Lepton AI Review (2026): Features, Pricing & Verdict
Lepton AI was, for a brief and brilliant window, one of the most respected high-performance AI inference platforms in the market — and this review has to open with the reason its score is what it is: the original standalone Lepton AI product no longer exists as an independent service. Founded in 2023 by an exceptionally credentialed team — Yangqing Jia, the creator of the Caffe deep-learning framework and a key contributor to PyTorch and ONNX, formerly a VP at Alibaba and an engineer at Meta, alongside co-founder Junjie Bai — Lepton built a cloud-native platform for running AI applications on GPUs with genuine speed and elegance. It offered serverless, OpenAI-compatible inference for open-source models, an acclaimed open-source Python framework called Photon for building and deploying AI services, and clean primitives for dev pods, batch jobs and autoscaling inference endpoints, competing directly with Together AI and Fireworks. Its viral “Search with Lepton” demo showed off the infrastructure in under 500 lines of code. Then, in April 2025, NVIDIA acquired the company in a deal reported at several hundred million dollars — a remarkable outcome for a roughly 20-person startup, and a strong validation of the team. But acquisition brought consolidation: the original Lepton AI product has been sunset, existing customers were migrated to NVIDIA’s NIM inference microservices, and the brand was relaunched in June 2025 as NVIDIA DGX Cloud Lepton — a fundamentally different product (a GPU-compute marketplace aggregating capacity across many clouds) currently in preview. So this is not a review recommending you sign up for Lepton AI, because you no longer can. It’s an honest account of what Lepton was, what replaced it, and — importantly — where to go instead if you wanted what Lepton originally offered.
- Status
- Acquired by NVIDIA (April 2025). Original standalone product sunset; customers migrated to NVIDIA NIM. Rebranded as NVIDIA DGX Cloud Lepton (a different GPU-marketplace product, in preview). Open-source Photon framework continues
- Was best for
- Developers who wanted fast, serverless, OpenAI-compatible open-model inference with a clean Python deployment framework — now better served by the alternatives listed below, or evaluate the NVIDIA DGX Cloud Lepton successor
- Platform (original)
- Cloud-native AI inference platform — serverless LLM inference API, Photon open-source framework, Inference Endpoints, Batch Jobs, Dev Pods, FastGPU; optimised serving engine
- Successor
- NVIDIA DGX Cloud Lepton — a GPU-compute marketplace aggregating capacity from 20+ cloud providers (CoreWeave, Lambda, Nebius, Crusoe, AWS, Azure and more) via one interface; currently in preview, no free tier
- Vendor
- Lepton AI (founded 2023 by Yangqing Jia — creator of Caffe — and Junjie Bai). Now part of NVIDIA
What Was Lepton AI?
Lepton AI was a cloud-native, high-performance platform for building, deploying and serving AI applications — an “AI cloud” designed to make running models on GPUs fast, efficient and developer-friendly. It launched in 2023 with unusually strong technical credentials: co-founder and CEO Yangqing Jia created Caffe, one of the foundational deep-learning frameworks, made key contributions to PyTorch and ONNX, and had served as a VP of Technology at Alibaba (leading its AI and data platforms) after an earlier stint at Meta; co-founder Junjie Bai brought deep experience from Alibaba Cloud and Meta. That pedigree translated into a platform respected for its engineering quality despite the company raising only a modest $11M seed round. Lepton’s offering centred on a few well-designed pieces. Photon, its flagship open-source framework (Apache 2.0, the leptonai Python package), let developers wrap any model or AI service into a deployable unit — a “photon” — and push it to the cloud with a single command, using the same abstraction locally and in production. On top of that sat serverless Inference Endpoints (persistent, autoscaling model-serving), Batch Jobs (for training and one-off workloads that run to completion), Dev Pods (interactive GPU-attached development environments), and an OpenAI-compatible LLM Inference API that served popular open models (Llama, Mistral, Mixtral, Qwen and others) — putting Lepton in direct competition with Together AI and Fireworks as a high-performance open-model inference provider. Its optimised serving engine used techniques like dynamic batching, quantisation and speculative decoding to push throughput and cut latency, and the viral “Search with Lepton” demo — a conversational AI search engine built in under 500 lines of code — became a widely-shared showcase of how quickly you could build real applications on the platform. In short, Lepton was a genuinely good, technically elegant inference platform. The reason it sits in this review’s lower score band has nothing to do with the quality of what it built and everything to do with the fact that, as an independent product you could adopt, it no longer exists — which is the honest and decisive factor for anyone evaluating it today.
