
June's updates bring NVIDIA B200 and B300 Blackwell Ultra GPU supply to the marketplace, plus new templates, guides, and platform improvements.

## NVIDIA B200 and B300 GPUs Now Available

Fresh GPU supply is live on Vast.ai, with the powerful NVIDIA B300 now available. Built on the Blackwell Ultra architecture, the B300 features 288GB of HBM3e memory, 8 TB/s of memory bandwidth, and 20,480 CUDA cores - designed for large-scale inference, reasoning models, distributed training, and any workloads that keep running out of VRAM. See up-to-date [B300 pricing](https://vast.ai/pricing/gpu/B300), and rent by the hour or reserve long-term.

We've also added hundreds of B200 GPUs to the marketplace. Rent individual instances or use [Vast.ai Serverless](https://vast.ai/products/serverless) to automatically scale workloads without managing infrastructure yourself.

## Platform Updates: Serverless, Billing, and Security Fixes

We've continued our platform updates with several improvements this month. Serverless metrics reporting has been revamped with increased granularity and new metrics, and serverless UI performance has been significantly improved to accommodate high instance count endpoints.

In addition, when you create a machine report, that machine now stays out of your search results for longer, so your results stay more relevant to you.

On the fixes side: we've improved support for manual invoices on billing transaction history exports, fixed some SSH key issues with key editing and stale data, addressed OAuth login and two-factor authentication (2FA) issues, and made a number of security improvements.

## New Templates

This month, we've added a template for [Qwen3.6 35B A3B](https://vast.ai/model/qwen36-35b-a3b). It's the first open-weight model in the Qwen3.6 series, combining a hybrid Gated DeltaNet and Gated Attention architecture with sparse Mixture-of-Experts (MoE) routing and a vision encoder optimized for agentic coding, frontend workflows, and repository-level reasoning.

Serverless templates for Qwen3.6 and Gemma 4 are now available. The Qwen3.6 template supports a 35B-total, 3B-active model.

We've also updated the vLLM template to include TurboQuant support and made some maintenance updates to the ComfyUI template.

## Our Commitment

We remain committed to keeping high-performance AI infrastructure accessible and affordable for everyone. With fresh B200 and B300 GPU supply live on Vast.ai, we're continuing to expand the resources available to our users while making it easier to build, train, fine-tune, and deploy AI workloads at scale.

Need help? Reach out anytime at [support@vast.ai](mailto:support@vast.ai) or join our [Discord](https://discord.gg/vast) server to connect with the community and keep up with our latest updates.

## Change Log

### New Features

- NVIDIA B200 and B300 GPUs are now available.
- Creating a machine report now removes the machine from your results for a longer duration.
- Revamp of serverless metrics reporting: increased granularity and new metrics.
- Significant improvement in the performance of serverless UI to accommodate high instance count endpoints.
- Billing page UI improvements.
- Search page performance improvements.
- B300 support.

### Issues Resolved

- Improved support for manual invoices on billing transaction history export.
- Fixed instance SSH key issues with key editing and stale data.
- OAuth login issues.
- Various 2FA fixes.
- Security improvements.

### New Templates

- [Qwen3.6 35B A3B](https://vast.ai/model/qwen36-35b-a3b): the first open-weight model in the Qwen3.6 series, built on direct community feedback and focused on stability and real-world utility. It combines a hybrid Gated DeltaNet and Gated Attention architecture with sparse Mixture-of-Experts routing and a vision encoder for unified multimodal reasoning.
- Serverless templates for Qwen3.6 and Gemma 4.

### Updated Templates

- vLLM template: TurboQuant support.
- ComfyUI maintenance.

### New Guides

- Step-by-step instructions for moving workloads from [Salad Cloud to Vast.ai](https://docs.vast.ai/examples/migrations/salad-to-vast), including instance setup and cost comparison.
- How to generate [video with LTX-2.3 and ComfyUI](https://docs.vast.ai/examples/video-generation/ltx-video-2-3) on Vast.ai.
- A practical guide to fine-tuning open-source LLMs on Vast.ai GPU instances using the [Axolotl](https://docs.vast.ai/examples/ai-ml-frameworks/axolotl-fine-tuning) framework.
- Deploy and run [NVIDIA Nemotron 3 Super](https://docs.vast.ai/examples/text-generation/nemotron-3-super) on Vast.ai for high-performance inference.

### Top Blog Posts

- [All-in-One App Studio: Your Complete Creative AI Toolkit in a Single Container](https://vast.ai/article/all-in-one-app-studio)
- [How Much Money Can You Earn Renting Out Your GPU on Vast.ai?](https://vast.ai/article/how-much-money-can-you-earn-renting-out-your-gpu-on-vast-ai)
- [TurboQuant Explained: How It Reduces LLM Memory by 5x and Speeds Up Inference](https://vast.ai/article/turboquant-explained-llm-memory-inference)
- [How to Use Unsloth Studio for Fast and Affordable AI Fine-Tuning](https://vast.ai/article/how-to-use-unsloth-studio-for-fast-affordable-ai-fine-tuning)


Vast.ai's June 2026 product update brings NVIDIA B200 and B300 Blackwell Ultra GPU supply to the marketplace, plus new templates, guides, and platform improvements.

June 2026 Product Update


With AI came the technology to transform entire industries, but the infrastructure to run it was suddenly constrained. Hyperscalers secured much of the available advanced GPU capacity for their own needs, leaving many startups, research labs, and smaller enterprises scrambling for compute.

Demand for GPUs surged so high that pricing quickly became prohibitively expensive. In early 2024, renting four NVIDIA H100-powered instances from a hyperscaler for just one month could [cost](https://journal.uptimeinstitute.com/neoclouds-a-cost-effective-ai-infrastructure-alternative/) nearly $300,000.

This opened up a gap in the market, and into it stepped an entirely new type of cloud provider: the [neocloud](https://vast.ai/hosting/neoclouds).

## What Is a Neocloud? Definition and Core Models

Born out of GPU scarcity and cost barriers, as well as hyperscaler allocation bottlenecks, neoclouds are AI-first cloud infrastructure providers that specialize in GPU-as-a-Service (GPUaaS) and high-performance computing (HPC) workloads. Neoclouds serve a market that's [projected](https://www.abiresearch.com/blog/leading-neocloud-companies) to grow from $42B last year to $250B+ by 2030.

