NVIDIA DGX Cloud Lepton is a GPU brokerage marketplace: you request compute, Lepton allocates it from a curated set of neoclouds (CoreWeave, Lambda, and others), and NVIDIA handles billing. That single-vendor allocation layer is where the friction starts. Teams that want pricing transparency, specific neocloud relationships, regions outside the Lepton network, or simply don't want NVIDIA sitting between them and their compute provider are looking for alternatives. Here are 10.
Why Teams Evaluate Lepton Alternatives
Lepton has real advantages: NVIDIA's ecosystem backing, a growing neocloud partner network, and a unified API across providers. But three friction points keep coming up.
NVIDIA-controlled allocation. Lepton adds NVIDIA's brokerage margin on top of neocloud rates and controls which providers enter the network. That is a conflict of interest when NVIDIA also sells hardware to those providers. Teams that need CoreWeave directly, or a provider not in Lepton's curated list, have no path forward through Lepton.
Opaque pricing. Published Lepton rates bundle the underlying neocloud price with brokerage overhead. You cannot compare Lepton pricing directly against going to CoreWeave or Lambda directly without doing your own math. For teams modeling infrastructure cost at scale, that opacity is a real blocker.
Curated provider set. The Lepton marketplace does not include all neoclouds. Teams needing specific regional data centers, GPU SKUs not yet in the Lepton catalog, or provider relationships they have built independently are locked out.
1. Spheron: Best Overall Alternative
H100 SXM5: $5.01/hr on-demand / $2.91/hr spot | H200 SXM5: $5.92/hr on-demand / $3.31/hr spot | B200 SXM6: $9.36/hr on-demand / $2.68/hr spot | Per-minute billing | No contracts
Pricing fluctuates based on GPU availability. The prices above are based on 04 Jun 2026 and may have changed. Check current GPU pricing → for live rates.
Spheron and Lepton are both multi-provider GPU marketplaces, but they solve the problem differently. Lepton is NVIDIA's own brokerage layer with a curated neocloud partner set. Spheron is an open, decentralized marketplace not controlled by any hardware vendor, with verified live pricing published on a public API.
| Spheron | DGX Cloud Lepton | |
|---|---|---|
| Ownership | Decentralized marketplace | NVIDIA-operated brokerage |
| Pricing model | Verified live pricing via public API | Brokerage rate on top of neocloud |
| Provider network | 5+ providers | NVIDIA-curated neocloud set |
| Contract required | No | Varies by allocation size |
| Bare metal | Guaranteed | Depends on underlying neocloud |
| API | Public, no auth for pricing | NVIDIA developer account required |
| Vendor lock-in | None | NVIDIA controls allocation layer |
Spheron GPU pricing (live, 04 Jun 2026):
| GPU | On-Demand | Spot |
|---|---|---|
| H100 SXM5 | $5.01/hr | $2.91/hr |
| H100 PCIe | $2.01/hr | N/A |
| H200 SXM5 | $5.92/hr | $3.31/hr |
| B200 SXM6 | $9.36/hr | $2.68/hr |
| A100 80G | $1.48/hr | $1.19/hr |
| GH200 PCIe | $3.02/hr | N/A |
Pricing fluctuates based on GPU availability. The prices above are based on 04 Jun 2026 and may have changed. Check current GPU pricing → for live rates.
Three differences matter most when comparing Spheron to Lepton:
No brokerage overhead. Spheron's marketplace connects you directly to provider capacity with transparent pricing. There is no NVIDIA intermediary adding margin on top of the underlying rate. What you see in the API is what you pay.
No vendor conflict. NVIDIA sells hardware to the same providers it brokers through Lepton. That creates structural tension between acting as a neutral marketplace and protecting NVIDIA's hardware partner relationships. Spheron has no such conflict.
Bare-metal access is guaranteed. Lepton's bare-metal guarantee depends on which underlying neocloud takes your job. Spheron's provider partnerships are specifically for bare-metal access, so you get root SSH, full VM control, and custom driver installation regardless of which backend fills your request.
H100 SXM5 on Spheron starts at $2.91/hr spot, compared to brokerage-dependent rates through Lepton that vary by partner and allocation. For H200, H200 on Spheron is available at $3.31/hr spot. For teams evaluating Blackwell, B200 capacity on Spheron is available on spot at $2.68/hr.
For a full multi-cloud comparison, see our Shadeform alternatives guide.
2. Shadeform
H100 from ~$2.01/hr | API-first multi-cloud brokerage | 30+ partner clouds
Shadeform is also a multi-cloud GPU marketplace and brokerage layer, but it is not NVIDIA-controlled. You get a unified API across 30+ partner clouds with transparent routing. Unlike Lepton, Shadeform does not apply NVIDIA's brokerage margin and is not restricted to NVIDIA's curated neocloud set.
What it does well: Clean API across dozens of providers. Good catalog coverage for teams that want programmatic multi-cloud access without building their own provider integrations. No NVIDIA vendor relationship to navigate.