What It Offered (and What Happened)
The Photon framework and developer experience
Photon was Lepton’s most distinctive and enduring contribution, and it’s the one piece of the original offering that lives on. It was an open-source Python framework that let developers define an AI service — a model, a pipeline, an application — as a self-contained, deployable unit and run it identically on a laptop or in Lepton’s cloud, deploying with a simple command like lep photon push. This “build once, run anywhere” abstraction, combined with a clean developer experience, made Photon genuinely pleasant to work with and earned it a solid following (thousands of GitHub stars). Crucially for anyone assessing Lepton’s legacy today, the open-source Photon framework and its leptonai package remain maintained and available on GitHub under Apache 2.0 (with releases continuing into 2026), as does the team’s gpud open-source GPU-monitoring tool — so the open-source technical contributions survive the acquisition even though the hosted commercial platform does not. If you valued Lepton primarily for its open framework rather than its managed service, that part is still usable.
Serverless inference and GPU primitives
The commercial heart of Lepton was its managed infrastructure: an OpenAI-compatible serverless inference API for open-source LLMs, autoscaling Inference Endpoints for deploying your own models, Batch Jobs for training and bulk workloads, and Dev Pods for interactive GPU development — all backed by an optimised serving engine tuned for high throughput and low latency. This was a direct, credible competitor to Together AI and Fireworks, and by many accounts a technically strong one, with competitive pricing (its FastGPU offering undercut on cost) and a clean API. This is precisely the part of Lepton that has been discontinued: following the NVIDIA acquisition, the standalone serverless inference service and hosted endpoints were wound down and existing customers were migrated to NVIDIA’s NIM (NVIDIA Inference Microservices) ecosystem. So the specific capability most people came to Lepton for — easy, fast, serverless open-model inference behind a developer-friendly API — is no longer available under the Lepton name as it was, which is why this review points readers who want that capability to the living alternatives listed at the end.
The NVIDIA acquisition and DGX Cloud Lepton
In April 2025, NVIDIA acquired Lepton AI in a deal reported at several hundred million dollars — an extraordinary outcome for a roughly 20-person company that had raised only $11M, and one of the more notable acqui-hires in AI infrastructure (NVIDIA’s second such consolidation of an inference platform, after OctoAI in late 2024). The founders joined NVIDIA, with Junjie Bai taking a senior director role in Physical AI. In June 2025, NVIDIA relaunched the brand as NVIDIA DGX Cloud Lepton — but it’s important to understand that this is a different kind of product from the original. Rather than a serverless inference platform, DGX Cloud Lepton is a GPU-compute marketplace: it aggregates GPU capacity from more than twenty cloud providers (including CoreWeave, Lambda, Nebius, Crusoe, and major hyperscalers) and exposes them through a single unified interface, letting developers access GPUs across many clouds, deploy with Dev Pods, Batch Jobs and NIM-based Inference Endpoints, and select regions for data sovereignty. It’s a strategically interesting product — NVIDIA positioning itself as an aggregator and orchestration layer across the GPU cloud ecosystem rather than competing head-on with the hyperscalers — but as of this writing it is in preview, has no publicly documented free tier, and serves a different use case (multi-cloud GPU access and orchestration) than the original Lepton’s easy serverless inference. So the honest summary of “what happened” is: the team and technology were absorbed into NVIDIA at a strong valuation, the original developer-facing inference product was retired, and the brand now lives on as a preview-stage GPU marketplace that is not a like-for-like replacement.
