Skip to content
Bastion Token Factory is now in preview — open-weight models through one OpenAI-compatible API. Join the waitlist →

The AI Cloud built on real infrastructure

Build and scale AI on capacity Bastion has already secured — power, data centers and GPUs locked in with close partners, delivered as clusters and tokens.

From power to tokens. On your terms

Capacity
with dates attached

AI-ready capacity in the United States and Europe, each site with a named power source and delivery plan.

Raw power
No surprises

Bare-metal NVIDIA B300 and GB300 systems, non-blocking InfiniBand or RoCE and 72-hour burn-in before handover.

Any user
Any workload

Pre-training on dedicated clusters, post-training and fine-tuning, or per-token inference through Token Factory.

GPUs
locked in

GPU allocations locked in with close partners and institutions, reserved and scheduled before you sign. One counterparty, one agreement.

Preview

Introducing Bastion Token Factory

Call leading open-weight models — OpenAI gpt-oss, Mistral, Microsoft Phi, IBM Granite and more, all under permissive Apache 2.0 or MIT licenses — through one OpenAI-compatible API, served on GPU capacity Bastion has secured in U.S. and EU data centers. Now in preview, with indicative pricing from $0.03 per million tokens.

gpt-oss-120b
OpenAI · Reasoning · 128K context · Apache 2.0
Preview pricing
$0.15 / 1M input tokens
Output$0.60 / 1M
Batch (24h)−50%
Data retentionZero by default
Token Factory · Preview

Run open models at production scale

Access leading open models, optimize inference, choose how you deploy and adapt models with your own data — all through one platform, on infrastructure Bastion operates.

  • OpenAI-compatible — switch in two lines
  • Serverless or dedicated endpoints
  • LoRA, full fine-tuning and DPO
  • Batch inference at 50% off
  • U.S. or EU-only data residency
  • Zero data retention by default
from openai import OpenAI

client = OpenAI(
    base_url="https://api.bastion-cloud.com/v1",
    api_key=os.environ["BASTION_API_KEY"],
)

resp = client.chat.completions.create(
    model="openai/gpt-oss-120b",
    messages=[{"role": "user",
               "content": "Plan a 2,048-GPU B300 rollout."}],
    stream=True,
)
for chunk in resp:
    print(chunk.choices[0].delta.content or "", end="")
gpt-oss-120b · OpenAIgpt-oss-20b · OpenAIMagistral Small · Mistral AIMistral Small 3.2 · Mistral AIDevstral Small · Mistral AIGranite 3.3 8B · IBMPhi-4 · MicrosoftFLUX.1 [schnell] · Black Forest LabsWhisper large-v3 · OpenAI
13Open-weight models, all Apache 2.0 or MIT licensed
$0.03Indicative preview price per 1M input tokens
2 linesTo switch from the OpenAI SDK
US · EUData residency zones

Open-weight models, permissively licensed

O
OpenAIgpt-oss · Whisper · Apache 2.0 / MIT
MA
Mistral AIMistral · Magistral · Devstral · Apache 2.0
MS
MicrosoftPhi-4 · E5 · MIT
IBM
IBMGranite · Apache 2.0
O
OpenAIgpt-oss · Whisper · Apache 2.0 / MIT
MA
Mistral AIMistral · Magistral · Devstral · Apache 2.0
MS
MicrosoftPhi-4 · E5 · MIT
IBM
IBMGranite · Apache 2.0
O
OpenAIgpt-oss · Whisper · Apache 2.0 / MIT
MA
Mistral AIMistral · Magistral · Devstral · Apache 2.0
MS
MicrosoftPhi-4 · E5 · MIT
IBM
IBMGranite · Apache 2.0
O
OpenAIgpt-oss · Whisper · Apache 2.0 / MIT
MA
Mistral AIMistral · Magistral · Devstral · Apache 2.0
MS
MicrosoftPhi-4 · E5 · MIT
IBM
IBMGranite · Apache 2.0
BF
Black Forest LabsFLUX.1 [schnell] · Apache 2.0
S
SnowflakeArctic Embed · Apache 2.0
N
Nomic AINomic Embed · Apache 2.0
+
Your modelBring your own weights
BF
Black Forest LabsFLUX.1 [schnell] · Apache 2.0
S
SnowflakeArctic Embed · Apache 2.0
N
Nomic AINomic Embed · Apache 2.0
+
Your modelBring your own weights
BF
Black Forest LabsFLUX.1 [schnell] · Apache 2.0
S
SnowflakeArctic Embed · Apache 2.0
N
Nomic AINomic Embed · Apache 2.0
+
Your modelBring your own weights
BF
Black Forest LabsFLUX.1 [schnell] · Apache 2.0
S
SnowflakeArctic Embed · Apache 2.0
N
Nomic AINomic Embed · Apache 2.0
+
Your modelBring your own weights