A neocloud is an AI-first cloud infrastructure provider that specializes in GPU compute for machine learning training and inference workloads. Unlike traditional hyperscalers, neoclouds are purpose-built for AI, offering faster provisioning, transparent pricing, and configurations optimized for high-performance computing at significantly lower cost.

Neoclouds are an emerging type of AI-first cloud infrastructure provider that emphasizes GPUs, faster networking, and optimized storage for the intensive demands of generative AI workloads. Inference, training, and other machine learning algorithms require more advanced infrastructure to keep their data pipelines filled. Otherwise, resources become idle and turnaround time suffers.

Unlike hyperscalers, which offer general-purpose infrastructure for a wide variety of workloads, neoclouds typically focus on AI. Many provide GPU compute through simpler pricing models. Another advantage: faster provisioning and configurations that are optimized for AI training and inference.

But not all neoclouds are built the same way. Here's a look at the three business models that neoclouds operate under.

### 1. Own Infrastructure

Larger neoclouds like CoreWeave and Nscale build AI infrastructure in their own data centers. This requires massive capital investment in the millions to billions of dollars and involves depreciation risk for aging hardware, but it also offers full-stack control. Neoclouds that invest in their own infrastructure can provide earlier access to next-gen GPUs and support larger-scale enterprise deployments.

### 2. Colocation Strategy

Under the colocation model, neoclouds take a hybrid approach. They partner with data centers to lease established infrastructure, reducing upfront capital expenditure. High-density, GPU-optimized colocation environments allow neoclouds to scale quickly without building facilities from scratch.

### 3. Asset-Light Marketplace

With this approach, neoclouds like Vast.ai aggregate GPU capacity without owning infrastructure. The marketplace model connects distributed compute providers with customers who need flexible, cost-effective access without long-term commitments, while the platform itself acts as the coordination layer.

## Neocloud Benefits for Enterprise AI: Cost, Speed, and Scale

No matter their business model, neoclouds offer numerous benefits to organizations working with AI.

First and foremost, better pricing. Because they run leaner and have less overhead, neoclouds can [deliver](https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/the-evolution-of-neoclouds-and-their-next-moves) up to 85% cost savings compared to hyperscaler GPU instances. They also monetize differently. Rather than complex, layered billing, neoclouds tend to offer transparent, per-second pricing where you only pay for what you use.

Second, specialized infrastructure. Neoclouds are purpose-built for AI, so they can deliver faster provisioning with infrastructure optimized specifically for model training and inference.

Third, greater elasticity and scalability. AI workloads can be unpredictable. Neoclouds give organizations the ability to scale GPU resources up or down dynamically, so they get exactly what they need when they need it and avoid overprovisioning. Some neoclouds, like Vast.ai, even support [serverless autoscaling](https://vast.ai/article/vast-ai-serverless-automated-gpu-scaling) for workloads with variable demand.

Finally, better support for AI-specific compute requirements. AI workloads demand high-bandwidth interconnects, large memory pools, and massive parallel processing capacity. Because neoclouds are built around these requirements, they're often better suited for GPU-intensive workloads than general-purpose cloud environments.

Neoclouds like Vast.ai are also built to meet enterprise-grade security and compliance needs, so organizations don't have to choose between affordability and peace of mind.

However, not all neoclouds operate under the same economic or operational constraints.

## The Limitations of Traditional GPU Cloud Providers

Traditional neoclouds that own or lease infrastructure face a few major challenges:

- **Rapid price erosion:** Each new GPU generation decreases the value of older chips, with the price of a GPU hour [declining](https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/the-evolution-of-neoclouds-and-their-next-moves) by up to 50% within five years.
- **Thin margins:** After depreciation, power, and labor costs, gross profit margins are often quite slim.
- **Difficulty differentiating:** To stay competitive, neoclouds may offer more than just raw compute, but developing higher-margin AI services can put them in direct competition with hyperscalers.

These pressures mean that traditional neoclouds have to maintain high utilization rates and constantly reinvest in new hardware while also navigating how to differentiate their services successfully.

The marketplace model, on the other hand, avoids these challenges.

## How GPU Compute Marketplaces Solve What Traditional Neoclouds Can't

By aggregating distributed GPU capacity rather than owning hardware, marketplace-based neoclouds eliminate the millions in upfront investment that would otherwise be needed, and never have to manage an aging fleet of depreciating hardware. There's also often little to no direct competition for the massive enterprise contracts that infrastructure owners depend on.

This means that the marketplace model is able to provide the same advantages as other neoclouds, including flexibility, elasticity, and on-demand access, while offering the most competitive rates possible.

One important byproduct of this approach is democratized compute. Organizations that can't afford higher rates can still leverage the latest hardware and pursue AI development.

### Neocloud vs. Hyperscaler: Key Differences

<div className="article-table-scroll neocloud-comparison-table">
  <table aria-label="Neocloud vs. hyperscaler key differences">
    <thead>
      <tr>
        <th>Neocloud</th>
        <th>Hyperscaler</th>
      </tr>
    </thead>
    <tbody>
      <tr>
        <td>GPU first</td>
        <td>CPUs first, with GPUs as add-ons</td>
      </tr>
      <tr>
        <td>Built for fast networking speed natively</td>
        <td>Mixed networking speeds</td>
      </tr>
      <tr>
        <td>Geared toward bare metal or thin VMs</td>
        <td>Virtualization overhead</td>
      </tr>
      <tr>
        <td>Simplified pricing</td>
        <td>Complex a la carte pricing</td>
      </tr>
      <tr>
        <td>Quick provisioning</td>
        <td>DIY provisioning or managed services</td>
      </tr>
    </tbody>
  </table>
</div>

## The Bottom Line

Both traditional neoclouds and hyperscalers carry overhead that drives up costs. Marketplace-based neoclouds like Vast.ai offer a better path forward, delivering the GPU compute your AI and HPC workloads need at prices that make experimentation and innovation possible for teams of any size.

