
# Inside GTC 2026: A Bright Future for AI and GPU Computing

NVIDIA's flagship conference, [GTC](https://www.nvidia.com/gtc/), concluded in San Jose with much fanfare. We attended and want to share our key takeaways and what they mean for the future of AI compute. GTC's sessions generated a buzz in the industry, ranging from NVIDIA CEO Jensen Huang's keynote, which introduced the powerful (and controversial) DLSS 5, to announcements of new hardware technologies and the shift from generative to agentic systems. The industry event sparked excitement about the future of AI and GPU computing.

## AI Infrastructure Becomes the Product

The focus has shifted as greater attention is now being paid to AI inference versus yesteryear's training. As AI systems mature and vie for market saturation, the need for sustained computing power grows, and with it, AI infrastructure. NVIDIA and others look to reshape the narrative around AI datacenters, framing them as "AI Factories" where training and inference converge and full-stack systems bring cost efficiency and reliability.

## From Generative to Agentic AI

With the announcement of NemoClaw, a secure, enterprise-grade wrapper for the agentic framework OpenClaw, NVIDIA has broadened the market. NemoClaw was built to address the issue of trust in AI agent adoption, making it safer for companies to utilize agentic systems at scale. This is a crucial development for founders as infrastructure will increasingly need to support long-lived processes with memory and retrieval for 24/7 persistent workloads.

## The Evolution of AI in Gaming

The latest iteration of NVIDIA's gaming tech, DLSS 5, was announced during the show and raised eyebrows in more ways than one. Some critics quickly panned it as "AI slop" but behind the graphical changes that are sure to improve over time lies the "Real-Time Neural Rendering" architecture. The advancement expands possibilities for AI use in creative fields as developers will now be able to precisely control AI rendering parameters. As barriers to AI use in creative projects fall and pressure for workflow efficiency rises, demand for GPU compute power in creative professions will grow.

## Changes in the Global AI Market

Jensen Huang had a lot to say at this year's GTC but some select comments about the broader market and NVIDIA's renewed growth within China stood out. "Our supply chain is getting fired up," Huang said, revealing that NVIDIA is in the process of restarting manufacturing of H200 chips for the Chinese market. After receiving licenses from the U.S. government to once again sell the chips in China, NVIDIA will begin fulfilling orders from Chinese companies. This will have a broader impact on the GPU market as demand for advanced chips expands once again.

## Powering the Next Era of AI Compute

GTC 2026 brought together major and emerging players who showcased new technologies, shared industry developments, and made predictions for the future of AI. What comes next will be defined by how efficiently and broadly AI compute can be deployed.

At Vast.ai, we value the fair distribution of AI computing resources. Our goal is to help you access the infrastructure you need to build, scale, and compete. [Speak to our team](https://vast.ai/contact-sales) today to learn how we can support your AI workloads.


Vast.ai attended NVIDIA GTC 2026 and shares key takeaways on what they mean for the future of AI compute.

Inside GTC 2026: A Bright Future for AI and GPU Computing


# Running Nemotron-Cascade-2 on Vast.ai

NVIDIA released [Nemotron-Cascade-2-30B-A3B](https://huggingface.co/nvidia/Nemotron-Cascade-2-30B-A3B), a reasoning model that achieves gold-medal-level scores on international math and programming competitions. It does this with 30B total parameters and only 3B active per token - 20x fewer than competing models - outperforming both Qwen3.5-35B-A3B and larger open-weight alternatives.

The model fits on a single 80 GB GPU at about 60 GB in BF16 with no quantization or multi-GPU setup required. It includes built-in chain-of-thought reasoning, tool-integrated reasoning for math and code, and an OpenAI-compatible API.

This guide walks through deploying the model on [Vast.ai](https://vast.ai) with vLLM. Vast.ai offers A100 and H100 GPUs at 3-5x less than traditional cloud providers, with no contracts or minimums.

