On 31 August 2026, AMD, Cisco and HUMAIN, which is backed by Saudi Arabia's Public Investment Fund, announced that AI infrastructure using AMD Instinct MI355X GPUs, EPYC CPUs and Cisco Silicon One networking had entered production in Saudi Arabia and begun providing training and inference compute to customers in the Kingdom and elsewhere. AMD described the platform as open infrastructure for “sovereign AI”, emphasising local operation, data location, model customisation, and control over deployment and governance. Independent reporting by The National confirmed that the system was operational while also noting that its core GPUs, CPUs and networking are supplied by AMD and Cisco. The announcement's plans to begin deploying up to 250 megawatts in 2027 and reach one gigawatt by 2030 concern later stages; they should not be confused with the capacity already operating.
The question raised by this event is not whether Saudi Arabia now “owns AI”, nor whether foreign products are inherently incompatible with autonomy. The real question is whether technological autonomy has been achieved when compute is physically local, a local institution operates it, and data and deployment rules can also be controlled locally. My judgement is that local compute is an important material condition of autonomy but not a sufficient one. Technological autonomy is not another name for owning equipment, nor is it achieved merely by moving dependency from a foreign cloud into a domestic data centre. It is an institutional capacity to keep choosing, testing, replacing and revising technical relationships.
Four things must be distinguished. Infrastructure location tells us where computation occurs. Legal jurisdiction tells us which rules may apply. Operational control tells us who can allocate resources, configure systems and permit models to run. Technological autonomy asks a different question: if suppliers, software stacks, licences, maintenance conditions or external relationships change, can the actor continue operating, alter course in accordance with its own purposes and assume responsibility for the decision? The first three can strengthen the fourth, but they do not automatically produce it. A locally operated data centre may offer strong data-residency guarantees and extensive day-to-day control while remaining dependent on a small number of external providers for critical chips, drivers, network firmware, maintenance knowledge or upgrade paths.
This does not require autonomy to mean self-sufficiency. Every complex technical system is built through cooperation across organisations and borders. If autonomy required independent production of every component, almost no real actor would qualify. Dependency does not itself cancel autonomy. What matters is whether dependencies can be identified, negotiated, substituted and subjected to effective constraints. An institution can rely on partners and remain autonomous when it retains genuine exit options, conversion capacity, technical understanding and final decision authority. Conversely, if a critical service can be interpreted, repaired, authorised or terminated only by one external party, the practical field of choice may remain narrow even when the equipment is local.
In their peer-reviewed 2020 article on digital sovereignty, Julia Pohle and Thorsten Thiel argue that the concept is used by states, companies and individuals to express different claims to self-determination and therefore cannot be treated as a technical label with a settled meaning. They connect state-level claims to a capacity to make autonomous decisions about digital infrastructure and technology deployment, while warning that digital sovereignty often operates as political discourse rather than an accomplished organisational fact. In a peer-reviewed 2024 paper in Ethics and Information Technology, Huw Roberts further distinguishes descriptive control from normative authority. His position is that identifying who controls a technology does not yet answer whether that control is legitimate or responsive to the interests of those affected.
Together, these analyses expose two dimensions that the phrase “sovereign AI” can leave implicit. One is capability. Can the operator audit the system, understand failures, migrate workloads, replace suppliers, maintain critical software and preserve essential services when external supply is interrupted? The other is legitimacy. Who may decide how data is used, which objectives models serve, who receives access and which risks are accepted? Can affected people know, challenge and revise those decisions? Localisation without these capabilities may leave sovereignty at the geographical level. Control without an accountable decision structure may instead reduce sovereignty to concentrated power.
From the perspective of Sustenesis Theory, autonomy is not a static property possessed by an object. It is a structure formed in relationships and maintained over time. Difference first requires the boundaries of action to be distinguishable. We need to know who owns the equipment, who controls software updates, who holds keys and logs, who can terminate service and who bears the consequences of failure. If these differences are hidden by the phrase “local deployment”, the actual distribution of control cannot be assessed.
Constraint does not mean limitation in a generic sense. It refers to the conditions that make some forms of action possible and others impossible. Chip supply, software licences, interface standards, energy, expertise, contracts, audit rights and legal responsibility are all constraints in this case. Technological autonomy is not the absence of constraints. It exists when an actor can identify them, choose among more than one viable path and alter the arrangements that determine its field of action.
Sustained Coherence requires this capacity for choice to persist through feedback, failure and changing relationships. A system running on the day of its launch does not by itself show that its operator has acquired lasting autonomy. Stronger evidence would include demonstrated portability across platforms, the independent inspectability of software and models, alternative routes for critical components, locally held maintenance and correction expertise, and governance decisions that can be publicly tested. Sustained coherence is not simply keeping equipment switched on. It is maintaining the relationship among purpose, knowledge, control and correction as circumstances change.
On this account, the 31 August launch should be understood precisely as a real increase in local AI infrastructure and operational control. It places part of the capacity for training, inference, data residency and resource allocation within Saudi Arabia and provides a wider field of action than complete reliance on remotely delivered foreign services. At the same time, the public materials do not disclose the platform's capacity for chip substitution, firmware audit, software forking, continuity during supply interruption or independent maintenance. Nor is there enough information to judge how its governance is accountable to those affected. The present evidence therefore supports the claim that conditions for technological autonomy have increased, not that technological autonomy has been completed.
The boundary of this judgement is also clear. It is not a ranking of national capability, it does not deny that international cooperation can constitute reliable autonomy, and it does not treat domestic manufacture as the sole criterion. The judgement could change if verifiable evidence emerges of migration tests, open audits, alternative supply, independent operations, accountable governance and continued functioning during disruption. Technological autonomy is not measured by how many GPUs are located in one place. It concerns whether a community of action can retain the capacity to understand, choose, revise and take responsibility within unavoidable relations of dependency.
References
AMD, AMD, Cisco and HUMAIN Expand Saudi Arabia’s AI Infrastructure as AMD Instinct Systems Go Live, 31 August 2026
https://ir.amd.com/news-events/press-releases/detail/1298/amd-cisco-and-humain-expand-saudi-arabias-ai-infrastructure-as-amd-instinct-systems-go-live
The National, Saudi Arabia expands AI capacity with AMD and Cisco, 31 August 2026
https://www.thenationalnews.com/future/technology/2026/08/31/saudi-ai-humain-leap-amd-cisco/
Julia Pohle and Thorsten Thiel, Digital sovereignty, Internet Policy Review, 17 December 2020
https://policyreview.info/concepts/digital-sovereignty
Huw Roberts, Digital sovereignty and artificial intelligence: a normative approach, Ethics and Information Technology, 18 October 2024
https://link.springer.com/article/10.1007/s10676-024-09810-5
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.