Scored Categories
Status & Pricing
| Offering | Status | Notes |
|---|---|---|
| Original Lepton AI platform | Sunset / discontinued | Serverless inference API, hosted endpoints and dev/batch primitives wound down after the NVIDIA acquisition. No longer available for signup; existing customers migrated to NVIDIA NIM |
| NVIDIA DGX Cloud Lepton (successor) | Preview | A GPU-compute marketplace aggregating 20+ cloud providers. Marketplace pricing (each provider sets its own on-demand/reserved rates). No publicly documented free tier. Sign up via NVIDIA’s developer site |
| Photon framework (open source) | Continues — free | The leptonai Python package and Photon framework remain maintained on GitHub (Apache 2.0); free to use. The gpud GPU-monitoring tool likewise remains available |
| Recommended alternatives | Available now | For the serverless open-model inference Lepton originally offered: Together AI, Fireworks AI, Baseten, or RunPod (see related reviews below) |
Strengths (of what it was)
- Exceptional founding team — Yangqing Jia (creator of Caffe, PyTorch/ONNX contributor, ex-Alibaba VP) and Junjie Bai
- Elegant Photon open-source framework — clean “build once, run anywhere” developer experience
- Fast, high-performance serving engine (dynamic batching, quantisation, speculative decoding)
- OpenAI-compatible serverless inference for popular open models; competitive pricing
- Clean GPU primitives — Inference Endpoints, Batch Jobs, Dev Pods
- Validated by a major NVIDIA acquisition — a strong outcome and endorsement of the tech
- Open-source Photon (leptonai) and gpud tools remain maintained and usable
- Successor (NVIDIA DGX Cloud Lepton) is a credible, strategically interesting product
Weaknesses (for adopting it today)
- The original standalone product is sunset — you cannot sign up for Lepton AI as reviewed
- Existing customers were migrated to NVIDIA NIM — a forced platform change
- The successor (DGX Cloud Lepton) is a different product — a GPU marketplace, not easy serverless inference
- Successor is in preview with no publicly documented free tier
- Absorbed into NVIDIA — no independent roadmap; direction set by NVIDIA’s strategy
- Not a like-for-like replacement path for those who wanted Lepton’s original hosted inference
- Anyone evaluating “Lepton AI” today should plan around alternatives instead
Verdict: 5.7 / 10 — A Great Platform, Acquired and Absorbed
Lepton AI earns a 5.7, and it’s essential to be clear about what that number does and doesn’t say. It is not a judgement on the quality of what Lepton built — by that measure the platform was genuinely excellent, created by one of the strongest technical teams in AI infrastructure (the creator of Caffe among them), with an elegant open-source framework, a fast serving engine, and a serverless inference product that competed credibly with Together AI and Fireworks. Had this review been written in early 2025, Lepton would have scored in the low-to-mid eights. The 5.7 instead reflects a single decisive reality for anyone evaluating it today: the original Lepton AI product no longer exists as something you can adopt. NVIDIA’s April 2025 acquisition — a strong outcome that validated the team and technology — led to the standalone platform being sunset, existing customers migrated to NVIDIA NIM, and the brand relaunched as NVIDIA DGX Cloud Lepton, a different, preview-stage GPU-compute marketplace rather than the easy serverless inference service Lepton was known for. A tool review has to score what a reader can actually use, and on that basis Lepton sits in the “discontinued and absorbed” band — a touch above a pure shutdown, because the acquirer is as credible as they come, the successor product is real and strategically interesting, and the open-source Photon framework lives on and remains genuinely useful. But it is well below the score any active, adoptable inference platform would earn, because you cannot go and build on the original Lepton today. So the practical verdict is simple. If you’re reading this because you were considering Lepton AI for serverless open-model inference, don’t — that product is gone; instead use Together AI, Fireworks AI or Baseten, all of which offer the fast, developer-friendly hosted inference Lepton pioneered and are thriving. If you specifically want a multi-cloud GPU marketplace, evaluate NVIDIA DGX Cloud Lepton on its own merits, bearing in mind it’s still in preview. And if it was Photon’s open-source framework you valued, that remains available on GitHub. Lepton’s story is ultimately a success — a small, brilliant team building something good enough that NVIDIA bought it — but for a prospective user in 2026, it’s a platform to learn from and move on from, not one to adopt.
Frequently Asked Questions
Can I still sign up for and use Lepton AI?
