AI Concentration SurfaceView the portfolio through eight exposure lenses

Critical input covers accelerators, energy, network, specialist software, model weights, proprietary data and scarce expertise. Provider and ownership traces the direct vendor, parent, investor, exclusive partner, subcontractor and common corporate group. Different contract names can still resolve to one source of capital, control or emergency decision.

Infrastructure and geography records cloud, region, availability zone, data centre, jurisdiction, network path and recovery site. Model and lineage joins foundation models, derivatives, safety layers, fine-tunes, version families and shared synthetic or training-data ancestry. Two suppliers can package the same base behaviour and therefore inherit a correlated refusal, bias or adversarial weakness.

Data and knowledge includes aggregators, licences, ontologies, retrieval services, labelling operations and feedback loops. Workflow and control includes identity, orchestration, policy, tool gateways, monitoring, evaluation, incident, evidence and approval services. Internally owned gateways remain concentration points if every important workflow must pass through them and no usable bypass exists.

Human and operating capacity traces integrators, managed reviewers, support, assurance specialists, internal key persons and transition staff. Commercial and legal joins minimum commitments, bundles, correlated renewals, common terms, cross-default, change rights and exit assistance. The eight lenses describe one surface because loss often propagates from one lens into several others.

Dependency Overlap MatrixAggregate by common component and provider family

Rows are important business services and consequential AI capabilities. Columns are direct suppliers, parents, nth parties, foundation-model families, cloud regions, data sources, control services and specialist teams. Each cell records criticality, volume, affected population, consequence, substitute, recovery time, data exposure and evidence confidence. The matrix does not award a generic concentration percentage. It shows which columns can disable several important rows together.

Normalise names before aggregation. A software vendor, managed service and internal application may all use the same model family. Two cloud contracts may share a corporate group or region. A local tool may depend on one open-source runtime and maintainer community. Parent relationships, exclusive partnerships and subcontractors turn apparently separate columns into a provider-family view.

Evidence confidence matters. Known sharing can be stress-tested. Unknown lineage or subcontracting lengthens detection and decision because responders do not know which services are exposed. The matrix therefore treats an unverified dependency cell as uncertainty, not an empty or favourable cell. Owners must seek evidence, narrow the capability or accept the uncertainty explicitly.

Concentration Loss FunctionDescribe common loss with six separate terms

Breadth counts affected capabilities, business services, populations, regions and control functions. Consequence asks what those services do and what harm follows shared impairment. Ten low-impact drafting tools may matter less than one common evaluator approving three consequential decision systems. Spend and count become useful only after outcome criticality is attached.

Duration runs until safe recovery, not until the provider reports availability. Backlog, evidence reconstruction, revalidation and human readiness can outlast the technical incident. Substitutability covers qualification time, reserved capacity, rights, data movement, control compatibility and cutover. A named alternative that cannot accept load or pass cases inside tolerance is not a substitute.

Amplification captures shared data, model lineage, decisions and feedback. One biased data feature can influence several models; one evaluator blind spot can approve the same harm; correlated outputs can reinforce a market or operational action. Uncertainty reflects incomplete direct and nth-party evidence. It widens the plausible affected set and delays containment. These terms are kept distinct because a low average cannot erase an intolerable consequence or unknown recovery path.

Common-Mode Scenario DeckTest more than outage

The first card is outage or capacity withdrawal. The second is security or software-supply-chain compromise. The third is a model, safety-policy or API update. The fourth is a legal, licence, jurisdiction or data-rights restriction. The fifth is a price, credit, acquisition or commercial shock. These can all reach a portfolio that remains technically available.

The remaining cards are correlated model error, refusal, bias or adversarial weakness; a common evaluation or control blind spot; and loss of specialist, integrator or reviewer capacity. Each exercise starts at the shared dependency, traverses the overlap matrix and records affected outcomes, authority changes, evidence, fallback, substitute time and recovery. An outage dashboard cannot answer a licence dispute, grader blind spot or model-family behaviour change.

Severe but plausible does not mean maximal imagination. Use provider notices, dependency evidence, incident history, contractual events, research and control findings to construct credible shocks. Run combinations where coupling is real, such as a security defect that also withdraws capacity, or a model update that increases human review during a staffing shortage. The objective is to reveal common paths and decision latency, not to produce theatrical disaster.

