Open source picks up the work VMware and the AI boom left behind

The newest piece of open source infrastructure news is also one of the smallest in code terms: a single shared library. On 3 November, The Register reported that cloud architect Andrija Panić has built OpenVDDK, an independent replacement for VMware's Virtual Disk Development Kit, or VDDK.

VMware describes VDDK as "a set of libraries and utilities that enable software developers to create applications that can access and manipulate virtual disks." In August, the company restricted downloads of the kit to partners. Customers lost access. So did the migration tools that quietly depended on it.

Panić told The Register he wrote the library in Go, with a thin C layer. That layer exposes the same C interface, library name and symbol versions as VMware's code. He said he built it from public documentation and black-box testing, and that it contains no VMware code. It reads disks. It cannot write them.

Why one library matters to a migration market

The point is not the library. It is what loads it. "Many tools that have nothing in common except VMware use the same small library to read a virtual machine's disks: virt-v2v and nbdkit, the importers in Forklift/MTV, Proxmox VE and Apache CloudStack, and countless in-house scripts," Panić told The Register by email. "When public VDDK downloads ended, all of them became harder to install and run."

That is a fair summary of how enterprise infrastructure ages. It does not happen through a single dramatic deprecation. It happens through a download page going quiet. Panić framed his motive plainly. "I didn't think it was fair that people's ability to move or protect their own data should depend on a download page."

He told The Register he had tested OpenVDDK on vSphere versions 6.7 U3, 7.0.3, 8.0.3 and 9.1.1. Lab tests on the last version, reading through vCenter on a standalone ESXi host, produced matching results for every disk read. He also said independent testers had reported success with Forklift (VMware to KubeVirt, cold and warm), MTV on OpenShift, a Proxmox VE pipeline and custom nbdkit-based tools.

Then the caveat, which he volunteered. Reports of production use are "not independently audited customer endorsements or a guarantee of suitability for every production environment." That is the honest position for a project this young. It is worth repeating before anyone plans a migration around it.

He is not alone in the gap. The Register notes that OpenNebula's OneSwap tool claims a 96 percent success rate for KVM migrations. Nutanix has adjusted its Move tool to work without VDDK, and Red Hat is working on a fix with a published workaround.

The other kind of release: an operating system for agents

Two days earlier, on 1 November, a different project used the same word, "crew," to describe something less mundane. Friday, published on GitHub, runs Claude Code and Codex side by side on a local machine as a small team of coding agents. It is open source and it is early. It is also a sign of where release activity has moved: away from kernels and desktop environments, toward orchestration layers for models.

The same pattern shows up across the week's releases. On 3 November, a developer published Octomind, described as an open source AI harness built to save tokens and written in static Rust. Also on 3 November, NVIDIA posted auto-ontology, an open source repository. On 28 October, The Register reported that Cisco had open sourced Project Quarry. It is a behaviour detection engine that watches AI agents for sequences of actions that look benign one by one and malicious together.

Cisco's Amy Chang, head of AI threat and security research, gave The Register an exclusive interview on the project. She compared the problem to credit card fraud detection: a small test charge, then another purchase, then "now I'm gonna buy the PS5." Quarry takes the whole agent session into account, she said, because "one tool call may seem in-scope until you realize that that next step is taking that information and sending it to an attacker-controlled inbox or server."

"A policy can produce a finding when the trace shows a sensitive file read followed within a defined window by an external send," Cisco researchers wrote in a blog post shared with The Register.

That is a security product, not an operating system. But it is the same genre of release: a shared language, a rule format, a reference implementation, published before the market has settled on standards.

What the older releases tell us about the pattern

Not every open source release in the last week is about AI. On 27 October, a project called Zen appeared, a system-wide ad blocker for Windows, macOS and Linux. It installs a local proxy and a locally generated root certificate so it can filter HTTPS traffic, including from desktop apps. It is MIT licensed, and its site lists 33 contributors. On 1 November, a developer released Medo, a 4K media editor for Haiku OS in under 1.44MB. On 2 November, OpenCoaster arrived: a clean-room, browser-based theme park simulation game inspired by 1999 to 2004 era roller-coaster tycoon titles, with 19 rides and 13 stalls.

