In this briefing
- 01A local project places geography, language and equipment within one deployment boundary
- 02A customer control plane connects access, operations and compliance
- 03Hardware-backed trusted execution extends the boundary to models in use
- 04An air-gapped network makes continued operation a separate acceptance item
- →What to watch next
- ↗Sources and verification
- HPE identifies the servers, graphics processors, network equipment, hosting location and multilingual agents in a South African deployment. HPE does not disclose the system scale, power, price or independent performance results.
- IBM places a customer-operated control plane, storage, networking and observability within one validated architecture. Enterprises can therefore distinguish data that is held locally from a system that the customer actually controls.
- NVIDIA extends the protection boundary to models and data in use, while Teradata treats an environment with no external connection as a separate product form. Partner plans, Public Preview and future capabilities still cannot be treated as completed production delivery.
A local project places geography, language and equipment within one deployment boundary
On 22 October, HPE announced that the South African AI platform Mzansi Mind had selected HPE ProLiant Compute DL380a Gen12 servers. The servers use NVIDIA RTX PRO 6000 Blackwell Server Edition graphics processors. HPE Juniper Networking QFX5120 switches connect the servers, storage and agent compute. The platform is hosted at an Equinix colocation data centre in Johannesburg. The platform supports South Africa's official languages and provides agents for finance, legal work, human resources, customer service and management.
The announcement identifies product models, network components, the hosting city and the application scope, but the announcement does not provide server numbers, memory capacity, power, prices, supply lead times, time to first token, throughput or concurrency data. HPE also does not provide independent test results. The available evidence therefore establishes that Mzansi Mind has selected and deployed specified infrastructure. The available evidence cannot establish that the arrangement has produced a repeatable cost or performance benchmark.
For enterprise adoption, a local deployment needs to state where data is held, which equipment forms the system, how the network connects the components and which languages the business uses. A defined hosting location can narrow the geographical boundary, but one location cannot by itself prove that an enterprise has secured performance, cost, security and operational guarantees. Enterprise acceptance still needs separate checks of capacity, power, service levels, failure recovery and independent performance results.
A customer control plane connects access, operations and compliance
On 22 October, IBM published a validated deployment architecture that combines IBM Sovereign Core software with IBM Fusion infrastructure. IBM says that the architecture supports AI workloads and traditional workloads on premises. IBM also says that Red Hat OpenShift provides a tested and repeatable implementation foundation. The validated configuration covers installation, network compatibility, storage integration and observability.
IBM identifies a customer-operated control plane as an important part of the architecture. The control plane governs access, workloads and compliance processes. IBM also describes some regional sovereign support as an optional service, and IBM states that the service remains subject to availability and scope. IBM does not disclose prices, performance results, hardware quantities or a complete regional list in the announcement, so the validated architecture cannot replace capacity and procurement checks for a particular project.
For enterprise adoption, data residency describes only the location of data, while a customer control plane also describes who can configure the system, inspect operating state and make changes. An enterprise can define control-plane ownership, administrator identities, log retention, patch responsibility and regional support as separate acceptance items. A supplier-validated combination can reduce integration uncertainty, but supplier validation cannot by itself prove that every customer environment meets local regulatory or internal-control requirements.
Hardware-backed trusted execution extends the boundary to models in use
On 21 October, NVIDIA announced the Confidential Bare Metal deployment option. NVIDIA says that the arrangement can protect data in transit, measure platform integrity and enforce hardware-level access restrictions. NVIDIA also says that enterprises can use the protection in bare-metal, container and virtual-machine environments. NVIDIA has extended support to Grace Blackwell NVL72 and HGX and DGX Rubin NVL8 systems based on x86, and NVIDIA has incorporated the capabilities into DSX OS.
NVIDIA's announcement also lists partnerships at different stages of maturity. Cohere plans to integrate the capabilities, while Google Cloud is extending support and other partners have expressed an intention to adopt them. These statements establish that hardware-backed trusted execution is entering product roadmaps for model platforms and cloud services, but the statements cannot establish that every partner has completed integration or opened the capability to every customer. The announcement does not provide independent performance-overhead tests, prices or power data.
For enterprise adoption, the country or data centre that contains a system cannot by itself establish whether a model is protected while it runs. Enterprises need separate checks of boot integrity, key control, administrator privileges, remote attestation, virtualisation boundaries and performance overhead. Hardware-level restrictions can supplement software permissions, but enterprises still need configuration records and repeatable tests to prove that the relevant controls operate on the target hardware, target runtime and target service model.
An air-gapped network makes continued operation a separate acceptance item
On 20 October, Teradata announced two sovereign deployment forms. Teradata Factory with AI Studio places data, vectors, models and agents in an on-premises environment. Teradata Sovereign Cloud on Google Distributed Cloud air-gapped addresses an air-gapped environment that has no external connection and is operated by the customer. Teradata separates sovereignty into data sovereignty, operational sovereignty and digital sovereignty. This separation directly concerns the control plane and technical dependencies that sit beyond geographical location.
Teradata distinguishes the state of the two capabilities. Teradata describes the expanded Factory capabilities as available in the first quarter of 2027. Teradata describes Sovereign Cloud on Google Distributed Cloud air-gapped as Public Preview. The announcement does not provide prices, actual throughput, recovery times or a complete regional scope. The available evidence therefore establishes that the supplier is designing products for operation without a network connection, but the evidence cannot establish that a preview product has completed every production-environment validation.
For enterprise adoption, a fully isolated environment turns model updates, vulnerability remediation, log export, licence validation and knowledge-base synchronisation into separate operating processes. Enterprises need to verify that a system continues to operate without a network connection. Enterprises also need to verify how controlled media carry updates and audit material into or out of the environment. Public Preview can support a bounded evaluation, but Public Preview cannot replace formal supply, service levels or recovery exercises.
What to watch next
- Will HPE and Mzansi Mind publish the system scale, power, prices, supply lead times and independent performance results for different language and agent workloads?
- Will IBM and NVIDIA provide customer-repeatable control-plane audits, platform-integrity evidence and performance-overhead tests?
- Will Teradata publish the general-availability date, regions, upgrade process, offline maintenance method and service levels after Public Preview?
Sources and verification
Golden Data has edited this briefing from the public materials listed above. The original sources govern facts and figures. The enterprise relevance sections are Golden Data editorial analysis and do not constitute an endorsement of any third-party product.
← Back to AI Daily Briefing