For teams evaluating [GPU cloud](https://vast.ai/products/gpu-cloud) providers, from scrappy AI startups to enterprise ML teams, understanding the neocloud landscape is the first step toward smarter infrastructure decisions. Stop overpaying for GPU compute. Start building on Vast.ai today.


Neoclouds are AI-first GPU cloud providers built for machine learning workloads. Learn how they work, the three business models, and how they compare to hyperscalers.

What Is a Neocloud? The Business Model Explained


The NVIDIA GeForce RTX 5090 launched in January 2025 – and over a year later, it's still surprisingly difficult to get your hands on one.

Despite retail prices officially starting at $1,999 MSRP, limited availability and strong demand from both gamers and AI developers means stock sells out quickly and real-world pricing has pushed higher.

As NVIDIA's most powerful consumer GPU, the RTX 5090 is undoubtedly an impressive machine. Is it worth the steep price tag, though? That depends on more than just raw power – let's take a look.

### RTX 5090: The Consumer Powerhouse

For starters, here are some of the official specs for the RTX 5090:

-   **Architecture:** Powered by NVIDIA's cutting-edge Blackwell architecture.
-   **Gaming Chip:** Featuring the next-gen GB202 gaming chip.
-   **Memory:** 32GB of GDDR7 memory with a blazing-fast 1,792GB/sec bandwidth, supported by a super-wide 512-bit memory interface for even faster data transfer.
-   **CUDA Cores:** 21,760 cores for high-performance graphics workloads, complementing the 680 Tensor cores that accelerate AI inference and training.
-   **L2 Cache:** Expanded to 98MB, up from 73MB on the RTX 4090, improving memory efficiency and reducing latency.
-   **Display Support:** PCIe Gen 5 with DisplayPort 2.1b connectors supporting up to 8K at 165Hz.
-   **Power Requirements:** 575W total graphics power (TGP), with a 1000W power supply recommended for optimal performance.

Another bonus: the RTX 5090 has a redesigned **Founders Edition** with a three-slot design and a dual flow-through fan configuration – so it's more compatible with tighter builds, a remarkable shift from its predecessors.

For reference, here's a quick overview of the RTX 5090:

<table>
    <thead>
        <tr>
            <th>Spec</th>
            <th>RTX 5090</th>
        </tr>
    </thead>
    <tbody>
        <tr>
            <td><strong>CUDA Cores</strong></td>
            <td>21,760</td>
        </tr>
        <tr>
            <td><strong>Ray Tracing Cores</strong></td>
            <td>170 (4th-gen)</td>
        </tr>
        <tr>
            <td><strong>Tensor Cores</strong></td>
            <td>680 (5th-gen)</td>
        </tr>
        <tr>
            <td><strong>Base Clock</strong></td>
            <td>2,017 MHz</td>
        </tr>
        <tr>
            <td><strong>Boost Clock</strong></td>
            <td>2,407 MHz</td>
        </tr>
        <tr>
            <td><strong>L2 Cache</strong></td>
            <td>98,304KB</td>
        </tr>
        <tr>
            <td><strong>ROPs</strong></td>
            <td>176</td>
        </tr>
        <tr>
            <td><strong>VRAM</strong></td>
            <td>32GB 28Gbps GDDR7</td>
        </tr>
        <tr>
            <td><strong>VRAM Interface</strong></td>
            <td>512-bit</td>
        </tr>
        <tr>
            <td><strong>Memory Bandwidth</strong></td>
            <td>1,792 GB/s</td>
        </tr>
        <tr>
            <td><strong>Interface</strong></td>
            <td>16x PCIe 5.0</td>
        </tr>
        <tr>
            <td><strong>Power Connectors</strong></td>
            <td>1 x 16-pin (PCIe 5.0 / ATX 3.0) native connector; includes 4 x 8-pin adapter</td>
        </tr>
        <tr>
            <td><strong>Total Graphics Power (TGP)</strong></td>
            <td>575W</td>
        </tr>
    </tbody>
</table>

(For more specifications – and to see how the RTX 5090 stacks up against its younger siblings in the RTX 50-series lineup as well as the RTX 4090 – check out the [chart at the bottom of this page](#additional-specifications).)

#### RTX 5090: AI and Professional Workloads

The RTX 5090 is well-suited for high-performance AI and professional workloads. Its **680 Tensor cores** and **32GB of GDDR7** memory efficiently accelerate deep learning, 3D rendering, and other compute-intensive tasks. With **98MB of L2 cache** and a **512-bit memory interface**, the RTX 5090 delivers faster data access and reduced latency – crucial for AI training, scientific computing, and content creation.

The **5th-gen Tensor cores** boost AI inference efficiency in frameworks like TensorFlow and PyTorch, while **4th-gen RT cores** enhance real-time ray tracing for animation, visual effects, CAD work, and more.

For demanding AI and video workloads, the RTX 5090 more than meets the challenge. In Procyon's AI XL (FP16) test, the RTX 5090 was **40% faster** than the RTX 4090, and in PugetBench's DaVinci Resolve video processing test, it was **12% faster**.

#### Next-Gen Gaming Features

The biggest leap in technology for this GPU comes from DLSS 4's new **Multi Frame Generation**, an AI-driven frame interpolation technique exclusive to the RTX 50 series. It uses AI to generate up to three additional frames between traditionally rendered ones, significantly boosting frame rates with minimal visual artifacts.

Further, the RTX 5090 introduces a range of AI-driven enhancements in what's called **neural rendering**, designed to push in-game realism to new levels. **RTX Neural Shaders** improve texture compression and deliver film-quality lighting and shading in real time, while **RTX Neural Faces** uses generative AI to create lifelike skin, hair, and facial details.

Building on the power of DLSS 4, these neural rendering techniques integrate directly into the graphics pipeline, leveraging NVIDIA’s Tensor cores to enhance visual fidelity. As game support continues to expand, these technologies demonstrate the potential for deeper shadows, realistic translucency in materials, and an overall richer environment.

Another breakthrough is **Mega Geometry**, a feature that uses RT cores to significantly boost the number of triangles in ray-traced scenes. This allows game engines to maintain full geometric detail without sacrificing performance, resulting in more realistic object depth, shadows, and fine details.

