Skip to content
AI NEWS JOURNAL
SiliconANGLE AIAI NEWS / JOURNAL

Sovereign AI: Where control becomes an architectural requirement

We believe sovereignty is becoming an architectural requirement for firms operating globally, especially in highly regulated industries. The business question for buyers is how much control an organization needs over the systems that produce its intelligence, and what happens when that control is ceded to others. U.S. companies supply most of the cloud and frontier […] The post Sovereign AI: Where control becomes an architectural requirement appeared first on SiliconANGLE.

UPDATED 12:50 EST / NOVEMBER 21 2026 Sovereign AI: Where control becomes an architectural requirement BREAKING ANALYSIS by David Vellante, Amit Eyal Govrin and John Furrier We believe sovereignty is becoming an architectural requirement for firms operating globally, especially in highly regulated industries. The business question for buyers is how much control an organization needs over the systems that produce its intelligence, and what happens when that control is ceded to others. U.S. companies supply most of the cloud and frontier artificial intelligence technology enterprises depend on. Buyers need to understand the resulting jurisdictional, operating and commercial risks, and put in place compensating controls to mitigate them. This reality applies for U.S. firms operating abroad and even those in the US, especially for regulated firms. Welcome to Breaking Analysis #334: Sovereign AI: Where control becomes an architectural requirement. Access the full slide deck In this episode we examine the issues around where data lives, who holds the keys, which model touches the data, who operates the system and who is exposed to the bill. Our premise is that sovereignty requires deliberate risk decisions, compensating controls and proof that those controls work. The appropriate architecture involves tradeoffs and will depend on the workload and the consequences of losing control. Today we draw on three sources of input as shown below: First, Amit Govrin’s buyer’s guide, which evaluates the documented controls across hyperscaler offerings and buyer territories. Second, a study we commissioned with our partners at Qualitate that brings the experiences of eleven senior technology decision-makers in regulated and critical-infrastructure organizations. Third, John Furrier’s conversation with Amit at SC26, which explains the reasoning behind our framework. The buyer’s guide independently measures vendors’ documented claims. The in-depth interviews show what buyers care about and what they’ve tested. theCUBE interview helps connect those findings to architecture and business risk. The survey sample is small but deep, so we use it to understand decisions and expose gaps. Why a sovereign AI framework is needed Let’s start by hearing from Amit with why he developed a practical and uniform framework to measure organizational control over AI, data and local risk: In-depth interviews reveal gaps in sovereignty posture and incumbent advantages The first set of takeaways from our survey shows a gap between buyer priorities and operational reality. Legal exposures featured prominently in purchasing decisions within the sample. Yet seven of eleven respondents report no testing of disconnected systems, key-revocation or exit scenarios. So the hard evidence for practical operational control is somewhat shaky given the high levels of concern about losing that very control. As shown below, buyers also tend to add controls to the platforms they already use. Only four of eleven have deployed or piloted a distinct sovereign or dedicated offering. Cost and the burden of adding another environment was the primary reason that explained such behavior. As seen in the quote above, a European information technology and telecom buyer says deep support escalations “can still in some cases involve personnel or systems outside the EU, particularly for deep technical issues.” What’s the message there? Our take is that the support protocol must become a fundamental part of the architecture review; and that has to account for emergency and privileged access. The key point as Amit often stresses is sovereignty is not a purity test. Organizations can certainly live with such an exposure during an emergency. Our message is simply be aware of the condition, evaluate the risk and if necessary put in compensating controls or live with the exposure. Risk determines spending behavior but regulations lack clarity It’s early days, so some customers are waiting for clearer sovereign-specific mandates or enforcement guidelines from local governments before expanding deployment and testing sovereignty. In other words, they’re leaning on existing privacy and cybersecurity laws and practices. We’d caution, however, that this type of uncertainty only strengthens the case for preparing acceptance tests sooner rather than later. Costs heavily influence the scope of sovereignty investments. Buyers justify stronger controls for sensitive workloads while accepting different levels of exposure for lower-risk