## Model Overview

| Property | Value |
| --- | --- |
| Developer | [NVIDIA](https://research.nvidia.com/labs/nemotron/nemotron-cascade-2/) |
| Model | Nemotron-Cascade-2-30B-A3B |
| Architecture | Hybrid Mamba-Attention MoE |
| Total Parameters | 30B |
| Active Parameters | 3B per token |
| Training | Cascade RL post-training from Nemotron-3-Nano base |
| License | [NVIDIA Open Model License](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-nemotron-open-model-license/) |
| HuggingFace | [nvidia/Nemotron-Cascade-2-30B-A3B](https://huggingface.co/nvidia/Nemotron-Cascade-2-30B-A3B) |

## Deploy Nemotron-Cascade-2 on Vast.ai with vLLM

The model requires a single 80 GB GPU (A100 or H100) with about 150 GB disk for weights and cache.

### Prerequisites

- A [Vast.ai](https://cloud.vast.ai) account with credits
- The Vast CLI installed and configured:

```bash
pip install --upgrade vastai
vastai set api-key <YOUR_API_KEY>
```

Generate an API key to secure your model endpoint:

```bash
export NEMOTRON_API_KEY=$(openssl rand -hex 24)
echo $NEMOTRON_API_KEY  # save this somewhere
```

### Find a GPU

Search for the cheapest single 80 GB GPU with direct port access:

```bash
# -o dph sorts results by price (dollars per hour, cheapest first)
vastai search offers "gpu_ram>=80 num_gpus=1 direct_port_count>=1 cuda_vers>=12.4 cuda_vers<13" -o dph
```

### Launch the Instance

Pick an instance ID from the search results and deploy with vLLM:

```bash
vastai create instance <INSTANCE_ID> \
  --image vllm/vllm-openai:latest \
  --env '-p 8000:8000' \
  --disk 150 \
  --onstart-cmd "vllm serve nvidia/Nemotron-Cascade-2-30B-A3B \
    --host 0.0.0.0 --port 8000 \
    --max-model-len 32768 \
    --api-key $NEMOTRON_API_KEY \
    --trust-remote-code"
```

This pulls the model from HuggingFace and starts an OpenAI-compatible API server on port 8000. Initial startup takes several minutes while the model downloads.

Once the instance is running, get your connection details:

```bash
vastai show instances --raw | jq '.[] | select(.id == <INSTANCE_ID>) | {id, actual_status, public_ipaddr, ports}'
```

You can also find the IP and port mapping in the [Vast console](https://cloud.vast.ai/instances/).

## Call the API

Send a request to test the model:

```bash
curl http://<VAST_IP>:<PORT>/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $NEMOTRON_API_KEY" \
  -d '{
    "model": "nvidia/Nemotron-Cascade-2-30B-A3B",
    "messages": [{"role": "user", "content": "Find all integers n such that n^2 + 2n + 4 is divisible by 7."}],
    "max_tokens": 2048
  }'
```

The model thinks before it answers - you will see its step-by-step reasoning in the response, followed by the final answer. This chain-of-thought reasoning is what drives its competition-level performance on math and code.

## Cleanup

When you're done, destroy the instance to stop charges:

```bash
vastai destroy instance <INSTANCE_ID>
```

## Conclusion

Nemotron-Cascade-2 delivers reasoning performance that previously required models 20x its size, and it runs on a single GPU. That combination of capability and efficiency makes it a practical option for math, code, and general reasoning workloads. With Vast.ai, you can have it running in minutes at a fraction of what traditional cloud providers charge.

## Resources

- [Nemotron-Cascade-2-30B-A3B on HuggingFace](https://huggingface.co/nvidia/Nemotron-Cascade-2-30B-A3B)
- [Nemotron-Cascade-2 Research Page](https://research.nvidia.com/labs/nemotron/nemotron-cascade-2/)
- [Nemotron-Cascade-2 Paper (arXiv:2603.19220)](https://arxiv.org/abs/2603.19220)
- [Vast.ai Documentation](https://docs.vast.ai/)
- [vLLM Documentation](https://docs.vllm.ai/)


Deploy NVIDIA's Nemotron-Cascade-2-30B-A3B on Vast.ai with vLLM and run a powerful reasoning model on a single 80GB GPU.

Running Nemotron-Cascade-2 on Vast.ai


The world’s smallest desktop AI supercomputer is now available in the form of the NVIDIA DGX™ Spark. Originally [announced in January 2025](https://www.theverge.com/news/798775/nvidia-spark-personal-ai-supercomputer) under the name Digits, NVIDIA CEO Jensen Huang praised Spark as a tool for data scientists, AI researchers, and students “to engage and shape the age of AI.”

