# What is private AI inference?

**Private AI inference means running an AI model so that nobody but you can read your prompts and the answers, not the cloud provider, not the model provider and not the company serving the model. The model runs inside a trusted execution environment (TEE), a sealed area of the CPU and GPU that keeps data encrypted from the machine's own operator. The hardware signs a report of what is running, so you can verify the privacy instead of trusting a policy.**

NEAR AI Cloud is a private AI inference API: open-weight models run in TEEs behind an OpenAI-compatible API, and every deployment can be verified with a hardware attestation.

## Key points

- A typical AI API encrypts data in transit and at rest, but the provider can read it while the model runs. Private inference closes that gap.
- The privacy comes from hardware: Intel TDX confidential virtual machines and NVIDIA GPUs in confidential computing mode.
- Attestation reports are signed by keys rooted in Intel and NVIDIA, not in the provider, so you can check them yourself.
- Private inference runs open-weight models. It works with the OpenAI SDK by changing the base URL and key.
- TEEs make attacks far harder, not impossible. Only models that run in a TEE are private.

## Is ChatGPT private? Is it encrypted?

ChatGPT encrypts your conversations in transit and at rest ([OpenAI](https://openai.com/security-and-privacy/)). That protects them from outsiders. It does not protect them from OpenAI: the service holds the keys and processes your conversations in readable form. It is not end-to-end encrypted.

On the Free, Plus and Pro plans, OpenAI says it may use your content to train its models unless you turn off "Improve the model for everyone" ([OpenAI Help Center](https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance)).

## How private inference works

1. **Your request is encrypted on your machine.** You call the API over HTTPS, as with any OpenAI-compatible API. The encrypted connection ends inside the gateway's own enclave, not at a load balancer the provider's staff can read.
2. **The model runs in sealed hardware.** The gateway forwards the request to a model running in an Intel TDX confidential virtual machine with an NVIDIA GPU in confidential computing mode. Memory stays encrypted, so the host, the cloud provider and NEAR AI cannot read the prompt or the answer.
3. **The hardware signs what is running.** Each enclave produces an attestation report: a statement of the exact hardware and code it runs, signed by keys rooted in Intel and NVIDIA, not in NEAR AI. Responses can carry a signature from a key that never leaves the enclave.
4. **You check the proof.** You verify the report against Intel's and NVIDIA's attestation services, compare the code measurement with the published source, and check the response signature.

## Private inference vs a typical AI API vs self-hosting

- **Who can read your prompts.** Typical API: the provider. Private inference: nobody but you, because the operator cannot see into the enclave. Self-hosted: you and whoever runs your servers.
- **What you rely on.** Typical API: a privacy policy and a contract. Private inference: hardware isolation you can verify. Self-hosted: your own security.
- **Setup.** Typical API and private inference: an API key. Self-hosted: GPUs, model serving and upkeep.
- **Models.** Typical API: closed and open models. Private inference and self-hosted: open-weight models.

Keeping data in one cloud region is not the same thing. The region decides where your data sits. It does not decide who can read it.

## How to verify it

- **In the browser.** The [Verification Center](https://cloud.near.ai/verify) shows a model's hardware attestation and checks a reply's signature, with no code.
- **In your code.** Request a fresh attestation report with your own random nonce, then check it with the [verification guides](https://docs.near.ai/cloud/verification) and the open-source verifier.
- **From the source.** The enclave software, the [Private ML SDK](https://github.com/nearai/private-ml-sdk), is open source. The attestation includes a measurement of the exact code running.

## What private inference does not protect

- **TEEs are not magic.** Researchers have published attacks on TEEs before. TEEs make reading your data far harder and more costly. They do not make attacks impossible.
- **Only models in a TEE are private.** On NEAR AI Cloud, models marked Private TEE or Private 3P TEE run in enclaves. Models marked Incognito run at their provider: NEAR AI hides your account, but there is no hardware guarantee.
- **Your own systems still matter.** What you log, store or send elsewhere is up to you.

