TL;DR
TL;DR: Shadeform is a YC-backed GPU marketplace that aggregates 8+ cloud providers behind one API, letting you search live H100/A100 availability, launch instances in your own account or theirs, and pay a 5% fee only on ShadeCloud deployments.
Source and Accuracy Notes
- Product site: shadeform.ai
- Platform: platform.shadeform.ai
- Original launch: Show HN, 62 points, August 2023 (objectID 37164864)
- Founders: Ed, Zach, and Ronald (ex-AWS and Azure)
The product has been live for nearly three years at this point, and the GPU-marketplace category it helped pioneer has only gotten more crowded. Pricing and provider counts below reflect what the team shared in the Show HN thread and on the public site; for current numbers always check shadeform.ai.
What Is Shadeform?
Shadeform is a “single plane of glass” for GPU compute. Instead of opening accounts with Lambda, RunPod, CoreWeave, AWS Capacity Blocks, and a handful of regional H100 operators, you get one dashboard and one REST API that talks to all of them. The founders describe it as moving away from “all-encompassing hyperscalers” toward specialized clouds for high-performance use cases.
The product solves three problems that anyone who has tried to rent an H100 in 2024–2026 will recognize:
- Fragmented inventory. A100s and H100s are available, but spread across dozens of niche providers with different provisioning models, IAM systems, and pricing pages.
- Reservation friction. When a particular GPU is sold out at one provider, you have to manually check the next one. Shadeform’s reservation system polls real-time availability and launches an instance the moment stock returns.
- Vendor lock-in. Once you build IaC around one provider’s APIs and quotas, switching costs grow fast. Shadeform gives you a portable layer on top so you can move workloads between clouds without rewriting pipelines.
Why It Matters Now
The Show HN launch was in August 2023, right when H100 scarcity was peaking. Three years later, scarcity has shifted from “H100s everywhere” to “the new generation is locked up the day it’s announced.” B200 and MI300 capacity now follows the same pattern H100s did in 2023. A multi-cloud abstraction layer is the obvious response, and Shadeform has had runway to mature.
The second reason it matters: inference economics. As more teams self-host open models to control cost and data, the ability to spin up a temporary 8xH100 node for a fine-tuning job, then tear it down, is much more useful on a marketplace that prices per-hour than on a reserved hyperscaler contract.
How the Marketplace Actually Works
The two deployment modes are the cleanest part of the design:
# Mode 1: bring your own cloud (BYOC) — free
# Add your cloud credentials (AWS, GCP, Azure, etc.) in the Shadeform dashboard.
# Shadeform's control plane provisions instances in your own account.
# Mode 2: ShadeCloud — 5% platform fee
# Launch directly in Shadeform's pooled capacity.
# No need to negotiate with multiple hyperscalers.In BYOC mode, you keep your existing enterprise discounts, committed-use pricing, and IAM boundaries. Shadeform just acts as a multi-cloud orchestrator. The 5% fee only applies to ShadeCloud deployments, which is competitive against hyperscaler markups on spot-like SKUs.
Repo-Specific Setup Workflow
Step 1: Sign up and create an API key
Head to platform.shadeform.ai, create an account, and generate an API key from the dashboard. The key is what your scripts and CI will use to launch instances.
export SHADEFORM_API_KEY="sk-..."Step 2: Search live GPU availability
The marketplace exposes a search endpoint that returns real-time pricing and stock across all connected providers. This is the killer feature when you need an H100 today and don’t want to bounce between eight pricing pages.
curl -s https://platform.shadeform.ai/api/v1/instances \
-H "Authorization: Bearer $SHADEFORM_API_KEY" \
-G --data-urlencode "gpu_type=H100" \
--data-urlencode "region=us-east" | jq '.data[] | {provider, hourly_price, available}'You’ll get back a list of providers, their current hourly price, and stock status. The same call works for A100, L40S, RTX 4090, and the newer B200 / MI300 SKUs as they come online.
Step 3: Launch an instance
Pick a provider and launch. The API normalizes the launch call, so you don’t have to learn Lambda’s CLI, RunPod’s SDK, and the AWS capacity-blocks JSON syntax.
curl -s -X POST https://platform.shadeform.ai/api/v1/instances \
-H "Authorization: Bearer $SHADEFORM_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"provider": "lambda",
"gpu_type": "H100",
"gpu_count": 8,
"region": "us-east",
"image": "runpod/pytorch:2.4.0-py3.11-cuda12.4.1-devel"
}'Within a minute or two you get back an SSH endpoint or a remote-desktop URL, depending on the image you selected. Shadeform’s reservation mode handles the “wait until stock returns” case automatically.