Powered by a 240 watt power supply that uses a standard wall outlet and with a golden asteroid-like textured casing, it looks more like a high-end PC than an AI supercomputer. Only an inch wider and longer than the M4 Mac Mini, Spark measures just 6” x 6” x 2” and weighs just 2.6 pounds. This compact size makes it all the more impressive that it can fuel enterprise AI development with ease. If you’re looking for a mini AI companion to run local AI models, the NVIDIA DGX Spark is the best-looking, most-talked-about option on the market. But does it deliver on the hype?

### **Inside the NVIDIA DGX Spark**

Spark runs on a 20-core ARM64 CPU and a NVIDIA GB10 [Grace Blackwell](https://vast.ai/article/report-nvidia-developing-new-blackwell-based-ai-chip-for-china) GPU. It’s capable of handling AI models with up to 200 billion parameters at a million billion calculations per second. Its 128GB of unified memory can be seamlessly allocated to both its CPU and GPU without static partitioning. Plus, its stock 4TB NVME.M2 SSD hard drive provides more storage than many of its closest competitors.  
With four USB-C ports and an HDMI port, you can plug it into an array of peripheries, monitors, or even external GPUs. You can also connect two Sparks together and run them simultaneously to double your performance.

#### NVIDIA DGX Spark Technical Specifications

<table class="w-full border-collapse border border-gray-300">
    <thead>
        <tr class="bg-gray-100">
            <th class="border border-gray-300 px-4 py-2 text-left font-semibold">
                Specification
            </th>
            <th class="border border-gray-300 px-4 py-2 text-left font-semibold">
                NVIDIA DGX Spark
            </th>
        </tr>
    </thead>
    <tbody>
        <tr>
            <td class="border border-gray-300 px-4 py-2 font-semibold">
                Architecture
            </td>
            <td class="border border-gray-300 px-4 py-2">
                NVIDIA Grace Blackwell
            </td>
        </tr>
        <tr class="bg-gray-50">
            <td class="border border-gray-300 px-4 py-2 font-semibold">GPU</td>
            <td class="border border-gray-300 px-4 py-2">
                Blackwell Architecture
            </td>
        </tr>
        <tr>
            <td class="border border-gray-300 px-4 py-2 font-semibold">CPU</td>
            <td class="border border-gray-300 px-4 py-2">
                20 core Arm, 10 Cortex-X925 + 10 Cortex-A725 Arm
            </td>
        </tr>
        <tr class="bg-gray-50">
            <td class="border border-gray-300 px-4 py-2 font-semibold">
                CUDA Cores
            </td>
            <td class="border border-gray-300 px-4 py-2">
                Blackwell Generation
            </td>
        </tr>
        <tr>
            <td class="border border-gray-300 px-4 py-2 font-semibold">
                Tensor Performance<sup>1</sup>
            </td>
            <td class="border border-gray-300 px-4 py-2">1 PFLOP</td>
        </tr>
        <tr class="bg-gray-50">
            <td class="border border-gray-300 px-4 py-2 font-semibold">
                System Memory
            </td>
            <td class="border border-gray-300 px-4 py-2">
                128 GB LPDDR5x, coherent unified system memory
            </td>
        </tr>
        <tr>
            <td class="border border-gray-300 px-4 py-2 font-semibold">OS</td>
            <td class="border border-gray-300 px-4 py-2">NVIDIA DGX™ OS</td>
        </tr>
        <tr class="bg-gray-50">
            <td class="border border-gray-300 px-4 py-2 font-semibold">
                System Dimensions
            </td>
            <td class="border border-gray-300 px-4 py-2">
                150 mm L x 150 mm W x 50.5 mm H
            </td>
        </tr>
    </tbody>
</table>

See full specs at [NVIDIA.com](https://www.nvidia.com/en-us/products/workstations/dgx-spark/).

