Dora NodaOn June 30, 2026, the company that taught the industry what API-driven bare metal could feel like deleted its last servers. Equinix Metal — born as Packet, bought for $335 million, rebranded as the future of the edge — is gone, and the exodus it triggered is the largest forced migration in bare-metal history. If you ran on Metal, your servers are already terminated. If you automated against its API, your provisioning code now points at a dead endpoint with documentation that disappears on September 30.
Here is the short version for anyone still mid-migration: the shutdown happened on a corporate timeline, not a technical one; five credible destinations are competing for the refugees; and the teams hurting most are the ones whose automation was baked against Metal's API instead of sitting behind an abstraction. The rest of this post shows the milestone timeline, the destination map, and the Cluster API provisioning story for each door — plus the lesson a self-hosted platform should take from watching a provider API sunset.
The shutdown in one timeline
Equinix gave the industry more than 18 months' warning, communicated in a customer letter from Chief Business Officer Jon Lin and Chief Sales Officer Mike Campbell. The wind-down ran in phases, and two deadlines are still in the future as of this writing:
| Date | Milestone |
|---|---|
| Jan 2020 – Mar 2020 | Equinix announces ($335M) and closes the Packet acquisition; Packet becomes "Packet, an Equinix company" |
| 2023 | Packet founder Zac Smith leaves to co-found Datum; Metal loses its original leadership |
| Nov 2024 | Sunset announced; Metal enters maintenance mode — no new features, only performance, security, and stability fixes |
| Feb 28, 2026 | docker-machine-driver-metal archived; the ecosystem starts shedding integrations |
| Jun 30, 2026 | Service ends; remaining servers permanently deleted; final billing runs issued |
| Sep 30, 2026 | User documentation removed; Metal support services retired (account teams only after this) |
| Dec 31, 2026 | Console access ends for Owner-role users (invoices and history until then); console taken offline Jan 1, 2027 |
Two traps in that timeline deserve emphasis because both cost real money. First, Equinix warned that customers must delete IP addresses, fabric virtual connections, and VLANs themselves — failure to do so means continued charges "even after all servers have been terminated." Second, any Fabric interconnection associated with a Metal service keeps billing until explicitly deleted or disconnected. The servers die on their own; the network objects do not.
If you held Metal resources, audit your Fabric bill before the console goes dark — after December 31, your visibility into what you are still paying for gets much worse.
The Kubernetes integrations died on the same schedule. The Metal Cluster API provider (CAPP) is deprecated and now lives under kubernetes-retired/cluster-api-provider-packet; the Equinix Metal cloud provider (cloud-provider-equinix-metal) is deprecated and unmaintained. Anyone whose fleet lifecycle ran through CAPP is not just migrating servers — they are migrating the control plane that made the servers.
How a $335 million bet became 1.25% of revenue
The arc is worth thirty seconds because it explains why the shutdown was a business decision no migration discount could reverse. Equinix bought Packet to push into edge computing and smaller markets, with then-CEO Charles Meyers pitching bare metal as the on-ramp for enterprise hybrid multicloud. Six years later, Metal accounted for just 1.25% of Equinix's revenue, according to CFO Keith Taylor on the earnings call preceding the shutdown announcement. A rounding-error business unit inside a colo giant selling AI infrastructure does not get a second act.
And Equinix has been explicit about where its attention went instead: a vendor-neutral Distributed AI Hub for inference at the edge, a "smart automation" connection-management service called Fabric Intelligence, and Fabric Geo Zones for data-sovereignty enforcement at the network layer. Metal was not killed by a failed product — developers who used it mostly loved the API. It was killed by portfolio math. That distinction matters, because portfolio math is exactly the kind of risk no SLA protects you from: the API worked fine until the quarter it did not need to exist.
Where ex-Metal users go: five doors
Three rivals published explicit Metal-migration guides and are actively competing for refugees: Megaport (via Latitude.sh), OVHcloud, and OpenMetal. Two more destinations — Hetzner and Vultr — belong on any shortlist for teams that provision through Cluster API. Here is the map, with the two columns that actually decide a migration: what replaces Metal's network automation, and what replaces CAPP.
| Destination | Model | Network replacement | CAPI story | Best fit |
|---|---|---|---|---|
| Latitude.sh + Megaport | Global CPU/GPU bare metal (~5-second deploys) on Megaport's NaaS fabric | Megaport private connectivity to clouds, partners, 1,000+ DCs — the closest Fabric analog | No maintained CAPI provider; API + Terraform, or BYOH into your own CAPI management cluster | Teams whose Metal usage was really "compute next to great interconnect" |
| OVHcloud Bare Metal 2026 | Dedicated servers (AMD EPYC 9005 Scale line) across EU, CA, US, APAC | vRack private networking across regions and datacenters | No maintained CAPI provider for dedicated; API + Terraform, or self-run provisioning | EU-centric fleets that want one vendor for dedicated + cloud + managed K8s |
| OpenMetal | OpenStack-API bare-metal clouds | Neutron: API-driven networks, floating IPs, security groups | CAPO (Cluster API Provider OpenStack) — a maintained, first-class CAPI provider | Teams that want Metal-style API control with an open-source exit hatch |
| Hetzner | Dedicated servers (Robot) + Hetzner Cloud side by side | Cloud Networks + Load Balancers; vSwitch for dedicated; Robot API | CAPH for Hetzner Cloud (maintained); dedicated via API/Terraform or self-run provisioning | Cost-driven fleets comfortable splitting stateless cloud + dedicated iron |
| Vultr | Cloud + Bare Metal under one API across its global regions | VPC 2.0 across regions, firewall and LB APIs | CAPI provider listed in the Cluster API Book; bare metal behind the same vendor API | Teams that want one throat to choke for cloud, bare metal, and CAPI |
A few honest footnotes on that table. First, nobody replicates Metal exactly: the combination of a single global API, per-server BGP, and Fabric interconnect in one invoice was the product, and it is gone. Every destination recovers two of the three and asks you to rebuild the third. Second, the "best fit" column hides a variable that flips answers: fleet size. A three-server lab and a three-hundred-server edge deployment do not want the same door, and the right question is not "which is best" but "which API will my automation target for the next five years." Third, pricing is deliberately absent from the table — Metal refugees report that re-quoting is the easy part; re-wiring is the work. Which brings us to provisioning.