Where it falls short: API-first model requires integration before you can deploy anything. Bare-metal access depends on the underlying provider, so it is not guaranteed. Hardware availability varies by partner stock.
Best for: API-driven teams who want multi-cloud GPU access without NVIDIA's brokerage overhead and can handle the integration work upfront. See our dedicated Spheron vs Shadeform comparison for a detailed head-to-head.
3. SkyPilot
Open-source | No GPU supply of its own | Multi-cloud job orchestration
SkyPilot is an open-source multi-cloud orchestrator, not a marketplace or brokerage. It does not source compute itself. You configure SkyPilot with credentials to your own provider accounts (Spheron, AWS, Lambda, etc.), and it handles job placement, cost optimization, and spot fallback across them. Where Lepton is a closed allocation system, SkyPilot is a transparent optimizer you control.
What it does well: Policy-driven job submission across any provider you configure. Automatic spot fallback across clouds when a provider runs dry. Cost optimizer that picks the cheapest available instance matching your spec. Open-source, no vendor lock-in by design.
Where it falls short: SkyPilot does not supply GPUs, so you need provider accounts set up. For teams wanting a single-platform solution, SkyPilot adds orchestration complexity. It is not a drop-in replacement for Lepton's managed inference tier.
Best for: Teams with existing provider accounts (or those adding Spheron as a backend) who want automated multi-cloud cost optimization without NVIDIA controlling the routing. For a full walkthrough, see our SkyPilot multi-cloud GPU orchestration guide.
4. CoreWeave
HGX H100 8-GPU node: ~$49/hr (~$6.15/hr per GPU) | InfiniBand networking | Enterprise onboarding
CoreWeave is one of the providers Lepton routes to. Going direct removes the NVIDIA brokerage layer and lets you negotiate capacity terms directly. CoreWeave runs Kubernetes-native infrastructure with InfiniBand networking and large-cluster support up to 256+ GPUs.
What it does well: Best-in-class InfiniBand networking for distributed training. Large cluster availability. Enterprise SLAs with guaranteed uptime. Early access to new hardware generations through NVIDIA's partnership.
Where it falls short: On-demand HGX H100 pricing is high at ~$6.15/hr per GPU. New users go through enterprise onboarding with credit checks. Competitive reserved pricing requires multi-month contracts.
Best for: Large enterprises and frontier AI labs training models at scale with multi-year compute budgets who need dedicated infrastructure. Not for teams looking for flexible on-demand access or spot pricing. For a direct comparison, see our Spheron vs CoreWeave guide.
5. Crusoe
H100 on-demand: ~$2.98/hr | Sustainable compute focus | Direct neocloud
Crusoe is a direct neocloud that operates its own infrastructure with a focus on low-carbon compute using stranded and flared natural gas. It is another provider that Lepton can route to. Going direct gives you a cleaner billing relationship without brokerage overhead and a predictable SLA from a single provider.
What it does well: Competitive H100 on-demand pricing. Direct provider relationship with no intermediary. Sustainability positioning matters for teams with carbon reporting requirements. Clean API and straightforward provisioning.
Where it falls short: Smaller catalog than aggregator platforms. Limited regional coverage compared to hyperscalers. Less hardware variety (focused on H100/A100 core workloads). For a detailed comparison against Spheron, see our Spheron vs Crusoe guide.
Best for: Teams that want a direct neocloud relationship with a sustainability angle and predictable pricing on H100 workloads, without Lepton's NVIDIA brokerage layer in the middle.
6. Nebius
H100: $2.95/hr | GDPR-compliant EU infrastructure | Growing Blackwell catalog
Nebius (built by the former Yandex Cloud team) runs its own infrastructure across Europe and a US cluster in Kansas City. Like Crusoe, it is a provider that Lepton could route to, and going direct removes the brokerage layer.
What it does well: GDPR-compliant EU infrastructure for teams with data residency requirements. Growing catalog with H100, H200, and HGX B200 available. Direct provisioning without a sales process on standard configurations.
Where it falls short: H100 on-demand at $2.95/hr is on the higher end compared to marketplace spot options. Quota-based access for H100 and H200 at scale requires an approval process. EU-centric infrastructure adds latency for non-EU teams.
Best for: EU teams with GDPR compliance requirements who want a direct neocloud relationship. For a full comparison, see our Spheron vs Nebius guide.
7. RunPod
H100 SXM: ~$2.69/hr (on-demand) | Community Cloud and Secure Cloud tiers | Serverless + dedicated
RunPod is an AI-focused GPU cloud with two tiers: Community Cloud (independent hosts, cheaper) and Secure Cloud (managed infrastructure). Unlike Lepton's brokerage model, RunPod gives you a direct product experience with fast self-serve onboarding.
What it does well: Wide GPU selection across multiple tiers. Per-second billing. Large template library for PyTorch, vLLM, ComfyUI, and stable diffusion. Serverless GPU endpoints for teams that need auto-scaling inference.
Where it falls short: Community Cloud instances can go offline mid-job (independent hosts). Standard instances run in containers, not bare metal. Support quality varies between tiers.