The above features position the RTX 5090 as a powerhouse for next-generation gaming visuals... but there's a catch.

##### Caveat: Performance Claims vs. Real-World Results

The way that NVIDIA presents its performance numbers can be misleading. The company leans heavily on Multi Frame Generation-boosted FPS (frames per second) to compare these new GPUs against older ones.

Frame Generation and Multi Frame Generation (MFG) are powerful tools, but they aren't universal fixes. Their impact varies by game – some titles see massive FPS gains with little visual trade-off, while others may introduce artifacts or latency.

In addition, **not all games will support these features**, as DLSS adoption depends on individual developers. A GPU's raw performance still matters. MFG-inflated benchmarks can make performance gains seem more significant, but Frame Generation still needs a strong base frame rate to be effective. In CPU-limited games, or titles that don't support DLSS 4, the real-world performance gap may be far smaller.

Ultimately, while the RTX 5090 excels at AI-driven rendering, **real-world results will depend on game support and how well these features are implemented.**

### Worldwide RAM Shortage and Price Increases

Due to explosive growth in AI data centers and tech giants continuing to invest heavily in AI accelerators, there's an ongoing worldwide RAM shortage that shows no signs of slowing down. Some forecasts <a href="https://www.theverge.com/ai-artificial-intelligence/914672/the-ram-shortage-could-last-years" target="_blank" rel="noopener noreferrer">predict</a> that shortages could last until 2030.

It's no surprise that high demand and limited supply lead to rising prices. With the RTX 5090, though, there may be another factor at play. DeepSeek's rapid rise in popularity has sparked a huge <a href="https://www.theinformation.com/articles/deepseek-spurs-black-market-demand-for-nvidia-gaming-chips-in-china" target="_blank" rel="noopener noreferrer">surge</a> in demand for NVIDIA's gaming GPUs in China, as these cards can be repurposed to run DeepSeek's AI models – and provide a workaround for U.S. export restrictions on advanced GPUs.

See our guide on [how to run DeepSeek R1 on a GPU cloud](https://vast.ai/article/deepseek-r1-open-source-disruptor-or-overhyped-upstart).

With a global shortage of RAM and both gamers and AI developers vying for the same hardware, the strain on supply could push RTX 5090 prices even higher in the near future. Most buyers currently face inflated market prices ranging from $2,500 to over $4,000. And with NVIDIA set to <a href="https://wccftech.com/nvidia-passes-another-300-hit-to-rtx-5090-board-partners-due-to-vram-shortage/" target="_blank" rel="noopener noreferrer">increase prices</a> further on account of VRAM sourcing costs, those numbers could skyrocket to over $5,000.

### Final Thoughts: The Power of the RTX 5090 – Without the Hassle

The RTX 5090 remains the most powerful consumer GPU available today. For cutting-edge 4K gaming with DLSS 4 and Multi Frame Generation, as well as demanding AI and professional workloads, it delivers unmatched performance in a relatively accessible package.

At the same time, it may only be worth purchasing if you really want the best of the best and you're willing to shell out for it. The high price tag makes it a significant investment for the average user.

In the meantime, why wait? Skip the delays and high costs – [rent an RTX 5090](https://cloud.vast.ai/?gpu_option=RTX%205090) on Vast.ai for instant access to top-tier performance. Given its limited availability and its 575-watt power draw, our affordable rental options are the most practical way to access the RTX 5090's full potential without the upfront investment.

[Try the 5090 GPU](https://cloud.vast.ai/?gpu_option=RTX%205090) for yourself today.

