Your models stay yours
Private to your organization, and callable only with your API keys. A GPU serves one organization at a time, and its memory is wiped before it changes hands.
SecurityTwo ways to run a model.
Deploy your own
Any model from Hugging Face, 8B to 1T. It scales from zero in seconds, and you pay for the GPU-seconds it serves.
How deploying works- Cold start, 8B model
- 0.2 s
- Cold start, 1T model
- 3.4 s
- Time to first token, p50
- 30 ms
- GPUs, scaled from zero
- ∞
Pay for the traffic, not the peak.
38% less than dedicated GPUs
- Boardwalk $131k/mo
- Dedicated GPUs $212k/mo
64 H200s
A daily traffic cycle, against dedicated GPUs held for the peak at published list prices, September 2026.
Works with any application.
SDK
from openai import OpenAI
client = OpenAI(
base_url="https://api.boardwalk.cloud/v1",
api_key="bw_live_…",
)
# Your model, deployed from Hugging Face.
response = client.chat.completions.create(
model="my-model",
messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)ModelsAll models
Your data. Your models.
Zero data retention
Prompts and completions are never stored. We keep token counts.
Privacy and accessNews and research
Research from the lab on making inference cheaper and faster, at every size.
See all
Nick CrewsQuestions and answers.
Have another question? Talk to usWhat is Boardwalk?
An elastic inference platform. Call open models from the catalog per token, or deploy any model from Hugging Face, from 8B to a trillion parameters. It scales from zero to as many GPUs as your traffic needs, in seconds, and you pay for the GPU-seconds it serves.
Which models can I deploy?
Any standard LLM on Hugging Face, public or private: dense or mixture-of-experts, bf16, FP8, INT4, NVFP4 or any standard quantized format, and LoRA adapters. We serve exactly the checkpoint you deploy. Choose vLLM or SGLang, the GPU and the context length, or take our defaults. How deploying works.
How fast are cold starts?
Fast enough to scale to zero. Your weights wait in host memory, so a new replica is on a GPU in seconds: the first token in 0.2 s for an 8B model and 3.4 s for a trillion-parameter one. You pay for that load, never for a boot. What to expect.
How does it scale?
Replicas follow your traffic, from zero to as many GPUs as it needs, and back to zero when it stops. A replica keeps its GPU for a hold after its last request, 5 seconds by default and up to 300, so the next request finds it warm. How serving works.
How does reserved capacity work?
Reserved capacity keeps GPUs allocated to your organization, shared by all your models and billed even while idle. Traffic beyond your reservation uses shared GPUs, billed for the time it holds them. Open Capacity in the console to see which GPU classes offer reservations. How reservations are billed.
What do I pay for?
Catalog models bill per token at published rates. A deployment bills its GPU time: per second for on-demand GPUs, its hold included, and by the hour for reserved capacity, plus storage for its weights. Deploying is free. Reserved capacity and the hold are billed even while idle. Prepaid credits start at $5 and never expire. See pricing.
Is my model isolated?
Yes. A GPU serves one organization at a time, your model runs in its own engine processes, and memory is wiped before a GPU changes hands. Your deployments are private to your organization, and we never store prompts or completions. Privacy and access.
Which APIs do you support?
OpenAI Chat Completions, Completions and Responses, plus Anthropic Messages. Use their SDKs with Boardwalk’s base URL and your API key. Available endpoints and features depend on the model. Explore the API.