Runcrate

Train, deploy and scale AI.

[Runcrate @runcrate]

GPUs and clusters don't have to break your bank. Spawn the cheapest on-demand instances or request a cluster. Choose Runcrate today.

[Runcrate @runcrate]

@arthurlge 56M is tiny, an A100 is overkill. A single RTX 4090 is $0.57/hr on Runcrate right now, and you only pay while it's running, so a quick run costs cents. If you try it, would you write it up? Curious how much faster it is than the Mac.

[Runcrate @runcrate]

@lipko Agree. Most of the hourly GPU market isn't frontier training, it's A100s and L40S doing fine-tuning and inference for small teams. None of that shows up when people tie the CapEx story to two labs.

[Runcrate @runcrate]

New Compute Report: did GPT-6 "Astra" really edit Qwen's weights, or just write Python to get the math right? If you run open-weight models, which of those happened changes what the result tells you. https://t.co/zuxzhLXUWX

[Runcrate @runcrate]

@YingjunWu The rent line decides it. Read just now on Runcrate: H200 is $4.63/hr on demand and $3.94 on a 12 month prepay, about 15% off. B300 is $13/hr. Did you price it at on-demand or term rates?

[Runcrate @runcrate]

Hermes Agent's plugin catalog now lists a Runcrate provider (v0.22.6): open models through one OpenAI-compatible endpoint, picked from Hermes. Thanks @Teknium and Nous Research for merging it. https://t.co/JR5S6hFoda

[Runcrate @runcrate]

Runcrate, the AI cloud for agents, is now in Claude's connector directory. Ask in plain words: check GPU prices and stock, launch an instance, run a job that stops billing when its command ends, schedule a run. Each sign-in is capped to one budget. https://t.co/5bpKaGLfPH

[Runcrate @runcrate]

@SemiAnalysis_ The rent input decides most of this. Read just now on Runcrate: H200 from $4.63/hr, B300 $13/hr on demand. So a B300 only wins if it pushes about 2.8x the H200's tokens per hour on V4.1 Flash. Worth rerunning the break-even at today's rates, not list prices.

[Runcrate @runcrate]

@guilleflorvs the 30 cents only holds if those 56 B200s stay busy. one office timezone means the fleet sits half idle overnight. the teams that keep a blended number that low push evals, backfills and batch jobs into the dead hours, otherwise the real per token cost quietly doubles.

[Runcrate @runcrate]

@daniel_koss Rental side, read at 10:30 UTC on Runcrate: H100 from $3.55/hr, H200 $4.63, B200 $7.67, B300 $13. B300 was $9.74 on Saturday. Record GPU prices and a falling stock can both be true when the margin sits with whoever locked the multi-year contracts.

[Runcrate @runcrate]

@silicondata @ArtificialAnlys Curious whether the indices split single GPUs from full 8x nodes. On our marketplace right now the same H100 runs $3.55 to $7.79/hr depending on config and host, so one blended number can hide a 2x spread.

[Runcrate @runcrate]

@Abbyxvx87 Rental side data point for the 10x volume, 90% price math: the cheapest H100 on Runcrate went from $3.55 to $2.83/hr between Saturday and today. Spot moves 20% in two days, which is why the neocloud case lives in multi-year contracts, not list prices.

[Runcrate @runcrate]

@ArtificialAnlys @silicondata Useful metric. One caution from the marketplace side: spot moves fast. On Runcrate the cheapest H100 is $2.83/hr today, $3.55 on Saturday, and B300 went the other way, $9.74 to $12.35. Throughput per $/hr can flip inside a week.

[Runcrate @runcrate]

Monday GPU board, live on Runcrate at 15:40 UTC:

H100 from $2.83/hr, down from $3.55 on Saturday
H200 from $4.63/hr, 12 listings, up to 8x
B200 from $7.93/hr
B300 from $12.35/hr, up from $9.74

H100 is the buy this week.
https://t.co/DGeBydDeGO

[Runcrate @runcrate]

Runcrate is now a built-in provider in open-source tools people use every day: claude-mem, airi, Fabric, Docker's docker-agent, genoffice, DeepTutor and more. Pick Runcrate from the provider list, add a key, run open models. Thank you to every maintainer who merged it.

[Runcrate @runcrate]

@alex_verem Owning the weights doesn't have to mean owning the hardware. A ~30B open model in FP8 is about 30GB of weights, so a single 80GB card serves it with room left for KV cache. Rent that by the hour when the laptop runs out, and it's still your version of the model.

[Runcrate @runcrate]

@outliercapx Live read at 12:30 UTC on Runcrate (I work there): H100 from $3.55/hr across 14 listings, while B200 is down to 2 listings, both full 8x nodes at $7.93/GPU-hr. B200 tightness shows up as missing listings before it shows up in an index.

[Runcrate @runcrate]

B200 on Runcrate: 5 listings from $7.15/hr this morning, zero right now.

H200 still has 13 options from $4.63/hr. If your model fits in 141GB, that's the card you can actually get today.

https://t.co/E0Dsp3NWqF

[Runcrate @runcrate]

@jeremychone @cGPU_Official Thanks for adding us, Jeremy.

[Runcrate @runcrate]

@pyth0ngiant Sorry for the slow reply on this one, by the way.

[Runcrate @runcrate]

@pyth0ngiant A full 8x B200 node is live on demand on Runcrate right now. If you need one reserved for a longer term, DM us the term and start date and we'll take it from there.