Built by infrastructure people

Hyperscale capacity planning and data center delivery experience. One console for clusters, jobs and tokens.

Bastion Console · cluster-01● CONCEPT
GPUs
2,048 B300
UTILIZATION
94.2%
IT LOAD
4.31 / 5 MW
TOKENS/S
1.8M
pretrain-moe-400bslurm · 1,536 GPU · running
rl-post-train-v7k8s · 256 GPU · running
tf-endpoint/gpt-oss-120bdedicated · 64 GPU · serving
Facility PUE (7d)1.14 · coolant 28.4 °C

Product concept with illustrative sample data.

24/7 operationsMonitoring, incident response and node auto-remediation under SLA.
Clusters in hours72-hour NCCL/HPL burn-in completed before handover.
Slurm + KubernetesTrain and serve on one fleet, without repartitioning.
Compliance by designCounterparty, export-control and security review built in.

One delivery chain, five services

Use the whole chain or only the layer you need — from a reserved megawatt to a single API call.

Preview · Token Factory

Managed inference for leading open models

Pay per token, or reserve dedicated endpoints on GPU capacity Bastion has locked in. OpenAI-compatible, with post-training and batch built in.

ServerlessDedicatedBatch −50%LoRA · DPOUS · EU

AIDC sourcing & leasing

High-density, AI-ready capacity matched to your power, latency and residency needs.

Secured GPU allocations

B300, GB300 NVL72 and Vera Rubin capacity, locked in through committed allocations with partners.

Cluster deployment

Liquid-cooled racks, rail-optimized InfiniBand or RoCE, burn-in and acceptance.

Managed operations

Single-tenant clusters run 24/7 under SLA, with phased expansion rights.

Not sure where to start?

Tell us your workload and timeline. We'll map it to the right layer and share options under NDA.

Talk to an expert

GPU capacity, locked in

Bastion locks in GPU supply ahead of demand through committed allocations with a close network of hardware, infrastructure and institutional partners. Initial phases run on HGX B300; later phases can move to GB300 NVL72 or Vera Rubin as they ship — without re-contracting.

Locked-in allocations

GPU supply locked in with OEM and channel partners ahead of demand, so capacity is in place before you need it.

Partner-backed capital

Institutional partners back the capacity alongside Bastion, so customers get dedicated compute without carrying GPU capital.

Delivery certainty

Power, racks and GPUs are matched and scheduled together, with delivery dates locked in before you sign.

Available

HGX B300

Blackwell Ultra, 8-GPU nodes, air or liquid cooled.

Available

GB300 NVL72

72-GPU NVLink domain, rack-scale liquid cooling.

Roadmap

Vera Rubin

Next-generation platform for later expansion phases.

Fabric & storage

400G / 800G

InfiniBand or Spectrum-X RoCE, with PB-scale parallel storage.

AI-ready capacity in two regions

Owned and partner capacity across the United States and Europe. Site and partner details are shared only under NDA.

UNITED STATES
580 MW
and growing

Capacity across multiple U.S. regions for large-scale training and inference, with U.S. data residency.