## Who uses it

- **Code assistants**, whose prompts carry proprietary source code.
- **Regulated teams** in health, legal and finance that handle client data.
- **Security teams** that analyze exploits with [uncensored models](https://near.ai/uncensored-ai-models) without exposing findings.
- **Products built on AI.** [Brave Leo runs on private inference from NEAR AI Cloud](https://near.ai/blog/near-ai-cloud-brings-private-inference-to-brave-leo).

## Use it through the NEAR AI Cloud API

Create an account, create an API key, and call the OpenAI-compatible endpoint at `https://cloud-api.near.ai/v1`. The current private models, their privacy tier and prices are listed at [near.ai/models.md](https://near.ai/models.md). Get a key at [cloud.near.ai](https://cloud.near.ai).

## FAQ

### What is private AI?

Private AI is AI that does not expose your data to anyone else while it works. It means running a model on your own hardware, or using a private inference API that runs the model inside sealed hardware (a TEE) and lets you verify that nobody else can read your prompts.

### Is ChatGPT encrypted?

In transit and at rest, yes. OpenAI says your content is encrypted between you and OpenAI and when it is stored. It is not end-to-end encrypted: OpenAI holds the keys, so its systems process your conversations in readable form.

### Is ChatGPT private?

Not in the sense of private inference. Your conversations are readable on OpenAI's systems. On the Free, Plus and Pro plans, OpenAI says it may use your content to train its models unless you turn off "Improve the model for everyone". Business plans are not used for training by default.

### What is the difference between private AI and ChatGPT?

With ChatGPT you trust OpenAI's policies about who can read your conversations. With private AI inference, the model runs in hardware the provider cannot see into, and you can verify it with a hardware attestation. Private inference runs open-weight models, so the models differ too.

### What is a trusted execution environment (TEE)?

A TEE is an area of a processor that keeps code and data isolated and encrypted from the rest of the machine, including its operating system and operator. NEAR AI Cloud uses Intel TDX on the CPU and NVIDIA confidential computing on the GPU.

### What is confidential AI?

Confidential AI is AI that runs on confidential computing hardware, so data stays encrypted while it is being processed, not only in transit and at rest. Private AI inference is confidential AI applied to running models.

### How do I verify that inference is private?

Ask the service for an attestation report, check its signatures against Intel's and NVIDIA's attestation services, and compare its code measurement with the published source. NEAR AI Cloud has a browser Verification Center and open-source verifiers that do these checks.

### Is private AI inference slower?

Slightly. Confidential computing adds encryption work. NVIDIA's confidential computing mode runs on the GPU itself, and NEAR AI's documentation describes the latency overhead as minimal. Test your own workload to measure the difference.

### Can I use private AI inference with the OpenAI SDK?

Yes. NEAR AI Cloud has an OpenAI-compatible API, so you change the base URL and API key and keep your code.

### Are all NEAR AI Cloud models private?

No. Models marked Private TEE run in NEAR AI's own enclaves, and models marked Private 3P TEE run in a third party's attested enclave. Other models are marked Incognito: NEAR AI hides your account from the provider, but there is no hardware privacy guarantee. Each model page shows its tier.

## Sources

- NEAR AI, [Private inference documentation](https://docs.near.ai/cloud/private-inference)
- NEAR AI, [Verification documentation](https://docs.near.ai/cloud/verification)
- Intel, [Trust Domain Extensions](https://www.intel.com/content/www/us/en/developer/tools/trust-domain-extensions/overview.html)
- NVIDIA, [Confidential computing on H100 GPUs](https://developer.nvidia.com/blog/confidential-computing-on-h100-gpus-for-secure-and-trustworthy-ai/)
- OpenAI, [Security and privacy](https://openai.com/security-and-privacy/), checked November 20, 2026
- OpenAI, [How your data is used to improve model performance](https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance), checked November 20, 2026

By Alejo Escriva, reviewed by Sergey Astretsov. Updated November 20, 2026.

Canonical page: https://near.ai/private-ai-inference