### **Should You Buy the NVIDIA DGX Spark?**

Following a delayed launch, Spark was released earlier this month. It is [available on](https://www.nvidia.com/en-us/products/workstations/dgx-spark/) [NVIDIA.com](http://NVIDIA.co) and other retailers for $3,999. However, Spark was initially slated to retail for only $2,999 — the price of the similarly spec’d ASUS Ascent GX10. Though the Ascent comes with only a 1TB hard drive and is far less inspired design-wise. Beyond its delayed release and initial sticker shock, early reviews have been mixed.

Many early adopters praise Spark’s ability to run larger AI models and champion the GP10’s ability to outperform many PC GPUs. Additionally, remote access through tools like Tailscale have made Spark even more accessible. Using it as a companion to developers’ existing PCs and other devices, you can work with your Spark from anywhere on earth.

However, Spark is being criticized for not reaching its listed capabilities. Though rated for 240 watts, noted programmer and video game developer [John Carmack reported](https://www.pcgamer.com/hardware/graphics-cards/og-doom-legend-john-carmack-says-his-nvidia-dgx-spark-box-has-thermal-throttling-issues-raising-doubts-over-the-viability-of-nvidias-ai-apu-for-laptop-pcs/) less than half of that. As a result, he suggested it only delivers about half of its advertised performance. With wide reports of latency, throttling, and reboot issues, some shoppers have voiced reservations about buying it for more data-intensive projects.

For now, Spark’s full range of ideal use cases have yet to be explored and its ecosystem is still evolving. Recent [news that open source platforms](https://blogs.nvidia.com/blog/open-source-aerial-ai-native-6g/) like NVIDIA Sionna Research Kit and Aerial Testbed are available for the DGX network could provide Spark with native tools to accelerate its capabilities. It has shined brightest for agentic AI and reasoning models. For a mini AI supercomputer, it does a lot — it may just not be enough for everyone.

### **Bringing Powerful Compute Within Reach**

Spark's steep price tag makes it an expense that may be hard to justify. While on-premise, local AI development sounds great, Spark may be too niche for wide adoption at its price point. But the tiny supercomputer does nonetheless reflect a push toward making advanced AI computing more widely available.

Here at Vast.ai, we share that goal of making AI development accessible. Our mission is to democratize AI by making powerful GPUs available to anyone, anywhere — without the upfront cost of purchasing and maintaining costly hardware. With our cloud GPU rental platform, you can leverage machines that exceed DGX Spark with affordable, on-demand compute power designed to scale with your needs.

After all, democratizing AI starts with access. On Vast, you'll save 5–6X compared to traditional cloud providers. Find your compute and [get started](https://vast.ai/pricing) today\!


Explore the NVIDIA DGX Spark — the world’s smallest desktop AI supercomputer — its specs, performance, reviews, and how Vast.ai offers a more scalable alternative.

The NVIDIA DGX Spark: The New Desktop AI Supercomputer That Fits in Your Hand


NVIDIA is [reportedly](https://www.reuters.com/world/china/nvidia-working-new-ai-chip-china-that-outperforms-h20-sources-say-2025-08-19/) working on a new China-specific AI chip that will be more powerful than the aging HGX H20 model the company is currently authorized to sell in China.

Recent reports [suggest](https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-responds-to-reports-that-its-h20-gpu-for-china-is-ending-production-next-gen-b30a-green-light-up-to-the-united-states-government-according-to-ceo-jensen-huang) that the H20 – originally developed as a scaled-down alternative to the A100 and H100 that were banned from export to China – is nearing the end of its production run. Its successor: the tentatively named B30A chip.

Based on the latest Blackwell architecture, the B30A will feature a single-die design delivering roughly half the raw computing power of NVIDIA's flagship B300 accelerator, according to sources. The new chip will include high-bandwidth memory and NVLink support for fast data transfer between processors – both of which are also present in the Hopper-based H20.

NVIDIA aims to send test batches of the B30A chip to Chinese customers as early as sometime this month, sources said.

The move comes following remarks from President Donald Trump suggesting that more advanced NVIDIA chips could be cleared for sale in China. However, regulatory approval remains uncertain. There are ongoing concerns about granting China broader access to U.S. AI technology – which could weaken the United States' competitive edge in the sector.

At the same time, analysts warn that restricting China too heavily risks driving developers toward rivals like Huawei. NVIDIA argues that it's important to retain market share in China – which generated 13% of the company's revenue in the past year – so that its software ecosystem stays relevant in one of the world's largest AI markets.