TrainingInferenceU.S. residency
EUROPE
300 MW
and growing

Capacity for sovereign and regulated workloads, with training and inference data kept inside the EU.

Sovereign AIEU-only inferenceGDPR

Capacity figures are indicative and subject to definitive agreements. Map markers are illustrative only and do not indicate actual site locations. Site and partner details are shared only under NDA and with partner consent.

Reliable capacity at scale

US · EUtwo regions with local data residency
Locked inGPU allocations secured with partners
72 hrsburn-in before every cluster handover
10+ yrssite-partner agreements for land and power

A simple structure behind complex infrastructure

Site partners bring land and power. Hardware and institutional partners lock in GPU supply with Bastion. Bastion brings it together and operates it — you get dedicated compute, or tokens, under one agreement.

Site partners

Provide land, power and shell under long-term agreements of 10+ years, with electricity billed on actual use.

Customers

Receive dedicated compute under one service agreement with phased expansion rights — or consume tokens on demand.

Built for real production workloads

Whether you need a 10,000-GPU training cluster or a billion tokens a day, the same infrastructure sits underneath.

Foundation-model developers

Large dedicated training clusters with a path to multi-phase expansion on next-generation GPUs.

AI products & agents

Ship on Token Factory — open models via API — then move hot paths to dedicated endpoints.

Internet & technology companies

Regional capacity for inference and fine-tuning, placed to meet latency, data and regulatory needs.

Quant & research institutions

Isolated, high-performance environments with predictable capacity and strict access controls.

Sovereign & regulated workloads

EU-resident training and inference in European facilities, with no data leaving the region.

Enterprises & integrators

Reserve capacity, deploy private models and integrate with existing identity, networking and billing.

From requirement to production

For Token Factory, skip straight to step five: create an API key and start calling models today.

1

Requirements

Capacity, region, GPU platform, network design and target go-live date.

2

Site match & NDA

Shortlisted sites shared under NDA, capacity held against a letter of intent.

3

Contract & allocation

Service agreement signed and GPU allocation confirmed with our partners, with delivery dates attached.

4

Deployment

Rack integration, fabric build-out, burn-in and customer acceptance.

5

Operate & scale

Managed operations under SLA; more phases and Token Factory capacity as needed.

Why Bastion

01

Capacity with dates attached

Each campus has a named power source, capacity figure and delivery plan.

02

Locked-in supply

GPU allocations are locked in with close hardware, infrastructure and institutional partners, with capacity and delivery dates fixed in advance.

03

Two continents, one roadmap

Capacity in the United States and Europe, delivered on one roadmap with regional data residency.

04

Hyperscale DNA

Capacity planning and data center delivery experience from the world's largest clouds.

The team behind Bastion

Operators who have planned capacity, built data centers and shipped AI platforms at hyperscale.

CM
Founder & CEO

Carl Morris

Former capacity-planning specialist at AWS and Oracle. Sets company strategy and leads capacity strategy, land and power sourcing, and site-partner relationships.

TS
Chief Technology Officer

Tong (Jessy) Shen

AI and internet technology specialist. Leads the Neo Cloud platform and Token Factory, engineering, and technical customer partnerships.

SM
Chief Operating Officer

Si Meng

Data center planning and construction specialist. Leads AIDC development, engineering, project execution and site operations.

Company information

Legal name
Bastion Cloud Services, Inc.
Entity type
Delaware corporation
Incorporated
November 2026
Headquarters
Irvine, California
Operating regions
United States · Europe
General enquiries
Headquarters

Irvine

Corporate Park
Irvine, CA 92606
United States

Silicon Valley office

Cupertino

Apple Park area
Cupertino, CA 95014
United States

Discuss a requirement

Tell us about your capacity, GPU or inference needs. Please don't send confidential technical data before an NDA is in place.

By submitting, you agree to our privacy policy. We reply within one business day.

Start with a token.
Scale to a cluster.