Spheron vs SF Compute: GPU Marketplace vs Cluster Market Compared

SF Compute is a GPU market focused on cluster-scale workloads for AI researchers and teams that think in nodes. Spheron is a GPU marketplace designed for any AI team that wants instant access to bare-metal GPUs at competitive per-GPU rates. They serve overlapping needs, both offer H100 and H200 access without hyperscaler pricing, but the buying experience, hardware range, and pricing model are meaningfully different.

SF Compute was founded in 2023 by Alex Gajewski and Evan Conrad after they signed an inflexible 12-month GPU contract that required buying more capacity than their AI audio model startup actually needed. To avoid wasting that excess compute, they organized a shared GPU arrangement with other startups. Around 170 AI companies joined the collective within weeks, and that accidental marketplace became the foundation of SF Compute. Gajewski has since left the company; Conrad continues as CEO. The company raised a $40 million Series A in late 2025, co-led by DCVC and Wing Venture Capital at a $300 million valuation. By early 2026, standard VM-based GPU access became available via self-serve sign-up: create an account, fund it, and buy cluster time via CLI or buy page. Their about page describes a more selective model with NDA-based onboarding for enterprise cluster arrangements; for standard VM nodes (h100v), the process is self-serve. Its DNA is still research-grade and cluster-focused. Spheron, by contrast, is a commercial GPU marketplace aggregating capacity from multiple data center providers and pricing it per GPU with instant web-based deployment.

If you need a single H100 for a fine-tuning job this afternoon, or a dozen RTX 5090s for inference, the platforms diverge sharply. Here's the full comparison.

Quick Comparison

GPU pricing fluctuates based on availability and market demand. Prices shown are as of 13 Mar 2026. Always verify current rates at Spheron's pricing page and sfcompute.com.

How Each Platform Works

Spheron: Aggregated GPU Marketplace

Spheron aggregates bare-metal GPU capacity from vetted data center providers worldwide into a single marketplace. Sign up, add a payment method, pick a GPU from the catalog, and deploy, the whole process takes under five minutes through the web interface. The marketplace model means multiple providers compete on price for the same GPU types, keeping rates structurally lower than single-provider platforms. You pay per GPU per hour with no minimum commitment.

Spot pricing is available on most GPU types, cutting costs by 20–60% on interruptible workloads. For teams running training runs that can checkpoint and resume, spot pricing on Spheron dramatically reduces monthly GPU spend. For a full breakdown of what's available, see our GPU rental catalog.

SF Compute: GPU Cluster Market

SF Compute operates as a market for buying time on GPU clusters. It was originally built by founders who signed a 12-month GPU contract requiring more capacity than their own AI startup needed, then organized a collective to share that compute with other teams, a structure that grew into an open market when demand took off. Today you sign up on their website, fund your account, and purchase compute primarily via their CLI tool, though a dashboard buy page is available for basic orders.

The key structural difference: SF Compute sells in nodes. Each node contains 8 GPUs. Via their web buy page, the minimum order is 4 nodes (32 GPUs) for 4 hours; the CLI tool allows more flexible orders, potentially as small as one node for one hour subject to availability. This design optimizes for multi-node cluster training runs where you're buying 64, 128, or 512 GPUs at once, not for single-GPU fine-tuning or inference tasks.

SF Compute's documented node types are h100v (VM-based H100 nodes) and h100i (Kubernetes-based H100 nodes with 3.2Tb/s InfiniBand). A third type, h200ki (Kubernetes-based H200 nodes), was previously referenced but is no longer listed in SF Compute's current documentation; H200 in Kubernetes cluster configuration may be available on request via sales. VM nodes (h100v) are the default self-serve option and spin up in approximately 5 minutes without InfiniBand (InfiniBand for VMs is targeting Q2 2026). Kubernetes cluster nodes (h100i) offer InfiniBand and are available on request via sales contact.

SF Compute's available hardware is narrower: H100 SXM5 is the standard catalog. H200 SXM5 is listed on their buy page with hourly billing options, but as of March 2026 shows no active market pricing (displaying $0 on their prices page across all recorded dates); in practice, H200 access typically requires a dedicated or longer-term arrangement. B300 is on their roadmap for Q2 2026 (contact for pricing). B200 is not in their standard catalog but is available on request via their buy page.