NVIDIA is also preparing a second Blackwell-based China-specific chip called the RTX6000D, designed primarily for AI inference tasks, which is expected to start shipping this month. With weaker specs than the H20 and bandwidth just under U.S. export limits, it's set to be a compliant, lower-cost alternative within NVIDIA's lineup.

For those outside China, however, these region-specific models highlight the stark difference between scaled-down chips and the unrestricted performance available elsewhere.

### On-Demand Access to Powerful GPUs, Anytime

While NVIDIA adapts its offerings to meet specific market regulations, users elsewhere don't have to work around restricted versions.

On [**Vast.ai**](http://Vast.ai), you can access a wide range of powerful GPUs – from the [RTX 5090](https://vast.ai/article/nvidia-geforce-rtx-5090-specs-everything-you-need-to-know) to the [A100](https://vast.ai/article/nvidia-rtx-4090-vs-a100-two-powerhouses-two-purposes), [H200](https://vast.ai/article/nvidia-h100-vs-h200-two-hopper-based-heavyweights), [B200](https://vast.ai/article/nvidia-h200-vs-b200-comparing-datacenter-grade-accelerators), and more – on demand, at a fraction of the cost of traditional cloud providers. Our cloud GPU rental platform makes it easy. Just choose the hardware you need, spin up an instance in seconds, and pay only for the compute you use.

Get started with Vast today, and enjoy flexible, cost-effective GPU compute that scales with your needs\!


NVIDIA is developing the B30A AI chip for China, a Blackwell-based successor to the H20 with half the power of the B300. Learn what it means for AI exports, competition, and how you can still access unrestricted GPUs like the H200, B200, and A100 on Vast.ai.

Report: NVIDIA Developing New Blackwell-Based AI Chip for China


If you're working at the cutting edge of AI and high-performance computing (HPC), NVIDIA's H200 and B200 GPUs are likely already very much on your radar. They're each designed to handle the massive compute demands of datacenter-class workloads – such as training trillion-parameter models and running high-speed inference at scale.

While both GPUs deliver incredible power, it's crucial to understand the differences between the two in order to make the right choice for your needs.

Today, we'll explore how the **H200** and **B200** measure up against each other and how each fits into real-world deployments. Which one is right for you? Let's take a look.

### NVIDIA H200: Hopper-Based Heavyweight

Building on the foundation laid by the H100, the NVIDIA H200 is based on the same Hopper architecture while bringing substantial improvements in memory capacity and bandwidth. This means fewer bottlenecks, an advantage that becomes increasingly important as you move further into enterprise scale.

The H200 nearly doubles the memory of its predecessor, boasting **141 GB of HBM3e memory** and **4.8 TB/s of bandwidth** – delivering about **1.4X faster** data access. MLPerf benchmarks using Llama 2 70B demonstrated the H200's power: it reached over 31,000 tokens per second, about **45% faster** than the H100.

With the same **fourth-gen NVLink**, **PCIe Gen5**, **MIG support**, and full Tensor Core and Transformer Engine capabilities, the H200 is a solid next step for existing Hopper-based deployments. You don't have to rethink your entire infrastructure; you just get more horsepower.

To sum up, here are some of the most notable features of the H200:

-   **141 GB of HBM3e memory at 4.8 TB/s** – almost double the memory of the H100 with 1.4X the bandwidth.

-   **Up to 1.9X improved LLM inference performance**, with more tokens per second and higher throughput for gen AI workloads compared to its predecessor.

-   **Lower energy use and total cost of ownership** (TCO) over the H100 thanks to improved energy efficiency and thermal management.

-   **An upgrade path exists on qualified HGX/MGX platforms**, so that little to no reconfiguration is required.

In short, the H200 represents the very best of the Hopper-based lineup. When it was introduced in late 2024, it set a new standard for memory-intensive and high-throughput workloads in datacenter GPUs – and it's still one of the most powerful GPUs on the market today.

The B200, on the other hand, is an entire generational leap forward.