<span id="additional-specifications"></span>

#### Additional Specifications:

<div className="article-table-scroll">
<table className="w-full border-collapse border border-gray-300">
    <thead>
        <tr className="bg-gray-100">
            <th className="border border-gray-300 px-4 py-2 text-left font-semibold">
                Spec
            </th>
            <th className="border border-gray-300 px-4 py-2 text-center font-semibold">
                RTX 5090
            </th>
            <th className="border border-gray-300 px-4 py-2 text-center font-semibold">
                RTX 5080
            </th>
            <th className="border border-gray-300 px-4 py-2 text-center font-semibold">
                RTX 5070 Ti
            </th>
            <th className="border border-gray-300 px-4 py-2 text-center font-semibold">
                RTX 5070
            </th>
            <th className="border border-gray-300 px-4 py-2 text-center font-semibold">
                RTX 4090
            </th>
        </tr>
    </thead>
    <tbody>
        <tr>
            <td className="border border-gray-300 px-4 py-2 font-semibold">
                Architecture
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">GB202</td>
            <td className="border border-gray-300 px-4 py-2 text-center">GB203</td>
            <td className="border border-gray-300 px-4 py-2 text-center">GB203</td>
            <td className="border border-gray-300 px-4 py-2 text-center">GB205</td>
            <td className="border border-gray-300 px-4 py-2 text-center">AD102</td>
        </tr>
        <tr className="bg-gray-50">
            <td className="border border-gray-300 px-4 py-2 font-semibold">VRAM</td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                32GB GDDR7
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                16GB GDDR7
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                16GB GDDR7
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                12GB GDDR7
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                24GB GDDR6X
            </td>
        </tr>
        <tr>
            <td className="border border-gray-300 px-4 py-2 font-semibold">
                VRAM Speed
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">28Gbps</td>
            <td className="border border-gray-300 px-4 py-2 text-center">30Gbps</td>
            <td className="border border-gray-300 px-4 py-2 text-center">28Gbps</td>
            <td className="border border-gray-300 px-4 py-2 text-center">28Gbps</td>
            <td className="border border-gray-300 px-4 py-2 text-center">21Gbps</td>
        </tr>
        <tr className="bg-gray-50">
            <td className="border border-gray-300 px-4 py-2 font-semibold">
                VRAM Bus Width
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">512</td>
            <td className="border border-gray-300 px-4 py-2 text-center">256</td>
            <td className="border border-gray-300 px-4 py-2 text-center">256</td>
            <td className="border border-gray-300 px-4 py-2 text-center">192</td>
            <td className="border border-gray-300 px-4 py-2 text-center">384</td>
        </tr>
        <tr>
            <td className="border border-gray-300 px-4 py-2 font-semibold">
                Bandwidth
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                1792GB/s
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                960GB/s
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                896GB/s
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                672GB/s
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                1.01TB/s
            </td>
        </tr>
        <tr className="bg-gray-50">
            <td className="border border-gray-300 px-4 py-2 font-semibold">
                Streaming Multiprocessors (SMs)
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">170</td>
            <td className="border border-gray-300 px-4 py-2 text-center">84</td>
            <td className="border border-gray-300 px-4 py-2 text-center">70</td>
            <td className="border border-gray-300 px-4 py-2 text-center">48</td>
            <td className="border border-gray-300 px-4 py-2 text-center"></td>
        </tr>
        <tr>
            <td className="border border-gray-300 px-4 py-2 font-semibold">
                Tensor Cores
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">680</td>
            <td className="border border-gray-300 px-4 py-2 text-center">336</td>
            <td className="border border-gray-300 px-4 py-2 text-center">280</td>
            <td className="border border-gray-300 px-4 py-2 text-center">192</td>
            <td className="border border-gray-300 px-4 py-2 text-center">512</td>
        </tr>
        <tr className="bg-gray-50">
            <td className="border border-gray-300 px-4 py-2 font-semibold">
                Ray Tracing Cores
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">170</td>
            <td className="border border-gray-300 px-4 py-2 text-center">84</td>
            <td className="border border-gray-300 px-4 py-2 text-center">70</td>
            <td className="border border-gray-300 px-4 py-2 text-center">48</td>
            <td className="border border-gray-300 px-4 py-2 text-center">128</td>
        </tr>
        <tr>
            <td className="border border-gray-300 px-4 py-2 font-semibold">
                Boost Clock
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                2407 MHz
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                2617 MHz
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                2452 MHz
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                2512 MHz
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                2520 MHz
            </td>
        </tr>
        <tr className="bg-gray-50">
            <td className="border border-gray-300 px-4 py-2 font-semibold">
                Total Graphics Power (TGP)
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">575W</td>
            <td className="border border-gray-300 px-4 py-2 text-center">360W</td>
            <td className="border border-gray-300 px-4 py-2 text-center">300W</td>
            <td className="border border-gray-300 px-4 py-2 text-center">250W</td>
            <td className="border border-gray-300 px-4 py-2 text-center">450W</td>
        </tr>
        <tr>
            <td className="border border-gray-300 px-4 py-2 font-semibold">
                Recommended Power Supply
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">1000W</td>
            <td className="border border-gray-300 px-4 py-2 text-center">850W</td>
            <td className="border border-gray-300 px-4 py-2 text-center">750W</td>
            <td className="border border-gray-300 px-4 py-2 text-center">650W</td>
            <td className="border border-gray-300 px-4 py-2 text-center">850W</td>
        </tr>
        <tr className="bg-gray-50">
            <td className="border border-gray-300 px-4 py-2 font-semibold">
                Launch Date
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                Jan. 30, 2025
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                Jan. 30, 2025
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                Feb. 20, 2025
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                Mar. 5, 2025
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">
                Sept. 20, 2022
            </td>
        </tr>
        <tr>
            <td className="border border-gray-300 px-4 py-2 font-semibold">
                Launch Price
            </td>
            <td className="border border-gray-300 px-4 py-2 text-center">$1,999</td>
            <td className="border border-gray-300 px-4 py-2 text-center">$999</td>
            <td className="border border-gray-300 px-4 py-2 text-center">$749</td>
            <td className="border border-gray-300 px-4 py-2 text-center">$549</td>
            <td className="border border-gray-300 px-4 py-2 text-center">$1,599</td>
        </tr>
    </tbody>
</table>
</div>

Post originally written March 7, 2025, updated May 21, 2026.


NVIDIA’s RTX 5090 remains hard to find as AI and gaming demand push prices higher. Review the specs, performance features, and rental options on Vast.ai.

NVIDIA GeForce RTX 5090: Everything You Need to Know


120,000+ AI developers are actively searching for compute. Your hardware is what they're looking for.

Whether you're running a single RTX 4090 in your home office or managing a datacenter full of H100s, your idle compute could be generating revenue for you right now.

So, how much money can you actually make renting out GPUs on Vast.ai? The short answer: it varies.

## What Hosts Are Earning on Vast.ai

Your earnings depend on your specific hardware, uptime, reliability, rental type, and how your prices compare to similar listings. That said, as a host, here's a snapshot of what you might expect:

- **Consumer GPUs** like the RTX 5090 typically earn $0.30-$0.60 per GPU-hour, depending on your rental type and your reliability score. This score is a measure of uptime and connection stability. A four-GPU 5090 rig running at 80% utilization can generate $700-$1,400 per month. Use the [host earnings calculator](https://vast.ai/hosting/calculator) to estimate your specific hardware.
- **Datacenter hardware** commands much higher rates. H100s and H200s bring in $2.15-$4.00+ per GPU-hour. If you operate a GPU farm, you can turn dedicated infrastructure into a profitable compute business. Vast.ai even offers financing and will match you with enterprise clients looking for reserved compute, so you can presell capacity and have paying customers from day one.

For current market rates across the Vast.ai platform, check the public [pricing page](https://vast.ai/pricing). Live GPU prices are typically about 25% above what hosts earn.

Once you log in and register at least one machine with Vast.ai, your host dashboard will show detailed market stats with occupancy rates and more, allowing you to optimize your pricing strategy and better estimate potential earnings.

## How Vast.ai Works for Hosts

From hobbyists at home to datacenters and GPU farms, Vast.ai supports every scale. This flexibility is what makes the platform work for hundreds of providers already earning steady revenue.

Many GPU rental platforms lock you into contracts and dictate pricing structures. Some only work with datacenters, requiring certifications and minimum machine counts before you can even list.

Vast.ai is different. You set your rates, select your rental terms, and decide exactly how much bandwidth and storage to offer, whether you're running one GPU or one hundred. *Note: hosting requires a separate account from your client account; a single account cannot be used for both.*

To accommodate your needs, Vast.ai supports three rental models, each designed for different workloads and revenue strategies.

