Edgeless Systems
Edgeless Systems runs Llama 70B in encrypted cloud for regulated firms, reports 50% savings
Edgeless Systems, a German security software company, built Privatemode AI, a cloud service that runs Meta's open-weight Llama 70B so that even the operator cannot read customer prompts. In a Meta case study the company reports a 6 months faster launch and 50% savings compared with on-premises AI projects. All results are self-reported.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
6
months
Faster time to launch versus on-premises AI projects (self-reported, Edgeless Systems via Meta case study)1
50%
Savings compared with on-premises AI deployments (self-reported, Edgeless Systems via Meta case study)1
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Companies in heavily regulated sectors such as healthcare and banking were largely shut out of cloud AI tools. Those tools did not meet their strict rules on protecting sensitive data. The usual alternative was to run AI on the company's own servers, which is slow and expensive.
What they built
Edgeless Systems, a German company that makes security software, built Privatemode AI. Edgeless describes it as the first end-to-end cloud service for large language models that uses confidential computing. That is a technique where data stays encrypted even while a chip processes it. The service offers a chat app and an API that other software can call. Edgeless's press release says it launched in February 2025.
The service runs Meta's Llama 3.1 70B, later upgraded to Llama 3.3 70B. Llama is an open-weight model, meaning its files are published so a company can run it on its own platform. Edgeless chose it because it is open, multilingual and can be self-hosted.
How they did it
Edgeless used a compressed (quantized) version of the model so that it fits on a single NVIDIA H100 graphics chip, which confidential computing required. The service is hosted on bare-metal servers from Scaleway.
A small program on the customer's side encrypts each prompt and decrypts each answer locally. On the server, the model runs inside a confidential container, and the service checks the integrity of each step. The aim is that prompts are never exposed in plain text to the cloud provider or to Edgeless.
Results
Edgeless reports, through Meta's case study, that companies can launch secure AI assistants and chatbots 6 months faster than with on-premises AI projects. It also reports 50% savings against on-premises deployments and says thousands of employees have been onboarded ("1,000s" in the original).
Meta labels these results as self-reported and not identifiably repeatable. The sources name no customers.
What to copy
Treat data protection as the design brief. If your lawyers or regulators rule out standard cloud AI, check whether a confidential-computing service can meet the rules before you commit to building your own infrastructure.
Lessons for your project
If data sensitivity blocks cloud AI, test whether confidential computing (hardware that keeps data encrypted even while it is processed) removes the blocker before paying for an on-premises build.
Pick a model size to fit the hardware that your security design requires. Edgeless chose a compressed 70B Llama so it ran on a single GPU.
Plan model upgrades from the start. Edgeless moved from Llama 3.1 to Llama 3.3 on the same service.
Ask what baseline a savings claim compares against. Here it is generic on-premises AI projects, not a measured cost of any one customer.
Caveats
- The only source for the figures is a Meta case study built on Edgeless Systems' own statements. It says the results are self-reported and not identifiably repeatable. No independent source confirms them.
- The 6 months and 50% figures compare Privatemode AI with on-premises AI projects in general. The case study does not say how they were measured, nor which customers were involved.
- The case study is undated and names no customers. The deployed date shown is the launch date in Edgeless Systems' press release, which says Llama 3.3 was available at launch. The case study says an earlier Llama 3.1 version came first, but gives no date for it.
- No cost, team size or build timeline for Privatemode AI is published in these sources.
Sources
- LLM service uses confidential computing to deliver secure cloud AI applications
Meta (Llama case studies) · Undated · Vendor case study · llama.com
- Launch of Privatemode AI
Edgeless Systems · 2025-02-19 · Press release · edgeless.systems
Summarized in our own words from the sources above. Spotted an error? See corrections.