Models
The FSB notes that financial institutions deploying LLMs generally use pre-trained, third-party models. In the Bank of England and FCA’s 2026 survey, three providers held 54% of reported model API and hosting relationships.
Where dependence on a few AI providers builds up, what DORA and financial stability bodies say about it, and how to reduce it.
AI concentration risk is the exposure created when an organization, or a whole sector, depends on a small number of providers for the models, cloud capacity, chips or data behind its AI.
If one of them fails, changes its terms or is compromised, many AI systems are affected at once and substitutes are hard to find quickly.
of reported cloud and compute relationships for AI at UK financial firms sit with the three largest providers (2026)
Source: Bank of England and FCA, 2026
of reported model API and hosting relationships sit with the three largest providers (2026)
Source: Bank of England and FCA, 2026
critical ICT third-party providers on the list the ESAs published under DORA on November 18, 2025
Source: European Supervisory Authorities
The FSB’s 2024 report on AI names three layers: specialised hardware, cloud services and pre-trained models. A fourth is easy to miss.
The FSB notes that financial institutions deploying LLMs generally use pre-trained, third-party models. In the Bank of England and FCA’s 2026 survey, three providers held 54% of reported model API and hosting relationships.
The same survey found three providers behind 79% of reported cloud and compute relationships.
The FSB says the markets for accelerated computing chips and cloud services are dominated by a limited number of entities.
FSB report, 2024Ten SaaS tools can sit on one model provider. You see ten contracts and carry one dependency.
Fourth-party AI risk| Source | What it says |
|---|---|
| DORA, Article 3(29) | ICT concentration risk is exposure to one or several related critical providers, such that their failure could endanger a financial entity’s critical or important functions or the financial stability of the Union |
| DORA, Article 29(1) | Before contracting, consider whether the provider is not easily substitutable, and whether you would hold several arrangements with the same or closely connected providers. Weigh the alternatives |
| DORA, Article 31 | The European Supervisory Authorities designate critical ICT third-party service providers, using criteria that include systemic impact and substitutability, and publish the list yearly |
| FSB sound practices (October 2026) | Sound Practice 12, third-party AI risk management, asks for a focus on performance, transparency, data quality, supply chain and concentration risks, and business continuity |
For each AI system, list the model provider, host and data sources, including those behind your vendors. AI‑BOMs that Cranium AI generates from code repositories list the models and technologies your own applications call.
Find the providers that appear most often, and mark the ones that are both critical and hard to substitute.
In the 2026 UK survey, 37% of cloud and compute relationships were both high criticality and hard to substitute. Only 35% of that group had tested exit plans.
Keep prompts, evaluations and data portable, so a second model can take over a critical workflow.
Give the risk committee a provider-level view, and revisit it before each new AI contract.
Straight answers, kept current. Last updated .
Ask us anything →It is a dependency that becomes a risk when the work it supports is critical and the provider is hard to replace. DORA’s test is a useful one outside finance too: is the provider easily substitutable, and how many separate arrangements lead back to it? A single provider for low-stakes drafting is a commercial choice. A single provider behind every critical decision needs an exit plan.
DORA is written for ICT services used by EU financial entities and has applied since January 17, 2025. An AI model API or hosting service that qualifies as an ICT service is assessed like any other: Article 28(4) requires concentration risk to be considered before contracting, and Article 29 sets out what to weigh for services supporting critical or important functions.
Fourth-party risk is about visibility and control: suppliers you depend on but hold no contract with. Concentration risk is about correlation: many dependencies leading to the same provider. They meet when several vendors share one model provider or cloud, which is why mapping fourth parties is the first step in measuring concentration.
Primary sources for the figures and frameworks cited on this page. Last verified .
See how Cranium AI discovers, observes, governs, secures and proves every AI model, agent and vendor in your enterprise.
Cookie Policy · Privacy Policy
Global Privacy Control detected. Opt-out request honored: marketing tags stay off in this browser.