One cloud formodels, GPUs,and agents

Inference, compute, and agent runtimes behind a single API. Ship on day one.

Start Building

Talk to Us

TRUSTED BY

Halcyon

Tidepool

Kilobyte

Quorra

Openroute

Fishbowl

Hygo

Gizmo

Simular

Wizely

MODEL APIS

LLMImageAudioVideoVision

Model["BM-Sonnet-3"]

200+models

200mslatency

99.5%uptime

1Serverless Model APIs

Reach 200+ models through one endpoint. Nothing to provision.

Text, image, audio and video, all served on demand. Swap a model name and keep the same request shape. You are billed per token, never per idle hour.

Explore All Models

AAtlas 4 Pro$1.74/Mt Input ·
$3.48/Mt Output
1M ContextLLM

VVega M2$0.30/Mt Input ·
$1.20/Mt Output
200K ContextLLM

LLumen 5.1$1.40/Mt Input ·
$4.40/Mt Output
200K ContextLLM

CCorvid K2$0.95/Mt Input ·
$4.00/Mt Output
256K ContextLLM

EEmber 31B$0.14/Mt Input ·
$0.40/Mt Output
256K ContextLLM

SSable 397B$0.60/Mt Input ·
$3.60/Mt Output
256K ContextLLM

AAtlas 4 Pro$1.74/Mt Input ·
$3.48/Mt Output
1M ContextLLM

VVega M2$0.30/Mt Input ·
$1.20/Mt Output
200K ContextLLM

LLumen 5.1$1.40/Mt Input ·
$4.40/Mt Output
200K ContextLLM

CCorvid K2$0.95/Mt Input ·
$4.00/Mt Output
256K ContextLLM

EEmber 31B$0.14/Mt Input ·
$0.40/Mt Output
256K ContextLLM

SSable 397B$0.60/Mt Input ·
$3.60/Mt Output
256K ContextLLM

Base_url

["API.BLUEMETAL.DEV/YOUR-ENDPOINT"]

Operational

P99

53ms

Throughput

1,252 tok/s

Requests

19,421

Latency — last 40 requests

2Dedicated Endpoints

Private capacity. Predictable latency. No noisy neighbours.

Pin a model to compute that only you touch. Throughput stays flat as traffic climbs, so your p99 stops moving when it matters most.

Get Started

AGENT SANDBOX

Agent["CODING AGENTS"]

Coding agent · activeSandbox runtime

Run test suite · pytestQUEUED

Write fix · patch appliedRUNNING

Identify bug · null pointer line 84DONE

Read codebase · src/api/routes.pyDONE

startup

~200ms

isolation

full

billing

per second

status

running

1Agent sandbox

Isolated runtimes, built for agents that take real actions.

Not a notebook and not a container you have to babysit. A disposable environment where an agent can run commands, call tools and hit models — sealed off, every single time.

Get Started

GPU CLOUD

GPU[FLAGSHIP]

GPU model

NVIDIA H100Running

8× GPUs

GPU memory

80 GB HBM3×8

Instance specs

vCPU

Memory

960

Storage

2000

1GPU Instances

Full-control GPU machines, live in seconds.

Train, fine-tune or serve on hardware you own for the duration. Dedicated devices, root access, and performance that does not drift because of someone else's job.

Get Started

JobQueuedRunningComplete

Allocating GPU resourcesAllocating

12%

allocated

auto

duration

0.1s

cost

$0.0001

idle time

$0.00

2Serverless GPU

Hand us the job. We find the hardware.

Nothing to provision and nothing idling. Capacity scales up while the queue is deep and falls to zero the moment it drains — you pay for execution only.

Get Started

Cluster["CLUSTER-01"]

Cluster-01 · 6 nodesNVLink · GPUDirect RDMA · PCIe

Node-01

51%

Node-02

79%

Node-03

86%

Node-05

89%

Node-06

65%

Node-07

81%

GPU8× NVIDIA H200GPU memory141 GB HBM3e per GPU1.128 TB total

Nodes6 / 6InterconnectNVLink 4th Gen — 900 GB/sNetwork400 Gb/s RDMA

3Bare Metal

Peak throughput, with no hypervisor in the way.

Reserved physical clusters for large training runs and high-volume inference. NVLink inside the node, RDMA between them, and the whole machine answering to you.

Get Started

Why BlueMetal

Built for AI from day one. Designed for what you are actually shipping.

Get Started

Others

BlueMetal

Better price-performance

Up to 50% below the major clouds — the result of owning the infrastructure, not of cutting corners.

99.99%

Uptime SLA

<50ms

P50 latency

Global

PoPs

Operational

Built for production reliability

Low latency and high throughput that hold their shape under sustained, real-world load.

Agent runtimeSANDBOX

Model APIs200+ MODELS

GPU infrastructureBARE METAL

One platform for the full AI stack

Model APIs, GPU capacity and agent runtimes sit behind a single account and a single bill.

Serverless API

Dedicated endpoints

GPU control

Scale with your workload

Begin on the shared API and graduate to reserved clusters without rewriting a line.

Seeing latency spikes in prod — can you help debug?2 MIN AGOOn it — pulling your logs now. Config issue, fix incoming.JUST NOW

Dedicated support when it matters

Direct access to engineers who run this infrastructure themselves, not a ticket queue.

Built with BlueMetal

Render Arena

Compare several models side by side

Sandbox Bot

Run an open-source agent with one command

Deploy an Orchestrator

Multi-step agents on BlueMetal sandboxes

Host Your Own Stack

Bring your own images and templates

Testimonials

Don’t take our word for it.

Halcyon

“New checkpoints show up on BlueMetal within a day of release, which means our users get them before we have finished reading the paper. That cadence is hard to find anywhere else.”

MRM. RiveraCo-Founder & CTO

Gizmo

“Integration took an afternoon. The API behaves the same whether we send ten requests or ten thousand, so scaling our study tools has been a non-event.”

PCP. CastellanosCo-Founder and CEO

Fishbowl

“Our speech models sit on BlueMetal GPUs and simply stay up. We spend our time on the model instead of on capacity planning and driver versions.”

SLS. LindqvistCo-Founder & Chief Scientist

Kilobyte

“Agentic coding punishes slow inference. BlueMetal gives us consistent throughput across several models at once, and the team tunes for the workloads we actually run.”

AMA. MoreauHead of Partnerships

What's New

ANNOUNCEMENT

AAtlas 4 Pro

Available on BlueMetal now

LLM$1.60/3.38 in/out MTokens | 1,048,576 Context

ANNOUNCEMENT

LLumen 3.0 Flash

Available on BlueMetal now

LLM$0/0 in/out MTokens | 262,144 Context

ANNOUNCEMENT

VVega M2 Venti