Those are hobbyist and small-team projects, and they matter to the people who use them. But they are not where the money and the argument are.

The economic question underneath all of this is whether publishing an inference engine as open source pays. A research post from userecurrence.ai on 3 November asks exactly that: is hosting an open source inference engine profitable? The post does not give its answer, so treat the question as open. What is not open is the direction of travel. An arXiv paper posted on 3 November examines open source 5G RAN platforms on performance and capabilities, which is the telecom version of the same debate.

There is also a legal AI angle. A developer published casecite on 28 October, describing it as a rebuilt, open sourced version of what Harvey raised 500M for. That is a claim from the project's own repository, not an audit, and should be read as such.

The infrastructure the software is being written for

All of this open source activity is happening against a backdrop of extraordinary physical buildout. The buildout is where the friction is.

On 3 November, Tom's Hardware reported that xAI is using eight converted jet engine turbines to power a Mississippi data center. The permits allow a 25 percent higher limit for volatile organic compounds. The New York Times, as cited by Tom's Hardware, reports that the Southern Environmental Law Center has challenged Mississippi's decision to approve the turbines. The group argues xAI has not guaranteed the turbines meet emissions standards for pollutants including formaldehyde. It also accuses the company of undercounting projected emissions. The same group previously sued xAI over gas generators used without required permits.

Also on 3 November, Tom's Hardware reported that Envision Energy has completed what it calls the world's first commercial data center deployment of an integrated 800-volt DC power system. The site is its Galaxy AI data center in Ulanqab, Inner Mongolia. The system converts 10kV medium-voltage AC directly to an 800V DC bus. It uses a solid-state transformer and integrated battery storage with millisecond-level coordination to absorb GPU load swings. Envision says the campus will scale to 3GW and eventually support up to one million AI accelerators. Those are the company's figures, from its press release.

The naming of these facilities is now a political act. On 3 November, The Verge reported that OpenAI has begun calling its data centers "compute campuses." The term appeared in an announcement launching its Grid Protectors Program, starting with a site in Effingham County, Georgia. In roughly 900 words, The Verge counted two uses of "data center" against far more of "compute campus." Axios had reported on Monday that OpenAI planned the rebrand after the midterms. Holly Elmore of PauseAI US told The Verge: "Rebranding them 'compute campuses' may fool listeners once or twice but it's not the name 'data center' that is unpopular, it's the giant, polluting construction projects." Tom's Hardware also covered the rebrand on 2 November, noting reports that OpenAI sees the 2028 presidential election becoming an AI issue.

OpenAI's government-facing moves continued on 3 November. CNBC reported that the company hired former US Army Secretary Dan Driscoll for a national security role. The Decoder, citing The Wall Street Journal and OpenAI's announcement, said the title is "Senior Counselor for National Security Modernization." Driscoll oversaw nearly one million soldiers and more than 265,000 civilian employees, and submitted his resignation in late August 2026.

The components crunch reaches the desktop

Meanwhile the hardware side of the AI buildout is showing up in an unlikely dataset. On 3 November, Tom's Hardware reported that Valve's October Steam Survey shows the share of systems with 32GB of RAM fell by 5.16 percent. That came after 32GB was the most popular configuration only a month earlier. The share of gamers with 1TB or more of total storage also shrank by nearly 5 percent, and 8GB and 16GB systems rose. Tom's Hardware links the shift to the AI-driven components crunch, with RAM the hardest-hit segment.

That is the context in which open source releases should be read. Software is getting cheaper to write and easier to publish. The machines it runs on are getting more expensive and more contested. A library that restores a migration path matters precisely because the alternative, staying put, now costs more.