The Sovereign AI Paradox

Why owning the infrastructure is not the same as controlling the technology.

A workload-level framework for control in a globally interdependent stack. Sovereignty is judged by whether critical workloads can be governed, sustained, evaluated, and moved under plausible disruption — not by asset location, ownership, or a sovereign label.

Alaaeddin Al-Jallad · AIGNITRON Inc.DOI 10.5281/zenodo.23708820 →CC BY 4.0Evidence cut-off 27 Oct 2026

Cohere’s Sovereign Gamble

Canada’s proposed C$1.6B investment tests whether a national AI champion can expand strategic choices without becoming a new single point of dependence.

COMPANY CASE STUDY

Can Canada’s AI champion deliver strategic autonomy?

From the Cohere–Aleph Alpha combination and private enterprise deployment to public equity, exit rights, and the dependencies that remain across the AI stack.

Six countries on the evidence ladder

The cases were chosen to contrast policy models along two axes — position in the stack, and funding/delivery model. They are snapshots as of 27 October 2026, not rankings. Independent stress-test evidence (Rung 5) is essentially absent for all six.

STACK POSITION

Frontier exporters

United States and China shape the dependencies others face — chips, cloud, and models.

STRATEGIC AUTONOMY

Autonomy builders

France and Canada → fund domestic model champions and compute programmes, with imported hardware.

ACCESS & CAPITAL

Access-first & platforms

India expands shared, subsidised compute access; the UAE deploys sovereign capital with foreign technology partners.

Country Position in the stack Reported facts (source grade) Highest rung supported Caveat
🇺🇸 United StatesFrontier-stack exporter — chips, cloud, models. Private-led, with state export promotion.
  • Executive Order 14320 (23 Jul 2025) created the American AI Exports Program for full-stack export packages G1
  • Commerce announced implementation in October 2025 G1
Rung 1
policy instrument in force
Package-level delivery not reviewed; leverage does not remove upstream bottlenecks in fabrication and equipment.
🇨🇳 ChinaFrontier-stack exporter and self-reliance builder. State-coordinated.
  • State Council “AI Plus” Opinions (26 Aug 2025): deep AI integration across six areas; adoption targets >70% in key sectors by 2027 and 90% by 2030 G1 / G4
Rung 1
policy directive
Implementation and hardware self-reliance outcomes not assessed; English-language evidence on implementation is uneven.
🇫🇷 FranceStrategic-autonomy builder — domestic model developer, hardware imported. Mixed state funding plus private debt and customers.
  • €655M additional France 2030 AI funding announced 16 Jun 2026 G1
  • Mistral raised $830M in debt for 13,800 NVIDIA chips for a site near Paris G4; NVIDIA states the site is operating G3
  • BNP Paribas renewed its Mistral partnership for three years G3; SNCF reports a group assistant for 100,000+ staff G3
Funding: Rung 1
Mistral site: Rung 4
as reported by the vendor
Operation rests on company statements; chip counts differ across sources; no independent stress-test; hardware remains NVIDIA-supplied.
🇨🇦 Canada →Strategic-autonomy builder — compute programme, domestic model developer. State programme with private proponents.
  • Canadian Sovereign AI Compute Strategy: up to CAD $2B over five years from Budget 2024 G1
  • AI Sovereign Compute Infrastructure Program offers ~CAD $890M; applications closed 1 Jun 2026 G1
  • Cohere released Command A under Apache 2.0 on 20 May 2026 G3
Rung 1
programme funding; model release available to users
No award or commissioning evidenced in the sources reviewed; an open licence does not supply hardware, operations, or evaluation.
🇮🇳 IndiaAccess-first developer — shared, subsidised compute. State-led, globally sourced hardware.
  • IndiaAI Mission reports 45,000 shared GPUs as of June 2026 G1
  • 93.18 lakh (~9.32M) GPU-hours sanctioned for 237 projects by August 2026 G1
Rung 4
for access, government-reported
Sanctioned hours are not verified consumption; GPUs are globally sourced, so access expansion is not upstream self-sufficiency.
🇦🇪 United Arab EmiratesCapital and infrastructure platform. Sovereign capital with foreign technology partners.
  • Microsoft invested $1.5B in G42 (2024) G3; a 200 MW expansion expected to start coming online before end of 2026 G3
  • Stargate UAE plans 1 GW with 200 MW expected live in 2026 G3
  • Reuters reported in October 2026 that the 5 GW campus design is being revised after regional attacks G4
Rung 1–2
commissioning not evidenced
Forward-looking capacity; physical-security redesign illustrates that continuity depends on siting and defence, not only ownership.
Evidence grades: G1 primary legal/governmental document · G2 peer-reviewed or institutional report · G3 company statement · G4 journalism. Announcements are evidence of stated plans, not of performance. Excluded for scope, not relevance: the EU (supranational unit of analysis), other Gulf states, the UK, Japan, South Korea, and countries in Africa and Latin America. The Canada case has its own dedicated companion analysis — see the Canada deep-dive.