activity. Migration effort, retraining and the cost of exiting all affect that decision. At the very least, this means customers should assess their sovereign exposure, much in the same way they would do so for cybersecurity. AI inference introduces another gotcha. Nine of 11 respondents require inference inside the sovereign or regional boundary. Buyers report that model availability can lag inside those environments. As such, the architectural choice has to make a potential tradeoff between local jurisdiction risk and model capability for a specific workload. The question buyers are asking: Is it worth stepping outside the lines to get a business result? The answer depends on the degree of risk and the magnitude of the business value. Sometimes it’s worth just paying the parking fine. High-level view of cloud sovereign scores We’re not going to go deep into the buyer’s guide – we did that last week and we’d encourage you to read the full report. What we will share is one high-level comparison from the buyer’s guide as shown below. Google LLC leads all measured cloud vendors with a score of 3.45 across all regions and tests. Google is followed by Oracle Corp. at 3.25. IBM Corp. and Microsoft Corp.’s Azure score 3.0. AWS scores 2.7 – its score suffers from not having the ability to have a customer-run air-gapped stack without vendor management. Huawei Technologies Co. Ltd. and Alibaba Group Holding Ltd. came in at 1.2. Now to be clear, this is a point-in-time index from the Nov. 14 revision. It averages each supplier’s best live offering across five buyer regions. The framework has five pillars (territorial, legal, operational, technical and financial), 17 yes-or-no tests across four nonlegal dimensions, plus a separate legal evaluation. As well, a low score at the highest level doesn’t mean a vendor won’t be chosen. We found many situations, for example, where Amazon Web Services Inc. was the choice despite its lower score because of its excellent quality and incumbent position. We found similar dynamics with Azure as a ubiquitous cloud and often a convenient option. Similar dynamics are in play within IBM and Oracle customer bases. That said, we’re assessing the architecture and the vendor claims to help buyers make better decisions. Customer-run deployments generally improve the scores because the methodology rewards controls the buyer can exercise independently. Missing documentation also affects the grades. We would use the previous chart to start an evaluation, then examine the actual offering available in the buyer’s jurisdiction against your specific workload requirements and local laws. Five key questions around sovereignty These findings bring us back to five key questions: Where does the data live, including logs and recovery copies? Who holds the keys, and who can actually decrypt? Which model touches the data, through which endpoint? Who operates the system, including support escalations? Who sees the invoice, controls the service relationship and absorbs the cost of exiting? Listen below to John and Amit connect operational controls and money flows to sovereignty. Those dependencies extend well beyond the physical location of the data center. Technical evals alone won’t determine sovereignty – US Cloud Act is a ‘gotcha’ Let’s dig into some more survey data. Seven of eleven buyers say a locally incorporated subsidiary does not completely clarify the local jurisdiction question. This is a buyer assessment, and it suggests the need for much deeper legal analysis beyond technical evaluations. Under the U.S. CLOUD Act, a vendor subject to U.S. jurisdiction can be required to produce data in its possession or control, regardless of storage location. Corporate relationships and actual control do matter here. Customer-held keys may limit outside access, but key ownership does not necessarily ensure legal immunity. Buyers also need to know what stops working if and when access is revoked. The fuzzy gap is captured by a South American bank’s executive director for AI, talking about the absence of testing: “That’s probably the weakest point in our whole sovereign story.” Our takeaway is that legal review and operational testing serve different purposes and both are critical. AI inference where ROI lives but it’s also populated with landmines Coming back to the AI inference issue. As we showed earlier, nine of 11 buyers require inference to take place inside their sovereign or regional boundary. One Latin American financial-services buyer said: “If it calls out to a U.S. endpoint, it doesn’t count no matter what the contract says.” This means you have to trace the end-to-end request cycle, from prompt, retrieval, model usage, outputs and even logs. An application that gets green lit can still depend on another service which lives in the execution sequence. The test is what happens to the data during its actual use. For sensitive workflows, buyers should compare a self-hosted open-weight model architecture with an approved frontier service and evaluate the cost and quality of the outcome. Open weights create optionality but the frontier often produces