Shadeform

NVIDIA Blackwell Ultra B300 GPUs are now live on Shadeform. Blackwell Ultra brings a suite of improvements to the already dominating Blackwell architecture: ☑️ 288 GB vs 180 GB HBM3E memory (+60%) ☑️ 8 TB/s vs 7.7 TB/s memory bandwidth (+4%) ☑️ 15 PetaFLOPS vs 8 PetaFLOPS FP4 compute (+88%) With Shadeform, you get on-demand access to the latest GPU architectures as soon as they're released to the market. See the link in the comments or head over to our website to deploy.

3 years ago I was sitting in my house in North Carolina staying up all night drawing crazy diagrams about a unified infrastructure layer for compute and cloud services backed by an ecosystem of cloud marketplaces. It's incredibly energizing when projects that start from the same spark converge and jointly provide real value to the world. I am so glad Shadeform can power SkyPilot's many users with reliable, affordable compute and help build this new cloud ecosystem at scale today. Upwards and onwards!

Shadeform is now live on SkyPilot! We're excited to announce that Shadeform is now officially integrated with SkyPilot, the open-source framework for running AI workloads across clouds. With this integration, you can now access Shadeform's marketplace of 20+ vetted GPU cloud providers directly through SkyPilot's unified interface. Simply specify `infra: shadeform` in your task configuration, and SkyPilot will automatically search across our providers to find the optimal GPU for your workload at the best price. Why this matters: ☑️ Automatic price and availability optimization across our entire marketplace ☑️ Seamless provisioning and job management ☑️ Access to thousands of on-demand GPUs without quota restrictions ☑️ One command to leverage 20+ cloud providers To get started, just install SkyPilot, grab your Shadeform API key, and you're ready to launch jobs with better pricing and availability than ever before. Big thanks to the SkyPilot team for building such an incredible framework. See the link in the comments for our integration guide.

Shadeform has partnered with Hydra Host to make their high-performance bare metal GPU infrastructure deployable in two clicks on our marketplace. Hydra brings highly competitive pricing across both B200 and H200 GPUs, demonstrated reliability, and adds to our continuously growing bare metal capacity. If you’re looking for bare metal B200 or H200 GPUs at a highly competitive price, deploy Hydra’s offerings on-demand on the Shadeform marketplace.

Partner Spotlight: Shadeform x Hydra Host Hydra’s high-performance bare metal GPUs are now available to deploy on Shadeform’s GPU Cloud Marketplace. Through this partnership, developers and ML teams can explore and deploy Hydra’s infrastructure in two clicks. It’s a simple and transparent way to get powerful bare metal compute without the layers of abstraction that come with cloud. This partnership works because both teams value clarity and performance. Shadeform (YC S23) makes it easy to discover and compare options across the cloud market, and Hydra brings the reliable, competitively priced bare metal that makes scaling possible. Together, we’re helping builders train models faster and run workloads the way they were meant to, directly on hardware built for the job.

While Deepseek's V.3.2-Exp (DSA) is impressive with Sparse Attention, Memory Efficiency, however, it would be overstatement to say that it can smoothly run on a single H200 Node. Here is my take As our H100 inventory is sold out, had to shop else where for H200/B200 ➤ Platform - H200 + Ubuntu24 + Nvidia Driver 580 + CUDA 13 ➤ Parallelism: Tried -dp 8 --enable-expert-parallel → hit NCCL/TCPStore “broken pipe” issues ➤ Stable bring-up: -tp 8 (Tensor Parallel across 8 H200s) ➤ Long Warmup: Long FP8 GEMM warmups + CUDA graph capture on first run (subsequent restarts are much faster due to cache) ➤ Client validation: One-file OpenAI-compatible Python script; plus lm-eval for GSM8K ➤ Grafana: Dashboard parameterized with $model_name = deepseek-ai/DeepSeek-V3.2-Exp ➤ Cloud Provider: Shadeform (YC S23) ➤ Total Cost: $54/2 hours Thanks to Ed Goode Ronald Ding for the GPU credits. https://lnkd.in/gsCvBqnw

We're seeing companies in the AI landscape quickly embracing stablecoin payments on Stripe. Like Vercel, which now lets developers pay for v0 in stablecoins. Or Shadeform (YC S23), a GPU cloud marketplace, which also accepts stablecoins. For their global, high-volume users, ACH wasn’t an option and credit cards meant high fees and frequent declines. Stablecoins now account for 20% of their payment volume, and have boosted overall revenue by 10%. Costs are lower, too: Shadeform pays just 1.5% to accept stablecoins vs. 4.5% for international cards. Stablecoins are proving to be more than a crypto curiosity -- they’re becoming a practical way to monetize AI offerings globally.

Accepting stable coin payments has been a huge unlock to increase Shadeform's margins and dramatically speed up settlement times. Thanks for making it easy Stripe! Check out their recent case study: https://lnkd.in/gs7YmmAc