The evidence ladder

Claims about AI capacity are routinely made at different levels of maturity. The ladder makes the level explicit: a claim sits on the highest rung for which the cited evidence meets the stated test. Rungs are ordinal — never averaged, scored, or summed.

Rung 1 Announced
A public statement of intent or funding exists.
Ministerial announcement, call for applications, press release. Does not establish any binding commitment or delivery.
Rung 2 Contracted
A binding agreement or award exists.
Signed contribution agreement, financing close, purchase order. Does not establish construction, commissioning, or access.
Rung 3 Commissioned
Capacity is built and energised.
Commissioning notice, regulator or operator record. Does not establish accessibility to users or sustained use.
Rung 4 Operational
Users can access and use it for the stated service.
Usage, utilisation, queue and uptime data; named users. Does not establish resilience under disruption or exit.
Rung 5 Independently stress-tested
A third party has observed performance under a defined disruption and exit test.
Failover exercise, migration trial, audit report. Does not establish fitness beyond the tested scenario.

The five-part diagnostic

Applied to one named workload and one stated disruption. Each dimension records evidence, its grade, and its ladder rung — plus two cross-cutting requirements: independent evaluation and service adequacy.

01

Control

Who can approve changes, revoke access, inspect behaviour, suspend service, apply updates, and maintain the system?

02

Continuity

Can the service stay within agreed levels during a defined shock — API withdrawal, outage, export restriction, loss of operator?

03

Substitutability

Is there a named, tested alternative — and what does moving cost in time, labour, quality, and integration changes?

04

Data authority

Who sets purposes and access rules for prompts, outputs, embeddings, logs, telemetry, and backups — and do the actual flows match?

05

Economic viability

Is the capability sustainable over its lifecycle — cost per completed task, utilisation, labour, energy, and refresh?

+2

Cross-cutting

Independent evaluation without exclusive reliance on the supplier; service adequacy for the languages and communities served.

Non-compensatory decision rule: each requirement is met, unmet, or undetermined — missing or conflicting evidence yields undetermined, never met. No dimension offsets another; any mandatory unmet or undetermined requirement prevents an unqualified assurance conclusion. The output is a requirement-by-requirement profile and a versioned decision record — not a national score or a binary sovereign label.

Build, buy, lease, adapt, share

Countries and institutions should choose among five approaches by workload, not ideology. Options can coexist — each must leave a tested path to change.

BUILD

When demand is steady, the workload is critical, and ownership adds measurable control or resilience. Include power, cooling, staff, security, refresh, and utilisation in the business case.

BUY

Through an API when capability is needed quickly, demand varies, and data can lawfully go to the provider. Define a fallback for essential services.

LEASE

Cloud capacity when flexibility matters more than owning hardware. Assess proprietary services, egress fees, migration effort, and interoperability.

ADAPT

Open weights locally when an evaluated model meets the task at acceptable total cost — accounting for licensing, hardware, security updates, and skills.

SHARE

Trusted infrastructure when one organisation or country cannot justify the cost alone — with agreements on crisis allocation, legal access, and continuity if a partner withdraws.

THE TEST

Does the choice match the workload — and leave a credible, tested path to change suppliers or technical direction?


Policy recommendations (abridged)

  1. Classify workloads before buying infrastructureInventory by sensitivity, public impact, availability needs, and data rights.
  2. Map dependencies across the full service pathHardware, software, model, data, provider, operator, legal authority, network, energy.
  3. Procure options, not labelsRequire named operational authority, audit rights, portability, and a tested exit plan.
  4. Fund capacity against measured demand and accessReport the gap between announced, contracted, commissioned, and accessible capacity.
  5. Invest selectively in options and peopleEvaluation capacity, systems engineering, data governance, and multi-platform skills.
  6. Publish outcomes and preserve legitimate authorityLifecycle cost, reliability, language performance, and routes for affected users to challenge outputs.

Three observations from the cases

  1. Announcements dominateIndependent evidence (Rung 5) is essentially absent for all six cases.
  2. Dependencies persistExpanded access (India), capital partnership (UAE), and domestic champions (France, Canada) each leave hardware, software, and operational dependencies in place.
  3. Physical continuity mattersThe UAE reports show a facility's siting and defence can be a first-order sovereignty variable the ownership framing omits.
Status & limitations: independent working paper; targeted non-systematic review; sole-author coding; cases illustrate the ladder — they do not rank countries. Full paper: The Sovereign AI Paradox, Alaaeddin Al-Jallad, AIGNITRON Inc., v1.0, October 2026, CC BY 4.0 — Zenodo record.