Pricing: What You Actually Pay

Pricing is where the gap between these platforms is notable. SF Compute's H100 SXM5 on-demand rate starts from $1.64/GPU/hr per their prices page as of March 2026, with market averages between $1.49–$1.68/GPU/hr over Feb–Mar 2026. Their buy page shows the Richmond zone at $1.75/hr; the Yerba zone lists at $5.00/hr, illustrating that zone and availability drive significant price variation. Spheron's H100 dedicated pricing as of 13 Mar 2026 starts at $2.01/hr per GPU, in a similar range to SF Compute's market rates. Spheron's clear cost advantage is spot pricing: H100 spot instances from $0.80/hr per GPU, roughly 54% cheaper than SF Compute's Richmond zone rate for interruptible workloads.

Note: GPU pricing fluctuates over time based on supply and demand. The figures below reflect rates as of 13 Mar 2026.

Verify current Spheron rates at www.spheron.network/pricing and SF Compute rates at sfcompute.com before committing.

For dedicated H100, both platforms are in a similar range: Spheron from $2.01/hr per GPU vs. SF Compute from $1.64/hr (Richmond zone $1.75/hr) as of 13 Mar 2026. Where Spheron pulls ahead is spot pricing: H100 spot instances from $0.80/hr per GPU, roughly 54% cheaper than SF Compute's Richmond zone dedicated rate. For training runs that can checkpoint and resume, this gap is substantial.

For teams buying clusters at the 32-node or larger scale, it's worth getting a direct quote from both platforms, as SF Compute may have competitive pricing for committed large-scale research runs that isn't fully reflected in public documentation.

The minimum purchase difference matters for small teams. Spheron charges per GPU per minute with no minimums. If you need one H100 for six hours, you pay ~$12.06 on Spheron dedicated ($2.01/hr × 6 hrs), or ~$4.80 on spot ($0.80/hr × 6 hrs for interruptible workloads). On SF Compute via CLI, the smallest practical order is one node (8 GPUs); the same six-hour run costs roughly $84.00 ($1.75/GPU/hr × 8 GPUs × 6 hrs), even though you only need one GPU. Via SF Compute's web buy page, the minimum order is 4 nodes (32 GPUs) for 4 hours, which makes small single-GPU workloads impractical compared to Spheron's per-GPU pricing.

GPU Availability and Hardware Selection

SF Compute's standard self-serve catalog is focused: H100 SXM5 80GB is the primary VM option. H200 SXM5 141GB is listed on their buy page with hourly billing options, but shows no active market pricing as of March 2026 (showing $0 on their prices page); H200 access in Kubernetes cluster configuration is not currently listed as a standard node type in SF Compute's documentation and may be available on request via sales. B200 and B300 are not in the standard catalog (the buy page directs users to contact SF Compute for these); B300 is on their roadmap for Q2 2026. That leaves gaps if you need different GPU profiles on standard on-demand terms.

Spheron's marketplace includes 30+ GPU SKUs with varying performance and price points:

• H100 SXM5: flagship training hardware, from $2.01/hr dedicated, from $0.80/hr spot
• H200 SXM5: high-memory training, from $4.54/hr dedicated, from $1.72/hr spot
• GH200: high-memory training, from $1.97/hr dedicated
• RTX 5090: efficient inference and fine-tuning, from $0.69/hr dedicated
• RTX PRO 6000: professional workstation GPU for inference and fine-tuning, from $1.65/hr dedicated, from $0.72/hr spot
• A100 80GB: versatile training and inference, from $1.04/hr dedicated, from $0.45/hr spot
• L40S: inference-optimized, from $0.72/hr dedicated
• B200, B300: next-gen hardware available on the platform (B200 from $2.25/hr)

This breadth matters for teams that need to match hardware to workload rather than fitting every job into an H100. A fine-tuning job on a mid-size model may be more economical on an RTX 5090 or A100 at Spheron's rates than on an H100 SXM5 cluster. See our guide on the best NVIDIA GPUs for LLMs for a deeper look at matching GPU specs to model requirements.

Who Is SF Compute For?