The Cluster API provisioning map
CAPP's retirement is the sharpest edge of this migration. Teams that ran clusterctl against Metal did everything right architecturally — they put Cluster API between themselves and the vendor — and they still have a migration, because the provider itself is dead code. The good news is that their migration is a provider swap, not a rewrite. The shape of that swap depends on the door chosen above:
- OpenMetal → CAPO. The cleanest swap on the list. CAPO is one of the most mature CAPI providers, and an OpenStack-API endpoint is an OpenStack-API endpoint — the Cluster, MachineDeployment, and KubeadmConfig shapes transfer with new credentials and flavors. If your CAPP manifests were tidy, this is days, not months.
- Hetzner Cloud → CAPH. Also clean, via Syself's maintained Hetzner provider. The caveat is scope: CAPH covers Hetzner Cloud, not Robot-managed dedicated servers. Fleets that ran everything on Metal single-sourced now face a split — CAPH for cloud capacity, something else for the iron.
- Vultr → the Vultr CAPI provider. The Cluster API Book lists a Vultr provider, which makes Vultr the only door where cloud, bare metal, and a CAPI provider share one vendor relationship. Verify bare-metal coverage in the provider before committing — "has a CAPI provider" and "provisions the servers I buy" are different claims.
- Latitude.sh / OVHcloud dedicated → no native provider. This is the honest gap. Both expose solid APIs and Terraform providers, but neither has a maintained CAPI infrastructure provider for dedicated servers. The paths are: BYOH (Cluster API's Bring-Your-Own-Host agent on machines you provision out-of-band), or a self-run provisioning layer underneath CAPI.
- The self-run layer: Tinkerbell (CAPT) or Metal3 (CAPM3). Here the story comes full circle. Tinkerbell — the open-source provisioning stack born inside Packet — lives on as a CNCF sandbox project with its own CAPI provider, CAPT, and production adopters provisioning EKS Anywhere clusters on bare metal. Metal3 pairs with Ironic for the same role. Either lets you own the provisioning API the way Metal once rented it to you: DHCP/iPXE, hardware inventory, and OS imaging behind a Kubernetes-native workflow. The price is operating it — Rufio, Hegel, and the rest of the Tinkerbell microservices are yours to run.
Note what the team that skipped CAPI entirely faces by comparison: every metal_device Terraform resource, every BGP session script, every Fabric automation becomes vendor-specific dead code with documentation that vanishes September 30. Archive what you have — export runbooks, pin provider versions, screenshot the console — while the artifacts still exist.
The lesson and the checklist
The lesson is not "don't trust vendors." Every fleet trusts vendors for something — silicon, transit, power. The lesson is narrower and more actionable: a provider API is a dependency with a corporate sunset timeline, and it belongs behind an abstraction you control. CAPP users are doing a provider swap. Direct-API users are doing a rewrite. Same shutdown, different migration, and the difference was decided years ago when someone chose where the provisioning calls live.
That is the exact bet a Cluster-API-based platform makes structurally: CAPH today, another provider tomorrow, with the Cluster and MachineDeployment shapes surviving the swap. The provider is a plugin, not a foundation. Metal's death is the clearest possible demonstration of why — the biggest brand in API-driven bare metal went from acquisition press release to deleted servers in six years, and no amount of API quality changed the outcome.
If you are still unwinding Metal exposure, the checklist in priority order:
- Before September 30: export everything — docs you reference, support tickets with topology answers, Terraform state showing what existed. Documentation and support retire together.
- Now: audit Fabric interconnections, VLANs, and reserved IPs for orphaned billable objects. This is the only item on the list that charges you while you procrastinate.
- Before December 31: download invoices and billing history from the console with an Owner-role account. After January 1, the console is gone.
- This quarter: pick the door and prove the provisioning path — a CAPO/CAPH cluster, a BYOH pilot, or a Tinkerbell lab — before the institutional memory of how Metal behaved fades from your team.
- Structurally: put the next provider API behind CAPI (or at minimum Terraform + an internal module), so the next sunset letter starts a swap, not a rewrite.
Packet proved developers would rent bare metal like cloud. Equinix proved a great API is not a moat against portfolio math. The teams that internalize both lessons — API-speed provisioning, behind an abstraction that survives any single vendor — are the ones for whom June 30 was an inconvenience rather than an emergency.
Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own. Its Cluster-API-based provisioning treats every provider as a plugin, so a sunset letter starts a swap, not a rewrite. Star the repo on GitHub or deploy your first app today.