### NVIDIA B200: The New Generation

Built on NVIDIA's advanced Blackwell architecture, the B200 features **192 GB of ultra-fast HBM3e memory** with **8.0 TB/s of memory bandwidth.** Those are some staggering numbers, even compared to the H200's respectable 141 GB at 4.8 TB/s.

The Blackwell architecture also introduces **fifth-gen Tensor Cores** and **dual Transformer Engines**, which are particularly useful for AI workloads with long context windows and heavy token parallelism. The B200 scales much more efficiently than the H200 for the latest and most advanced model classes.

In fact, scaling is a major area where the B200 stands out. With **fifth-gen** **NVLink**, its inter-GPU bandwidth reaches **1.8 TB/s** within a node – compared to Hopper's 900 GB/s – with cross-node scaling supported via the NVLink Switch System.

Of course, this may not matter to you very much if you're only using a single card. At cluster scale, however, or in multi-GPU systems like the DGX B200 (which connects eight B200 GPUs together), the improvements in distributed training efficiency really make a difference.

The B200 truly sits at the frontier of performance. Its raw capacity and speed can support next-gen inference pipelines and AI workloads at hyperscale. According to NVIDIA, it delivers **3X the training performance** and **15X the inference performance** of the previous generation.

These gains do come at a cost in power draw. The B200 has a max thermal design power (TDP) of **1000 W** compared to the H200 at 700 W – which means it also requires far more robust cooling solutions. But the extra power consumption may just be worth it for the boost in performance.

To recap, a few advantages of the B200 include:

-   **192 GB of HBM3e memory at 8.0 TB/s** – ultra-high capacity that's well suited for trillion-parameter models and allows training of AI models _entirely in GPU memory_.

-   **Fifth-gen Tensor Cores and second-gen Transformer Engines** for higher efficiency in FP8, FP16, and mixed precision performance, especially in long-context and multi-modal workloads.

-   **NVLink 5 interconnect** with up to **1.8 TB/s** GPU-to-GPU bandwidth for multi-node scaling.

-   **Fine-grain micro-tensor scaling** that optimizes performance and accuracy, enabling 4-bit floating point (**FP4**) AI.

The B200 was built to shatter performance records. As you can imagine, it has an eyewatering price tag to match. Nonetheless, it can be helpful to compare the B200 and H200 based on specs alone.

### H200 vs. B200: Key Specs Compared

Here's a side-by-side look at some key features of the H200 and B200:

