# Introducing Next-Gen Training and Inference at Scale on Lambda Instances with NVIDIA Blackwell

August 12, 2025• 4 min read

[NVIDIA Blackwell GPUs](https://www.nvidia.com/en-us/data-center/technologies/blackwell-architecture/) are now available as 8x [Lambda Instances](https://lambda.ai/service/gpu-cloud) On-Demand, featuring the powerful [NVIDIA HGX™ B200](https://www.nvidia.com/en-us/data-center/hgx/) in addition to our trusted lineup.

Built for the frontier of AI development, NVIDIA Blackwell instances accelerate training and inference for trillion-parameter foundation models, while maintaining the simplicity and flexibility on-demand users expect. Whether you're training state-of-the-art LLMs or deploying custom fine-tunes in production, NVIDIA HGX™ B200 delivers massive performance gains with a familiar developer-friendly experience.

## Why NVIDIA HGX B200?

NVIDIA Blackwell GPUs represent a major step ahead of the [NVIDIA Hopper](https://www.nvidia.com/en-us/data-center/technologies/hopper-architecture/) generation:

- **Up to 2.25x the FP8 throughput** of NVIDIA HGX H100  
- **3x faster training performance** for LLMs  
- **15x faster inference throughput** for real-time applications

With 180GB of blazing-fast HBM3e memory per GPU and new [FP4 support](https://lambda.ai/blog/lambda-1cc-fp4-nvidia-hgx-b200), the NVIDIA Blackwell architecture is tailor-made for modern AI workloads.

## Now Available On-Demand

Lambda users can instantly launch 8x NVIDIA Blackwell GPUs on-demand and pay only for what you use.

Here’s what you get out of the box:

- **8x NVIDIA Blackwell GPUs**  
- **1.4 TB of unified GPU memory**  
- **2900 GiB of system memory**  
- Starting at **$4.99/GPU-hr**

## Built for Real Work, Not Just Benchmarks

This isn’t a testbed. It’s production-ready infrastructure for enterprises training frontier-scale models and running high-throughput inference pipelines.

You’ll benefit from:

- **Immediate availability** on Lambda Cloud, no waitlists or commitments.
- **Simplified model hosting**: Host your entire model on a single node for minimal latency and maximal throughput.
- **Transparent pricing and billing**: Usage-based, no surprises

## NVIDIA Blackwell Architectural Innovations

There are six revolutionary technologies inside the new Blackwell architecture that enable organizations to build and run real-time inference on trillion-parameter large language models. Along with being the basis of the world’s most powerful chip, which is packed with **208 billion transistors**, Blackwell includes a **second-generation Transformer Engine, fifth-generation [NVIDIA NVLink](https://www.nvidia.com/en-us/data-center/nvlink/) ™ interconnect, advanced confidential computing capabilities, a dedicated decompression engine, and a RAS Engine** that identifies potential faults that may occur early on to minimize downtime.

#### Quick Code Peek: Training Llama on Lambda's B200 instances with FP16 precision

Here’s what launching a training job on Lambda’s new NVIDIA HGX B200 instances looks like in practice. A simple DeepSpeed-based launch script for training a Llama-style model on 8x Blackwell GPUs with FP16 support.

Start by spinning up 8xB200 instance, access it via `CLI` in JupyterLab, and follow the steps below:

```
# Activate the virtual training environment
source ~/.venvs/train/bin/activate

# Clone the GitHub repository
git -C ~/transformers remote -v

# Navigate to the Transformer repo
cd ~/transformers

# Confirm whether the Python/DeepSpeed virtual environment is being used and that the required libraries were imported
which python; which deepspeed
python -c "import deepspeed, evaluate; print('OK', deepspeed.__version__)"

# Install HuggingFace’s Python package for model evaluation
python -m pip install -U evaluate datasets

# Ensure presence of the DeepSpeed config file
test -f /home/ubuntu/ds_config_b200_fp16.json && echo "DS config found"

# Confirm that the tiny training dataset exists
ls -lh /home/ubuntu/toy.txt

# Launch training with the virtual environment’s DeepSpeed binary
~/.venvs/train/bin/deepspeed --num_gpus=8 \
  --module examples.pytorch.language-modeling.run_clm \
  --model_name_or_path TinyLlama/TinyLlama-1.1B-Chat-v1.0 \
  --train_file /home/ubuntu/toy.txt \
  --validation_split_percentage 1 \
  --do_train \
  --fp16 \
  --per_device_train_batch_size 8 \
  --gradient_accumulation_steps 4 \
  --num_train_epochs 1 \
  --block_size 1024 \
  --output_dir /home/ubuntu/checkpoints/tinyllama-b200-fp16 \
  --deepspeed /home/ubuntu/ds_config_b200_fp16.json

# Verify that the training workload ran successfully by looking for success messages in the last 200 lines
tail -n 200 -f /home/ubuntu/checkpoints/tinyllama-b200-fp16/runs/*/events* 2>/dev/null || true
ls -l /home/ubuntu/checkpoints/tinyllama-b200-fp16

grep -E "Running training|Saving model checkpoint|Process .* exits successfully" -n training.log 2>/dev/null || true
```

And here's a matching DeepSpeed config for NVIDIA HGX B200:

```
{  
  "fp16": { "enabled": true },  
  "zero_optimization": { "stage": 2 },  
  "gradient_accumulation_steps": 4,
  "train_micro_batch_size_per_gpu": 8,
  "gradient_clipping": 1.0
}
```

This snippet shows how Lambda’s infrastructure and NVIDIA HGX B200 deliver scalable, cost-efficient LLM training. With 8x Blackwell GPUs and FP16 precision, B200 cuts epoch times by up to 2x versus H100s, powering faster pretraining and fine-tuning with zero hassle.

## How to Get Started

1. Head over to [Lambda On-Demand Cloud](https://lambda.ai/service/gpu-cloud)  
2. [Sign In](https://auth.lambdalabs.com/u/login/identifier?state=hKFo2SBaNU05VVVJQWtiRVRXU1NOTXZaVUE1a1VVdDNyU1V4daFur3VuaXZlcnNhbC1sb2dpbqN0aWTZIENPX25XYm9fTXhMUTN1SzJtVS1TY01kZmtScUd5NzZso2NpZNkgdFZ3U1o5QVk0S21RcUJDbTAyYXRkaThXVEtKMHlWV00) to your account  
3. On the instances page, click on **Launch Instance**.  
4. Select **Instance Type** → 8xB200.  
5. Choose **Region** → Select **Filesystem** → Select **Firewall** Rulesets  
6. Click on **Confirm** and launch your instance.

No long term contracts. No paperwork. Just compute.

For teams doing serious LLM work, or looking to do more with less, NVIDIA HGX B200 On-Demand is here to unlock your next milestone.