Docker writes its own hypervisor, and inherits every bug in it

"That means we own the full stack, and we can tune every part of the engine for container workloads specifically," Docker senior product managers Deanna Sparks and Colin Hemmings wrote in the post announcing the new backend ^[7]. Ownership is also liability. The Docker Engine sits inside the VMM, and what the VMM has to wrangle underneath it is the host's local networking and filesystems ^[12]. That is where the performance claims live, and it is where regressions surface on one company's VPN or one developer's disk layout. The backends being displaced are shared with a large population outside Docker; the replacement's field history, by Docker's own account, is Docker Sandboxes, the isolated environment it built for AI agents and used as the first test bed ^[6].

An agent sandbox and a working tree with a dependency directory on a bind mount are not the same I/O profile. The specific promise is about the inner loop: "When you're in an edit-compile-test loop, you'll see improvements every single build" ^[8]. That is a testable statement, and the beta is the cheap window in which to test it.

The population it lands on is smaller than the framing suggests, and more overlapping. Docker's 2025 State of Application Development report puts local-machine work at about 36% of Docker development ^[9], leaving 64% happening somewhere other than a laptop ^[1]. Inside that local group, Linux is used by 53% of developers, macOS by 51% and Windows by 47% ^[10]. Those add to 151%, or roughly 1.5 platforms per developer ^[2], which is the actual argument for parity: the divergence between backends is absorbed by one person moving between two machines, not merely compared across a team.

The sequencing is awkward. The beta covers the two smaller shares, and Linux, the largest single one at 53%, has nothing to try until general availability ^[4]. Docker's stated destination is one VMM natively driving every platform it supports ^[13], so the old arrangement is not a place anyone gets to stay.

For paid seats the money is trivial: professional plans run $11 a month, or $9 billed annually ^[11], which is $108 per seat per year ^[3]. The cost is the engineering hour it takes to get a baseline. A timed cold start, a timed incremental build and resident memory after an idle period, recorded on one real repository under both backends while both are still selectable, is the difference between a beta bug report with numbers attached and a support thread after GA. Docker says the previous split of virtualization technologies produced minor inconsistencies as work moved between platforms ^[14]. One new backend will produce its own, at least early on; the difference is that all of them will now be Docker's to fix.