Tenstorrent and ConfidentialMind Partner to Deliver Sovereign AI on Open Hardware
Galaxy Blackhole validated for ConfidentialMind's on-premises AI stack at $160,000 per server

Tenstorrent and ConfidentialMind announced a joint partnership on Monday to sell a validated, bundled sovereign AI stack that runs entirely within a buyer's own infrastructure — no cloud dependency, no proprietary GPU architecture, and no data leaving the customer's physical environment. The combination pairs Tenstorrent's Galaxy Blackhole server, a 32-chip RISC-V AI system that reached general availability in April 2026, with ConfidentialMind's self-hosted platform covering LLM inference, vector database retrieval, role-based access control, and observability. The joint stack is targeted at governments, universities, defense agencies, and regulated enterprises across Europe, the Middle East, Asia-Pacific, and the Americas that face binding legal obligations about where AI processing can occur.
What Galaxy Blackhole Puts on the Floor — and Why RISC-V Changes the Procurement Equation
The Galaxy Blackhole server is a 6U air-cooled chassis housing 32 Blackhole ASICs. Each chip pairs programmable RISC-V control cores with dedicated Tensix tensor-math tiles, giving the server an aggregate 23 PFLOPS at FP8 precision, 1 TB of GDDR6 memory running at 16 TB/s bandwidth, and 6.2 GB of on-chip SRAM operating at 2.9 PB/s. Scale-out is handled over standard 400G Ethernet, not a proprietary interconnect like NVLink, meaning buyers can expand capacity without committing to a vendor-locked fabric. Four systems assembled into a supercluster start at $640,000.
The RISC-V instruction-set architecture matters beyond benchmark tables. RISC-V is an open, royalty-free ISA maintained by RISC-V International, meaning the processor design is auditable by the customer — not possible with NVIDIA's closed CUDA-compatible architectures or AMD's x86-based accelerators. Tenstorrent has also open-sourced TT-Metalium, its kernel-level programming library, under the Apache 2.0 license. For sovereign buyers with procurement rules against proprietary instruction sets or vendor-controlled software stacks — a category that includes some governments subject to US GPU export controls — a RISC-V-based option with a fully open-source software stack is structurally differentiated, not just philosophically.
ConfidentialMind Supplies the Software Layer That Self-Hosting Actually Requires
Deploying a large language model on private hardware is straightforward. Operating one at enterprise scale — with document retrieval, audit logging, authentication, and cost visibility — is not. ConfidentialMind's platform addresses the gap with four integrated components: an OpenAI-compatible inference server for running open-weight models behind a standard API, a built-in vector database enabling retrieval-augmented generation without a third-party service, role-based access control with authentication, and an observability layer for logging and monitoring inference activity. A Streamlit-based playground surfaces document analysis and writing assistants directly to end users. The platform deploys on-premises, in a private cloud, in a customer-managed VPC, or at the edge, and previously carried validation on AMD GPU infrastructure through Vultr and Swedish cloud provider Elastx. The Tenstorrent validation extends that to RISC-V hardware. Among ConfidentialMind's existing European deployments are the University of Helsinki, LUT University, and Centria University of Applied Sciences.
EU AI Act Enforcement Created the Procurement Urgency
The timing is not arbitrary. On August 2, 2026, the EU AI Act's transparency obligations under Article 50 became enforceable, requiring AI providers and deployers to disclose when users are interacting with AI systems and to label AI-generated content — with fines of up to €15 million or 3 percent of global annual turnover for violations. The EU AI Office's full investigative and enforcement powers for general-purpose AI models also activated on that date. Separate obligations for high-risk AI systems under Annex III are on a later timeline, with standalone requirements taking effect in December 2027. GDPR obligations have long restricted personal data from leaving the European Economic Area in ways that cloud AI APIs frequently violate in practice — US cloud providers hold roughly 70 percent of the European cloud market, and infrastructure operated under US jurisdiction remains subject to US CLOUD Act requests regardless of where the servers sit physically. Self-hosted AI resolves those obligations at the network layer: data processed on the customer's own hardware never traverses a foreign vendor's infrastructure.
Tenstorrent has been building a sovereign AI customer base through most of 2026. Japanese operator ai& deployed more than 120 Galaxy systems as the foundation of Japan's largest sovereign AI compute environment in June 2026. Abu Dhabi-based Infinia Technologies signed a partnership in January 2026 to deploy Galaxy hardware for GCC government workloads. The Republic of Cyprus, UAE enterprise AI provider AIREV, and drug discovery platform JapanFold have all followed. The ConfidentialMind partnership adds a validated software stack to what has until now been primarily a hardware and infrastructure channel — giving buyers a single procurement path rather than a hardware-and-separately-sourced-software integration project.
Software Ecosystem Maturity Remains the Meaningful Limitation
The honest caveat is software. Tenstorrent claims that roughly 90 percent of models on HuggingFace run on Galaxy hardware without modification, and published company benchmarks from the TT-Deploy event in May 2026 show DeepSeek-R1 671B achieving more than 350 tokens per second per user across a 16-system cluster. Those figures have not been independently replicated by third parties. TT-Metalium, while open source and genuinely capable, is a significantly younger ecosystem than NVIDIA's CUDA stack — practitioners evaluating rapid deployment timelines for uncommon model architectures will encounter more friction than they would on GPU hardware with a decade of library support behind it. ConfidentialMind's prior validations on AMD infrastructure give it meaningful experience porting across hardware, which reduces but does not eliminate that integration risk. No customer deployments of the joint stack have been publicly disclosed yet.
The partnership also does not eliminate the security perimeter question entirely. A self-hosted inference environment still presents an application-layer attack surface — a remote code execution vulnerability in an inference server used by a separate operator was disclosed the week before this announcement. Physical control of the hardware is a necessary but not sufficient condition for security.
The next benchmark for the stack's traction will be the first named enterprise or government customer deployments, expected in the coming months. For procurement teams operating under data-residency mandates, the more consequential near-term question is whether TT-Metalium's model compatibility claims hold under production workloads — and whether independent benchmark replication supports the inference throughput numbers that make the $160,000-per-server economics compelling against NVIDIA alternatives. Tenstorrent closed a $1.5 billion Series E at a $6.2 billion post-money valuation in October 2026, giving it substantial capital to pursue commercial expansion.



