Private ChatGPT,
Model in Intel TDX, French Company
Same $20/mo as ChatGPT Plus. French controller, model in an Intel TDX enclave. No training on your data. Our relay forwards your prompts to it in clear.
For founders, lawyers, doctors, journalists, researchers, and anyone who refuses to paste sensitive prompts into a US chatbot that trains on them by default.
01, THE PRIVACY PROBLEM
ChatGPT Plus is convenient. It is not private.
Most users assume that paying $20/month for ChatGPT Plus buys them privacy. It does not. The consumer plan defaults to using your conversations for model training, runs on US infrastructure under FISA 702 jurisdiction, and provides no hardware confidentiality. Microsoft, as the underlying compute provider, has commercial access to the inference layer. None of this is illegal, hidden, or surprising, it is documented in OpenAI's terms, but it makes ChatGPT Plus a poor fit for any conversation that touches client data, medical information, financial details, or confidential business strategy.
Default training on your prompts
OpenAI uses consumer ChatGPT conversations to improve models unless you manually opt out. Most users never check the toggle.
US jurisdiction and FISA 702
OpenAI servers operate primarily in the US. Section 702 of the Foreign Intelligence Surveillance Act allows compelled access to non-US persons' data.
No hardware confidentiality
OpenAI runs on Microsoft Azure VMs. Operators with infrastructure-level access can technically inspect inference memory. There is no Intel TDX enclave separating you from them.
No native RGPD Article 28 DPA on consumer plans
A consumer ChatGPT Plus subscription is not a B2B contract. There is no DPA signed by OpenAI as your processor, that is reserved for ChatGPT Enterprise.
02, HOW VOLTAGEGPU PLUS IS DIFFERENT
Model in Intel TDX, French controller, no training defaults.
VoltageGPU Plus sends every conversation to a model running inside an Intel TDX enclave at our US inference sub-processor (SCCs). The CPU encrypts that memory with keys it holds, so the host cannot read it. Your prompts reach our servers over TLS and are forwarded to the model in clear. We are a French company (VOLTAGE EI, SIREN 943 808 824); our web host and our inference provider are US companies, listed in our sub-processor register. RGPD Article 28 Data Processing Agreement is native on every plan. We do not train on your prompts on any plan, ever.
Model in Intel TDX
The model runs in an Intel TDX enclave at our inference provider: the CPU encrypts its memory and that host cannot read it. Our relay handles your prompts in clear while forwarding them.
Attestation of the model instances
GET /v1/attestation returns the TDX quotes of the instances serving a model, on your own nonce, verifiable against Intel. It does not prove that a given reply came from them.
French controller, listed sub-processors
VOLTAGE EI, French company, SIREN 943 808 824, is your counterparty and the customer database is EU-hosted. Inference runs inside Intel TDX trust domains at a US inference provider named in the DPA.
Article 28 DPA on every plan
RGPD-compliant Data Processing Agreement is the default contract, not an enterprise upgrade.
No training on your data
Prompts and replies go through our relay to the model in Intel TDX; they are never used for training.
256K context on flagship
Qwen3.5-397B-TEE handles entire books, repositories, or contract bundles in one conversation.
03, COMPARISON
VoltageGPU Plus vs ChatGPT Plus, Microsoft Copilot, Mistral Le Chat
| FEATURE | VoltageGPU Plus | ChatGPT Plus | Copilot | Mistral Le Chat |
|---|---|---|---|---|
| Monthly price (consumer plan) | $20/mo (Plus) | $20/mo | $30/mo | €14.99/mo |
| Default training on user prompts | Never | Yes, opt-out required | No (Microsoft 365) | No |
| Model memory encrypted by the CPU (TDX / SEV-SNP) | Yes, Intel TDX at our inference provider | No | No | No |
| Primary jurisdiction | France (EU) | United States | United States | France (EU) |
| US companies in the data path | Yes: our web host and inference provider (listed) | Yes | Yes | No |
| Native RGPD Article 28 DPA | Native, no negotiation | Enterprise only | Available (Microsoft DPA) | Available |
| Maximum context window | 256K (Qwen3.5-397B-TEE) | 128K (GPT-4o) | 128K | 128K |
| Attestation of the model instances | Yes, on your nonce (GET /v1/attestation) | No | No | No |
Sources: OpenAI Terms of Use, Microsoft 365 Copilot Privacy and Compliance, Mistral Le Chat documentation. Comparisons reflect default consumer plan behavior at time of writing.
04, MODELS AVAILABLE INSIDE THE ENCLAVE
Two models, each in an Intel TDX enclave.
Qwen3.5-397B-A17B-TEE
256K context
Long-document conversations, code, analysis.
DeepSeek-V3.2-TEE
163K context
Step-by-step reasoning for math, code, planning. Comparable to OpenAI o1. Available on Pro.
05, PRICING
Same $20/mo as ChatGPT Plus. Or free for 50 messages.
PLUS
2,000 messages on Qwen3.5-397B-TEE (256K context). Model in Intel TDX. Native RGPD DPA.
Get PlusPRO
5,000 requests a month on Qwen3.5-397B-TEE plus DeepSeek-V3.2-TEE reasoning, 10 seats. Priority support.
Talk to sales06, FAQ
Frequently asked questions
Is VoltageGPU Plus a real ChatGPT alternative or a thin wrapper?
It is a complete alternative. We do not call OpenAI. Inference runs on open-weight Qwen and DeepSeek models inside Intel TDX enclaves at our inference provider. Same $20/month entry price as ChatGPT Plus.
Does ChatGPT Plus train on my conversations by default?
On the consumer plan, OpenAI uses your prompts for model training unless you manually disable it under Settings > Data Controls > Improve the model for everyone. Most users never touch that setting. VoltageGPU has no such toggle because we never train on user prompts on any plan.
What does Intel TDX actually protect against?
TDX uses keys held by the CPU to encrypt the memory of a virtual machine (a trust domain) at runtime, so the host running the model, hypervisor included, cannot read it. It does not cover our relay: your prompt reaches our servers over TLS and is forwarded to the enclave in clear.
Can I verify that my chat actually ran inside a TDX enclave?
Partly. GET /v1/attestation returns the Intel TDX quotes and NVIDIA GPU evidence of the instances serving a model, bound to a nonce you choose, and you verify them yourself (docs.voltagegpu.com/inference). It proves those instances are TDX guests, not that a given chat reply came from one of them; chat replies carry no attestation.
What about latency? TEE compute used to be slow.
We have not published our own latency or overhead measurement. A third-party benchmark of NVIDIA Hopper confidential mode reports an overhead below 7% for most typical LLM queries (arXiv 2409.03992).
How is this different from Mistral Le Chat?
Le Chat is also EU-based and does not train on prompts, which is great. The difference is where the model runs: here, inside an Intel TDX enclave at our inference provider, whose memory that host cannot read. Our relay still handles your prompts in clear while forwarding them.
Try Private ChatGPT free
50 messages per month, model in Intel TDX. No credit card.