Qualified Substitution ClockMeasure the time to a safe alternative, not a procurement name

The clock starts with detection: recognise the common dependency and affected set. Decision follows: an accountable owner selects restriction, fallback or transition. Access covers contractual rights, credentials, environment and reserved capacity. Migration moves data, configuration and state. Validation runs representative, harmful, failure and recovery cases against declared thresholds.

Human readiness prepares reviewers, operators, support and incident roles. Authority confirms that the substitute may lawfully and operationally serve the population. Cutover moves bounded traffic and reconciles in-flight effects. Stabilisation proves capacity, monitoring, evidence and recovery through an operating window. The complete clock, not the provider's technical onboarding estimate, is compared with the service impact tolerance.

Termination rights do not create emergency capacity. A market-wide incident may send many customers to the same alternative. Reserved capacity, current credentials, data rights and rehearsed operations matter. Qualification also expires as models, infrastructure, policies, people and contracts change. A warm substitute is maintained, not listed once.

AI Concentration Capacity BudgetLimit shared consequence, not suppliers by quota

The budget limits how many important services, consequential decisions, customer populations, regions and control functions may share one dependency or provider family. Each dimension stands separately. One common evaluator across several high-impact decisions can breach the control-function limit even when model and cloud exposure remain diverse. One model family serving low-consequence drafting may remain inside budget despite high application count.

A threshold crossing triggers an action, not an automatic vendor ban. The action may be diversification, warm substitution, ring-fencing, capacity reservation, narrower populations, independent evaluation or board acceptance. Provider scale may justify deliberate concentration when evidence shows stronger geographic resilience and the organisation retains usable fallbacks. The owner records why fragmentation would worsen recovery, which common modes remain, and when the acceptance expires.

Budgets require enterprise governance. Business units cannot see exposures created by unrelated contracts and shared nth parties. Procurement, architecture, resilience, risk and operations maintain the aggregated matrix and scenario evidence. Unknown cells consume attention and may consume budget because missing evidence enlarges the plausible failure domain.

Capacity reservation and transition ownershipTurn the portfolio map into commitments that exist before failure

A board decision is incomplete if it merely acknowledges a red column. Every material concentration needs an operating disposition. For diversification, name the services that move first, the capacity reserved at the alternative, the evidence required for qualification and the executive who can authorise cutover. For ring-fencing, state the maximum population, transaction and control exposure, the fallback mode and the signal that closes new admission. For deliberate concentration, record the provider advantage, monitoring, contractual evidence, restricted continuity mode and expiry of acceptance.

Capacity reservation is specific. It identifies workload shape, region, data and identity prerequisites, minimum and burst demand, activation time, price, competing-customer assumptions and evidence that the provider can honour the reservation during a common event. A paper right to buy service later is weaker than a periodically exercised allocation. Internal fallbacks need the same realism: named staff, protected access, current runbooks, safe queues and authority to operate under pressure.

Transition ownership spans the whole clock. Procurement owns usable rights and assistance; architecture owns dependency resolution and the target path; data owners approve lawful movement and lineage; model and product owners run outcome evidence; security and risk authorise the bounded substitute; operations own cutover, backlog and stabilisation; business owners decide customer continuity and residual loss. One incident commander coordinates the event, but no technical team can silently accept outcome or population risk for the enterprise.

Supplier-failure triggers are written before stress. They include missed capacity, unannounced model or subcontractor change, licence restriction, repeated evidence failure, degraded support, control disagreement and a substitution clock outside tolerance. Each trigger maps to admission freeze, restricted mode, alternative activation, customer communication, executive escalation or portfolio stop. Customer continuity distinguishes work that can queue, move to manual handling, operate read-only, narrow to a protected cohort or cease safely. Concentration governance becomes service governance through pre-funded capacity, rehearsed authority and explicit customer effects, not a diagram reviewed at renewal.

Ten worked casesMap, aggregate, stress, time and decide

1. Five vendors, one model family

The Surface reveals model-lineage sharing behind five software contracts. The Matrix lights one column across customer services. Breadth and consequence are high; an update scenario changes refusal behaviour together. The clock shows that two alternatives exist but neither has current cases or reserved capacity. The budget is breached. Decision: qualify one warm substitute immediately and diversify the most consequential workflows once it passes.

2. Three models, one infrastructure spine

Two external models and one internal model share a cloud region, identity plane and vector service. Infrastructure and control lenses dominate the Matrix. An outage plus identity-revocation scenario disables all paths. Substitution exceeds impact tolerance because state and credentials cannot move. The capacity budget rejects the claim of independence. Decision: ring-fence authority and establish a bounded fallback before calling the portfolio multi-provider.