<table class="w-full border-collapse border border-gray-300">
    <thead>
        <tr class="bg-gray-100">
            <th class="border border-gray-300 px-4 py-2 text-left font-semibold">
                Feature
            </th>
            <th class="border border-gray-300 px-4 py-2 text-left font-semibold">
                H200
            </th>
            <th class="border border-gray-300 px-4 py-2 text-left font-semibold">
                B200
            </th>
        </tr>
    </thead>
    <tbody>
        <tr>
            <td class="border border-gray-300 px-4 py-2 font-semibold">
                Architecture
            </td>
            <td class="border border-gray-300 px-4 py-2">Hopper</td>
            <td class="border border-gray-300 px-4 py-2">Blackwell</td>
        </tr>
        <tr class="bg-gray-50">
            <td class="border border-gray-300 px-4 py-2 font-semibold">
                Memory
            </td>
            <td class="border border-gray-300 px-4 py-2">141 GB HBM3e</td>
            <td class="border border-gray-300 px-4 py-2">192 GB HBM3e</td>
        </tr>
        <tr>
            <td class="border border-gray-300 px-4 py-2 font-semibold">
                Memory Bandwidth
            </td>
            <td class="border border-gray-300 px-4 py-2">4.8 TB/s</td>
            <td class="border border-gray-300 px-4 py-2">8 TB/s</td>
        </tr>
        <tr class="bg-gray-50">
            <td class="border border-gray-300 px-4 py-2 font-semibold">
                CUDA Cores
            </td>
            <td class="border border-gray-300 px-4 py-2">16,896</td>
            <td class="border border-gray-300 px-4 py-2">16,896 x2</td>
        </tr>
        <tr>
            <td class="border border-gray-300 px-4 py-2 font-semibold">
                Tensor Cores
            </td>
            <td class="border border-gray-300 px-4 py-2">528 (4th Gen)</td>
            <td class="border border-gray-300 px-4 py-2">528 x2 (5th Gen)</td>
        </tr>
        <tr class="bg-gray-50">
            <td class="border border-gray-300 px-4 py-2 font-semibold">
                Boost Clock
            </td>
            <td class="border border-gray-300 px-4 py-2">1.98 GHz</td>
            <td class="border border-gray-300 px-4 py-2">1.98 GHz</td>
        </tr>
        <tr>
            <td class="border border-gray-300 px-4 py-2 font-semibold">
                Transformer Engine
            </td>
            <td class="border border-gray-300 px-4 py-2">Single</td>
            <td class="border border-gray-300 px-4 py-2">Single (2nd gen)</td>
        </tr>
        <tr class="bg-gray-50">
            <td class="border border-gray-300 px-4 py-2 font-semibold">
                Form Factor
            </td>
            <td class="border border-gray-300 px-4 py-2">SXM</td>
            <td class="border border-gray-300 px-4 py-2">SXM</td>
        </tr>
        <tr>
            <td class="border border-gray-300 px-4 py-2 font-semibold">
                Interconnect
            </td>
            <td class="border border-gray-300 px-4 py-2">
                NVLink 4 (900 GB/s) per GPU
            </td>
            <td class="border border-gray-300 px-4 py-2">
                NVLink 5 (1.8 TB/s) per GPU
            </td>
        </tr>
        <tr>
            <td class="border border-gray-300 px-4 py-2 font-semibold">
                Precision Formats
            </td>
            <td class="border border-gray-300 px-4 py-2">
                FP8, FP16/BF16, TF32, FP32, FP64, INT8
            </td>
            <td class="border border-gray-300 px-4 py-2">
                FP4, FP6, FP8, FP16/BF16, TF32, FP32, FP64, INT8
            </td>
        </tr>
        <tr class="bg-gray-50">
            <td class="border border-gray-300 px-4 py-2 font-semibold">
                Multi-Instance GPUs
            </td>
            <td class="border border-gray-300 px-4 py-2">Up to 7</td>
            <td class="border border-gray-300 px-4 py-2">Up to 7</td>
        </tr>
        <tr>
            <td class="border border-gray-300 px-4 py-2 font-semibold">
                Max Thermal Design Power (TDP)
            </td>
            <td class="border border-gray-300 px-4 py-2">700 W</td>
            <td class="border border-gray-300 px-4 py-2">1000 W</td>
        </tr>
        <tr class="bg-gray-50">
            <td class="border border-gray-300 px-4 py-2 font-semibold">
                Recommended Power Supply
            </td>
            <td class="border border-gray-300 px-4 py-2">1100 W</td>
            <td class="border border-gray-300 px-4 py-2">1400 W</td>
        </tr>
    </tbody>
</table>

Choosing the right GPU often comes down to more than just specs, however.

### Use Cases: Which GPU Should You Choose?

While both the H200 and B200 deliver game-changing performance capabilities, they each have distinct advantages depending on your needs.

#### You should consider the H200 if you prefer:

-   **A high-performance GPU** ideal for fine-tuning LLMs and high-throughput inference at the enterprise level without getting into exascale territory.

-   **A proven workhorse** that balances strong performance with lower TDP and cost.

-   **Being able to seamlessly upgrade** from existing Hopper deployments without overhauling infrastructure.

#### Or the B200 might be the right choice if you need:

-   **To push model boundaries** and handle long-context and/or multi-modal foundation model training at scale.

-   **The absolute maximum throughput and efficiency** for multi-GPU or multi-node clusters.

-   **A future-proof investment** for infrastructure built to support the next generation of cutting-edge AI and AGI.

### Conclusion

At **Vast.ai**, we know that the choice between GPUs like the H200 and B200 can depend on a lot more than just performance. Access and cost are important factors, as well.

The H200 and B200 represent the very best of NVIDIA's current lineup, and they're among the most powerful and expensive GPUs in the world today. Buying them outright is undoubtedly out of reach for most teams.

That's what we're here for. With Vast, you can rent high-end GPUs like the [A100](https://vast.ai/article/H100-vs-A100-Comparing-two-Powerhouse-GPUs), [H100](https://vast.ai/article/nvidia-h100-vs-l40s-power-meets-versatility), [H200](https://vast.ai/article/nvidia-h100-vs-h200-two-hopper-based-heavyweights), and soon the B200, with on-demand pricing and spot instances – and pay only for what you use.

