vast.ai (@vast

vast.ai

744 posts

vast.ai

@vast_ai

Vast is the infrastructure layer where AI agents autonomously design, procure, and optimize their own compute.

California, USA

Joined August 2018

Affiliates Replies Articles Media

Pinned
vast.ai
@vast_ai
Jul 6
Vast.ai is hiring. We're growing. Fast. Agent-ready AI infrastructure. Profitable, venture-backed. New roles across engineering, research + DevRel, mostly deep systems + GPU work. On-site in SF or LA. vast.ai/jobs
vast.ai reposted
Aaron Lemke
@aaronlemke
Jul 10
Sol from @OpenAI one-shotted a multi-tentacled creature morphology and RL training harness, launched and monitored the run on @vast_ai on an H200 and created a @threejs + @webgl_webgpu browser environment for the harvested policy to crawl around in. I woke up and this guy was
00:00
vast.ai
@vast_ai
Jul 9
Token-based API pricing charges you per word your agent thinks. Rent GPUs and self-host instead, and you pay for compute time, not tokens. You control the efficiency. Why agents cost more than chatbots (and what to do about it):
Why AI Agents Cost More Than Chatbots (and How to Architect Around It)
From vast.ai
vast.ai
@vast_ai
Jun 30
NVIDIA Rubin is official: 6 chips, rack-scale design, datacenter-as-one-computer. 50 petaFLOPS NVFP4. 260 TB/s per-rack interconnect. First rack-scale confidential computing. Rumored: 2.3kW per GPU, 22 TB/s memory bandwidth (unconfirmed). Full breakdown 👇
NVIDIA Rubin Platform: Everything We Know So Far
From vast.ai
vast.ai
@vast_ai
Jun 27
One embedding model. Multiple vector sizes. No retraining. Matryoshka embeddings let you tune retrieval cost vs. quality at runtime. 768 dims for accuracy, 64 dims for speed. Same model handles both. Here's how:
Matryoshka Vector Embeddings: Flexible Embeddings for Cost-Efficient AI Systems
From vast.ai
vast.ai
@vast_ai
Jun 24
You can now rent GPUs from inside your coding agent. Vast ships as a plugin for Claude Code, Codex, and Cursor. Search, launch, monitor, and tear down GPU instances without leaving your editor — tell the agent what you want, or run a slash command directly. Safe defaults keep it
vast.ai
@vast_ai
Jun 24
running large model inference on h200 80GB HBM3e for $2.89/hr if you catch the p10. Other clouds averaging $8.46 for the same GPU.
vast.ai
@vast_ai
Jun 23
people keep saying h100s are holding steady but they dropped 8% this week. $2.15 median, $1.87 at p10. Supply is tight at 59 cards though
vast.ai
@vast_ai
Jun 19
rtx 5090 at $0.50/hr median with 1,244 available. Prices down 6% from last week.
vast.ai
@vast_ai
Jun 18
deepseek v3 on 8xH200, $3.42/hr. 211 instances available right now across the fleet
vast.ai
@vast_ai
Jun 17
We built a fully functional coding agent in under 100 lines of Python. No framework. No magic. Just the model requesting actions and your program executing them. Runs on a single RTX 5090 with Ollama. Full code + breakdown:
How to Build an Agent on Vast.ai in Less Than 100 Lines of Python
From vast.ai
vast.ai
@vast_ai
Jun 17
A100 80GB median is $1.05/hr, bottom decile hits $0.67/hr. 141 cards live. Solid option if you're doing multi-GPU training runs on a budget.
vast.ai
@vast_ai
Jun 16
NEW Installer Wizard for Hosts This TUI offers a new guided terminal setup that replaces the CLI script. It walks you through each step, validates your driver, hardware, ports, and speed before installing, and gives plain-English fixes instead of stack traces. Find the guided
GIF
vast.ai
@vast_ai
Jun 16
B300 has 2x the memory of H200 (288 GB vs 141 GB) and 8 TB/s bandwidth. But is it actually worth the 1,400W TGP for your workload? We break down H200 vs B300: specs, use cases, and when Blackwell Ultra is the right call 🧵
NVIDIA B300 vs. H200: Is Blackwell Ultra Worth the Upgrade?
From vast.ai