Best for: Teams that want fast self-serve deployment and don't need guaranteed bare-metal access. Developers building prototypes or smaller production inference endpoints who want broad template coverage. For a broader comparison, see our RunPod alternatives guide.
8. Vast.ai
H100 PCIe from ~$1.87/hr on marketplace | Peer-to-peer | Interruptible instances
Vast.ai is a decentralized peer-to-peer GPU marketplace where independent hosts list hardware and renters choose based on price, location, and host rating. The model creates some of the lowest floor prices available across the market.
What it does well: Lowest prices when supply is strong. Huge hardware variety from RTX 3090 through H100 and H200. Interruptible instances for cost-sensitive batch workloads.
Where it falls short: No uptime SLAs; unverified hosts can go offline. Hardware quality and networking vary significantly by host. No enterprise support. Standard instances run as Docker containers.
Best for: Price-sensitive teams running interruptible batch workloads who can tolerate host-level variability and don't need guaranteed uptime.
9. FluidStack
H100: ~$2.10/hr | A100 80G: ~$1.30/hr | Multi-region API-driven marketplace
FluidStack is a GPU cloud marketplace that aggregates supply from data centers worldwide through a unified API. Like Lepton, it is an aggregation layer, but it is not NVIDIA-controlled. FluidStack focuses on enterprise and research teams running large-scale training jobs with competitive pricing on H100 and A100.
What it does well: Clean API with straightforward provisioning. Competitive pricing on H100 and A100. Multi-region availability across North America, Europe, and Asia. Volume pricing for teams with predictable workloads.
Where it falls short: Narrower catalog than RunPod or Spheron. Blackwell (B200, B300) availability lags dedicated marketplaces. Support primarily through email and ticketing rather than live channels.
Best for: Teams that want an API-driven multi-provider marketplace with competitive H100/A100 pricing without NVIDIA's brokerage layer. For a full comparison, see our FluidStack alternatives guide.
10. Hyperstack
H100 PCIe: $1.90/hr | H100 SXM: $2.40/hr | GDPR-compliant EU and North America infrastructure
Hyperstack is NexGen Cloud's GPU cloud platform with infrastructure in Europe and North America. For teams with EU data residency requirements, it offers a price advantage over Nebius and a direct provider relationship without Lepton's brokerage overhead.
What it does well: GDPR-compliant infrastructure. VM hibernation for pausing workloads. InfiniBand for H100 NVLink cluster configs. Competitive on-demand pricing at $1.90/hr for H100 PCIe. No sales process required for standard configs.
Where it falls short: Smaller company with less market recognition. Documentation and support ecosystem are still maturing. Limited catalog compared to larger marketplaces.
Best for: EU teams with GDPR requirements who want direct-provider pricing without NVIDIA's brokerage overhead. For a broader market comparison, see our Hyperstack alternatives guide.
Decision Guide
| Your situation | Best choice |
|---|---|
| Want open, non-NVIDIA-controlled marketplace | Spheron |
| Need NVIDIA-curated SLA-backed allocation | DGX Cloud Lepton |
| Multi-cloud orchestration, bring your own accounts | SkyPilot |
| Direct neocloud, large clusters | CoreWeave |
| EU data residency | Hyperstack or Nebius |
| Lowest price, interruptible jobs | Vast.ai |
| Serverless + training in one account | RunPod |
| Sustainable compute focus | Crusoe |
Spheron is an open GPU marketplace with live, verified pricing across 5+ providers and no NVIDIA brokerage overhead. Deploy H100, H200, or B200 without the allocation opacity.
H100 SXM5 on Spheron → | Check GPU availability → | View live pricing →
Frequently Asked Questions
NVIDIA DGX Cloud Lepton is NVIDIA's own GPU marketplace product that brokers compute capacity from vetted neoclouds (CoreWeave, Lambda, and others) through a single API and unified billing layer. It sits between operators and the underlying providers, handling allocation and billing but not owning the hardware itself.
Three main reasons: NVIDIA controls the allocation layer, which introduces lock-in on the vendor that also supplies the hardware. Pricing transparency is limited because Lepton applies its own brokerage margin on top of underlying neocloud rates. And the provider set is curated, so teams wanting specific neoclouds or regional coverage outside Lepton's network are blocked.
Yes. Spheron is a decentralized GPU marketplace that aggregates supply from 5+ providers with live, verified pricing exposed on a public API. Unlike Lepton, Spheron is not controlled by NVIDIA, so there is no conflict of interest between the marketplace operator and the hardware vendor. Pricing is transparent and competitive: H100 from $2.01/hr on-demand.
SkyPilot is an open-source multi-cloud orchestrator, not a marketplace or a brokerage. It does not source GPU supply itself. You configure SkyPilot with credentials to your own provider accounts, and it handles job placement and cost optimization across them. It pairs well with Spheron as a provider backend.
Spheron publishes all pricing through a public API (app.spheron.ai/api/gpu-offers) with no registration required. Pricing updates in real time based on available supply. No contact sales, no quote request, no brokerage margin applied by an intermediary.