After looking at the VM behind Meta’s Muse, I then got curious about what OpenAI’s dots run on. So I asked my dot to inspect its own environment.
Here’s what that session reported:
- 9 logical CPUs visible, with CPU affinity covering cores 0–8
- AMD EPYC 9V74 Processor, reporting about 2.6 GHz
- Hypervisor flag present in /proc/cpuinfo
- 9.73 GiB of RAM, with no swap
- A workspace filesystem reporting about 31.45 GiB total, with 29.67 GiB available
- Debian 13.6, Linux kernel 6.18.44
It’s powerful, but unlike the Muse VM, which was ridiculously open and I could SSH onto it, this one was locked down. It lacked general internet access, and I couldn’t get my dot to benchmark it since it couldn’t download sysbench.
Does OpenAI cripple the machine?
Yes.
With Muse, I ran a sustained workload and compared throughput over time. I can’t run that experiment set here.
My dot can run shell commands and use a Linux workspace, but filesystem writes and direct network access are restricted. And these measurements don’t tell me whether the same machine stays allocated between sessions, while Muse was happy to stay running for extended periods of time.
Due to the specs of this machine, I’m pretty sure it’s ephermeal and spun up only when the dot is directly working in it, unlike Meta’s. Meta was pretty proud to point out they give every 1 user 1 VM.