Accelerated Cloud Storage

@rrhoover Cc @TensorPool

Request for GPUs: A portfolio co is scaling fast and needs 400 H100s or 275 B200s asap for a one week reservation (potential to convert to a longer contract). LMK if you have a hookup. DMs open.

@SimonGoodman_ Classic

Slurm queue was empty most of the week, CVPR reviews were released yesterday, now cluster is at 100% utilization

Happy new year to all the developers!

Many small files are the bane of any storage system at scale

@jhleath @TurnerNovak @greptile Coderabbit works well for us

@TurnerNovak @greptile we have not! we've used a lot of different AI code review tools, and have mostly found them to be noisy because a lot of things that might look wrong turn out to be correct when you have more context. maybe this is an indictment around not using a monorepo in 2025

an interesting update: the team is starting to move away from AI coding completely (devin/claude/etc) because it's so much harder to review the AI code than writing things themselves

GPT 5.3 would be out next week if OpenAI used ACS 🤷‍♂️

Rsync and tmux are my op fr

rclone is one of the best storage developer tools ever made

@joshkmartinez how much data would it take to do your job?

how many gpus would it take to do your job??

@dylan522p It's free for now too

Just met a Claude Code user who has been the top daily user numerous times. He says Codex is better now.

FYI GPT 5.1 is free for the next few days, my Cursor bill is already too high

Wonderful presentation youtu.be/beWO7h7Ut44?si…

Great infrastructure should simplify the work of their users. Deleting code, removing complexity, removing unnecessary systems etc.

ACS is up and running fyi

Piece from our friends in @cerebral_valley for those who love reading instead of listening cerebralvalley.beehiiv.com/p/nebius-is-yo…

@iamgingertrash @dylan522p Has a good take on this. Cloud based inference is collocated with the data and can produce far more tokens per second than locally make it much more responsive.

It is curious that VC has now pivoted to on-device inference But just a year ago The same VC’s told me it was a fools errand

@LignoL23 @criccomini Rclone mount works pretty well

@criccomini This is also possible with rclone rclone s3 serve rclone.org/commands/rclon…

Someone just pointed me at this: github.com/versity/versit… Seems neat.

ACS is closed source for nearly all the software we produce. We would rather be honest and forthcoming with developer community.

GPU clouds like @TensorPool use ACS to cut storage costs, improve GPU utilization, and move training jobs seamlessly between clusters and providers.

One big surprise to our team coming from a big tech background is how much of an issue egress can be for multi-cloud GPU training setups. Using something like OCI across clouds can cause storage to cost more than the GPUs on an hourly basis. ACS doesn't charge for egress.