### 1. On-Demand Instances: Premium Pricing for Guaranteed Access

On-demand is your highest-earning option. Set a fixed hourly rate and an offer end date. When a client rents your machine, those terms lock into a rental contract that cannot be changed -- not by modifying the offer, not by unlisting the machine, not by any other action. Once a rental contract is created, its terms -- price, end date, hardware specs -- are fixed. Unlisting your machine prevents new rentals but does not affect any active contracts. You must keep the machine available through every contract's rental end date.

This option attracts clients who are willing to pay a premium for reliability and guaranteed access, but you don't have to pick just one rental type. You can maximize utilization by offering interruptible capacity alongside on-demand instances.

### 2. Interruptible Instances: Lower Rates, Higher Flexibility

Interruptibles work in a bidding system. As the host, you set a minimum interruptible price for a client to rent your machine. Clients set a bid price for their instance; the current highest bid is the instance that runs, with lower bids paused. On-demand rentals always take priority over all interruptible instances regardless of bid price. Higher bids instantly preempt lower ones, and on-demand rentals always take priority.

Interruptible instances appeal to clients who are willing to trade guaranteed uptime for 50-80% cost savings. You'll earn less per hour with interruptible instances vs. on-demand, but they fill gaps in time that would otherwise generate nothing, so most hosts offer both.

### 3. Reserved Instances: Long-Term Contracts at Discounted Rates

Reserved instances let clients secure long-term rentals at a reduced rate. If you choose to offer reserved instances, you set your maximum possible discount off your normal on-demand price. That discount scales with rental duration -- longer commitments unlock larger discounts, up to your set maximum. You can also set this value to zero to opt out of reserved pricing entirely.

The reserved model works well for predictable, stable revenue if you're willing to trade some margin for higher occupancy. Customers commit upfront to weeks or months of utilization. That means you spend less time managing short-term rentals.

## How to Stand Out as a Host and Maximize Revenue

Once you're listed, a number of strategies can help you stand out and make more money:

- **Price competitively at launch.** New hosts often list their hardware at slightly below market rates to attract initial rentals and build a good reliability score faster. You can always adjust pricing later.
- **Offer reserved instance discounts.** Long-term contracts at discounted rates encourage more stable utilization and recurring revenue.
- **Keep machines online.** Strong uptime maintains your reliability score, which reflects connection stability. Note that reliability and verification are tracked independently; reliability drops don't affect your verification status. To avoid disruption, plan maintenance around active rental contracts. For unplanned maintenance, use the CLI's `schedule maint` command, which notifies affected clients to save their work.
- **Get and stay verified.** Vast.ai's automated verification process evaluates your machine's network stability, operational health, and performance. Verification runs on a randomized schedule. High-end hardware like H100s and 8x4090 rigs are typically tested fastest. Verified machines rank higher in search results and attract more rentals.
- **Apply for data center certification.** If you operate professionally managed infrastructure, you can become a [Certified Data Center Partner](https://docs.vast.ai/host/datacenter-status), giving you the blue trust label, priority placement, [Secure Cloud](https://vast.ai/compliance) classification for enterprise compliance, and increased customer traffic.

## Getting Started as a Host on Vast.ai

Becoming a host on Vast.ai is more straightforward than you might think. Setup complexity varies by hardware. The [Host Setup guide](https://cloud.vast.ai/host/setup/) walks you through Ubuntu installation, NVIDIA drivers, network port configuration, and other requirements. It can take as little as an hour.

But you don't need to figure everything out alone. Vast.ai maintains active host-only [Discord](https://discord.gg/vast) channels where providers share troubleshooting advice, setup guidance, and optimization tips. Get peer support from operators who've already solved the problems you might encounter.

Once your machines are online, hosting becomes less hands-on. Simply maintain your hardware and let the earnings roll in. Or you can take a more active approach and gradually optimize pricing, experiment with different rental models, and expand into multi-GPU systems or dedicated GPU infrastructure over time.

The choice is yours.

## The Bottom Line

On Vast.ai, hosts get a level of operational control and pricing transparency that's rare in the GPU cloud market. Whether you're hosting one GPU or an entire cluster, the platform is designed to let you monetize and grow on your own terms.

Here's what that means for you: 120,000+ developers are actively searching for compute on Vast.ai. That demand continues to grow, and it's waiting for your hardware. The only question is whether you're ready to start earning.

Want to put your hardware to work? Check [current GPU Pricing Rates](https://vast.ai/pricing), explore the [Host Setup guide](https://cloud.vast.ai/host/setup/), or jump straight in and list your first machine today.


Learn how much you can earn renting GPUs on Vast.ai - from RTX 4090s to H100s. Compare rental types, pricing strategies, and how to maximize your hosting revenue.

How Much Money Can You Earn Renting Out Your GPU on Vast.ai?


This month's updates bring new templates and guides, major upgrades to two-factor authentication, a benchmarking tool to help you pick the right GPU for your workload, and a complete creative AI studio that runs in a single container.

## NVIDIA Cloud GPU Updates

We've introduced several platform-level improvements this month. Two-factor authentication (2FA) now supports authenticator apps like Google Authenticator and 1Password instead of relying solely on SMS. Additional upgrades include CLI 2FA support, automatic backup codes for account recovery, multiple authentication methods per account, and 2FA for team roles and API key permission groups.

> **Note:** Action is required if you use legacy SMS 2FA. Simply log in normally, go to the Settings page, and click "Regenerate" on the backup codes section to migrate to the new 2FA in less than a minute.

For developers, the Vast [pip package](https://pypi.org/project/vastai/) has been upgraded, and the SDK and CLI repositories are now merged into a single, easy-to-grab repo. We've also introduced an agent skill for coding assistants like Claude and Codex to drive our CLI more effectively.

## Highlighted Feature: Benchmark Before You Deploy

Choosing the right production GPU is now faster and easier, without any guesswork involved. Vast's [benchmarking CLI](https://docs.vast.ai/guides/serverless/choosing-gpus) will take your workload and test it on H100s, A100s, 5090s, 4090s, or any selection you define to compare performance and cost efficiency across GPU classes.

