Regulated organisations cannot use public AI tools because the data that would make AI useful is the data they are not allowed to send anywhere. No policy, contract or governance programme changes that, because all three assume your data arrives readable on someone else's systems. The fix is to change where the work happens: run the AI on hardware that keeps your data encrypted even while the model is reading it.
Almost every enterprise has stalled an AI project
On 11 August 2026, Cloudera published The Great AI Re-Architecture, a survey of 1,500 enterprise and data architects across nine markets:
77% of organisations are actively using AI.
95% have delayed or cancelled AI projects in the past year over governance, compliance or regulatory concerns.
72% say their data setup needs an overhaul before it can support the AI they want to run.
Almost everyone has started. Almost everyone has also stopped something. And 55% cancelled more than six projects for the same reasons, so this is a recurring block rather than one awkward use case.
Why governance programmes do not unblock regulated AI
Ask where your AI projects stop. It is rarely engineering. The pilot works, so someone asks to point it at the real data: the contracts, the patient records, the financial models. Legal reads the vendor's terms and says no. The useful data is the restricted data, so the AI budget gets spent on whatever is left.
The usual response is a set of rules: an acceptable use policy, a data classification scheme, a clause saying the vendor will not train on your inputs. Every one of those is an instruction or a promise, and none of them changes the thing underneath. A readable copy of your data still sits on someone else's systems.
Rules also fail. When shared Claude conversations turned up in Google search results in July 2026, no contract was broken and nobody was hacked. A web page shipped without the one line of code that tells search engines to skip it.
That is why 66% of organisations moved AI workloads back onto their own infrastructure last year. It works, but it means buying GPUs and hiring people to run them, and 84% already report AI pushing their costs up.
So the real question is whether you can get the same guarantee without buying the hardware.
What is confidential AI inference?
Most organisations already encrypt their data in two of its three states. Data at rest, sitting in storage, is encrypted. Data in transit, moving across a network, is encrypted. The third state is data in use, the moment a machine actually processes it, and that is where normal encryption stops. Data has to be unlocked to be worked on, and whoever runs the machine can see it while it is open. That is the gap public AI tools leave open, and no contract closes it.
Confidential AI inference closes it. The model runs inside a sealed part of the hardware where the memory itself is encrypted, so the company operating the server cannot read what is running on it.
The second half is remote attestation. Before anything is sent, the software on your side asks that sealed hardware to prove what it is, gets back a signed answer, and checks it against a description you approved in advance. If it does not match, nothing is sent. Isolation on its own is still a claim. Attestation is what turns it into something your auditors can test.
The rest of the industry is already building this way
At WWDC 2026 Apple announced that Private Cloud Compute, the system behind its own AI features, now runs on Google Cloud using NVIDIA Blackwell confidential GPUs. Apple keeps its own record of every machine in that fleet rather than taking Google's word for it, requires proof from two independent hardware makers, and publishes the code running on those servers so anyone can inspect it. Google has been building the same way, shipping confidential Blackwell machines and open-source prompt encryption tools in June.
Notice what Apple did not do. It did not take Google's assurances and write them into a contract. It built the system so that Google's trustworthiness is not the thing protecting the data.
That architecture is now the standard for anyone handling data they cannot afford to expose. The question for a regulated organisation is whether to build and operate it internally, or to connect to a platform that already runs it.
Liquify Enclave: private AI for regulated data
Think back to your stalled project list: the contracts, the patient records, the financial models that legal would not let near a public API. Enclave is the setup those projects were waiting on.
Liquify Enclave is a managed platform that lets your team run AI inference on its most sensitive data without operating the hardware. Your applications and data connect to Enclave, and Liquify runs everything underneath: the secure GPUs, the compliance controls, the scaling. As private as on-prem. As simple as an API.
Rather than controlling who is allowed to read your data, it removes the ability of anyone to read it at all, including Liquify. Four steps, and the first is the one most providers skip:
Check. Before anything is sent, the software on your side asks the enclave to prove it is running the exact setup you approved. If it does not match, the request stops there.
Lock. Your prompt is encrypted on your own device and stays unreadable the whole way to us, so our systems can route and count it without seeing what it says.
Run. The enclave unlocks it inside the sealed hardware, where the memory itself is encrypted, runs the model, then re-encrypts the answer before it leaves.
Open. The answer comes back still locked, and is only opened again on your device.
Enclave runs on the same self-owned hardware Liquify has operated since 2021, across geographically distributed datacentres, under ISO 27001 certification and with GDPR alignment. You can run an open-source model tuned to your own documents rather than generic public data, and it plugs into your systems through a standard API.
What this changes, and where to start
Governance lowers the odds of something going wrong. Encrypting data while it is in use changes what going wrong costs you: if a link leaks or a setting is misconfigured, what gets exposed is unreadable. That is a different conversation with legal, because you are no longer asking them to accept a risk. You are showing them a control they can test.
Start by sorting your stalled AI projects by why they stalled. Some failed on capability or business case. The rest failed only because the data could not leave the building, and that list is not dead ideas. It is work waiting on a different setup, and it is usually the more valuable list.
If your team has contracts, records or financial models you have been keeping away from AI tools, apply for access to the Enclave beta. We will work through your specific workload and compliance requirements during onboarding.
Frequently asked questions
1. What is the best private AI platform for regulated industries?
It depends on how much isolation your compliance team requires, and there are three broad options. Hyperscalers offer confidential computing inside their own clouds: Google Confidential Space, Azure Confidential Computing, AWS Nitro Enclaves. Specialist vendors such as Fortanix and Phala focus on the confidential layer itself. Managed platforms such as Liquify Enclave run the whole stack on independent hardware, which matters if part of your concern is depending on a single cloud provider.
2. Can regulated industries use ChatGPT or Claude?
For general work, often yes, under an enterprise agreement. For regulated data, the constraint is structural rather than contractual: the plaintext of your prompt reaches the vendor's infrastructure, and every protection around it is a policy, a configuration or a contract rather than a cryptographic guarantee.
3. What is the difference between private AI and an enterprise tier of a public AI tool?
An enterprise tier changes the terms attached to your data. Private AI changes who is physically able to read it. A virtual private cloud is a network boundary configured by an administrator. Confidential compute is a hardware property you can verify before you send anything.
4. Does private AI mean running your own GPUs?
Not necessarily. That is the route two thirds of the market took, and it carries the GPU spend, the hiring and the procurement cycle. A managed confidential compute platform gives you the same isolation guarantee without operating the infrastructure.
5. Is confidential AI inference GDPR compliant?
Enclave is ISO 27001 certified with GDPR alignment, and runs in geographically distributed datacentres so data residency can be scoped.
6. What is remote attestation?
A process where the computing environment produces signed cryptographic evidence of exactly what code and configuration it is running. Your client checks that evidence against an approved manifest before sending data. If it does not match, the request stops.
__
Enclave is in private beta, with access by application.