On average, **Vast users save 5–6X on GPU compute** compared to traditional cloud providers. You'll find a variety of options on our platform to fit your needs, so you can focus on what really matters: your work, not your budget.

Explore Vast today and enjoy enterprise-grade compute without enterprise-level costs.


The H200 and B200 GPUs represent the top end of NVIDIA's current lineup for AI and high-performance computing (HPC). Both are designed for serious, large-scale workloads – from training foundation models to running complex scientific simulations.

NVIDIA H200 vs. B200: Comparing Datacenter-Grade Accelerators


We're back with a new update from the team at **Vast.ai**\! This month, we've rolled out fresh hardware options, new video-generation tools, and continued improvements behind the scenes.

The big news is that our GPU lineup now includes the **NVIDIA DGX B200** as well as the **RTX PRO 6000 WS**\! These powerful machines open the door to even more demanding workloads, bringing you the compute power required for everything from multi-modal inference to precision-driven simulations.

We also recently reached an exciting milestone: we now have **over one thousand RTX 5090 GPUs available on Vast** – 1,180 and counting.

Here are the details about what's new this month\!

### NVIDIA Cloud GPU Updates

Whether you're training foundation models, rendering high-resolution video, or running inference, our latest hardware additions give you extra flexibility:

-   **NVIDIA DGX B200** – A universal AI supercomputer built on the Blackwell architecture, equipped with **eight GPUs** and **1.4 TB** of GPU memory. With massive bandwidth and compute capacity, it's well suited to handle virtually any enterprise AI workload.

-   **RTX PRO 6000 WS** – A high-end Blackwell-based workstation GPU featuring **96 GB** of GDDR7 memory and up to **4000 TOPS** of AI compute. It can easily handle massive datasets and multi-billion-parameter models while supporting seamless multi-app workflows.

### New Guide and Templates

We've focused on video generation this month, adding a couple of new templates and a guide to help you go from idea to clip as seamlessly as possible:

-   [**Guide to Video Generation**](https://docs.vast.ai/video-generation) – A walkthrough for setting up and using ComfyUI on Vast.ai, so you can create advanced stable diffusion pipelines (ideal for video generation workflows) with this powerful node-based interface.

-   **LTX Video (ComfyUI)** – A ready-to-use template that pairs the LTX Video model with ComfyUI, optimized for consumer-grade GPUs.

-   **Open-Sora** – A template for this open-source video generation model with a streamlined, user-friendly platform that lowers the barrier to efficient, high-quality video content creation.

And that's a wrap for June 2025\!

### Our Commitment

Whether you're looking to launch a B200 instance, experiment with video generation using our latest templates, or scale up with our expanding lineup of RTX 5090 GPUs, we're here to support your work at every stage. Our goal at **Vast.ai** is to make powerful compute accessible, flexible, and easy to use – so you can focus on building, training, and creating.

Need help? Reach out to our customer support team anytime at [support@vast.ai](mailto:support@vast.ai), or connect with the community on our Discord server for tips, troubleshooting, and more.

### Change Log

#### New Templates & Guides:

-   [Guide to Video Generation](https://docs.vast.ai/video-generation)
-   LTX Video (Comfy UI) – Video-Gen model optimized for consumer-grade GPUs.
-   Open-Sora – Democratizing efficient video production for everyone.

#### New GPUs:

-   DGX B200
-   RTX PRO 6000 WS

Plus we now have over one thousand 5090 GPUs available.


New DGX B200 and RTX PRO 6000 WS GPUs, over 1000 RTX 5090s available, and new video generation templates.

Posts about: Nvidia

Inside GTC 2026: A Bright Future for AI and GPU Computing

Running Nemotron-Cascade-2 on Vast.ai

The NVIDIA DGX Spark: The New Desktop AI Supercomputer That Fits in Your Hand

Report: NVIDIA Developing New Blackwell-Based AI Chip for China

NVIDIA H200 vs. B200: Comparing Datacenter-Grade Accelerators

June 2025 Product Update