The CLI rents one instance of each GPU type you specify, or a default set for a reasonable starting point if you don't specify any, and then runs the template's built-in benchmark workload, reports performance per dollar, and tears down each rental automatically when finished. Multiple GPU types can be tested in parallel, making it much faster to evaluate your options before committing to longer rentals or production traffic.

Because each run uses real instances and consumes account credit, you get measured performance and cost per hour on actual hardware, not estimates. This benchmarking tool is built for the moment you're staring at a pricing page, wondering whether an H100 is actually worth 4x the cost of a 5090. Sometimes it is, and sometimes it isn't. Now you can find out before you deploy.

## New Templates and Guides

We've rolled out another batch of templates and guides to support creative workflows, fine-tuning, experimentation, and multimodal reasoning.

The standout addition this month is the **All-in-One App Studio** template. It packs eight creative AI applications and a GPU-accelerated remote desktop with KDE Plasma and Blender into a single container, giving you a complete AI production environment without having to manage separate instances.

### New Templates

- [All-in-One Creative AI Studio](https://cloud.vast.ai/template/readme/1d77da4218e4f86b1d2d0fe755d9b953) - One GPU instance: image generation, video, music, voice, transcription, LoRA training, LLM fine-tuning, and KDE desktop with Blender, each activated only when you need it.
- [Unsloth Studio](https://cloud.vast.ai/template/readme/6444356291eabeb423440bb946ccc197) - No-code UI for running and fine-tuning 500+ open-source models with 70% less VRAM.
- [Kimi K2.6](https://vast.ai/model/kimi-k2.6) - 1T-parameter MoE, 32B active, multimodal agentic model with 256K context.
- [Qwen3.5 27B](https://vast.ai/model/qwen35-27b) and [Qwen 3.5B-A3B](https://vast.ai/model/qwen36-35b-a3b) - MoE vision-language models with unified multimodal reasoning and 256K context.
- [Gemma 4 26B-A4B](https://vast.ai/model/gemma-4-26b-a4b-it) and [Gemma 31B IT](https://vast.ai/model/gemma-4-31b-it) - Google's vision-language models with thinking mode and 256K context.
- [Autoresearch](https://docs.vast.ai/examples/ai-agents/autoresearch) - Let Claude Code run ML experiments for you, now in a [template](https://cloud.vast.ai/?ref_id=146798&creator_id=146798&name=Autoresearcher%20%28Claude%20Code%29).

### New Guides

- [Unsloth Studio Guide](https://docs.vast.ai/examples/ai-ml-frameworks/unsloth-studio)
- [CLI Hello World](https://docs.vast.ai/cli/hello-world)
- [SDK Hello World](https://docs.vast.ai/sdk/python/quickstart)

## Other Improvements

This update also brings fleet-wide mitigations for the recently disclosed CopyFail exploit. No action is required from users, and running workloads remain unaffected.

## Our Commitment

As other clouds are out of GPU capacity or charging a premium for what's left, we remain focused on making high-performance AI infrastructure available, affordable, and easy to deploy at scale.

Need help? Contact us anytime at [support@vast.ai](mailto:support@vast.ai) or join our [Discord](https://discord.gg/vast) server for tips, community support, and platform updates in real time.

## Change Log

### New Features

- [Benchmarking CLI](https://docs.vast.ai/guides/serverless/choosing-gpus) for GPU performance and cost comparison.
- [All-in-One App Studio](https://cloud.vast.ai/template/readme/1d77da4218e4f86b1d2d0fe755d9b953): eight creative AI tools in one GPU instance.
- Authenticator app support: use any TOTP app, including Google Authenticator, 1Password, and Microsoft Authenticator, instead of SMS.
- CLI 2FA: 2FA now works through the Vast CLI.
- Backup codes: automatically generate one-time recovery codes so you're never locked out.
- Multiple authentication methods: register more than one per account.
- 2FA for team roles and API key permission groups.

Set up and enable 2FA in [Settings](https://cloud.vast.ai/settings/) -> [Full setup guide](https://docs.vast.ai/guides/reference/account-settings#two-factor-authentication).

### Issues Resolved

- CopyFail exploit: fleet-wide mitigations rolled out with no impact on running workloads and no user action required.
- SSH key input validation, error message, and link fixes.
- Cloud copy fixes.
- Mobile UI fixes.
- Notifications fixes.
- General proactive security fixes.

### API Changes

- The Vast [pip package](https://pypi.org/project/vastai/) has been upgraded, and the SDK and CLI repos have been merged into a single, easy-to-grab repo. Use: `pip install vastai`.
- Install our agent skill to help Claude, Codex, etc., drive our CLI: `npx skills add vast-ai/vast-cli`.

### Updated Templates

- Wan2GP - migrated from provisioning installer to Docker.
- Whisper UI - migrated from provisioning to Docker image.
- FluxGym - Dockerized to prevent provisioning failures.
- vLLM - updated build actions for new CU130 image tag.

### New Blog Posts

- [Deploy LLM Inference Using Vast.ai Serverless](https://vast.ai/article/deploy-llm-inference-using-vast-ai-serverless)
- [Mistral Small 4 Just Dropped - Run It on Affordable H200s with Vast.ai](https://vast.ai/article/mistral-small-4-just-dropped-run-it-on-affordable-h200s-with-vast-ai)
- [Unsloth Studio Is Now Available on Vast.ai](https://vast.ai/article/unsloth-studio-now-available-on-vast-ai)


Vast.ai's May 2026 product update includes authenticator app support for 2FA, a benchmarking CLI for GPU selection, new creative AI templates, and platform security improvements.

May 2026 Product Update


It's been a busy few months at Vast.ai, with updates spanning performance and reliability improvements, new templates and guides, and a major step forward for serverless deployment workflows.

## NVIDIA Cloud GPU Updates

Managing large numbers of rented instances is now faster and more responsive, thanks to pagination and performance improvements when fetching instance data. Monitoring serverless workloads is easier as well, with improved serverless metric visualization so you can track usage and performance at a glance.

We've also updated the Host Setup page with clearer instructions for new hosts, and extended two-factor authentication via SMS to support additional countries.

## Highlighted Feature: Deploy Serverless GPU Endpoints from Python

You can now create, configure, and manage serverless GPU endpoints directly from code, without relying on the dashboard.