superior results which may warrant easing up on sovereignty requirements. Sovereign posture must be a continuous discipline Sovereignty needs ongoing attention because the architecture changes continuously. A model endpoint can move, a new model ships, a service contract changes or a new capability goes into production. Every change can affect exposure and risk. In this clip, Amit describes an AI-driven five-pillar framework for continuous assessment and management of data sovereignty and third-party risk. The approach brings vendor capabilities and claims, and the customer’s architecture, into an ongoing assessment. Let’s hear how he puts it: We see tremendous value in this framework. Our test for any assessment comes down to the evidence it produces. In other words, what’s the exposure, who has control, is there an accountable owner and ultimately, what’s the risk you’re absorbing? Our advice is that major changes to models, services or contracts should trigger a reassessment. Documented assurance from the vendor combined with technical tests should both be included in that process. Action items and recommendations Our recommendation is to make sovereignty evaluations at the workload level. Start with a core business process, identify the exposures and agree on the controls it requires based on the risk you’re willing to absorb. Importantly, for critical workloads, we suggest obtaining proof that those controls work in a safe sandbox environment and then push into production. This means you have to pay attention to privileged access. And you must test what happens when a key gets revoked and you have to recover. As we’ve talked about in previous episodes, verifying what the model did and the route of the execution path is something most organizations can’t observe today. Testing and telemetry are major gaps with most customers. Can you observe the route and exercise an exit or completely disconnected scenario if a workload requires it? Your financial assessment should include ongoing operation, migration, retraining, the cost of leaving and business disruption. And importantly, a business owner has to accept the risk, not an IT pro. For vendors, we believe the opportunity is substantial to document controls, establish predictable operating economics and provide credible evidence that can support your proposed solution. In many cases we’ve seen public documentation and evidence lack marketing claims. Sovereignty is becoming part of enterprise architecture because intelligence now depends on these relationships. Buyers need the ability to make an informed choice and demonstrate that the resulting system behaves as intended. Check out the Hyperscaler Sovereign Buyers Guide – it’s the first of its kind and very comprehensive. And do let us know if you you have any questions – we’re here to help. Support our mission to keep content open and free by engaging with theCUBE community. Join theCUBE’s Alumni Trust Network, where technology leaders connect, share intelligence and create opportunities. 15M+ viewers of theCUBE videos, powering conversations across AI, cloud, cybersecurity and more 11.4k+ theCUBE alumni — Connect with more than 11,400 tech and business leaders shaping the future through a unique trusted-based network Are you an AWS customer? Support SiliconANGLE financially by buying your AWS services from our Marketplace portal page and links: https://siliconangle.com/aws-marketplace/ Founded by tech visionaries John Furrier and Dave Vellante, SiliconANGLE Media has built a dynamic ecosystem of industry-leading digital media brands that reach 15+ million elite tech professionals. Our new proprietary theCUBE AI Video Cloud is breaking ground in audience interaction, leveraging theCUBEai.com neural network to help technology companies make data-driven decisions and stay at the forefront of industry conversations. CrowdStrike reportedly in talks to acquire cybersecurity provider Zscaler Three insights you may have missed from theCUBE’s coverage of the Dell Technologies New York Symposium Trump announces plan to tax US companies that build data centers abroad Anthropic researcher claims to have proven Riemann hypothesis using Claude Equifax cuts database firefighting after moving billing to Oracle on AWS Inference pushes AI to smaller sites, making the data layer the deciding factor CrowdStrike reportedly in talks to acquire cybersecurity provider ZscalerSECURITY - BY MARIA DEUTSCHER . 16 HOURS AGO Three insights you may have missed from theCUBE’s coverage of the Dell Technologies New York SymposiumAI - BY VICTORIA GAYTON . 17 HOURS AGO Trump announces plan to tax US companies that build data centers abroadPOLICY - BY MARIA DEUTSCHER . 18 HOURS AGO Anthropic researcher claims to have proven Riemann hypothesis using ClaudeAI - BY MARIA DEUTSCHER . 20 HOURS AGO Equifax cuts database firefighting after moving billing to Oracle on AWSCLOUD - BY JONATHAN ANTHONY . 21 HOURS AGO Inference pushes AI to smaller sites, making the data layer the deciding factor - BY SLOANE KALI FAYE . 22 HOURS AGO

Sources

Read original ↗

Related topics

← Continue to all updates