NVIDIA DGX Cloud Lepton Pricing 2026: Marketplace Cost vs Spheron

Search "DGX Cloud Lepton pricing" and you won't find a rate card, because there isn't one. Lepton is NVIDIA's GPU marketplace: you send a request, Lepton routes it to whichever partner provider has capacity, and that provider bills you at its own rate. On the same H100 SXM5 hardware, that has meant $3.99/hr per GPU through Lambda and $6.16/hr per GPU through CoreWeave, both current NVIDIA Cloud Partners in the Lepton network. "DGX Cloud Lepton pricing" isn't a number. It's a question with a different answer depending on which provider answers your request that week.

Spheron takes the opposite approach. One live, verified rate across 5+ providers, published on a public API with no account required to check it. This post breaks down how Lepton's marketplace pricing actually works, what you'd pay through its provider network on H100, H200, and B200 today, and how that compares to Spheron's published rates. For the flip side of this comparison, our DGX Cloud Lepton alternatives guide covers the platforms worth evaluating beyond Lepton itself.

What DGX Cloud Lepton Actually Is (And Why It Has No Published Price List)

NVIDIA announced DGX Cloud Lepton on May 18, 2025, at Computex, describing it as "an AI platform with a compute marketplace that connects the world's developers building agentic and physical AI applications with tens of thousands of GPUs, available from a global network of cloud providers" (NVIDIA Newsroom). Jensen Huang put the pitch plainly: "Together with our NCPs, we're building a planetary-scale AI factory."

Lepton launched with an initial roster of 10 NVIDIA Cloud Partners (NCPs): CoreWeave, Crusoe, Firmus, Foxconn, GMI Cloud, Lambda, Nebius, Nscale, SoftBank Corp., and Yotta Data Services (source: NVIDIA Newsroom, above). A month later at GTC Paris, NVIDIA added AWS and Microsoft Azure as its first large-scale cloud providers, plus 8 more EU-based providers including Mistral AI, Nscale, Firebird, FluidStack, Hydra Host, Scaleway, and Together AI (Constellation Research).

Here's the part that matters for anyone trying to price this out: NVIDIA's own DGX Cloud Lepton marketing pages don't publish a per-GPU rate anywhere. The official product page describes the marketplace model and directs developers to build.nvidia.com for serverless endpoints and NIM microservices, but there's no fee schedule, no hourly rate table, nothing you'd recognize as a price list. That's not an oversight. It's structural. Lepton doesn't set prices; its partner providers do, and each one prices independently.

How Lepton's Marketplace Pricing Works: Provider-Set Rates, Not an NVIDIA Rate Card

A key thing to understand up front: DGX Cloud Lepton is a routing layer, not a hardware owner. Unlike NVIDIA's original DGX Cloud, where NVIDIA rented GPUs from neoclouds and subleased the capacity itself, Lepton doesn't take ownership of any hardware. It routes workloads to partner providers, AWS and Azure included, and those providers bill you directly at whatever rate they've set (Tom's Hardware).

That design has a direct pricing consequence. When you provision compute through Lepton, you're not paying an NVIDIA rate. You're paying whichever NCP's rate card happens to apply to your allocation, region, and GPU model. Two teams requesting the same H100 SXM5 through Lepton on the same day could land on different providers and pay meaningfully different amounts, and neither would have chosen that provider directly.

The Provider Network Behind Lepton (CoreWeave, Lambda, Nebius, Crusoe, AWS, Azure, and More)

The Lepton network spans neoclouds built specifically for GPU workloads and two hyperscalers that added it as a distribution channel. Here's what each partner actually charges on its own current pricing page, independent of Lepton:

That's a roughly 3x spread on identical H100 SXM5 hardware, from Lambda's $3.99/hr floor to Azure's $12.29/hr ceiling, all through providers sitting inside the same Lepton marketplace. CoreWeave's current on-demand H100 rate ($6.16/GPU-hr) alone runs close to double Lambda's ($3.99/GPU-hr), which is the cleanest illustration of why "Lepton pricing" isn't a number you can quote. It's a function of which NCP happens to fill your request.

One more wrinkle worth flagging if you're comparison shopping CoreWeave directly: CoreWeave still runs a legacy "Classic" pricing page quoting HGX H100 at $4.76/hr and H100 PCIe at $4.25/hr, materially lower than the $6.16/hr current-page rate for the same hardware. If you're pricing CoreWeave as a Lepton NCP, make sure you're reading the current page, not the classic one.

Why NVIDIA's Own DGX Cloud Pricing Failed First ($36,999/Month vs Hyperscaler Cuts)

Lepton exists because NVIDIA tried the opposite model first, and it didn't survive contact with hyperscaler pricing. The original DGX Cloud, announced in March 2023 at GTC, rented an 8-GPU instance (H100 or A100 80GB, 640GB of GPU memory) for $36,999 per instance per month (NVIDIA Newsroom). Divided across 8 GPUs, that works out to roughly $4,625 per GPU per month, defensible pricing when H100s were genuinely scarce.