With the open beta of the Vast SDK deployment workflow, you can define Docker images, apt packages, autoscaling settings, and more entirely in Python. The goal is simple: one install, one deployment flow, and a much faster path from local code to a live serverless endpoint.

Where this gets especially powerful is the `@remote` programming model. Instead of building HTTP wrappers and managing infrastructure by hand, you can decorate a Python function and call it like local code while Vast handles the serverless execution layer underneath.

This is the fastest way to spin up serverless GPUs at Vast's price-performance point, and it's available now to all users as part of the open beta.

Everything you need is in the [deployment docs](https://docs.vast.ai/documentation/serverless/deployments) and [Python examples](https://docs.vast.ai/documentation/serverless/deployments/examples/square).

![Python serverless deployment example](/uploads/april-2026-serverless-sdk.png)

## New Templates and Guides

We also shipped a fresh round of templates, model pages, and deployment guides to help you get up and running faster across training, inference, multimodal workflows, and migrations.

### New Templates

- [Unsloth Studio](https://cloud.vast.ai/?ref_id=62897&template_id=2fba903698162c83cca06d926a016764) - LLM fine-tuning and inference environment.
- [GLM 5](https://vast.ai/model/glm-5) - 744B MoE model for agentic reasoning, coding, and tool use.
- [Kimi K2.5](https://vast.ai/model/kimi-k2.5) - Next-generation multimodal agentic model.
- [vLLM Omni](https://cloud.vast.ai/?ref_id=62897&template_id=10d5f684d6da7700379b16a7bcb74faa) - Multimodal inference engine.
- [Voicebox TTS](https://cloud.vast.ai/?ref_id=62897&template_id=c2325adb01ae87b2a3e8f6f68f88fabe) - Multi-model text-to-speech with voice cloning.
- [LTX 2.3 via ComfyUI](https://vast.ai/model/ltx-2-3) - Video generation workflow.

### New Guides

- [Budget-Friendly Alternative to Claude Code - Overnight Ralph Loop](https://docs.vast.ai/examples/ai-agents/overnight-ralph-loop)
- [OpenClaw AI Assistant with vLLM on Vast.ai](https://docs.vast.ai/examples/ai-agents/openclaw)
- [Autonomous AI Research with Autoresearch on Vast.ai](https://docs.vast.ai/examples/ai-agents/autoresearch)
- [Migrate from Runpod to Vast.ai](https://docs.vast.ai/examples/migrations/runpod-to-vast)

## Other Improvements

This update also brings a number of platform-level changes, including BitPay support for crypto payments, serverless OpenAI-compatible endpoints, pagination for large rented-instance result sets, and a long list of quality-of-life and reliability fixes.

## Our Commitment

As we continue to improve platform performance and expand support for deployment workflows, our focus remains on making AI infrastructure easier to use at scale.

Need assistance? Reach out anytime at [support@vast.ai](mailto:support@vast.ai) or join our [Discord](https://discord.gg/vast) server for tips, discussion, and the latest updates as they happen.

## Change Log

### New Features

- BitPay payment gateway for crypto payments, including Lightning Network support, flat fees, and refunds across hundreds of wallets including Coinbase.
- Serverless OpenAI API-compatible endpoints released.
- Serverless metric visualization improvements.
- 2FA SMS support expanded to many additional countries.
- Pagination and performance improvements when fetching large numbers of rented instances.
- Host Setup page updated with improved instructions for new hosts.

### Issues Resolved

- Local volume deletion issue resolved.
- Copy error visibility improved.
- Copy progress bar issue resolved.
- Disk quota issues resolved.
- Machines missing geolocation resolved.
- Open button token persistence after instance recreation improved.
- Missing DLPerf on the instance page resolved.
- Various security improvements.
- Various quality-of-life improvements.
- Various serverless edge-case scaling improvements.
- Host CLI schedule maintenance command issue resolved.

### Feature Changes

- Coinbase is sunsetting Coinbase Commerce on March 31, so payment support is moving accordingly.

### API Changes

- Upcoming breaking change: `get instances` will soon return paginated results, with responses limited to a maximum of 25 instances per request. Update API, CLI, and SDK integrations to handle pagination.
- Deprecated: `disable_bundling` for `search offers` will be removed from the API, CLI, and SDK.
- Serverless SDK updated for enhanced security.

### Updated Templates

- PyTorch
- Open WebUI
- Ollama
- Ostris Toolkit
- ComfyUI
- Llama.cpp (ssl)
- SGLang (onstart script fixed)
- Pinokio (v6.x)

### New GPUs

- RTX PRO 5000

### New Blog Posts

- [Affordable Claude Code Alternative: Run Autonomous Coding Agents on Vast.ai](https://vast.ai/article/affordable-claude-code-alternative-ralph-loop)
- [Vast.ai Named Among Fastest Growing Vendors by Ramp and Brex](https://vast.ai/article/ramp-brex-fastest-growing-vendor)
- [Liquid AI's LFM2 Just Dropped - Here's How to Run It on Vast.ai](https://vast.ai/article/liquid-ai-lfm2-run-on-vast-ai)
- [Running Qwen 3.5 Medium Models on Vast.ai](https://vast.ai/article/running-qwen-3-5-medium-models-on-vast-ai)
- [Vast.ai GPUs Can Now Be Rented Through SkyPilot](https://vast.ai/article/vast-ai-gpus-can-now-be-rentend-through-skypilot)
- [7 Popular Model Templates for AI Workloads on Vast.ai Right Now](https://vast.ai/article/popular-templates-for-ai-workloads-on-vast-ai)
- [Running Nemotron-Cascade-2 on Vast.ai](https://vast.ai/article/running-nemotron-cascade-2-on-vast.ai)


Vast.ai's April 2026 product update includes Python-based serverless deployments in open beta, new templates and guides, payment updates, and platform reliability improvements.

Posts about: Vast.ai

June 2026 Product Update

What Is a Neocloud? The Business Model Explained

NVIDIA GeForce RTX 5090: Everything You Need to Know

How Much Money Can You Earn Renting Out Your GPU on Vast.ai?

May 2026 Product Update

April 2026 Product Update