No — not the original Lepton AI product, and this is the most important practical takeaway of this entire review. Following NVIDIA’s acquisition of Lepton AI in April 2025, the original standalone platform — the serverless inference API, the hosted model endpoints, and the developer-facing cloud service that Lepton was known for — has been sunset and is no longer available for new signups. Existing customers were migrated to NVIDIA’s NIM (NVIDIA Inference Microservices) ecosystem rather than continuing on the original Lepton service. So if you go looking to create a Lepton AI account and deploy a model the way you could have in early 2025, you’ll find that path no longer exists. What does exist under the name is NVIDIA DGX Cloud Lepton, the June 2025 relaunch — but that’s a different product (a GPU-compute marketplace aggregating capacity across many cloud providers), it’s currently in preview, and it serves a different purpose than the original’s easy serverless inference. You can request access to that through NVIDIA’s developer site if multi-cloud GPU access is what you need. Separately, the open-source Photon framework (the leptonai Python package) that Lepton created remains available and maintained on GitHub under an Apache 2.0 licence, so if it was specifically the open-source framework you wanted to use, that part is still usable independently of any hosted service. But for the hosted Lepton AI platform as most people knew it: no, you can’t sign up for it anymore, and you should plan around one of the alternatives instead.
What should I use instead of Lepton AI?
The right replacement depends on which part of Lepton you wanted, but for the most common use case — fast, serverless, developer-friendly inference for open-source models — there are several excellent, actively-supported alternatives that do exactly what Lepton did, and in most cases you’ll have a smooth transition. Together AI is probably the closest like-for-like replacement: it’s a leading platform for serverless open-model inference and fine-tuning, with an OpenAI-compatible API, a broad catalogue of open models, competitive per-token pricing, and the option to scale up to dedicated endpoints and GPU clusters — essentially the category Lepton competed in, executed at scale. Fireworks AI is another strong, direct alternative, particularly known for very high-performance, low-latency inference of open models with an OpenAI-compatible API and fine-tuning support, making it ideal if speed is your priority. Baseten is the pick if you lean more toward deploying and serving your own custom or fine-tuned models in production, with a polished developer experience and strong performance engineering. If you specifically valued Lepton’s GPU primitives (dev pods, batch jobs) and want cheap, flexible, self-serve GPU compute rather than a managed inference API, RunPod is an excellent choice, offering pods, serverless GPU endpoints and clusters at low cost. And if what you actually want is the multi-cloud GPU marketplace that Lepton became, then NVIDIA DGX Cloud Lepton itself is the option to evaluate — just note it’s in preview. For the great majority of people who came to Lepton for easy hosted inference of open models, the practical recommendation is to start with Together AI or Fireworks AI: both are mature, well-supported, offer OpenAI-compatible APIs (so migration is straightforward), publish clear pricing, and are actively investing in the exact capabilities Lepton pioneered. All of these are covered in their own reviews on this site, linked below.
Was the NVIDIA acquisition good or bad for Lepton?
For the company, its founders and its investors, the NVIDIA acquisition was unambiguously a good outcome — one of the better ones a startup of Lepton’s size could hope for; for users of the original product, it was disruptive. Consider the company side first: Lepton AI was a roughly 20-person startup that had raised only about $11 million in seed funding, and NVIDIA acquired it in a deal reported at several hundred million dollars. That represents an enormous return and one of the more remarkable per-employee acqui-hire valuations in recent AI infrastructure history, a powerful validation of the team’s technical reputation (unsurprising given co-founder Yangqing Jia’s stature as the creator of Caffe and a key figure in the deep-learning framework world). The founders joined NVIDIA in senior roles, and their technology and expertise now sit inside the most important company in the AI hardware ecosystem — by any startup measure, a success. For the broader ecosystem, it’s strategically significant too: NVIDIA is assembling a software and orchestration layer above its hardware (Lepton followed its earlier OctoAI acquisition), and DGX Cloud Lepton’s marketplace model — aggregating GPU capacity across many clouds rather than competing with them — is a genuinely interesting play. For existing users of the original Lepton platform, however, the acquisition was disruptive in the way most infrastructure acquisitions are: the product they were using was sunset, and they had to migrate to NVIDIA’s NIM ecosystem or elsewhere, an unwanted switching cost regardless of the destination’s quality. This is the perennial risk of building on a promising but small startup’s platform — the better it is, the more likely it gets acquired, and acquisition often means consolidation or shutdown of the original product. So the honest answer is that the acquisition was a success for Lepton the company and a genuine endorsement of its technology, while being a discontinuation event for Lepton the product — which is exactly why, for someone evaluating it as a tool to adopt today, the outcome nonetheless lands as a low score. It’s a case study in how a great technical outcome for a startup and a poor availability outcome for its users can be the same event.