Not every IT obituary makes noise. When IBM decided to stop proposing the ehea driver for the Linux 7.3 kernel, no alarms went off in data centers. Yet what looks like a routine code cleanup is a litmus test for a pressure shaking the foundations of on-premise infrastructure: the tension between legacy hardware, official support, and the need for updated stacks to run modern workloads.

The driver in question controlled the 10Gb Ethernet interfaces of older IBM POWER platforms. Introduced when 10 gigabits per second marked the frontier of server connectivity, it is now viewed as a relic by parts of the same Linux ecosystem that once welcomed it. The seemingly anodyne news comes as the Linux kernel prepares to prune other historic drivers — PCMCIA cards, ISA buses, low-speed network controllers — in a slimming exercise that smacks of necessity rather than cruelty.

Behind the removal lies more than code hygiene. IBM, as the primary maintainer, has stopped allocating resources to test and update that driver. The vendor itself is signalling: that piece of hardware no longer has a place on the roadmap. Anyone managing POWER clusters with recent Linux operating systems and still using those cards now faces a fork: stick with older kernels and live with security patches that are not backported, or accelerate a migration to supported hardware.

This is the kind of decision well known to those working in regulated or air-gapped on-premise environments, where refresh cycles are slow due to certification constraints, capital costs, or simply because “it still works.” But the disappearance of a driver from a mainline kernel sends an unmistakable message: freezing on unsupported versions means increasing technical debt in a non-linear way. It’s not just about missing patches; the entire software ecosystem — from orchestration frameworks to security modules — moves forward taking recent and tested interfaces for granted.

If we widen the lens, the ehea affair says something deeper about the architectures that host artificial intelligence workloads and, more generally, data-intensive services. Networks at 10Gb, once coveted backbones, are now the bare minimum for distributed training environments or inference nodes that move large data volumes. The fact that a top-tier supplier declares end of support is an indicator of how fast the hardware substrate evolves, independent of the workload running on it. For those evaluating on-premise LLM deployments, it’s not just about how much VRAM to deploy or which quantization level to adopt; the challenge also involves the vendor-guaranteed longevity of every component, including networking.

Those who benefit from this scenario are providers of modern hardware and internal teams pushing for technological renewal: they can finally argue that maintaining the old stuff is no longer stinginess, but an operational risk. Those who lose are organisations with long amortisation cycles, often in manufacturing, healthcare or public administration, where POWER servers have delivered stability for years. For them, the Linux roadmap ceases to be a silent ally and becomes an external accelerator of decisions they would have preferred to postpone.

The episode, in its small way, illustrates how the open source model manages obsolescence: not through prohibitions, but through the drying up of maintenance. And when the maintainer itself raises the white flag, the message is final.