SF Compute is genuinely well-suited for a specific type of buyer:

• Teams running large-scale pretraining or fine-tuning: if you're buying 64+ GPUs at once, SF Compute's node-based model and cluster-optimized infrastructure is a natural fit
• AI researchers familiar with CLI workflows: SF Compute's tooling is designed around sf nodes create and SSH-based access, which experienced ML practitioners handle comfortably
• Organizations that prefer to schedule compute in advance: SF Compute's market model lets you reserve clusters for specific time windows, which works well for planned training runs
• Teams with H100 SXM5 requirements: if you specifically need H100 SXM5 in cluster configurations, SF Compute has this as its core offering; H200 SXM5 is listed on their buy page but currently has limited on-demand market availability

SF Compute's roots in the AI research community mean it has a reputation among AI safety organizations, academic labs, and open-source AI teams. If you're in that ecosystem, the platform's focus on research-grade workloads aligns with your use case.

Who Is Spheron For?

Spheron's design choices optimize for a different buyer:

• Teams that need a single GPU or small multi-GPU setups: per-GPU pricing with no node minimum is fundamentally more flexible and affordable for smaller workloads
• Commercial AI teams building products: Spheron's web UI, API access, and instant deployment match the pace of product development cycles
• Teams that need hardware SF Compute doesn't stock: RTX 5090 for inference, A100 for legacy model compatibility, L40S for high-throughput inference, or newer B200/B300 hardware
• Startups and researchers who want spot pricing: Spheron's H100 spot starts at $0.80/hr per GPU vs. $2.01/hr dedicated, a 60%+ reduction for checkpointable workloads; SF Compute also offers spot nodes, but Spheron's broader GPU catalog means spot pricing applies across 30+ SKUs including A100 (spot from $0.45/hr), RTX PRO 6000 (spot from $0.72/hr), and H200 (spot from $1.72/hr)
• Teams that want web-based management: Spheron's dashboard handles deployment, monitoring, and billing without requiring CLI proficiency

For context on how Spheron compares across the broader GPU cloud market, see our comparison with RunPod, Vast.ai, and CoreWeave.

The Interface and Workflow Difference

SF Compute's primary interface is a CLI tool. The workflow is:

1. Sign up and fund account on sfcompute.com
2. Install the sf CLI
3. Run sf buy -n [node-count] -d [duration] -s [start-time] (VMs) or sf nodes create --zone [zone] --duration [d] with your preferred parameters
4. SSH into your nodes via startup script

This is a fine workflow for experienced ML engineers who live in the terminal. The buy page provides a basic web interface for purchasing VM clusters, but cluster management is CLI-centric. Kubernetes clusters with InfiniBand require contacting SF Compute sales directly.

Spheron is fully web-UI driven with API access for automation. The workflow is:

1. Sign up at app.spheron.ai
2. Add a payment method
3. Browse the GPU catalog, select hardware, configure, and deploy
4. SSH access, monitoring, and billing available in-dashboard

For teams that want infrastructure management without becoming proficient in a platform-specific CLI, Spheron's approach removes that friction entirely.

Making the Right Choice

Bottom Line

SF Compute is a capable platform that grew from a GPU resale operation into a legitimate GPU market. Its strength is cluster-scale H100 workloads for teams comfortable with CLI tooling and node-based purchasing. If you're planning a multi-week pretraining run and buying 512 GPUs at a time, it's worth getting a quote.

For most AI teams, those building products, running fine-tuning jobs, doing inference, or exploring GPU options without a multi-node minimum, Spheron is the more practical choice. The per-GPU pricing model, 30+ hardware SKUs, spot instances, and web-based deployment remove the overhead and cost floor that SF Compute's cluster focus introduces.

The hardware breadth alone is decisive for many workloads. If your stack runs on RTX 5090 for cost-efficient inference, or you need A100 for specific model compatibility, Spheron's catalog covers it. SF Compute doesn't.

For teams looking for alternatives beyond SF Compute, our top GPU cloud providers list and GPU rental guide cover the full landscape.

Need bare-metal GPU access today without node minimums or CLI overhead? Spheron has H100, H200, RTX 5090, A100, and more: deploy in minutes.

Start deploying on Spheron →