3. Open models, shared runtime and maintainers

Three open-weight models depend on one inference runtime, accelerator supply and small maintainer community. Input, infrastructure and people columns overlap. A compromise scenario forces coordinated containment and rebuild. The loss function adds uncertainty because downstream versions are poorly recorded. The clock lacks patched capacity. Decision: qualify another runtime for critical services and stop treating open access as deconcentration.

4. One data aggregator across fraud, credit and collections

A licensing dispute removes a feature used by three decision systems. The data lens connects populations and feedback loops. The legal scenario produces immediate model drift and retraining work. An alternative feed exists but has different semantics and has not passed outcome tests. The clock is slow and the decision budget breached. Decision: diversify the critical data dependency now, preserve stable feature meaning and revalidate each affected decision.

5. One evaluator approves every release

Provider models differ, but one evaluator and benchmark govern release. The control column crosses the entire estate. A blind-spot scenario misses the same harmful behaviour everywhere and amplifies exposure through shared approval. A second measurement path can be built quickly. Decision: diversify the critical control now, retain human adjudication and require disagreement analysis before further consequential releases.

6. One safety-policy change across workflows

A foundation provider changes refusal behaviour used by service, compliance and operations tools. Model-family aggregation exposes the breadth. The scenario is operational rather than outage: systems answer, but legitimate work stops. A warm alternative meets ordinary cases yet lacks edge-case authority testing. Decision: qualify and rehearse the substitute, while ring-fencing populations whose safe fallback remains manual.

7. One integrator carries the operating memory

The same integrator runs prompts, incident support and exception practice for several critical capabilities. Human capacity and commercial terms overlap. A talent-loss scenario shows that internal owners cannot stabilise the services inside tolerance. Code and model substitutes do not shorten the clock. Decision: stop scaling, transfer runbooks through witnessed operation and build internal plus alternate transition capacity.

8. Provider scale improves resilience

A major cloud-model provider offers stronger regions, security operations and recovery than smaller alternatives. The portfolio remains concentrated, but geographic and operating evidence show that forced fragmentation would lengthen recovery and dilute expertise. The Matrix and scenarios remain transparent, reserved fallback covers essential work, and the budget stays within tolerance. Decision: accept monitored concentration for strategic advantage, with time-bounded review and a tested restricted mode.

9. Named alternative without emergency capacity

A contract lists a second provider, but it has no reserved capacity during a market-wide disruption. The substitution clock stalls at access. An outage scenario shows many customers competing for the same supply. Breadth and duration exceed tolerance, while termination rights add no capacity. Decision: stop expansion until access, validation, human readiness and cutover are evidenced, or narrow the population to the available fallback.

10. Common open-source compromise

A library used in models, gateways and evaluation tooling is compromised. The critical-input column links otherwise diverse products. The scenario requires shared inventory, containment, provenance, rebuild and release decisions. Uncertain versions amplify duration. A qualified patched path exists for high-consequence services. Decision: ring-fence exposure with capacity and fallback, coordinate evidence across the estate, and restore only after component and outcome regression passes.

Interactive decision artefactPortfolio Concentration Stress Lab

The Lab returns one posture from the same evidence used in the Surface, Matrix, loss logic, scenario deck, substitution clock and capacity budget. It does not prescribe provider quotas or compute a vendor concentration percentage.

Portfolio operating cadenceRefresh shared-dependency evidence before it goes stale

At onboarding, register direct suppliers, ownership, nth parties, model family, data, infrastructure, controls, people and terms against the important service. Require version, region, lineage, subcontractor, population and evidence confidence. A name without these fields cannot support aggregation.

Quarterly and at material change, normalise provider families and common components across the portfolio. Update the Matrix for model, cloud, data, evaluator, gateway, library, integrator, contract or ownership changes. New unknown cells receive owners and evidence dates. The enterprise team, not each business unit alone, assesses capacity use.

In scenario exercises, select at least one operational, security, behavioural, legal, commercial, control or talent card. Traverse affected rows, narrow authority, invoke fallbacks, run the substitution clock and capture actual decision, validation, cutover and stabilisation time. A substitute that misses tolerance triggers diversification, ring-fencing, capacity, restricted scale or acceptance.

At renewal and scale, compare shared loss with provider advantage. Do not fragment a strong resilient service merely to improve a supplier count. Do not accept concentration because switching is inconvenient. The board receives the top common columns, affected populations, scenario results, clock, capacity position, evidence gaps and the expiry of every strategic acceptance.