The scarcity didn't last. As supply caught up with demand, AWS cut H100 and A100 prices by as much as 45%, and other hyperscalers followed with their own reductions (Tom's Hardware). NVIDIA was competing directly against the same cloud providers it depended on to actually run DGX Cloud's underlying hardware, an awkward position to hold a premium in. Customers migrated to cheaper capacity, and DGX Cloud's flat-rate model lost its reason to exist.

Lepton is NVIDIA's answer to that failure: instead of owning inventory and setting a price, act as the traffic controller and let each partner set its own rate. NVIDIA keeps the developer relationship and the marketplace fee structure without carrying the pricing risk of owning hardware in a market that keeps getting cheaper.

What You'd Actually Pay Through Lepton's Network: H100, H200, B200 by Provider

Because Lepton itself sets no price, "what would I pay through Lepton" really means "what does whichever NCP fills my request charge for this GPU model." Based on each provider's own current, independently published pricing:

CoreWeave sells H100 exclusively as an 8-GPU HGX bundle at $49.24/hr total, so there's no single-GPU on-demand option through that NCP if your workload only needs one card. Lambda and other NCPs offer smaller configurations down to a single GPU, at a modest per-GPU premium over the 8-GPU rate.

The pattern holds across GPU generations, not just H100. B200 is either priced per-GPU (Lambda, around $6.69-$6.99) or entirely absent from the public rate card and pushed to a sales conversation (Crusoe). If you're trying to budget a B200 cluster through Lepton and your request routes to a provider that hasn't published Blackwell pricing, you're back to a quote request, the exact opacity a marketplace is supposed to remove. For a fuller cross-provider Blackwell breakdown, see our B200 cloud pricing guide.

Rates above reflect each provider's own current, independently published pricing pages as of 18 Jul 2026 and are not Lepton-specific quotes, since Lepton applies no separate published rate of its own. Provider pricing changes frequently; check each provider's page directly before budgeting.

There's also a supply-side detail that matters for anyone assuming Lepton on-demand rates reflect what a small team would actually pay. CoreWeave, one of Lepton's founding NCPs, disclosed on its Q1 2026 earnings call a weighted average new-capacity contract length of roughly 5 years and a $99.4 billion contracted revenue backlog (Investing.com transcript). Most of that provider's GPU-hours are already sold under long-term commitments, not sitting on a shelf at the published on-demand rate. A workload routed to a capacity-constrained NCP may not get the headline rate at all; it gets whatever's actually available that day.

Lepton vs Spheron: Marketplace Model, Pricing Transparency, and Total Cost

Spheron and Lepton share the basic shape: both are marketplaces that route demand to underlying GPU supply rather than owning every rack themselves. Our Spheron vs Shadeform comparison covers this marketplace-vs-marketplace framing in more depth for a different pairing, but the core question is the same one to ask of Lepton: who sets the price, and can you see it before you commit?

The gap isn't small. Spheron's H100 SXM5 on-demand rate of $2.54/hr, verified against the live API at the time of writing, undercuts every Lepton NCP rate in the table above, including Lambda's floor of $3.99/hr. Spot pricing tells a similar story on most GPU models: A100 80GB SXM4 spot on Spheron runs $0.85/hr against an on-demand floor of $1.80/hr, and B200 SXM6 spot sits at $2.74/hr against a $7.50/hr on-demand rate. H100 SXM5 spot inventory was thin at the moment we checked, pricing close to the on-demand rate rather than below it, a reminder that even a transparent marketplace's spot tier moves with real-time supply. Check current GPU pricing → before budgeting off any number in this post, ours included.

The structural difference matters as much as the number. When NVIDIA brokers your workload to a partner it also sells hardware to, there's an inherent tension between acting as a neutral marketplace and protecting those partner relationships. Spheron has no hardware vendor to protect. What you see on the public API is the rate you pay, not a starting point for a sales conversation once you've picked a GPU.

Which GPU Marketplace Fits Your Team

Choose based on what you actually need from the marketplace layer, not just the sticker price:

• You want a single number you can budget against without calling anyone. Spheron's rates are published live, no account or sales call required. H100 on Spheron is available at $2.54/hr on-demand as of this writing.
• You already run production workloads inside AWS or Azure and want to stay in that billing relationship. Lepton's inclusion of both hyperscalers means you can route through infrastructure you already have contracts and support relationships with, at the cost of paying that provider's standard rate ($6.88/hr on AWS, $12.29/hr on Azure for H100).
• You need NVIDIA-curated SLA-backed allocation at enterprise scale and have the procurement bandwidth to negotiate per-NCP terms. That's the use case Lepton is built for. It's not built for a five-person team trying to compare hourly rates in one sitting.
• You're evaluating both marketplace models before committing. Read our DGX Cloud Lepton alternatives roundup for a broader field of 10 platforms beyond just Spheron.

For anyone whose main question really is "what will this cost per hour," the honest answer for Lepton is: check whichever provider you get routed to. For Spheron, the answer is on the pricing page right now.

Pricing fluctuates based on GPU availability. The figures above are based on 18 Jul 2026 and may have changed. Check current GPU pricing → for live rates.

Comparing marketplace models before committing to a GPU vendor? Spheron publishes one live rate across 5+ providers with no brokerage margin and no sales call required.

H100 GPU pricing → | Check B200 availability → | Get started on Spheron →