Step 4: Tear down and bill
When your job is done, stop the instance from the dashboard or via the API. ShadeCloud usage is billed hourly with a 5% platform fee layered on the underlying provider’s price. BYOC usage is free — you pay the provider directly.
Deeper Analysis
Reservation system. The most underrated feature is the reservation mode. If the exact SKU you want is sold out, you queue a reservation and Shadeform launches it the moment capacity appears on any connected provider. For teams running scheduled fine-tuning jobs, this turns a multi-day search into a single API call.
Multi-cloud load balancing (roadmap). The team has signaled that multi-cloud load-balanced inference is on the roadmap. If they pull it off, you’d be able to point a serving endpoint at Shadeform and have it auto-distribute requests across whichever provider has the cheapest H100 spot price that hour. That’s the real “hardwareless GPU cloud” vision they described in the original launch.
BYOC vs ShadeCloud trade-off. Bring-your-own-cloud is the right default for any team with existing hyperscaler commitments or strict data-residency requirements. ShadeCloud is the right default for ad-hoc workloads and small teams who don’t want to manage a second cloud account. The 5% fee is reasonable but not free — compare it against whatever your current orchestration overhead actually costs in engineering time.
What’s missing. Three years in, the platform is still strongest on NVIDIA SKUs. AMD MI300 and Intel Gaudi support exists but is thinner. If your workload is specifically optimized for one of those accelerators, check the current provider list before committing.
Practical Evaluation Checklist
- [ ] Confirm your required GPU SKU is in stock across at least 2 connected providers (resilience against single-provider outages)
- [ ] Decide BYOC vs ShadeCloud based on existing cloud commitments
- [ ] Test the launch API with a small job before wiring it into production training pipelines
- [ ] Compare ShadeCloud’s 5% fee against the engineering time saved by not maintaining multi-cloud IaC
- [ ] Set billing alerts in the Shadeform dashboard for ShadeCloud deployments specifically
Security Notes
BYOC mode is the more security-sensitive deployment. You’re handing Shadeform credentials that can launch instances in your AWS, GCP, or Azure account, so the blast radius is real. Use a dedicated cloud sub-account or project with hard budget caps and SCPs that prevent the credentials from doing anything outside the GPU SKU family. Rotate the API key on a schedule.
ShadeCloud instances run in Shadeform’s pooled accounts, so the isolation model is closer to a traditional VM provider. Your data is on someone else’s hardware — make sure that matches your compliance posture before pushing customer data through it.
FAQ
Q: How does Shadeform make money if BYOC is free? A: The 5% fee on ShadeCloud deployments is the primary revenue stream. BYOC is free to keep the funnel wide and to make Shadeform useful for teams that can’t move workloads to a new cloud account.
Q: Which GPU types are supported? A: NVIDIA A100, H100, H200, L40S, RTX 4090, and the newer B200 SKUs as providers add them. AMD MI300 and Intel Gaudi support is thinner. Check the live availability endpoint for current stock.
Q: Is Shadeform a YC company? A: Yes. The founders went through YC W22 and launched publicly on Show HN in August 2023 with 62 upvotes. The product has been live and iterating since.
Q: How does this compare to RunPod or Lambda directly? A: RunPod and Lambda are individual GPU providers that you can also access through Shadeform. Going direct gives you a slightly lower per-hour price and tighter provider-specific tooling. Going through Shadeform gives you a single API, automatic multi-cloud fallback, and reservation handling.
Q: Can I run multi-node training jobs across providers? A: Inter-node bandwidth is the constraint, not the API. A single Shadeform-launched cluster lives inside one provider’s network fabric, so NCCL/RCCL collectives are fast. Coordinating a job that spans two providers is technically possible but the cross-cloud bandwidth makes it impractical for tightly-coupled training.
Conclusion
Shadeform has carved out a clear niche: a portable control plane for GPU compute across an otherwise fragmented provider landscape. Three years after launch, the BYOC + ShadeCloud split, the reservation system, and the unified API are all more relevant than they were in 2023. If your team is spending engineering time on multi-cloud GPU orchestration, it’s worth a pilot.
Next time, we’ll look at a complementary tool — a managed inference platform that targets the same audience from the deployment side. If you want a deep-dive on Shadeform’s API or a comparison with a specific competitor, drop a comment on the runany.dev blog.