Prem AI, led by founder and CEO Simone Giacomelli, said in a post on X that multi-GPU confidential computing is available in production through Fluso, the secure AI workspace Prem AI introduced this month for teams that cannot put sensitive work into ordinary AI tools.
Prem AI on X The announcement matters because it pushes Prem AI's core bet into a higher-stakes part of the stack. Single-model private inference is becoming table stakes for the confidential AI market. Multi-GPU confidential inference is a different sell: it is aimed at heavier workloads, larger open-source models, and agentic workflows that need more compute while still giving customers hardware-backed assurances about where data is processed.
Prem AI has been careful to frame the company around ownership, not just privacy. Giacomelli has argued that "enterprise intelligence must be owned, not rented," according to The Next Web. That is the line connecting Prem AI's infrastructure pages, its May 20 beta launch of Prem Confidential APIs, and its June 12 limited-access launch of Fluso: Prem AI is trying to make sovereign AI usable as a product, not just as a deployment architecture.
Prem AI's public materials describe Fluso as its own product, not as a separate outside partner. The June 12 Fluso launch post says Fluso runs on Prem Confidential API and Compute, and that every session can be cryptographically attested. That makes the cleaner reading of the announcement straightforward: Prem AI is using Fluso as the application layer for its confidential computing stack, rather than selling only raw infrastructure to security teams.
From private inference to private work
Prem AI's pitch is that regulated companies need AI systems that can operate on sensitive prompts, files, model outputs, and workflow context without exposing that material to the model provider, cloud operator, or infrastructure administrator. Its security page says Prem Confidential Compute processes requests inside hardware-isolated trusted execution environments, with client-side encryption, remote attestation, and no operator access to plaintext.
That is the infrastructure story. Fluso is the workflow story.
In its launch post, Prem AI described Fluso as an agentic workspace that connects enterprise tools, builds a knowledge graph called ChronoGraph, and automates work according to an organization's procedures. Prem AI says the target users include hospitals, banks, law firms, governments, wealth managers, founders, operators, and other knowledge workers handling sensitive material. The product starts at $19 per month for Pro, per the launch post.
The strategic move is packaging. A hospital, bank, or law firm may understand why confidential inference is safer than sending sensitive data to a general-purpose API. That does not mean its employees want to interact with enclave infrastructure directly. Fluso gives Prem AI a surface area where the privacy claim shows up as a workspace feature, with attestable sessions and deployment inside a customer's trust boundary.
That is also where the multi-GPU claim becomes important. Agents are compute-hungry because they do not just answer one prompt. They read, search, retrieve, draft, verify, and call tools across steps. If Prem AI can support confidential inference across multiple GPUs in production, Prem AI is trying to remove one of the standard trade-offs in private AI: either keep the data inside the perimeter and accept a weaker user experience, or use a more capable external system and accept exposure to a third-party platform.
What is verified, and what is not
Prem AI's Prem Enclave page says the system installs on existing GPU clusters, activates enclave-grade isolation at the silicon level, and uses a confidential hypervisor to orchestrate CPU and GPU clusters. Prem AI also says its stack supports deployment on-premises, in a customer's cloud account, or in a private cloud setup.
Prem AI's public security materials go further, saying the stack uses CPU technologies such as Intel TDX and AMD SEV-SNP, GPU enclaves on NVIDIA Hopper and Blackwell, hardware-encrypted PCIe and NVLink via NVIDIA Confidential Computing, and cryptographic attestation before data is transmitted. Those are company claims, but they are more specific than the short X post and help explain what Prem AI means by hardware-level security guarantees.
The missing details are still material. Prem AI has not published the exact hardware configuration behind the multi-GPU production claim, the number of production customers using it, revenue from confidential compute, or a named enterprise deployment tied to Fluso. Prem AI's security page also acknowledges boundaries: its proxy can observe metadata such as request timestamps, payload sizes, and API keys, and it explicitly notes theoretical side-channel risk if a TEE is compromised. That does not undercut the confidential computing approach; it puts the claim in the right box. Prem AI is selling verifiable reduction of infrastructure trust, not magic.
That distinction matters in a market where every vendor is trying to replace a legal promise with a technical proof. Tinfoil markets verifiably private AI with confidential computing and says it has production-ready multi-GPU private AI. Germany's Edgeless Systems launched Privatemode AI in 2025 as an end-to-end encrypted AI chat app and API using AMD EPYC CPUs and Nvidia H100 GPUs. Fortanix announced a Confidential AI product in March 2026 for secure enterprise inference using NVIDIA Confidential Computing.
Prem AI's differentiator is not that it is alone in confidential AI. It is that Giacomelli is trying to collapse infrastructure, APIs, and an AI workspace into one buying motion.
The financing context
Prem AI is making that move while trying to raise much larger capital. On June 18, The Next Web reported that Prem AI is seeking a $100 million Series A at a valuation of at least $500 million, expected to close in the third quarter of 2026. The Next Web also reported a prior $6.1 million bridge at a $200 million valuation and said Prem AI has 35 employees.
The confirmed earlier financing is smaller but notable. FinSMEs reported in April 2024 that Prem Labs raised a $14 million seed round, with backers including David Maisel, founder and former chairman of Marvel Studios, Fan Zhang, co-founder of Sequoia Capital China, Alan Lipschitz, founder of Incubeta, and others. FinSMEs identified Giacomelli as CEO and said Prem Labs offered clients ways to train and fine-tune generative AI models with on-premise deployment.
The timing is not accidental. If Prem AI is asking investors to underwrite a $500 million-plus valuation, Prem AI needs to show that private AI is not only a compliance feature. It needs to look like a platform shift: enterprises moving from rented model access toward owned, attestable intelligence running inside their own trust boundaries.
Fluso is the clearest version of that argument. It gives Prem AI a way to sell the outcome - work flowing through private agents - rather than the plumbing. The risk is that enterprise buyers will still demand proof that the plumbing holds under production workloads, with real multi-GPU configurations, audited deployments, and measurable cost or performance advantages.
For Giacomelli, who previously built in decentralized AI and crypto-adjacent infrastructure before Prem AI, the founder throughline is control. The multi-GPU Fluso announcement is an attempt to make that principle operational: private AI that can run serious workloads without asking a bank, law firm, or hospital to hand its most valuable context to someone else's API.