Alpha.sh
Reserve compute Sign in
Backed by HF0

The performance layer
for open intelligence.

We run open-weight models 2× faster and 30% cheaper on US & EU capacity reserved for you.

Reserve compute Sign in

Trusted by teams at the frontier

The Alpha.sh Platform

A unified inference platform for high-performance, portable compute, enabling full optimizations from GPU kernel to API endpoint.

2× faster

One codebase from GPU kernel to API endpoint, so every layer is tuned for the one below it. Text, image and video pipelines.

AI across GPUs, CPUs and ASICs

Same model, same codebase, running on NVIDIA, AMD, arm, Intel, TPU and Trainium. Switch hardware without touching your application.

30%+ lower cost at scale

Higher GPU utilization, faster runtime and hardware chosen per workload. Against generic serving on the same models, savings compound with volume.

A B α 1m tok 5m tok 10m tok

One coherent stack, from your request to the GPU

Most inference is assembled from parts that were never designed to hold an SLA together. Alpha.sh runs every layer, so the contract can cover all of them.

  • GatewayAdmits each request against its deadline, refuses what it cannot finish
  • OrchestrationDeployments, rollouts and capacity changes, with no downtime
  • ServingFrontier open models, scheduled for your p99, not average latency
  • Cache & transportReuses shared prefixes and moves context between GPUs
  • Trust & telemetryVerified nodes, SLA metrics exported to your stack
  • HardwareReserved GPU pools with headroom for your peak, our cloud or yours

Fast inference, deployed your way

Run leading open models or your own custom models in our managed cloud, in your VPC or fully self-hosted.

Inference solutions

Serverless endpoints

High-performance inference on shared pools. No infrastructure to manage, no long-term commitment. Per-token pricing.

Dedicated endpoints

Reserved GPUs across vendors, isolated to your organization, with the full SLA. Per-minute pricing.

Custom models

Bring your own fine-tuned or custom models. Deployed on optimized infrastructure, in our cloud or your VPC.

Get started 
from openai import OpenAI

client = OpenAI(
    # the only change
    base_url="https://alpha.sh/v1",
    api_key="ALPHA_API_KEY",
)

completion = client.chat.completions.create(
    model="deepseek-ai/DeepSeek-V4.1-Flash-1017",
    messages=[
        {
          "role": "user",
          "content": "Summarize this."
        },
    ],
)

print(completion.choices[0].message.content)

Deployment solutions

Our cloud

Fully managed by Alpha.sh. We hold the SLA and the security posture so you don't have to. Best token economics across heterogeneous hardware vendors.

Your cloud

Alpha.sh manages the control plane while inference runs in your VPC. You own the hardware, the data and the cloud credits.

Self-hosted

The Alpha.sh serving engine in a container on your own infrastructure. NVIDIA, AMD, arm, Intel, TPU and Trainium: your hardware, your cloud, your rules.

Built for enterprise workloads

The six questions every security, finance and platform review asks an inference vendor, answered with a document rather than a slide.

Reliability

99.9% uptime, p99 in the contract

Credits computed from our telemetry and applied automatically on the schedule in your agreement. You never file a claim.

SLA & credit schedule · on request
Data

Zero retention, region pinned

Prompts and completions are not stored. EU or US residency per endpoint, single-tenant pools on request, subprocessor list published.

Security & residency brief · on request
Observability

Every request measured, every metric exported

Latency, uptime and SLA outcome per endpoint in the console, exported over OpenTelemetry to your own stack. Breaches are computed from those numbers.

Metrics & export guide · on request
Lock-in

One line to adopt, one line to leave

OpenAI-compatible request and response shapes. Keep your router; send us the traffic that cannot fail.

Migration guide · on request
Cost

Same weights, smaller invoice

Savings come from deadline scheduling and utilization, not from swapping in a compressed copy. When we quantize, the endpoint says so, next to published evals.

Pricing · on request
Capacity

Reserved, with headroom for your peak

Capacity is pre-committed against your contract. No spot, no silent fallback to another provider's build.

Reserve capacity · on request

What you get

Published token prices

Read current rates in the console and API model list, with input, output and reported cache hits priced separately.

DeepSeek Flash on Alpha.sh

$0.60 per 1M output tokens

Your data stays yours

Zero retention, region pinned, 99.9% uptime and p99 latency in the contract.

DeepSeek API · V4.1 Flash 1017peak
$1.20
DeepSeek API · V4.1 Flash 1017off-peak
$0.60
Alpha.sh · V4.1 Flash 1017current
$0.60
Effective October 26, 2026 · one rate at every hour
USD per 1M output tokens. DeepSeek author rates, checked October 25, 2026.

Get started
with Alpha.sh

Send one production workload. In 7 days, see it on Alpha side by side with your stack.

Request a demo

A custom end-to-end deployment built around your models, hardware and performance goals.

  • Dedicated endpoints with 99.9% uptime and p99 in the contract
  • 30% cheaper, measured on your workload
  • Deploy in Alpha.sh cloud, your VPC or self-hosted
  • Evaluate with your own workloads, no commitment
Book a demoTalk with our engineering team
30 min · your workloads · ask us anything