they are. The organisations getting this right in 2026 are not the ones making a loud, total move in either direction; they are the ones placing each workload where its economics actually land.

For the steady, regulated, data-heavy slice — and especially anything with Canadian data-residency requirements — a single-tenant dedicated server is increasingly the rational default rather than the contrarian choice. That is the shift these trends add up to: not the end of cloud, but the end of putting everything in it by reflex.

One last piece of advice that the 2026 churn makes unavoidable: choose the provider as carefully as the architecture. The same market forces driving repatriation are also reshaping who offers bare metal, and even well-known services are reaching end-of-life — Equinix Metal, for one, is winding down in mid-2026, forcing its customers into an unplanned migration. A repatriation that lands you on an unstable provider has only traded one kind of risk for another. Look for a clear ownership structure, a track record measured in years rather than funding rounds, transparent pricing without egress traps, and a jurisdiction that matches your compliance needs. The point of moving to owned hardware is durability and control; pick a partner whose own durability you can count on.