What is VoltageGPU Confidential Chat?

A ChatGPT-style interface running inside Intel TDX hardware enclaves. Prompts and responses are encrypted in CPU memory at runtime and never leave the sealed Trust Domain. VoltageGPU cannot read user data, cannot log it, cannot train on it.

Do I need an account to try it?

No. Anonymous visitors get 3 free messages on Qwen3-32B-TEE without any signup, no credit card required. After the free trial, signup unlocks 50 free messages per month, Plus $20/mo unlocks 2000 messages with Qwen3-235B 262K context.

How is it different from ChatGPT Plus?

Same $20/mo price point but sealed in Intel TDX hardware. OpenAI trains on user data by default. VoltageGPU Plus cannot technically read user data — memory is encrypted with CPU-fused keys and even the hypervisor is outside the trust boundary. EU company (France, SIREN 943 808 824), native GDPR Article 28 DPA available.

Which TEE models are available?

3 models sealed in Intel TDX: Qwen3-32B-TEE (40K context, fast), Qwen3-235B-A22B-Instruct-2507-TEE (262K context, flagship, Plus/Pro), DeepSeek-R1-0528-TEE (163K context, reasoning model, Enterprise).

Can I use it for regulated work (legal, medical, finance)?

Yes. Beyond the generic chat, 8 specialized agents are available: Contract Analyst, Financial Analyst, Medical Records, Compliance Officer, Due Diligence, Cybersecurity, HR, Tax. GDPR Article 28, HIPAA, DORA, NIS2 ready. DPA available on request.

Yes — api.voltagegpu.com/v1 is OpenAI-compatible, pay-per-token starting at $0.15/M input tokens on Qwen3-32B-TEE. Drop-in replacement for the OpenAI SDK (change base_url).

How does fine-tuning stay private inside Intel TDX?

Training data is uploaded over TLS 1.3, then loaded directly into the Intel TDX trust domain where AES-256 memory encryption protects RAM at runtime. The LoRA training loop runs inside the enclave; weights are returned encrypted with your public key. VoltageGPU operators cannot read training data nor extract resulting weights.

Which base models can I fine-tune?

Qwen3-32B-TEE, Qwen3-235B-A22B-Instruct-2507-TEE, DeepSeek-R1-0528-TEE. LoRA only — base weights remain frozen and shared. Adapter sizes typically 50-500MB depending on rank.

Is this RGPD-compliant?

Yes. VOLTAGE EI is an EU controller (France, SIREN 943 808 824). Native RGPD Article 28 DPA available. Hardware-sealed processing means VoltageGPU is technically incapable of reading user training data — this satisfies the strictest interpretation of Article 32 confidentiality requirements.

Private LLM Fine-Tuning in Intel TDX — Provider-Blind LoRA Training

Sovereign agentic AI: train on your regulated data without anyone seeing it

VoltageGPU runs LoRA fine-tuning inside Intel TDX hardware enclaves. Your training data enters the enclave over TLS 1.3, AES-256 memory encryption protects it at runtime, the LoRA adapter weights come back encrypted with your public key. Even VoltageGPU operators cannot read your data nor extract your trained model. EU company (VOLTAGE EI, France), native RGPD Article 28 DPA, attestation evidence on every job.

Why fine-tune inside a TEE

Train on PII, PHI, financial records, legal documents without leaving sovereign control
Comply with RGPD Article 32, HIPAA Security Rule, DORA Article 5
Eliminate vendor lock-in — base weights stay open (Qwen3, DeepSeek)
Provider-blind: VoltageGPU is technically incapable of reading your data

Workflow (4 steps)

Upload encrypted training set (TLS 1.3 to enclave entry endpoint)
Intel TDX boots, AES-256 keys are CPU-fused, attestation report signed
LoRA training runs inside the trust domain on Qwen3 or DeepSeek base
Encrypted LoRA adapter returned with your public key — only you can decrypt

Compared to Azure OpenAI fine-tuning, AWS SageMaker, GCP Vertex

Hyperscaler fine-tuning runs in standard VMs. Operators can attach a debugger, inspect RAM, or be compelled by US authorities to do so under FISA 702 / CLOUD Act. Intel TDX makes memory inspection cryptographically infeasible — the threat model assumes a malicious hypervisor and still keeps your data sealed.

Related resources

Private Fine-Tuning in TDX

Early access

World's first · Fine-tuning inside Intel TDX hardware enclave

Fine-tune AI on your
most confidential data.

Your training data enters the TDX enclave encrypted. Training runs inside. You receive only the encrypted LoRA weights. Nobody — including VoltageGPU — ever sees your data or the resulting model.

For legal, medical, financial, and HR teams that need custom AI but cannot send sensitive data to any cloud provider.

ARCHITECTURE — PRIVATE FINE-TUNING IN TDX
STEP 1Your training data
Encrypted in transit (TLS 1.3)
STEP 2Intel TDX EnclaveCONFIDENTIAL
AES-256 memory encryption · ECDSA-attested · Provider-blind
STEP 3LoRA training inside enclave
Base model + your data → LoRA adapter weights
STEP 4Encrypted LoRA weights
Returned encrypted with your public key — we never see them
VoltageGPU never sees training data
ECDSA attestation per job
LoRA weights encrypted at rest

WHY THIS IS DIFFERENT

Provider-blind training

Training data never leaves the encrypted enclave — VoltageGPU infrastructure cannot access it.

You own the model

LoRA weights are returned encrypted with your RSA public key. Only you can decrypt them.

Cryptographic proof

Each training job produces an ECDSA attestation report you can independently verify.

Not fine-tuning-as-a-service

Unlike OpenAI, Cohere, or Together — we never store, log, or train on your training data.

Request early access

We're onboarding enterprise teams for a private beta. Response within 48 hours.

ESTIMATED PRICING

Starter$299/run

Up to 1K training examples · Qwen3-32B base · 1 LoRA export

Professional$899/run

Up to 10K examples · Qwen3-235B base · 3 LoRA exports + attestation PDF

EnterpriseCustom

Dedicated TDX cluster · unlimited runs · SLA · compliance reports

* Pricing not final. Based on GPU hours + Targon TDX compute.

USE CASES

LegalClassification

Legal document classifier

Fine-tune on your firm's proprietary contracts and case law without exposing it to any cloud provider.

HealthcareHIPAA

Medical entity extractor

Train a specialized NER model on HIPAA-protected patient records that stays fully within the enclave.

FinanceM&A

Financial risk scorer

Your internal deal data + LoRA training = a proprietary risk model your competitors can't replicate.

HRCompliance

HR policy analyzer

Fine-tune on confidential HR documents to detect policy violations — zero exposure to SaaS providers.

Want to see it running now?

Try the Contract Analyst — same TDX enclave, immediate access.

Open Playground

Fine-tune AI on yourmost confidential data.

Fine-tune AI on your
most confidential data.