BT H100 HGX
Hopper × 
Blackthorn
Coming soon
- Compute
- 8× H100 SXM
- GPU memory
- 640 GB HBM3
- Runtime
- bLLM, encrypted end to end
ARCIUM BLACKTHORN
Encrypted Superintelligence
Arcium Blackthorn™ turns the world's GPUs into one encrypted supercomputer. It runs the most powerful AI models with every prompt, file, and answer encrypted from the cloud, the operator, and even Arcium.
Learn more RequestComing soon
WHAT IT IS
The world's GPUs, as one encrypted supercomputer
Confidential AI has been rare because it meant trusting a breakable Intel or AMD chip wrapped in a product each data center had to build itself. Blackthorn replaces the chip with math no company controls, and runs on virtually any bare-metal NVIDIA GPU or HGX server already out there.
Confidential by design, not by promise
Every prompt, every answer, every model stays encrypted and invisible to everyone but you. Confidentiality you can prove, not a policy you take on faith.
No trade-offs
The same frontier models. The same speed as ordinary AI. Roughly the same price. Confidentiality becomes the default, not a premium add-on.
No new hardware
A simple software upgrade turns the millions of AI chips already running in datacenters into confidential AI, a 2,500× leap in confidential compute.
Data that could never touch AI, finally can
Hospitals, banks, law firms, and governments can finally put their most confidential information to work.
SECURITY
The most secure confidential AI on Earth
Every other option still leaves something to trust: the chipmaker, whose Intel or AMD enclave keeps getting broken, and the operator running the machine. Blackthorn replaces both with math no single company controls, so there's no chip to break and no operator to trust.
THE COMPUTE UNLOCK
A 2,500× leap, with no new hardware
Making a GPU confidential today is a whole product each data center must build around a security chip that keeps getting broken, which is why you can rent only about 2,000. Blackthorn runs on any bare-metal GPU or HGX server already out there, trusting no chip.
Addressable today
~2,000 rentable confidential GPUs — 0.04% of the world's
With Arcium Blackthorn
~5,000,000 GPUs turned confidential — 2,500× more
1 cell ≈ 2,000 GPUs · full grid ≈ 5,000,000
USE CASES
The data that could never touch AI, finally can
Blackthorn unlocks the most sensitive data on Earth, the kind that makes AI better, not just safer. It stays encrypted end to end and never leaves your control.
01 Everyday
Privacy-preserving AI for everyone
The problem
Every prompt, document, and photo you send to an AI is handed to whoever runs the model. For everyday people and businesses that means personal messages, finances, and customer data sitting exposed on someone else's servers.
How Arcium solves it
Blackthorn runs the same frontier models with your inputs encrypted end to end. Your prompts, files, and answers stay encrypted and invisible to the host, the operator, and even Arcium.
02 Healthcare
Healthcare AI
The problem
Patient records, genomics, and clinical notes are the most valuable data in medicine and the most locked away. Privacy law and risk keep them off every hosted model.
How Arcium solves it
Blackthorn runs inference and training over medical data that stays encrypted end to end. The hospital's data never leaves its control in the clear, and no host, operator, or Arcium ever sees it.
03 Finance
Financial AI
The problem
Trading signals, customer records, and transaction histories are a bank's edge and its liability. Sending them to a hosted model means handing over the crown jewels.
How Arcium solves it
Blackthorn keeps financial data encrypted through inference and training. Models run on the full book without the provider, the operator, or Arcium ever seeing a position or a customer.
04 Sovereign
Sovereign & Regulated AI
The problem
Governments, defense, and law firms hold the most sensitive data there is, and the strictest rules against moving it. Frontier AI has been out of reach.
How Arcium solves it
Blackthorn runs frontier models over sensitive and privileged data that stays encrypted, on hardware you control, with no operator or chipmaker to trust.
FULL UNITS
Request a Full Blackthorn Unit
Coming soonA Blackthorn Unit is dedicated confidential compute running bLLM, Arcium's confidential AI runtime, on the GPU class you choose. Bring your own open or confidential model and run encrypted inference, your data, weights, and outputs never visible to the host or to Arcium.
Hopper ×
Blackthorn
BT H200 HGX
Hopper ×
Blackthorn
Coming soon
- Compute
- 8× H200 SXM
- GPU memory
- 1,128 GB HBM3e
- Runtime
- bLLM, encrypted end to end
BT B200 HGX
Blackwell ×
Blackthorn
Coming soon
- Compute
- 8× B200 SXM
- GPU memory
- 1,440 GB HBM3e
- Runtime
- bLLM, encrypted end to end
ENCRYPTED MODEL API
The best open models, encrypted by API
Coming soonDon't want to bring your own model? Call the biggest, most capable open-weight models on Earth through the Blackthorn API. Every request runs encrypted inside a Blackthorn Unit, so your prompts, files, and outputs stay confidential, billed per token.
DeepSeek 
Moonshot AI 
Meta 
Z.ai 
Alibaba 
Mistral AI 
OpenAI 
Google DeepSeek Moonshot AI Meta Z.ai Alibaba Mistral AI OpenAI Google BRING YOUR OWN GPU
Run encrypted intelligence on your own GPUs
Coming soonAlready operate GPUs? Install bLLM and turn your own NVIDIA Hopper or Blackwell hardware into confidential AI. Same machines, same models, now encrypted end to end, with no security chip to trust.
Runs on hardware you already have
Any bare-metal NVIDIA Hopper or Blackwell GPU or HGX server. No new datacenter, no new chips.
A software switch, not a rebuild
Install bLLM and run, roughly ten times easier to deploy than chip-based confidential AI, with no security chip to trust.
Encrypted end to end
Every prompt, weight, and output stays confidential, encrypted even from you, the operator, by math no company controls.
You keep full control
Your hardware, your datacenter, your models. Blackthorn just makes them confidential.
RequestComing soon